A First Course in Systems Biology (2018)

A F IR S T COUR SE IN
SYSTEMS
BIOLOGY
SECOND
EDITION
To Ann,
Still the Hub of my Support System
A F IR S T COUR SE IN
SYSTEMS
BIOLOGY
SECOND
EDITION
Eberhard O. Voit
Garland Science Front cover image. The beautiful geometric shape of the fractal is
Vice President: Denise Schanck called self-similar because it has the same appearance at smaller
Senior Editorial Assistant: Katie Laurentiev and smaller scales. It reminds us of fundamental design features like
Assistant Editor: David Borrowdale feedback loops that we encounter at many organizational levels of
Production Editor: Georgina Lucas biological systems. Fractals are generated with nonlinear recursive
Illustrations: Nigel Orme models, and they are discussed with simpler examples in Chapter 4.
Copyeditor: H.M. (Mac) Clarke (Courtesy of Wolfgang Beyer under Creative Commons Attribution-
Typesetting: NovaTechset Pvt Ltd Share Alike 3.0 Unported license.)
Proofreader: Sally Huish
Indexer: Nancy Newman
Cover Design: Andrew Magee
© 2018 by Garland Science, Taylor & Francis Group, LLC
This book contains information obtained from authentic

and highly regarded sources. Reprinted material is
quoted with permission, and sources are indicated.
Reasonable efforts have been made to publish reliable
data and information, but the author and the publisher
cannot assume responsibility for the validity of all
materials or for the consequences of their use. All
rights reserved. No part of this publication may be
reproduced, stored in a retrieval system or transmitted
in any form or by any means – graphic, electronic, or
mechanical, including photocopying, recording, taping,
or information storage and retrieval systems – without
permission of the copyright holder.
ISBN 978-0-8153-4568-8
Library of Congress Cataloging-in-Publication Data

Names: Voit, Eberhard O., author.
Title: A first course in systems biology / Eberhard O. Voit.
Description: Second edition. | New York : Garland Science, 2017.
Identifiers: LCCN 2017017580 | ISBN 9780815345688 (alk. paper)
Subjects: LCSH: Systems biology. | Computational biology.
Classification: LCC QH324.2 .V65 2017 | DDC 570.1/13–dc23
LC record available at https://lccn.loc.gov/2017017580
Published by Garland Science, Taylor & Francis Group, LLC, an informa business,
711 Third Avenue, 8th Floor, New York NY 10017, USA,
and 2 Park Square, Milton Park, Abingdon, OX14 4RN, UK.
Printed in the United States of America
15 14 13 12 11 10 9 8 7 6 5 4 3 2 1
Visit our website at http://www.garlandscience.com

Preface
Hard to believe, but it is already time for the second edition! I am happy to report
that the first edition of A First Course in Systems Biology has met with great suc-
cess. The book has been a required or recommended text for over 70 courses
worldwide, and it has even been translated into Korean. So why should a new
edition be necessary after only five short years? Well, much has happened.
Systems biology has come out of the shadows with gusto. Research is flourishing
worldwide, quite a few new journals have been launched, and many institutions
now offer courses in the field.
While the landscape of systems biology has evolved rapidly, the fundamental
topics covered by the first edition are as important as they were five years ago
and probably will be several decades from now. Thus, I decided to retain the
structure of the first edition but have rearranged some items and added a few
topics, along with new examples. At Georgia Tech we have used the book to
teach well over 1000 students, mostly at the undergraduate level, but also for an
introductory graduate-level course. Most of the additions and amendments to
this new edition respond to feedback from these students and their instructors,
who have pointed out aspects of the material where more or better explanations
and illustrations would be helpful. New topics in this edition include: default
modules for model design, limit cycles and chaos, parameter estimation in
Excel, model representations of gene regulation through transcription factors,
derivation of the Michaelis-Menten rate law from the original conceptual
model, different types of inhibition, hysteresis, a model of differentiation,
system adaptation to persistent signals, nonlinear nullclines, PBPK models, and
elementary modes.
I would like to thank three undergraduates from my classes who helped me with
the development of some of the new examples, namely Carla Kumbale, Kavya
Muddukumar, and Gautam Rangavajla. Quite a few other students have helped
me with the creation of new practice exercises, many of which are available on
the book’s support website. I also want to express my gratitude to David
Borrowdale, Katie Laurentiev, Georgina Lucas, Denise Schanck, and Summers
Scholl at Garland Science for shepherding this second edition through the
review and production process.
It is my hope that this new edition retains the appeal of the original and has
become even better through the alterations and twists it has taken, large and
small.
Eberhard Voit
Georgia Tech
2017
vi Instructor resources WebsIte
Instructor Resources Website

The images from A First Course in Systems Biology, Second Edition are available
on the Instructor Site in two convenient formats: PowerPoint® and JPEG. They
have been optimized for display on a computer. Solutions to end-of-chapter
exercises are also available. The resources may be browsed by individual chap-
ters and there is a search engine. Figures are searchable by figure number, figure
name, or by keywords used in the figure legend from the book.
Accessible from www.garlandscience.com, the Instructor’s Resource Site

requires registration and access is available only to qualified instructors. To
access the Instructor Resource site, please email [email protected].
Acknowledgments
The author and publisher of A First Course in Systems Biology, Second Edition
gratefully acknowledge the contributions of the following reviewers in the
development of this book:
Guy Grant, University of Bedfordshire

Princess Imoukhuede, University of Illinois at Urbana-Champaign
Dimitrios Morikis, University of California at Riverside
Oliver Schildgen, University of Witten
Manuel Simões, University of Porto
Mark Speck, Chaminade University
Marios Stavridis, Ninewells Hospital & Medical School
Geraint Thomas, University College London
Floyd Wittink, Leiden University
Contents
chapter 1: biological systems 1 3.2 Small-World Networks 58

Reductionism and Systems Biology 5 Dependencies Among Network
Even Simple Systems Can Confuse Us 8 Components 62
Why Now? 10 3.3 Causality Analysis 62
Communicating Systems Biology 13 3.4 Mutual Information 62
The Task Before Us 16 Bayesian Reconstruction of Interaction
Networks 63
Exercises 17
3.5 Application to Signaling Networks 66
References 17
3.6 Applications to Other Biological
Further Reading 18
Networks 69
Static Metabolic Networks and Their Analysis 69
chapter 2: Introduction to 3.7 Stoichiometric Networks 70
Mathematical Modeling 19 3.8 Variants of Stoichiometric Analysis 73
Goals, Inputs, and Initial Exploration 24 3.9 Metabolic Network Reconstruction 73
2.1 Questions of Scale 24 3.10 Metabolic Control Analysis 74
2.2 Data Availability 25 Exercises 78
Model Selection and Design 26 References 80
2.3 Model Structure 27 Further Reading 82
2.4 System Components 30
2.5 Model Equations 35
chapter 4: the Mathematics of
2.6 Parameter Estimation 36
biological systems 83
Model Analysis and Diagnosis 37
Discrete Linear Systems Models 85
2.7 Consistency and Robustness 38
4.1 Recursive Deterministic Models 85
2.8 Exploration and Validation of
4.2 Recursive Stochastic Models 88
Dynamical Features 40
Discrete Nonlinear Systems 91
Model Use and Applications 43
Continuous Linear Systems 93
2.9 Model Extensions and Refinements 43
4.3 Linear Differential Equations 94
2.10 Large-Scale Model Assessments 45
4.4 Linearized Models 95
2.11 Questions of Design 46
Continuous Nonlinear Systems 100
2.12 Simplicity versus Complexity 47
4.5 Ad hoc Models 101
Exercises 49
4.6 Canonical Models 102
References 50
4.7 More Complicated Dynamical
Further Reading 50
Systems Descriptions 110
Standard Analyses of Biological
chapter 3: static network Models 51 Systems Models 110
Strategies of Analysis 52 4.8 Steady-State Analysis 110
Interaction Graphs 53 4.9 Stability Analysis 115
3.1 Properties of Graphs 54 4.10 Parameter Sensitivity 118
contents ix
4.11 Analysis of Systems Dynamics 119 6.13 Transcription Factors 188

Other Attractors 122 6.14 Models of Gene Regulation 190
4.12 Limit Cycles 123 Measuring Gene Expression 191
4.13 Chaotic Attractors 126 Localization of Gene Expression 194
Exercises 128 Outlook 196
References 132 Exercises 196
Further Reading 133 References 198
Further Reading 200
chapter 5: Parameter estimation 135

Parameter Estimation for Linear Systems 136 chapter 7: Protein systems 201
5.1 Linear Regression Involving a Chemical and Physical Features of Proteins 202
Single Variable 136 7.1 Experimental Protein Structure
5.2 Linear Regression Involving Several Determination and Visualization 206
Variables 138 An Incomplete Survey of the Roles and
Parameter Estimation for Nonlinear Systems 141 Functions of Proteins 208
5.3 Comprehensive Grid Search 143 7.2 Enzymes 209
5.4 Nonlinear Regression 145 7.3 Transporters and Carriers 211
5.5 Genetic Algorithms 146 7.4 Signaling and Messenger Proteins 214
5.6 Other Stochastic Algorithms 148 7.5 Proteins of the Immune System 215
5.7 Typical Challenges 149 7.6 Structure Proteins 216
Parameter Estimation for Systems of Current Challenges in Protein Research 218
Differential Equations 153 7.7 Proteomics 218
Structure Identification 160 7.8 Structure and Function Prediction 220
Exercises 161 7.9 Localization 222
References 166 7.10 Protein Activity and Dynamics 224
Further Reading 167 Exercises 226
References 228
Further Reading 230
chapter 6: Gene systems 169
The Central Dogma 169
chapter 8: Metabolic systems 231
Key Properties of DNA and RNA 171
Biochemical Reactions 232
6.1 Chemical and Physical Features 171
8.1 Background 232
6.2 Size and Organization of DNA 174
8.2 Mathematical Formulation of
6.3 Genes and Noncoding DNA 175 Elementary Reactions 234
6.4 Eukaryotic DNA Packing 178 8.3 Rate Laws 235
6.5 Epigenetics 178 Pathways and Pathway Systems 240
RNA 178 8.4 Biochemistry and Metabolomics 240
6.6 Messenger RNA (mRNA) 179 8.5 Resources for Computational
6.7 Transfer RNA (tRNA) 182 Pathway Analysis 241
6.8 Ribosomal RNA (rRNA) 182 8.6 Control of Pathway Systems 244
6.9 Small RNAs 183 Methods of Metabolomic Data Generation 246
6.10 RNA Viruses 184 8.7 Sampling, Extraction, and
Gene Regulation 185 Separation Methods 247
6.11 The lac Operon 186 8.8 Detection Methods 247
6.12 Modes of Regulation 187 8.9 Flux Analysis 249
x contents
From Data to Systems Models 250 chapter 11: Integrative Analysis of

8.10 Case Study 1: Analyzing Metabolism Genome, Protein, and
in an Incompletely Characterized Metabolite Data: A case
Organism 250 study in Yeast 303
8.11 Case Study 2: Metabolic Network On the Origin of Models 304
Analysis 251
A Brief Review of the Heat Stress
8.12 Case Study 3: Extraction of Dynamic Response in Yeast 306
Models from Experimental Data 251
11.1 The Trehalose Cycle 308
Exercises 252
Modeling Analysis of the Trehalose Cycle 310
References 254
11.2 Design and Diagnosis of a Metabolic
Further Reading 255 Pathway Model 310
11.3 Analysis of Heat Stress 312
chapter 9: signaling systems 257 11.4 Accounting for Glucose Dynamics 314
Static Models of Signal Transduction 11.5 Gene Expression 315
Networks 259 Multiscale Analysis 318
9.1 Boolean Networks 259 11.6 In Vivo NMR Profiles 318
9.2 Network Inference 261 11.7 Multiscale Model Design 320
Signal Transduction Systems Modeled with 11.8 The Trehalase Puzzle 324
Differential Equations 261 Concluding Comments 327
9.3 Bistability and Hysteresis 261 Exercises 328
9.4 Two-Component Signaling Systems 266 References 329
9.5 Mitogen-Activated Protein Kinase Further reading 330
Cascades 270
9.6 Adaptation 273
9.7 Other Signaling Systems 274 chapter 12: Physiological Modeling:
Exercises 278
the Heart as an example 331
Hierarchy of Scales and Modeling Approaches 332
References 279
12.1 Basics of Heart Anatomy 333
Further reading 281
12.2 Modeling Targets at the Organ Level 334
12.3 Modeling Targets at the Tissue Level 335
chapter 10: Population systems 283 12.4 Modeling Targets at the Cell Level 337
Population Growth 283
Simple Models of Oscillations 339
10.1 Traditional Models of
12.5 Black-Box Models of Oscillations 339
Population Growth 284
12.6 Summary of Black-Box Oscillation
10.2 More Complex Growth Phenomena 286
Models 342
Population Dynamics Under External
12.7 From a Black Box to Meaningful
Perturbations 288
Models 343
Analysis of Subpopulations 289
Electrochemistry in Cardiomyocytes 345
Interacting Populations 292
12.8 Biophysical Description of
10.3 General Modeling Strategy 292 Electrochemical Processes at the
10.4 Phase-Plane Analysis 292 Membrane of Cardiomyocytes 347
10.5 More Complex Models of 12.9 Resting Potentials and Action
Population Dynamics 297 Potentials 348
Exercises 299 12.10 Models of Action Potentials 350
References 301 12.11 Repeated Heartbeats 354
Further reading 302 Issues of a Failing Heart 355
contents xi
12.12 Modeling Heart Function and Failure 14.3 Design Principles 404
Based on Molecular Events 356 14.4 Operating Principles 406
Outlook for Physiological Multiscale Goal-Oriented Manipulations and Synthetic
Modeling 361 Design of Biological Systems 407
Exercises 362 14.5 Metabolic Engineering 407
References 365 14.6 Synthetic Biology 408
Further Reading 366 Case Studies of Synthetic Biological
Systems Designs 411
chapter 13: systems biology in Medicine 14.7 Elementary Mode Analysis in
and Drug Development 369 Metabolic Engineering 411
Are you Unique? 369 14.8 Drug Development 414
13.1 Biological Variability and Disease 369 14.9 Gene Circuits 415
13.2 Modeling Variability and Disease 370 The Future Has Begun 419
Personalized Medicine and Predictive Health 372 Exercises 419
13.3 Data Needs and Biomarkers 373 References 421
13.4 Personalizing Mathematical Models 374 Further Reading 423
The Drug Development Process 378
The Role of Systems Biology in chapter 15: emerging topics in
Drug Development 380 systems biology 425
13.5 Computational Target and Lead Emerging Applications 426
Identification 381 15.1 From Neurons to Brains 426
13.6 Receptor Dynamics 382 15.2 Complex Diseases, Inflammation,
13.7 Pharmacokinetic Modeling 385 and Trauma 428
13.8 Pathway Screening with Dynamic 15.3 Organisms and their Interactions
Models 390 with the Environment 432
13.9 Emerging Roles of Systems Biology Modeling Needs 435
in Drug Development 393 15.4 Multiscale Modeling 436
Exercises 394 15.5 A Data-Modeling Pipeline 437
References 395 Toward a Theory of Biology . . . or Several
Further Reading 396 Theories? 439
References 441
chapter 14: Design of biological systems 399 Further Reading 443
Natural Design of Biological Systems 400
14.1 The Search for Structural Patterns 400 Glossary 445
14.2 Network Motifs 402 Index 459
Biological Systems
1
When you have read this chapter, you should be able to:
• Describe the generic features of biological systems
• Explain the goals of systems biology
• Identify the complementary roles of reductionism and systems biology
• List those challenges of systems biology that cannot be solved with intuition
alone
• Assemble a “to-do” list for the field of systems biology
When we think of biological systems, our minds may immediately wander to the
Amazon rainforest, brimming with thousands of plants and animals that live with
each other, compete with each other, and depend on each other. We might think of
the incredible expanse of the world’s oceans, of colorful fish swimming through
coral reefs, nibbling on algae. Two-meter-high African termite mounds may come
to mind, with their huge colonies of individuals that have their specific roles and
whose lives are controlled by an intricate social structure (Figure 1.1). We may think
of an algae-covered pond with tadpoles and minnows that are about to restart yet
another life cycle.
These examples are indeed beautiful manifestations of some of the fascinating
systems nature has evolved. However, we don’t have to look that far to find biologi-
cal systems. Much, much smaller systems are in our own bodies and even within our
cells. Kidneys are waste-disposal systems. Mitochondria are energy-production sys-
tems. Ribosomes are intracellular machines that make proteins from amino acids.
Bacteria are amazingly complicated biological systems. Viruses interact with cells in
a well-controlled, systemic way. Even seemingly modest tasks often involve an
amazingly large number of processes that form complicated control systems
(Figure 1.2). The more we learn about the most basic processes of life, such as cell
division or the production of a metabolite, the more we have to marvel the incredi-
ble complexity of the systems that facilitate these processes. In our daily lives, we
usually take these systems for granted and assume that they function adequately,
and it is only when, for example, disease strikes or algal blooms kill fish that we
realize how complex biology really is and how damaging the failure of just a single
component can be.
We and our ancestors have been aware of biological systems since the beginning
of human existence. Human birth, development, health, disease, and death have
long been recognized as interwoven with those of plants and animals, and with the
environment. For our forebears, securing food required an understanding of sea-
sonal changes in the ecological systems of their surroundings. Even the earliest for-
ays into agriculture depended on detailed concepts and ideas of when and what to
2 Chapter 1: Biological Systems
Figure 1.1 Biological systems abound at

all size scales. Here, a termite mound in
Namibia is visible evidence of a complex
social system. This system is part of a larger
ecological system, and it is at once the host
to many systems at smaller scales. (Courtesy
of Lothar Herzog under the Creative
Commons Attribution 2.0 Generic license.)
plant, how and where to plant it, how many seeds to eat or to save for sowing, and
when to expect returns on the investment. Several thousand years ago, the Egyp-
tians managed to ferment sugars to alcohol and used the mash to bake bread. Early
pharmaceutical treatments of diseases certainly contained a good dose of supersti-
tion, and we are no longer convinced that rubbing on the spit of a toad during full
moon will cure warts, but the beginnings of pharmaceutical science in antiquity and
the Middle Ages also demonstrate a growing recognition that particular plant prod-
ucts can have significant and specific effects on the well-being or malfunctioning of
the systems within the human body.
In spite of our long history of dealing with biological systems, our mastery of
engineered systems far outstrips our capability to manipulate biological systems.
We send spaceships successfully to faraway places and predict correctly when they
will arrive and where they will land. We build skyscrapers exceeding by hundreds of
ABA PEPC
RCN1
Sph SphK Malate

NOS Arg NlA12 Nitrite NADPH
S1P OST1
NO
GCR1 GPA1 AGB1
PLC PIP2 NAD+ ADPRc GTP GC InsPK Figure 1.2 Diagram of a complicated
PLD PC NADPH Atrboh
system of molecules that coordinate the
DAG InsP3 cADPR cGMP InsP6 RAC1 PA ROS response of plants to drought. While the
details are not important here, we can see
that a key hormone, called abscisic acid
CIS ABH1
ROP2 (ABA), triggers a cascade of reactions that
Actin ABI1 pHc ultimately promote the closure of stomata
and thereby reduce water evaporation [1].
ROP10 ERA1 CalM H+ ATPase
Even a narrowly defined response like this
Ca2+ ATPase closure process involves a complicated
Ca2+c KEV Depolar
control system that contains a multitude of
molecules and their interactions. In turn, this
AnionEM system is just one component within a much
KAP KOUT
larger, physiological stress response system
(cf. Figure 1.7). (From Saadatpour A, Albert I
& Albert A. J. Theor. Biol. 266 [2010] 641–656.
AtPP2C closure
With permission from Elsevier.)
BIOLOGICAL SYSTEMS 3
times the sizes of the biggest animals and plants. Our airplanes are faster, bigger,
and more robust against turbulence than the most skillful birds. Yet, we cannot cre-
ate new human cells or tissues from basic building blocks and we are seldom able to
cure diseases except with rather primitive methods like cutting into the body or kill-
ing a lot of healthy tissue in the process, hoping that the body will heal itself after-
wards. We can anticipate that our grandchildren will only shake their heads at such
medieval-sounding, draconian measures. We have learned to create improved
microorganisms, for instance for the bulk production of industrial alcohol or the
generation of pure amino acids, but the methods for doing so rely on bacterial
machinery that we do not fully understand and on artificially induced random
mutations rather than targeted design strategies.
Before we discuss the roots of the many challenges associated with understand-
ing and manipulating biological systems in a targeted fashion, and our problems
predicting what biological systems will do under yet-untested conditions, we should
ask whether the goal of a deeper understanding of biological systems is even worth
the effort. The answer is a resounding “Yes!” In fact, it is impossible even to imagine
the potential and scope of advances that might develop from biological systems
analyses. Just as nobody during the eighteenth century could foresee the ramifica-
tions of the Industrial Revolution or of electricity, the Biological Revolution will
usher in an entirely new world with incredible possibilities. Applications that are
already emerging on the horizon are personalized medical treatments with minimal
side effects, pills that will let the body regain control over a tumor that has run amok,
prevention and treatment of neurodegenerative diseases, and the creation of spare
organs from reprogrammed stem cells. A better understanding of ecological systems
will yield pest- and drought-resistant food sources, as well as means for restoring
polluted soil and water. It will help us understand why certain species are threat-
ened and what could be done effectively to counteract their decline. Deeper insights
into aquatic systems will lead to cleaner water and sustainable fisheries. Repro-
grammed microbes or nonliving systems composed of biological components will
dominate the production of chemical compounds from prescription drugs to large-
scale industrial organics, and might create energy sources without equal. Modified
viruses will become standard means for supplying cells with healthy proteins or
replacement genes. The rewards of discovering and characterizing the general prin-
ciples and the specifics of biological systems will truly be unlimited.
If it is possible to engineer very sophisticated machines and to predict exactly
what they will do, why are biological systems so different and difficult? One crucial
difference is that we have full control over engineered systems, but not over biologi-
cal systems. As a society, we collectively know all details of all parts of engineered
machines, because we made them. We know their properties and functions, and we
can explain how and why some engineer put a machine together in a particular
fashion. Furthermore, most engineered systems are modular, with each module
being designed for a unique, specific task. While these modules interact with each
other, they seldom have multiple roles in different parts of the system, in contrast to
biology and medicine, where, for instance, the same lipids can be components of
membranes and have complicated signaling functions, and where diseases are
often not restricted to a single organ or tissue, but may affect the immune system
and lead to changes in blood pressure and blood chemistry that secondarily cause
kidney and heart problems. A chemical refinery looks overwhelmingly complicated
to a layperson, but for an industrial engineer, every piece has a specific, well-defined
role within the refinery, and every piece or module has properties that were opti-
mized for this role. Moreover, should something go wrong, the machines and facto-
ries will have been equipped with sensors and warning signals pinpointing problems
as soon as they arise and allowing corrective action.
In contrast to dealing with sophisticated, well-characterized engineered sys-
tems, the analysis of biological systems requires investigations in the opposite direc-
tion. This type of investigation resembles the task of looking at an unknown machine
and predicting what it does (Figure 1.3). Adding to this challenge, all scientists col-
lectively know only a fraction of the components of biological systems, and the spe-
cific roles and interactions between these components are often obscure and
change over time. Even more than engineered systems, biological systems are full of
sensors and signals that indicate smooth running or ensuing problems, but in most
Figure 1.3 Analyzing a biological system

resembles the task of determining the
function of a complicated machine
that we have never seen before. Shown
here as an example is the cesium fountain
laser table of the United States Naval
Observatory, which is used to measure time
with extreme accuracy. This atomic clock is
based on transitions in cesium, which have a
frequency of 9,192,631,770 Hz and are used
to define the second. See also [2].
cases our experiments cannot directly perceive and measure these signals and we
can only indirectly deduce their existence and function. We observe organisms,
cells, or intracellular structures as if from a large distance and must deduce from
rather coarse observations how they might function or why they fail.
What exactly is it that makes biological systems so difficult to grasp? It is cer-
tainly not just size. Figure 1.4 shows two networks. One shows the vast highway
system of the continental United States, which covers several million miles of major
(A)
Figure 1.4 The size of a network or

system is not necessarily correlated
with its complexity. (A) The network of
major highways in the continental United
States covers over 3 million square miles.
Nonetheless, its functionality is easy to
grasp, and problems with a particular road
are readily ameliorated with detours.
(B) The web of the European diadem spider
(Araneus diadematus) (C) is comparatively
small, but the functional details of this little
network are complex. Some lines are made
of silk proteins that have the tensile strength
of steel but can also be eaten and recycled
by the spider; other lines are adhesive due
to a multipurpose glue that may be sticky
(B) (C) or rubbery depending on the situation;
yet others are guide and signal lines that
allow the spider to move about and sense
prey. The creation of the web depends on
different types of spinneret glands, whose
development and function require the
complex molecular machinery of the spider,
and it is not yet clear how the instructions for
the complicated construction, repair, and use
of the web are encoded and inherited from
one generation to the next. ((A) From the
United States Department of Transportation.)
REduCTIOnISM And SYSTEMS BIOLOGY 5
(A) Figure 1.5 Biological phenomena are often

difficult to understand, because our minds
are trained to think linearly. (A) The return
on an investment grows (or decreases) linearly
with the amount invested. (B) In biology, more
is not necessarily better. Biological responses
often scale within a modest range, but lead
to an entirely different response if the input is
× 100 increased a lot.
$100 investment $120 return $10,000 investment $12,000 return
(B)
× 100
1 tablespoon of fertilizer 50 blossoms 100 tablespoons of fertilizer dead roses!
highways. It is a very large system, but it is not difficult to understand its function or
malfunction: if a highway is blocked, it does not take much ingenuity to figure out
how to circumvent the obstacle. The other network is a comparably tiny system: the
web of a diadem spider. While we can observe the process and pattern with which
Ms. Spider spins her web, we do not know which neurons in her brain are respon-
sible for different phases of the complicated web production process and how she
is able to produce the right chemicals for the spider silk, which in itself is a marvel
of material science, let alone how she manages to survive, multiply, and maybe
even devour her husband.
Biological systems often consist of large numbers of components, but they pose
an additional, formidable challenge to any analysis, because the processes that
govern them are not linear. This is a problem, because we are trained to think in
linear ways: if an investment of $100 leads to a return of $120, then an investment of
$10,000 leads to a return of $12,000. Biology is different. If we fertilize our roses with
1 tablespoon of fertilizer and the rose bushes produce 50 blossoms, a little bit more
fertilizer may increase the number of blossoms, but 100 tablespoons of fertilizer will
not produce 5000 blossoms but almost certainly kill the plants (Figure 1.5). Just a
small amount of additional sun exposure turns a tan into sunburn. Now imagine
that thousands of components, many of which we do not know, respond in such a
fashion, where a small input does not evoke any response, more input evokes a
physiological response, and a little bit more input causes the component to fail or
exhibit a totally different “stress” response. We will return to this issue later in this
and other chapters with specific examples.
REduCTIOnISM And SYSTEMS BIOLOGY

So the situation is complicated. But because we humans are a curious species, our
forebears did not give up on biological analysis and instead did what was doable,
namely collecting information on whatever could be measured with the best current
methods (Figure 1.6). By now, this long-term effort has resulted in an amazing list of
biological parts and their roles. Initially, this list contained new plant and animal
Figure 1.6 Collecting information is the

first step in most systems analyses. The
eighteenth-century British explorer Captain
James Cook sailed the Pacific Ocean and
catalogued many plants and animal species
that had never been seen before in Europe.
species, along with descriptions of their leaves, berries, and roots, or their body
shapes, legs, and color patterns. These external descriptions were valuable, but did
not provide specific clues on how plants and animals function, why they live, and
why they die. Thus, the next logical step was to look inside—even if this required
stealing bodies from the cemetery under a full moon! Cutting bodies open revealed
an entirely new research frontier. What were all those distinct body parts and what
did they do? What were organs, muscles, and tendons composed of? Not surpris-
ingly, this line of investigation eventually led to the grand-challenge quest of discov-
ering and measuring all parts of a body, the parts of the parts (. . . of the parts), as well
as their roles in the normal physiology or pathology of cells, organs, and organisms.
The implicit assumption of this reductionist approach was that knowing the building
blocks of life would lead us to a comprehensive understanding of how life works.
If we fast-forward to the twenty-first century, have we succeeded and assembled
a complete parts catalog? Do we know the building blocks of life? The answer is a
combination of yes’s and no’s. The catalog is most certainly not complete, even for
relatively simple organisms. Yet, we have discovered and characterized genes, pro-
teins, and metabolites as the major building blocks. Scientists were jubilant when
the sequencing of the human genome in the early years of this millennium was
declared complete: we had identified the ultimate building blocks, our entire blue-
print. It turned out to consist of roughly three billion nucleotide pairs of DNA.
The sequencing of the human genome was without any doubt an incredible
achievement. Alas, there is much more to a human body than genes. So, the race for
building blocks extended to proteins and metabolites, toward individual gene varia-
tions and an assortment of molecules and processes affecting gene expression,
which changes in response to external and internal stimuli, during the day, and
throughout our lifetimes. As a direct consequence of these ongoing efforts, our parts
list continues to grow at a rapid pace: A parts catalog that started with a few organs
now contains over 20,000 human genes, many more genes from other organisms,
and hundreds of thousands of proteins and metabolites along with their variants. In
addition to merely looking at parts in isolation, we have begun to realize that most
biological components are affected and regulated by a variety of other components.
The expression of a gene may depend on several transcription factors, metabolites,
and a variety of small RNAs, as well as molecular, epigenetic attachments to its DNA
sequence. It is reasonable to expect that the list of processes within the body is much
larger than the number of components on our parts list. Biologists will not have to
worry about job security any time soon!
The large number of components and processes alone, however, is not the
only obstacle to understanding how cells and organisms function. After all, modern
computers can execute gazillions of operations within a second. Our billions of
telephones worldwide are functionally connected. We can make very accurate
REduCTIOnISM And SYSTEMS BIOLOGY 7
predictions regarding a gas in a container, even if trillions of molecules are involved.

If we increase the pressure on the gas without changing the volume of the container,
we know that the temperature will rise, and we can predict by how much. Not so with
a cell or organism. What will happen to it if the environmental temperature goes up?
Nothing much may happen, the rise in temperature may trigger a host of physiologi-
cal response processes that compensate for the new conditions, or the organism may
die. The outcome depends on a variety of factors that collectively constitute a com-
plex stress response system (Figure 1.7). Of course, the comparison to a gas is not
H2O O2 ABIOTIC STRESS AND BIOTIC RESPONSES

PHOTOSYNTHESIS SA signaling JA signaling ET signaling
CO2
stomata SA JA ET
ETR1
NPR1 JAZ
EIN2
light ABA
photosynthesis WRKY MYC2 ERF
MVB DREB SA-responsive JA-responsive ET-responsive

genes genes genes
defence
Glu HXK1
AUX GA
ABA
CK
ethylene
pollinator
H2O CO
2
O2
light
VOC
ORGAN AND PLANT GROWTH
temperature
pathogens
AUX CK
nutrients, microbes CYCB

minerals KRP
water M
CDK
KEY:
transcription factors hormones CYCD

G2 G1 ANT
kinases carbohydrates CDKA
receptors enzymes S
CYCA DEL E2F RBR
other signaling proteins activation DP
CDKB CELL CYCLE
environmental interactions suppression
Figure 1.7 Stress responses are coordinated by systems at different levels of organization (cf. Figure 1.2). At the physiological level, the stress
response system in plants includes changes at the cellular, organ, and whole-plant levels and also affects interactions of the plant with other species.
(From Keurentjes JJB, Angenent GC, Dicke M, et al. Trends Plant Sci. 16 [2011] 183–190. With permission from Elsevier.)
quite fair, because, in addition to their large number, the components of a cell are not
all the same, which drastically complicates matters. Furthermore, as mentioned ear-
lier, the processes with which the components interact are nonlinear, and this per-
mits an enormous repertoire of distinctly different behaviors with which an organism
can respond to a perturbation.
EVEn SIMPLE SYSTEMS CAn COnFuSE uS

It is easy to demonstrate how quickly our intuition can be overwhelmed by a few E
nonlinearities within a system. As an example, let’s look at a simple chain of processes
and compare it with a slightly more complicated chain that includes regulation [3].
The simple case merely consists of a chain of reactions, which is fed by an external Input X Y Z
input (Figure 1.8). It does not really matter what X, Y, and Z represent, but, for the
sake of discussion, imagine a metabolic pathway such as glycolysis, where the input,
glucose, is converted into glucose 6-phosphate, fructose 1,6-bisphosphate, and pyru- Figure 1.8 The human brain handles
vate, which is used for other purposes that are not of interest here. For illustrative linear chains of causes and events very
well. In this simple pathway, an external
purposes, let’s explicitly account for an enzyme E that catalyzes the conversion of X input is converted sequentially into X, Y, and
into Y. Z, which leaves the system. The conversion of
We will learn in the following chapters how one can formulate a model of such a X into Y is catalyzed by an enzyme E. It is easy
pathway system as a set of differential equations. And while the details are not to imagine that any increase in Input will
important here, it does not hurt to show such a model, which might read cause the levels of X, Y, and Z to rise.
X = Input − aEX 0.5 ,

Y = aEX 0.5 − bY 0.5 , (1.1)
Z = bY 0.5 − cZ 0.5 .
Here, X, Y, and Z are concentrations, E is the enzyme activity, and a, b, and c are rate 1.0
constants that respectively represent how fast X is converted into Y, how fast Y is concentration X Z
converted into Z, and how quickly material from the metabolite pool Z leaves the Y
0.5
system. The dotted quantities on the left of the equal signs are differentials that
describe the change in each variable over time, but we need not worry about them
at this point. In fact, we hardly have to analyze these equations mathematically to 0
0 15 30
get an idea of what will happen if we change the input, because intuition tells us that time
any increase in Input should lead to a corresponding rise in the concentrations of
the intermediates X, Y, and Z, whereas a decrease in Input should result in smaller Figure 1.9 Simulations with the system
values of X, Y, and Z. The increases or decreases in X, Y, and Z will not necessarily be in (1.1) confirm our intuition: X, Y, and
exactly of the same extent as the change in Input, but the direction of the change Z reflect changes in Input. For instance,
should be the same. The mathematical solution of the system in (1.1) confirms this reducing Input in (1.1) to 75% at time
intuition. For instance, if we reduce Input from 1 to 0.75, the levels of X, Y, and Z 10 (arrow) leads to permanent decreases
decrease, one after another, from their initial value of 1 to 0.5625 (Figure 1.9). in X, Y, and Z.
Now suppose that Z is a signaling molecule, such as a hormone or a phospho-
lipid, that activates a transcription factor TF that facilitates the up-regulation of a
gene G that codes for the enzyme catalyzing the conversion of X into Y (Figure 1.10).
The simple linear pathway is now part of a functional loop. The organization of this
loop is easy to grasp, but what is its effect? Intuition might lead us to believe that the
positive-feedback loop should increase the level of enzyme E, which would result in E G TF
more Y, more Z, and even more E, which would result in even more Y and Z. Would
the concentrations in the system grow without end? Can we be sure about this pre-
diction? Would an unending expansion be reasonable? What will happen if we Input X Y Z
increase or decrease the input as before?
The overall answer will be surprising: the information given so far does not allow
us to predict particular responses with any degree of reliability. Instead, the answer Figure 1.10 Even simple systems may
depends on the numerical specifications of the system. This is bad news for the not allow us to make reliable predictions
unaided human mind, because we are simply not able to assess the numerical con- regarding their responses to stimuli.
Here, the linear pathway from Figure 1.8 is
sequences of slight changes in a system, even if we can easily grasp the logic of a
embedded into a functional loop consisting
system as in Figure 1.10. of a transcription factor TF and a gene G that
To get a feel for the system, one can compute a few examples with an expanded codes for enzyme E. As described in the text,
model that accounts for the new variables (for details, see [3]). Here, the results are the responses to changes in Input are no
more important than the technical details. If the effect of Z on TF is weak, the longer obvious.
EVEn SIMPLE SYSTEMS CAn COnFuSE uS 9
response to a decrease in Input is essentially the same as in Figure 1.9. This is not too
surprising, because the systems in this case are very similar. However, if the effect of
Z on TF is stronger, the concentrations in the system start to oscillate, and after a
while these oscillations dampen away (Figure 1.11A). This behavior was not easy to
predict. Interestingly, if the effect is further increased, the system enters a stable
oscillation pattern that does not cease unless the system input is changed again
(Figure 1.11B).
The hand-waving explanation of these results is that the increased enzyme activ-
ity leads to a depletion of X. A reduced level of X leads to lower levels of Y and Z,
which in turn lead to a reduced effect on TF, G, and ultimately E. Depending on the
numerical characteristics, the ups and downs in X may not be noticeable, they may
be damped and disappear, or they may persist until another change is introduced.
Intriguingly, even if we know that these alternative responses are possible, the
unaided human mind is not equipped to integrate the numerical features of the
model in such a way that we can predict which system response will ensue for a
specific setting of parameters. A computational model, in contrast, reveals the
answer in a fraction of a second.
The specific details of the example are not as important as the take-home mes-
sage: If a system contains regulatory signals that form functional loops, we can no
longer rely on our intuition for making reliable predictions. Alas, essentially all real-
istic systems in biology are regulated—and not just with one, but with many control
loops. This leads to the direct and sobering deduction that intuition is not sufficient
and that we instead need to utilize computational models to figure out how even
small systems work and why they might show distinctly different responses or even
fail, depending on the conditions under which they operate.
The previous sections have taught us that biological systems contain large num-
bers of different types of components that interact in potentially complicated ways
and are controlled by regulatory signals. What else is special about biological sys-
tems? Many answers could be given, some of which are discussed throughout this
book. For instance, two biological components are seldom 100% the same. They vary
from one organism to the next and change over time. Sometimes these variations are
inconsequential, at other times they lead to early aging and disease. In fact, most
(A)
2
X
concentration
TF, E, G
Z
Y
0
0 50 100
time
(B)
5.0
Figure 1.11 Simulation results

X demonstrate that the looped system
in Figure 1.10 may exhibit drastically
concentration
different responses. If the effect of Z on TF

2.5 is very small, the response is essentially like
that in Figure 1.9 (results not shown). (A) If
Y, Z, the effect of Z on TF is relatively small, the
TF, E, functional feedback loop causes the system
G
to go through damped oscillations before
0 assuming a new stable state. (B) For stronger
0 250 500 2000 2250 2500 effects of Z on TF, the system response is a
time persistent oscillation.
diseases do not have a single cause, but are the consequence of an unfortunate com-
bination of slight alterations in many components. Another feature that complicates
intuition is the delay in many responses to stimuli. Such delays may be of the order of
seconds, hours, or years, but they require the analyst to study not merely the present
state of a biological system but also its history. For instance, recovery from a severe
infection depends greatly on the preconditioning of the organism, which is the col-
lective result of earlier infections and the body’s responses [4].
Finally, it should be mentioned that different parts of biological systems may
simultaneously operate at different scales, with respect to both time and space.
These scales make some aspects of their analysis easier and some harder. Let’s begin
with the temporal scale. We know that biology at the most basic level is governed by
physical and chemical processes. These occur on timescales of the order of millisec-
onds, if not faster. Biochemical processes usually run on a scale of seconds to min-
utes. Under favorable conditions, bacteria divide every 20–30 minutes. Our human
lifespan extends to maybe 120 years, evolution can happen at the genetic level with
lightning speed, for instance, when radiation causes a mutation, while the emer-
gence of an entirely new species may take thousands or even millions of years. On
one hand, the drastically different timescales make analyses complicated, because
we simply cannot account for rapid changes in all molecules of an organism over an
extended period of time. As an example, it is impossible to study aging by monitor-
ing an organism’s molecular state every second or minute. On the other hand, the
differences in timescales justify a very valuable modeling “trick” [5, Chapter 5]. If we
are interested in understanding some biochemical process, such as the generation
of energy in the form of adenosine triphosphate (ATP) by means of the conversion
of glucose into pyruvate, we can assume that developmental and evolutionary
changes are so slow in comparison that they do not change during ATP production.
Similarly, if we study the phylogenetic family tree of species, the biochemical pro-
cesses in an individual organism are comparatively so fast that their details become
irrelevant. Thus, by focusing on just the most relevant timescale and ignoring much
faster and much slower processes, any modeling effort is dramatically simplified.
Biology also happens on many spatial scales. All processes have a molecular
component, and their size scale is therefore of the order of ångströms and nanome-
ters. If we consider a cell as the basic unit of life, we are dealing with a spatial scale
of micrometers to millimeters, with some exceptions such as cotton “fiber” cells
reaching the length of a few centimeters [6] and the afferent axons of nerve cells in
giraffes, reaching from toe to neck, extending to 5 meters [7, p. 14]. The sizes of typi-
cal cells are dwarfed by higher plants and animals and by ecosystems such as our
oceans, which may cover thousands of square kilometers. As with the different tem-
poral scales, and using analogous arguments, models of biological systems often
focus on one or two spatial scales at a time [5]. Nonetheless, such simplifications are
not always applicable, and some processes, such as aging and algal blooms, may
require the simultaneous consideration of several temporal and spatial scales. Such
multiscale assessments are often very complicated and constitute a challenging
frontier of current research (see Chapter 15).
WHY nOW?
Many of the features of biological systems have been known for quite a while, and,
similarly, many concepts and methods of systems biology have their roots in its
well-established parent disciplines, including physiology, molecular biology, bio-
chemistry, mathematics, engineering, and computer science [8–11]. In fact, it has
been suggested that the nineteenth-century scientist Claude Bernard might be con-
sidered the first systems biologist, since he proclaimed that the “application of
mathematics to natural phenomena is the aim of all science, because the expression
of the laws of phenomena should always be mathematical” [12, 13]. A century later,
Ludwig von Bertalanffy reviewed in a book his three decades of attempting to con-
vince biologists of the systemic nature of living organisms [14, 15]. At the same time,
Mihajlo Mesarović used the term “Systems Biology” and declared that “real
advance . . . will come about only when biologists start asking questions which are
based on systems-theoretic concepts” [16]. The same year, a book review in Science
WHY nOW? 11
envisioned “. . . a field of systems biology with its own identity and in its own right”
[17]. A few years later, Michael Savageau proposed an agenda for studying biologi-
cal systems with mathematical and computational means [5].
In spite of these efforts, systems biology did not enter the mainstream for several
more decades. Biology kept its distance from mathematics, computer science, and
engineering, primarily because biological phenomena were seen as too complicated
for rigorous mathematical analysis and mathematics was considered applicable only
to very small systems of little biological relevance. The engineering of biological sys-
tems from scratch was impossible, and the budding field of computer science con-
tributed to biology not much more than rudimentary data management.
So, why has systems biology all of the sudden moved to the fore? Any good detec-
tive will know the answer: motive and opportunity. The motive lies in the realization
that reductionist thinking and experimentation alone are not sufficient if complex
systems are involved. Reductionist experiments are very good in generating detailed
information regarding specific components or processes of a system, but they often
lack the ability to characterize, explain, or predict emergent properties that cannot
be found in the parts of the system but only in their web of interactions. For instance,
the emergence of oscillations in the example system represented by the equations
in (1.1) cannot be credited to a single component of the system but is a function of
its overall organization. Although we had complete knowledge of all details of the
model pathway, it was very difficult to foresee its capacity either to saturate or oscil-
late in a damped or stable fashion. Biology is full of such examples.
A few years ago, Hirotada Mori’s laboratory completed the assembly of a com-
plete catalogue of single mutants in the bacterium Escherichia coli [18]. Yet, the
scientific community is still not able to foresee which genes the bacterium will up-
or down-regulate in response to new environmental conditions. Another very chal-
lenging example of emergent system properties is the central nervous system. Even
though we understand quite well how action potentials are generated and propa-
gated in individual neurons, we do not know how information flows, how memory
works, and how diseases affect the normal functioning of the brain. It is not even
clear how information in the brain is represented (see also Chapter 15). Thus, while
reductionist biology has been extremely successful and will without any doubt
continue to be the major driving force for future discovery, many biologists have
come to recognize that the detailed pieces of information resulting from this
approach need to be complemented with new methods of system integration and
reconstruction [19].
The opportunity for systems biology is the result of the recent confluence and
synergism of three scientific frontiers. The first is of course the rapid and vast accu-
mulation of detailed biological information at the physiological, cellular, molecular,
and submolecular levels. These targeted investigations of specific phenomena are
accompanied by large-scale, high-throughput studies that were entirely infeasible
just a couple of decades ago. They include quantification of genome-wide expres-
sion patterns, simultaneous identification of large arrays of expressed proteins,
comprehensive profiling of cellular metabolites, characterization of networks of
molecular interactions, global assessments of immune systems, and functional
scans of nervous systems and the human brain. These exciting techniques are gen-
erating unprecedented amounts of high-quality data that are awaiting systemic
interpretation and integration (Figure 1.12).
The second frontier is the result of ingenuity and innovation in engineering,
chemistry, and material sciences, which have begun to provide us with a growing
array of technologies for probing, sensing, imaging, and measuring biological sys-
tems that are at once very detailed, extremely specific, and usable in vivo. Many
tools supporting these methods are in the process of being miniaturized, in some
cases down to the nanoscale of molecules, which allows diagnoses with minute
amounts of biological materials and one day maybe biopsies of individual, living
cells. Devices at this scale will allow the insertion of sensing and disease treatment
devices into the human body in an essentially noninvasive and harmless fashion
[20–22]. Bioengineering and robotics are beginning to render it possible to measure
hundreds or thousands of biomarkers from a single drop of blood. It is even becom-
ing feasible to use molecular structures, prefabricated by nature, for new purposes
in medicine, drug delivery, and biotechnology (Figure 1.13).
short-day long-day Figure 1.12 Modern high-throughput

methods of molecular biology offer data
1st 2nd 1st 2nd
in unprecedented quantity and quality.
0 4 8 12 16 20 0 4 8 12 16 20 0 4 8 12 16 20 0 4 8 12 16 20 As an example, the heat map shown here
represents a genome-wide expression
profile of 24-hour-rhythmic genes in the
mouse under chronic short-day (left two
panels) and long-day (right two panels)
conditions. (From Masumoto KM, Ukai-
Tadenuma M, Kasukawa T, et al. Curr. Biol.
20 [2010] 2199–2206. With permission from
Elsevier.)
24 h rhythmic genes
–2SD 0 +2SD
The third frontier is the co-evolution of mathematical, physical, and computa-

tional techniques that are more powerful and accessible to a much wider audience
than ever before. Imagine that only a few decades ago computer scientists used
punch cards that were read by optical card readers (Figure 1.14)! Now, there are
even specific computing environments, including Mathematica• and MATLAB•, as
well as different types of customized mark-up languages (XML), such as the systems
biology mark-up language SBML [23] and the mark-up language AGML, which was
developed specifically for analyzing two-dimensional gels in proteomics [24].
Before today’s much more effective computer science techniques were avail-
able, it was not even possible to keep track of the many components of biological
systems, let alone analyze them. But over the past few decades, a solid theoretical
and numerical foundation has been established for computational methods specifi-
cally tailored for the investigation of dynamic and adaptive systems in biology and
medicine. These techniques are now at the verge of making it possible to represent
and analyze large, organizationally complex systems and to study their emergent
properties in a rigorous fashion. Methods of machine learning, numerical mathe-
matics, and bioinformatics permit the efficient mining and analysis of the most use-
ful data from within an overwhelming amount of data that are not pertinent for
a given task. Algorithmic advances permit the simulation and optimization of
very large biological flux distribution networks. Computer-aided approximation
approaches yield ever-finer insights into the dynamics of complex nonlinear sys-
tems, such as the control of blood flow in healthy and diseased hearts. New mathe-
matical, physical, and computational methods are beginning to make it possible to
COMMunICATInG SYSTEMS BIOLOGY 13
Figure 1.13 “Protein cages” are

particles that have applications in
bionanotechnology and nanomedicine.
These particles are very interesting biological
heat-shock lumazine tobacco mosaic virus
ferritin building blocks because they self-assemble
protein cage synthase complex 20S intermediate
12–14 nm into a variety of different shapes. The
12–14 nm 16 nm (18 nm × 4 nm) brome mosaic virus 28 nm
features of these bionanoparticles can be
genetically manipulated and fine-tuned for
biomedical purposes, such as drug delivery,
gene therapy, tumor imaging, and vaccine
development. (From Lee LA & Wang Q.
Nanomedicine 2 [2006] 137–149. With
adenoassociated cowpea chlorotic cucumber permission from Elsevier.)
MS2 phage virus-2 mosaic virus mosaic virus cowpea mosaic virus
29 nm 29 nm 29 nm 30 nm 32 nm
turnip yellow papillomavirus human hepatitis murine

mosaic virus L1 capsid B virus polyomavirus
28 nm 27 nm 31 nm 48.6 nm
Figure 1.14 Advances in computer power,

accessibility, and user-friendliness over
the past 40 years have been tremendous.
Not too long ago, computer code had to be
fed manually into the computer with punch
cards. (Courtesy of Mutatis mutandis under
the Creative Commons Attribution-Share
Alike 3.0 Unported license.)
predict the folding of proteins and the binding between target sites and ligands.
These predictions, in turn, suggest insights into specific molecular interactions and
promise the potential of targeted drug interventions that minimize toxic side effects.
Motive and opportunity have met to make systems biology attractive and feasi-
ble. It has become evident that the relevant disciplines complement each other in
unique ways and that the synergism among them will revolutionize biology, medi-
cine, and a host of other fields, including biotechnology, environmental science,
food production, and drug development.
COMMunICATInG SYSTEMS BIOLOGY

It is not a trivial task to talk succinctly about 25,000 genes and their expression state
or about the many processes occurring simultaneously in response to a signal that a
cell receives at its outer surface. Our minds are ill equipped to characterize numeri-
cal relationships, let alone discuss complicated mathematical functions, especially
if these depend on many variables. If we did not have numbers, we would even have
problems describing everyday features such as temperature. Of course, we can say
that it is cold or hot, and we have a dozen or so adjectives in between. But if we need
more accuracy, common language ceases to be sufficient. Discerning 37.0°C from
38.4°C is not easy without a numerical scale, but it is necessary to have the tools to
describe the difference, for instance, because the former reflects normal body tem-
perature, while the latter is a sign of fever.
We may willingly or grudgingly accept the fact that we need mathematics, which
comes with its own terminology, but communication is a two-way process. If we
start talking about eigenvalues and Hopf bifurcations, we are almost guaranteed to
lose mainstream biologists, let alone laypeople. This is a real problem, because our
results must be conveyed to biologists, who are providing us data, and to the public
that pays for our research and has a right to reap the fruit of the enormous invest-
ment of resources going into science [25]. The only true solution to this challenge is
the bilingual education and nurturing of systems biologists who can translate bio-
logical phenomena into math and computer code and who can explain what it really
means for the biological system if the real part of an eigenvalue is positive [19].
Communication is not trivial even within biology itself, because specialization
has progressed so far that different fields such as molecular biology, immunology,
and nanomedicine have developed their own terminology and jargon. Let’s look at
this issue in the form of a parable from Indian folklore that describes six blind men
exploring an elephant (Figure 1.15). This story is quite old and usually ends in utter
confusion, but it is useful to analyze it further than is usually done. The story has it
that each of the blind men touched a different part of an elephant and came to a dif-
ferent conclusion concerning the object of his research. The man touching the side
thought he was touching a wall, the one feeling the leg concluded he was touching a
tree. The elephant’s trunk gave the impression of a snake, the tusk that of a pointed
scimitar, the tail felt like a rope, and the ear appeared to be like a large leaf or fan. It
is not difficult to see the analogy to a complex biological system like the onset of
Alzheimer’s disease. The first scientist found “the Alzheimer gene,” the second dis-
covered “a strong association between the disease and former head injuries,”
another scientist detected “problems with fatty acid metabolism in the brain,” and
yet another suggested that “aluminum in cookware might be the culprit.” As in the
case of the elephant, the scientists were right, to some degree.
Let’s analyze the elephant story a little further. The first problem among the six
blind men might have been the homogeneous pool of researchers. Including a female
or a child might have provided additional clues. Also, we have to feel sorry for the
Indian men for being blind. However, they were apparently not mute or deaf, so that a
little discussion among them might have gone a long way. While all six were blind, it is
furthermore fair to assume that they had friends with working vision, who could have
set them straight. They could have used not just their hands but also their other senses,
Figure 1.15 Information about isolated

parts of a system alone does not always
reveal the true nature of the system. An
old story of six blind Indian men trying to
determine what they touch is a parable for
the dangers of scientific silos and the lack of
good communication.
COMMunICATInG SYSTEMS BIOLOGY 15
such as smell. Do tree trunks really smell like elephant feet? Finally, they apparently
stayed in their one spot, thereby greatly limiting their experience base.
It is again easy to translate these issues into biology, especially when we think of
purely reductionist strategies. Instead of a homogeneous pool of biologists analyzing
biological systems, it is without doubt more effective to have a multidisciplinary team
including different varieties of biologists, but also physicists, engineers, mathemati-
cians, chemists, and smart people trained in the liberal arts or economics. Instead of
only focusing on the one aspect right in front of our nose, communication with oth-
ers provides context for singular findings. We don’t know whether the Indian men
spoke the same language, but we know that even if biologists, computer scientists,
and physicists all use English to communicate, their technical languages and their
views of the scientific world are often very different, so that communication may ini-
tially be superficial and ineffective. That is where multidisciplinary groups must
engage in learning new terminologies and languages and include interdisciplinary
translators. Just as the Indian men should have called upon their seeing friends,
investigators need to call in experts who master techniques that have not been
applied to the biological problem at hand. Finally, it behooves the trunk analyzer to
take a few steps and touch the tusk and the side. Established scientific disciplines
have in the past often become silos. Sometimes without even knowing it, researchers
have kept themselves inside these silos, unable or unwilling to break out and to see
the many other silos around, as well as a whole lot of space between them.
Systems biology does not ask the six blind men to abandon their methods and
instead to run circles around the elephant. By focusing on one aspect, the reduc-
tionist “elephantologists” are poised to become true experts on their one chosen
body part and to know everything there is to know about it. Without these experts,
systems biology would have no data to work on. Instead, what systems biology sug-
gests is a widening of the mindset and at least rudimentary knowledge of a second
language, such as math. It also suggests the addition of other researchers, assisting
the “trunkologist” and the “tuskologist” by developing new tools of analysis, by tell-
ing them in their language what others have found, by closing the gap between
trunks and tusks and tails.
One strategy for accomplishing this synergism is to collect the diverse pieces of
data and contextual information obtained by the six blind men and to merge them
into a conceptual model. What kind of “thing” could consist of parts that feel like a
snake, tree trunks, large walls, two scimitars, two fans, and a rope? How could an aber-
rant gene, former head injuries, and altered brain metabolism functionally interact
to result in Alzheimer’s disease? Well-trained systems biologists should be able to
develop strategies for merging heterogeneous information into formal models that
permit the generation of testable hypotheses, such as “tree-trunk-like things are
connected by wall-like things.” These hypotheses may be wrong, but they can never-
theless be very valuable, because they focus the scientific process on new, specific
experiments that either confirm or refute the hypothesis. An experiment could be to
walk along the “wall” as far as possible. Is there a “tree trunk” on the end? Are there
“tree trunks” to the left and to the right? Is there a “pointed scimitar” at one or both
ends? Is the “snake” connected to a “wall” or to a “tree trunk”? Does the “wall” reach
the ground? Each answer to one of these questions constrains further and further
what the unknown “thing” could possibly look like, and this is the reason that
refuted hypotheses are often as valuable or even more valuable than confirmed
hypotheses. “The wall does indeed not reach the ground!” Then, how is it supported?
By “tree trunks”?
The story tells us that effective communication can solve a lot of complex ques-
tions. In systems biology, such communication is not always easy, and it requires
not only mastering the terminology of several parent disciplines but also internal-
izing the mindset of biologists and clinicians on the one hand and of mathemati-
cians, computer scientists, and engineers on the other. So, let’s learn about biology.
Let’s study laboratory data and information and explore the mindset of biologists.
Let’s study graphs and networks with methods from computer science. Let’s see
how mathematicians approach a biological system, struggle with assumptions,
make simplifications, and obtain solutions that are at first incomprehensible to the
non-mathematician but do have real meaning once they are translated into the lan-
guage of biology.
THE TASK BEFORE uS

We have discussed the need to understand biological systems. But what does that
really mean? Generically, it means that we should be able to explain how biological
systems work and why they are constructed in the fashion as we observe them and
not in a different one. Second, we should be able to make reliable predictions of
responses of biological systems under yet-untested conditions. And third, we should
be able to introduce targeted manipulations into biological systems that change
their responses to our specifications.
This level of understanding is a tall order, and we will need many years to achieve
it even for a narrowly focused domain within the huge realm of biological systems.
An important component of the task is the conversion of actual biological systems
into computational models, because this conversion, if it is valid, allows us almost
unlimited and comparatively cheap analyses. The resulting models of biological sys-
tems come in two types. The first focuses on specific systems and includes all perti-
nent functional and numerical details—one might think of the analogy to a flight
simulator. The second type of model is intended to help us understand the funda-
mental, generic features of the organization of biological systems; here one might
think of elementary geometry, which offers us valuable insights into spatial features
of the world by dealing with ideal triangles and circles that do not really exist in
nature. The two model types point to two opposite ends of a large spectrum. The
former models will be large and complex, while the latter will be as reduced and
simple as feasible. In practice, many models will fall between these two extremes.
To pave the way toward these goals, this book is organized in three sections. The
first of these introduces in four chapters a set of modeling tools for converting bio-
logical phenomena into mathematical and computational analog and for diagnos-
ing, refining, and analyzing them. The second section describes in five chapters the
molecular inventories that populate biological systems, and the five chapters in the
third section are devoted to representative case studies and a look into the future.
The modeling approaches parallel two fundamental properties of biological sys-
tems, namely their static structure and their dynamics, that is, their changes over time.
For static analyses, we will characterize and rationalize how nature put particular sys-
tems together and which parts are directly or loosely connected with each other. We
will see that there are distinct types of connections and interactions. One important
difference is that some connections allow the flow of material from a source to a target,
while others serve the sole purpose of signaling the state of the system. In the latter
case, no material changes location. Like a billboard that is not changed whether hun-
dreds of people look at it or nobody at all, a signaling component may not be changed
when it sends a signal. It is also to be expected that some connections are crucial, while
others are of secondary importance. Finally, there is the very challenging question of
how we can even determine the structure of a system. What types of data do we need to
infer the structure of a system, and how reliable is such an inference?
The dynamics of a system is of the utmost importance, because all biological
systems change over time. Organisms traverse a life cycle during which they undergo
tremendous changes. Even a lowly yeast cell fills up with scars where it has given
birth to daughter cells, and once its surface is full, the cell slides into the sunset of
life. We can easily see these changes under a microscope, but an incomparably
larger number of changes remain hidden from our sight. Gene expression patterns,
amounts of proteins, profiles of metabolites, all of these change dramatically
between birth and death. In addition to normal changes throughout its lifetime,
every organism responds to fast and slow changes in the environment and adapts
rather quickly to new situations. Today we may observe and characterize the gene
expression network in a bacterium, but tomorrow it may already have changed in
response to some environmental pressures. Indeed, bacteria evolve so quickly that
the commonly used term “wild type” no longer has much meaning. Even more than
static aspects, dynamic aspects of biological systems mandate the use of computa-
tional models. These models help us reveal how fascinating living systems are, with
respect both to their overall efficiency and to the ingenuity with which dynamic
responses and adaptations are coordinated.
Whether static or dynamic, some model designs and analyses will be performed
in the bottom-up and others in the top-down direction. However, since we seldom
REFEREnCES 17
start at the very bottom, that is, with individual atoms, or at the very top with models
of complete organisms as they interact with their environment, most modeling
strategies in systems biology are in truth “middle-out,” to use Nobel Laureate Syd-
ney Brenner’s expression (cited in [26]). They begin somewhere in between the
extremes, maybe with pathways or with cells. Over time, they may be incorporated
into larger models, or they may become more and more refined in detail.
The second section of the book addresses the molecular inventories of biological
systems. Paralleling the biological organization of organisms, one chapter is devoted
to gene systems, one to proteins, and one to metabolites. A further chapter discusses
signal transduction systems, and the final chapter of this section describes features
of populations. Clearly, all these chapters are woefully incomplete and should not
be thought of as substitutes for real biology books. Their purpose is merely to pro-
vide brief overviews of the main classes of biological components that are the foun-
dation of all modeling strategies.
The third section contains case studies that in one way or another highlight
aspects of biological systems that are in some sense representative. One chapter
describes the coordinated stress response system in yeast, which operates simulta-
neously at the gene, protein, and metabolite levels. Another chapter provides a pre-
sentation of how very different models can be useful to focus attention on selected
aspects of a multiscale system, the heart. A third chapter indicates how systems
biology can contribute to medicine and drug development. The fourth chapter illu-
minates aspects of the natural design of biological systems and of the artificial
design of synthetic biological systems. Finally, the last chapter discusses emerging
and future trends in systems biology.
EXERCISES
1.1. Search the Internet, as well as different dictionaries, Suppose further that the activating path reacts much
for definitions of a “system.” Extract commonalities faster than the inhibiting path. What would be the
among these definitions and formulate your own consequence with respect to the effect of an increase
definition. in input (ABA) on output (stomata closure)? How
1.2. Search the Internet for definitions of “systems could a difference in speed be implemented in a
biology.” Extract commonalities among these natural cell? Does it matter how strong or weak the
definitions and formulate your own definition. activation and inhibition are? Discuss!
1.3. List 10 systems within the human body. 1.8. We have discussed that it is often difficult to infer the
structure of a biological system from data. Is it
1.4. Exactly what features make the system in Figure 1.10
possible that two different systems produce exactly
so much more complicated than the system in Figure
the same input–output data? If you think it is
1.8?
impossible, discuss and defend your conclusion.
1.5. Imagine that Figure 1.10 represents a system that has If you think the answer is affirmative, construct a
become faulty owing to disease. Describe the conceptual example.
consequences of its complexity for any medical
treatment strategy. 1.9. List and discuss features supporting the claim
that reductionism alone is not sufficient for
1.6. In Figure 1.2, are there control paths along which understanding biological systems.
ABA either activates or inhibits closure of stomata? If
so, list at least one path each. If both paths exist in 1.10. List those challenges of systems biology that cannot
parallel, discuss what you expect the effect of an be solved with intuition alone.
increase in ABA to be on stomata closure? 1.11. Discuss why it is important to create terminology
1.7. Imagine a control system like that in Figure 1.2, but and tools for communicating systems biology.
much simpler. Specifically, suppose there is only one 1.12. Assemble a “to-do” list for the future of systems
activating and one inhibiting path in parallel. biology.
REFEREnCES
[1] Li S, Assmann SM & Albert R. Predicting essential components [2] USNO. United States Naval Observatory Cesium
of signal transduction networks: a dynamic model of guard cell Fountain. http://tycho.usno.navy.mil/clockdev/cesium.html
abscisic acid signaling. PLoS Biol. 4 (2006) e312. (2010).
[3] Voit EO, Alvarez-Vasquez F & Hannun YA. Computational [15] Bertalanffy L von. General System Theory: Foundations,
analysis of sphingolipid pathway systems. Adv. Exp. Med. Biol. Development, Applications. G Braziller, 1969.
688 (2010) 264–275. [16] Mesarović MD. Systems theory and biology—view of a
[4] Vodovotz Y, Constantine G, Rubin J, et al. Mechanistic theoretician. In Systems Theory and Biology (MD Mesarović,
simulations of inflammation: current state and future ed.), pp 59–87. Springer, 1968.
prospects. Math. Biosci. 217 (2009) 1–10. [17] Rosen R. A means toward a new holism: “Systems Theory
[5] Savageau MA. Biochemical Systems Analysis: A Study of and Biology. Proceedings of the 3rd Systems Symposium,
Function and Design in Molecular Biology. Addison-Wesley, Cleveland, Ohio, Oct. 1966. M. D. Mesarović, Ed. Springer-
1976. Verlag, New York, 1968. xii + 403 pp” [Book review]. Science
[6] Kim HJ & Triplett BA. Cotton fiber growth in planta and in vitro. 161 (1968) 34–35.
Models for plant cell elongation and cell wall biogenesis. Plant [18] Yamamoto N, Nakahigashi K, Nakamichi T, et al. Update on
Physiol. 127 (2001) 1361–1366. the Keio collection of Escherichia coli single-gene deletion
[7] Kumar A. Understanding Physiology. Discovery Publishing mutants. Mol. Syst. Biol. 5 (2009) 335.
House, 2009. [19] Savageau MA. The challenge of reconstruction. New Biol.
[8] Wolkenhauer O, Kitano H & Cho K-H. An introduction to 3 (1991) 101–102.
systems biology. IEEE Control Syst. Mag. 23 (2003) 38–48. [20] Freitas RAJ. Nanomedicine, Volume I: Basic Capabilities. Landis
[9] Westerhoff HV & Palsson BO. The evolution of molecular Bioscience, 1999.
biology into systems biology. Nat. Biotechnol. 22 (2004) [21] Lee LA & Wang Q. Adaptations of nanoscale viruses and other
1249–1252. protein cages for medical applications. Nanomedicine 2 (2006)
[10] Strange K. The end of “naive reductionism”: rise of systems 137–149.
biology or renaissance of physiology? Am. J. Physiol. Cell [22] Martel S, Mathieu J-B, Felfoul O, et al. Automatic navigation
Physiol. 288 (2005) C968–C974. of an untethered device in the artery of a living animal using
[11] Voit EO & Schwacke JH. Understanding through a conventional clinical magnetic resonance imaging system.
modeling. In Systems Biology: Principles, Methods, and Appl. Phys. Lett. 90 (2007) 114105.
Concepts (AK Konopka ed.), pp 27–82. Taylor & Francis, [23] SBML. The systems biology markup language. http://sbml.org/
2007. Main_Page (2011).
[12] Bernard C. Introduction a l’étude de la médecine [24] Stanislaus R, Jiang LH, Swartz M, et al. An XML standard
expérimentale. JB Baillière, 1865 (reprinted by Éditions for the dissemination of annotated 2D gel electrophoresis
Garnier-Flammarion, 1966). data complemented with mass spectrometry results. BMC
[13] Noble D. Claude Bernard, the first systems biologist, and the Bioinformatics 5 (2004) 9.
future of physiology. Exp. Physiol. 93 (2008) 16–26. [25] Voit EO. The Inner Workings of Life. Cambridge University
[14] Bertalanffy L von. Der Organismus als physikalisches Press, 2016.
System betrachtet. Naturwissenschaften 33 (1940) [26] Noble D. The Music of Life: Biology Beyond Genes. Oxford
521–531. University Press, 2006.
FuRTHER REAdInG
Alon U. An Introduction to Systems Biology: Design Principles of Szallasi Z, Periwal V & Stelling J (eds). System Modeling in Cellular
Biological Circuits. Chapman & Hall/CRC, 2006. Biology: From Concepts to Nuts and Bolts. MIT Press, 2006.
Kitano H (ed.). Foundations of Systems Biology. MIT Press, 2001. Voit EO. The Inner Workings of Life. Cambridge University Press,
Klipp E, Herwig R, Kowald A, et al. Systems Biology in Practice: 2016.
Concepts, Implementation and Application. Wiley-VCH, 2005. Voit EO. Computational Analysis of Biochemical Systems: A Practical
Noble D. The Music of Life: Biology Beyond Genes. Oxford University Guide for Biochemists and Molecular Biologists. Cambridge
Press, 2006. University Press, 2000.
Introduction to
Mathematical Modeling 2
• Understand the challenges of mathematical modeling
• Describe the modeling process in generic terms
• Know some of the important types of mathematical models in systems biology
• Identify the ingredients needed to design a model
• Set up simple models of different types
• Perform basic diagnostics and exploratory simulations
• Implement changes in parameters and in the model structure
• Use a model for the exploration and manipulation of scenarios
At the core of any computational analysis in systems biology is a mathematical

model. A model is an artificial construct in the language of mathematics that repre-
sents a process or phenomenon in biology. Modeling is the process of creating such
a construct and squeezing new insights out of it. Mathematical models in systems
biology can take many forms. Some are small and simple, others large and compli-
cated. A model can be intuitive or very abstract. It may be mathematically elegant in
its streamlined simplicity or account for very many details with a large number of
equations and specifications that require sophisticated computer simulations for
their analysis. The key commonality among all good models in biology is their abil-
ity to offer us insights into processes or systems that we would not be able to gain
otherwise. Good models can make sense of a large number of isolated facts and
observations. They explain why natural systems have evolved into the particular
organizational and regulatory structures that we observe. They allow us to make
predictions and extrapolations about experimentally untested situations and
future trends. They can lead to the formulation of new hypotheses. Good models
have the potential to guide new experiments and suggest specific recipes for manip-
ulating biological systems to our advantage. Models can tell us how cell cycles are
controlled and why plaques appear in certain sections of our arteries and not in
others. Future models may help us understand how the brain works and prescribe
specific interventions in tumor formation that will be infinitely more subtle and less
crude than our current methods of radio- or chemotherapy.
20 Chapter 2: Introduction to Mathematical Modeling
Modeling in systems biology is really a special way of looking at the world.

Whether we just conceptualize a situation in the form of components and interac-
tions, or whether we actually set up equations and analyze them, we begin to see
how complicated phenomena can be conceptually simplified and dissected into
manageable submodels or modules, and we develop a feel for how such modules
interact with each other and how they might respond to changes. Be aware! Model-
ing may become a life-changing adventure.
An analogy for a mathematical model is a geographical map. This map is obvi-
ously very different from reality, and many details are missing. What do the houses
look like in some remote village? Are the lakes clean enough for swimming? Are the
people friendly? Then again, it is exactly this simplification of the complicated real
world that makes the map useful. Without being distracted by secondary details, the
map allows us to look at the network of major roads throughout an entire country
while sitting at home on our living room floor. The map provides us with distances
and highway types that let us estimate travel times. Altitudes on the map indicate
whether the area we want to drive through is hilly and could be cold even during the
summer months. Once we learn how to interpret the map, we can infer what the
countryside might look like. Empty spaces, like the Badlands in South Dakota, sug-
gest that living in the area might be difficult or undesirable for humans. The density
of cities and villages provides us with some idea of how easy it might be to find lodg-
ing, food, or gas. Clusters of cities along the ocean probably imply that living there is
desirable. Mathematical models have the same generic properties: they reflect some
aspects quite well, but barely touch on others or ignore them altogether. They allow
us to infer certain properties of reality but not others. The inclusion or exclusion of
certain features in models has a direct, important implication for the modeling pro-
cess. Namely, the key task in designing a suitable model is a crisp focus on the
aspects of primary interest and a representation of these aspects that retains their
most important properties. At the same time, extraneous or distracting information
must be minimized, and the model must be limited in complexity to permit analysis
with reasonable efficiency.
Modeling is a mathematical and computational endeavor that in some sense
resembles the fine arts (Figure 2.1). Basic math problems in calculus or linear alge-
bra have the correct solutions in the back of the book, or at least somebody,
Figure 2.1 Modeling resembles a fine

art. Modeling requires both mastery of
techniques and creativity.
INTRODUCTION TO MATHEMATICAL MODELING 21
somewhere, knows these unique, correct solutions. In contrast to this rigor and
crispness, many aspects of modeling allow a lot more flexibility. They permit differ-
ent approaches, and often there are multiple, quite distinct solutions that are all valu-
able and correct in their own right. Who is to say that an oil painting is better than a
watercolor? Or that an opera has higher artistic value than a symphony? Similarly, a
mathematical model has aspects and properties in different dimensions and its value
is not always easy to judge. For instance, is a very complicated, detailed model better
than a simple, yet elegant model?
The answers to these questions depend critically on the goals and purposes of
the model, and it is not just beneficial but indeed mandatory to specify the reasons
for setting up a model and to ask a lot of questions before we embark on actually
designing it. These questions must clarify the potential utilization of the model, its
desired accuracy, the degree of complexity we are willing to accept, and many other
details that we will touch on in this chapter. Indeed, it is prudent to invest quite a bit
of time and exploration in pondering these aspects before one begins selecting a
specific model structure and defining variables, parameters, and other model fea-
tures, because technical challenges often distract from the big picture.
As in the fine arts, the creation of a good model requires two ingredients: techni-
cal expertise and creativity. The technical aspects can and must be learned through
the acquisition of a solid repertoire of math and computer science skills. To make
theoretical and numerical analyses meaningful, one also has to dig sufficiently
deeply into biology in order to understand the background and terminology of the
phenomenon of interest and to interpret the model results in a correct and sensible
manner. As with the fine arts, the creative aspect of the process is more difficult to
learn and teach. For instance, it requires more experience than one might expect to
ask the right questions and reformulate them in such a way that they become ame-
nable to mathematical and computational analysis. For a novice, all data may look
equally valuable, but they really are not. In fact, it takes time and experience to
develop a feel for what is doable with a particular set of data and what is probably
out of reach. It is also important to learn that some data are simply not well suited for
mathematical modeling. This insight is sometimes disappointing, but it can save a
lot of effort and frustration.
The shaping and refining of specific questions mandates that we learn how to
match the biological problem at hand with appropriate mathematical and computa-
tional techniques. This matching requires knowledge of what techniques are avail-
able, and it also means that the modeler must make an effort to understand how
biologists, mathematicians, and computer scientists are trained to think. The differ-
ences in approach are quite pronounced in these three disciplines. Mathematicians
are encouraged to simplify and abstract, and there is hardly anything more appealing
than a tough problem that can be scribbled on the back of a napkin. Computer scien-
tists look at problems algorithmically, that is, by dissecting them into very many, very
tiny steps. Contrast that with biologists, who have learned to marvel at the complexity
and intricacies of even the smallest biological item, and you can imagine the mental
and conceptual tensions that may develop during the modeling process.
The formation of good questions requires decisions about what is really impor-
tant in the model, which features can be ignored, and what inaccuracies we are
willing to tolerate when we simplify or omit details. It is this complex of decisions on
the inclusion and exclusion of components, facts, and processes that constitutes the
artistic aspect of modeling. As in the fine arts, a good model may give an outsider the
impression “Oh sure, I could have done that,” because it is well designed and does
not reveal the hard work and failed attempts that went into it. A good degree of pro-
ficiency with respect to these decision processes can be learned, but, like everything
else, it requires practice with simple and then increasingly complex modeling tasks.
This learning by doing can be exhilarating in success and depressing in failure, and
while we have every right to enjoy our successes, it is often the failures that in the
end are more valuable, because they point to aspects we do not truly understand.
The computational side of modeling draws from two classes of methods: math-
ematical analysis and computer simulation. Most practical applications use a com-
bination of the two, because both have their own strengths. A purely mathematical
analysis leads to general solutions that are always true, if the given assumptions and
prerequisites are satisfied. This is different from even a large set of computer
simulations, which can never confer the same degree of certainty and generality.
For instance, we can mathematically prove that every real number has a real inverse
(4 has 0.25, −0.1 has −10), with the notable exception of 0. But if we were to execute
a mega-simulation on a computer, randomly picking a real number and checking
whether its inverse is real, we would come to the conclusion that indeed all real
numbers have a real inverse. We would miss zero, because the probability of ran-
domly picking zero from an infinite range of numbers is nil. Thus, in principle,
mathematics should always be preferred. However, in practical terms, the mathe-
matics needed to analyze biological systems becomes very complicated very
quickly, and even relatively simple-looking tasks may not have an explicit analytical
solution and therefore require computational analysis.
Independent of the specific biological subject area and the ultimate choice of a
particular mathematical and computational framework, the generic modeling pro-
cess consists of several phases. Each phase is quite distinct both in input require-
ments and in techniques. Broadly categorized, the phases are shown in Figure 2.2,
together with a coarse flow chart of the modeling process. Further details are pre-
sented in Table 2.1. The process starts with conceptual questions about the biologi-
cal problem, becomes more and more technical, and gradually returns to the
biological phenomenon under study. It is very important to devote sufficient effort to
each of these phases, especially the early ones, because they ultimately determine
whether the modeling effort has a good chance of succeeding.
One of the early choices addresses the explanatory character of the model. One
may choose a comparatively simple regression model, which correlates two or more
quantities to each other, without, however, suggesting a rationale for the correlation.
As an example, one may find that cardiovascular disease is often associated with
high blood pressure, but a regression model does not offer an explanation. As an
alternative, one could develop a much more complex mechanistic model of cardio-
vascular disease, in which blood pressure is a component. If formulated validly, the
model would explain the mechanistic connection, that is, the network of causes and
effects leading from high blood pressure to the disease. Even if we disregard the dif-
ficulties in formulating a comprehensive mechanistic model, the choice of model
type is not black and white. Regression models do not provide an explanation, but
they often make correct, reliable predictions. An explanatory model may also be
quite simple, but if it accounts for a large number of processes, it often becomes so
GOALS, INPUTS, scope, goals, objective

AND INITIAL
EXPLORATION data, information, prior knowledge
MODEL
SELECTION type of model
variables, interactions
MODEL DESIGN
equations, parameter values
consistency, robustness
MODEL ANALYSIS
validation of dynamics
AND DIAGNOSIS Figure 2.2 Flow chart of the modeling
exploration of possible behaviors process. The modeling process consists
of distinct phases, but is often iterative
and involves refinements in model design
and repeated diagnosis. See also Table 2.1.
hypothesis testing, simulation, discovery, explanation (Adapted from Voit EO, Qi Z & Miller GW.
MODEL USE AND
APPLICATIONS Pharmacopsychiatry 41(Suppl 1) [2008]
manipulation and optimization S78–S84. With permission from Thieme
Medical Publishers.)
INTRODUCTION TO MATHEMATICAL MODELING 23
TABLE 2.1: ISSUES TO PONDER DURING THE MODELING PROCESS

Goals, Inputs, and Initial Exploration
• Scope of the model, including goals, objectives, and possible applications
• Data needs and availability (types, quantity, quality)
• Other available information (nonquantitative heuristics, qualitative input from experts)
• Expected feasibility of the model
• Relevance and degree of interest within the scientific community
Model Selection
• Model structure
⚬ Explanatory (mechanistic) versus correlative (black box)
⚬ Static versus dynamic
⚬ Continuous versus discrete
⚬ Deterministic versus stochastic
⚬ Spatially distributed versus spatially homogeneous
⚬ Open versus closed
⚬ Most appropriate, feasible approximations
Model Design
• Model diagram and lists of components
⚬ Dependent variables
⚬ Independent variables
⚬ Processes and interactions
⚬ Signals and process modulations
⚬ Parameters
• Design of symbolic equations
• Parameter estimation
⚬ Bottom-up
⚬ Top-down
Model Analysis and Diagnosis
• Internal consistency (for example, conservation of mass)
• External consistency (for example, closeness to data and expert opinion)
• Reasonableness of steady state(s)
• Stability analysis
• Sensitivity and robustness analysis
• Structural and numerical boundaries and limitations
• Range of behaviors that can or cannot be modeled for (example, oscillations, multiple steady states)
Model Use and Applications
• Confirmation, validation of hypotheses
• Explanation of causes and effects; causes of failure, counterintuitive responses
• Best case, worst case, most likely scenarios
• Manipulation and optimization (for example, treatment of disease; avoidance of undesired
states or dynamics; yield optimization in biotechnology)
• Discovery of design principles
complicated and requires so many assumptions that it may no longer describe

or predict new scenarios with high reliability. In reality, most models are
intermediates—explanatory in some aspects and correlative in others.
In the following sections, we discuss each phase of the modeling process and
illustrate it with a variety of small, didactic “sandbox models” as well as a famous,
very old model describing the spread of infectious diseases. The process begins with
an exploration of the available information and the selection of a model that can
utilize this information and ultimately yield new insights (see Figure 2.2 and
Table 2.1). Often, one will go back and forth between these two aspects before a suit-
able model design is chosen. Once an appropriate model type is determined, the
actual model design phase begins. Components of the model are identified and
characterized. The resulting model is diagnosed, and, should troublesome aspects
emerge, one returns to the model design. A model that fares well in the diagnostic
phase is used for explanations, simulations, and manipulations. It often turns out
that model features are not optimal after all, requiring a return to the model design
phase.
GOALS, INPUTS, AND INITIAL EXPLORATION

The initial phase of model development consists of defining the purpose of the
model and of surveying and screening input data and contextual information. This
setting of goals, combined with an exploration of available information and the con-
text for the model, is arguably the most crucial of all phases. This assessment may
sound counterintuitive, because one might think that the most effort would go into
some complicated math. However, if insufficient consideration is given to this initial
exploration, the modeling process may derail right away, and one may not even
notice that something is not quite the way it was intended until the final phases of
the modeling process.
The main task of the input and exploration phase is to establish the specific pur-
poses and goals of the model, and to feel out whether the model may have a chance
to achieve its goals, given the available data and information. “Feeling out” does not
sound like rigorous mathematical terminology, but actually describes the task quite
well. It is often here where experience is needed the most: right at the beginning of
the effort!
2.1 Questions of Scale

The first important aspect of the exploration phase is to ask what are the most
appropriate scales for the model, with respect to both time and space, and also
with respect to the organizational level (Figure 2.3). These scales are typically tied
to each other, because processes at the organizational level of a cell occur on a
small spatial scale and usually run faster than processes at the scales of organisms
or ecological systems. Suppose we are interested in trees and want to develop a
mathematical model describing their dynamics. While that might sound like a fine
idea, it is far too unspecific, because trees may be studied in many different ways.
To model the growth of a forest or a tree plantation, the scale of organization is typi-
cally the individual tree, the spatial scale is in hectares or square kilometers, and
the corresponding timescale is most likely in years. Contrast this with the photo-
synthesis in a leaf or needle, which truly drives tree growth. That organizational
years
eco-
systems
human
days chicken organs
egg
typical plant,
animal cells
hours
bacteria
minutes
genes, Figure 2.3 Scales in biology span several
proteins, orders of magnitude in size and time. This
lipids
seconds plot shows typical sizes and time ranges.
Of course, human organs and ecosystems
contain small molecules and genes, so that
small
milliseconds molecules they really span multiple size- and time-
scales, which makes multiscale modeling
1 nm 1 µm 1 mm 1 cm 1m 1 km challenging and fascinating.
GOALS, INPUTS, AND INITIAL EXPLORATION 25
level is biochemical or physiological, and the corresponding timescale is that of

seconds or minutes. It is easy to imagine other time and organizational scales
between or outside these two scenarios.
While multiscale modeling is one of the challenging frontiers of systems biology
(see Chapter 15), the attempt to accommodate too many different scales at once is
very difficult and, at least at the beginning of model design, dangerous. Although
there are exceptions (see for instance Chapter 12), it is usually not really feasible to
account in the same model for the entire spectrum of processes, from the details of
photosynthesis or wood formation to the long-term growth of forests. Nonetheless,
the real strength of mathematical models lies in their ability to explain phenomena
at one level through the analysis and integration of processes at a lower level. Thus,
a good model, even of small size, may span two or maybe three levels of biological
organization. For instance, it may describe the formation of a leaf in terms of genetic
and biochemical processes. A model may also skip some levels. For instance, we
could try to associate long-term tree growth and final size with alterations in gene
expression. In this case, one might be able to translate the effects of gene modula-
tions into the tree’s growth rate, but a model of this type would not capture all the
biochemical and physiological causes and effects that occur between gene expres-
sion and the total amount of wood in the forest after 50 years.
The determination of suitable scales depends very significantly on the available
data that are needed to construct, support, and validate the model. If only yearly
biomass measurements of the tree stand are available, we should not even attempt
to model the biochemistry that deep down controls tree growth. Conversely, if the
model uses light conductance as main input data, good as they may be, this input
information would hardly support a model of the growth of trees in the forest over
many years.
As a different example, suppose it is our task to improve the yield of a fermenta-
tion process in which yeast cells produce alcohol. Typical experimental techniques
for this task are selections of mutants and changes in the growth medium. Thus, if
we intend to use a model to guide and optimize this process, the model must clearly
account for gene activities and their biochemical consequences, as well as for con-
centrations of nutrients in the medium and their uptake by cells.
2.2 Data Availability

The quantity of available data provides important clues for exploring options of
model selection and design. If a complicated model with many variables and
parameters is supported only by very scarce data, the model may capture these
data very nicely, but as soon as we try to apply it to a slightly different scenario,
such as wider spacing of the trees or a richer fertilization regimen, the model pre-
dictions may become unreliable or inaccurate. This over-fitting problem (that is,
using a model that is too complicated for the given data) always looms when only
one or a few datasets are available. A similar problem arises if the data show a lot of
variability, or noise, as it is often called in the field of data analysis. In this case,
many models may yield a good representation of the particular data, but predic-
tions regarding new data may be unreliable. Counterintuitive as it may sound, it is
often easier to model bad data than good data, because bad data grant us a lot more
latitude and potentially let us get away with an inferior model. Of course, this
inferior model may turn round and bite us later when the model is expected to fit
new data and simply doesn’t!
While crisp, numerical data are the gold standard for mathematical modeling,
we should not scoff at qualitative data or information that contains a fair amount of
uncertainty. For instance, if we are told that experiments show some output increas-
ing over time (even though the exact degree is not given), this piece of information
can provide a valuable constraint that the model must satisfy. Especially in a clinical
setting, this type of qualitative and semiquantitative information is prevalent, and
the more of it that is available, the more any model of the phenomenon is con-
strained. Expressed differently, this information can be extremely helpful for weed-
ing out models that could otherwise appear plausible. In a study of Parkinson’s
disease, we asked clinicians for concentrations of dopamine and other brain metab-
olites associated with the disease. While they could not give us exact values, they
provided us with rough, relative estimates of concentrations and metabolic fluxes

through the system, and this information, combined with literature data, allowed us
to set up a model that made surprisingly accurate predictions [1].
Finally, the exploration phase should ascertain whether experimental biologists
or clinicians are really interested in the project. If so, their interest will help drive the
project forward, and their direct insights and contacts with other experts can be
invaluable, because much biological knowledge is not published, especially if it is
semiquantitative or not yet one hundred percent supported by data and controls. It
is, of course, possible to base models solely on information found in the literature,
but if no experts are at hand, many pitfalls loom.
MODEL SELECTION AND DESIGN

To illustrate the process of model design, let’s consider one specific example that we
will carry through this chapter and revisit again in Chapter 4. This example concerns
an infectious disease, such as tuberculosis or influenza, that easily spreads through-
out a human population (Figure 2.4). Down the road, we would like to determine
whether it is beneficial to vaccinate all or some individuals, whether infected persons
should be quarantined, or whether the disease will eventually go away by itself
because people recover and acquire immunity. Immediately, many questions come
to mind. Do people die from the disease? If the death rate is small, can we ignore it in
the model? How does the disease get started in the first place? For how long are peo-
ple sick? How infectious are they? Are there individuals who show no symptoms yet
are infectious? Do all people in our target population have potential contact with all
other people? Do we need to account for air travel? Should we attempt to character-
ize waves of infection that start at the point of the first contact with an infected person
and move through certain geographical areas or the entire population? Does our dis-
ease model have to be accurate enough to predict whether a specific person will get
sick, or is our target the average infection level within a community or the nation?
Clearly, questions are unlimited in number and type, and some of them may be
important to ponder now while others are probably less relevant, at least at the
beginning.
For our purposes of studying a very complex phenomenon with as simple a
model as possible, let’s suppose the following:
• The infectious disease is spread by human-to-human contact.
• Although bacteria or viruses are the true vectors of the disease, we ignore them.
• Individuals may acquire and lose immunity.
• The primary questions are how fast the disease progresses, whether it disappears
without intervention due to immunity, and whether vaccination or quarantine
would be effective.
Figure 2.4 Mathematical modeling can be

a powerful tool for studying the spread
of bacterial diseases. Shown here is the
bacterium Mycobacterium tuberculosis, which
2 µm causes tuberculosis, one of the ongoing
scourges of humankind.
MODEL SELECTION AND DESIGN 27
• We do not intend to account for the spatial distribution of cases of infection

throughout some geographical area.
2.3 Model Structure

Once we have decided what the model is supposed to accomplish, we need to
determine its structure. Some decision points are given in Table 2.1. None of these
are really associated with different degrees of sophistication, usefulness, or quality
of the model. We are simply required to make decisions, again based on the data
and the purposes of the model.
The first decision targets the question of how much detail the model is expected
to explain. As mentioned before, the two extremes in this dimension are correlative
and explanatory models; most models have aspects of both. A correlative model
simply and directly relates one quantity to another. For instance, data may show that
the infectious disease is particularly serious among the elderly; specifically, the
number of infected individuals per 100,000 may increase with age (Figure 2.5). A
linear or nonlinear regression could even quantify this relationship and allow us to
make predictions about the average disease risk for a person of a certain age. How-
ever, even if the prediction were quite reliable, the model would not give us a hint as
to why older people are more likely to become infected. The model does not account
for health status, genetic predisposition, or lifestyle. Sometimes a prediction would
be wrong, but the model would not be able to explain why.
An explanatory model, by contrast, relates an observation or outcome of an
experiment to biological processes and mechanisms that drive the phenomenon
under investigation. To illustrate, imagine an explanatory model of cancer. Cancer
is associated with cells that are not dying when they should. Cell death normally
occurs through a process called apoptosis, whose control involves the tumor sup-
pressor protein p53, which in turn acts as a transcription factor for a number of
pertinent genes. The p53 protein becomes activated under many stress condi-
tions, such as ultraviolet radiation or oxidative stress. Pulling the pieces of infor-
mation together allows us to compose a conceptual or mathematical model in the
form of a chain of causes and effects: Suzie gets baked on the beach. She is exposed
to enough ultraviolet radiation to activate the p53 proteins in her skin. The pro-
teins change the normal activation of some genes. The genes cause some cells to
escape apoptosis and proliferate out of control. Suzie is diagnosed with mela-
noma. The model explains to some degree (although not completely, of course)
the development of cancer. This explanatory potential does not come for free:
explanatory, mechanistic models are usually much more complicated than cor-
relative models and require vastly more input in the form of data and assump-
tions. While explanatory models offer more insight, we should not discard
correlative models offhand. They are much easier to construct and often make
more reliable predictions than explanatory models, especially when assumptions
are uncertain and data are scarce.
The second decision refers to the question of whether something associated
with the phenomenon of interest changes over time. If the data show trends chang-
ing with time, we will probably prefer a dynamical model, which allows us to quan-
tify these changes. By contrast, in a static model, such as a regression model, only
Figure 2.5 Correlative and explanatory
the dependencies of one variable on others are explored. Box 2.1 contrasts static models serve different purposes.
The correlation between x and y in (A),
whether linear or nonlinear, permits robust
predictions of a value of y, given a new
(A) (B) value of x, although the model does not
y X5 explain the connection between x and y.
An example could be disease prevalence (y)
X3
and age (x) among the elderly. The simple
– explanatory model in (B) could describe a
metabolic pathway with two inhibitors (X5
X1 X2
and X6). It not only captures the (positive)
correlation between X3 and X6, but also
–
provides an explanation for why X3 increases
X4
if the inhibitor X6 of the other branch of the
x X6 pathway is increased.
BOX 2.1: STATIC AND DYNAMIC MODELS
Suppose it is our task to compute the volume of an Escherichia coli cell. In the typical understanding of dynamical models, one expects
Studying a photograph of an E. coli culture (Figure 1), we see that each that the current state of the object or system will influence its further
bacterium looks more or less like a hot dog with variable length and development. In the case of a bacterium, the growth in volume very
flattened ends, and since the volume doesn’t change if the bacterium often depends on how big the bacterium already is. If it is small, it
curves up, we can just study the shape when it is stretched out. Thus, a grows faster, and as it reaches its final size before it divides, growth
simple static model could be a cylinder with two “caps” at the end. The slows down. This “self-dependence” makes the mathematics more
caps look a little flatter than hemispheres and we decide to represent complicated. Specifically, one formulates the change in volume over
them mathematically by halved ellipsoids that have the same axis time as a function of the current volume. The change in volume
radius in two directions (a circular cross section) and a shorter axis over time is given in mathematical terminology as its time derivative
radius in the third direction (Figure 2). The formula for the volume of dV/dt. Thus, we come to the conclusion that volume-dependent
a whole ellipsoid is 43 πr1r2 r3, where r1, r2, r3 are the axis radii. Thus, if the growth should be formulated as
“straight” part of the cell is x μm long and has a diameter of y μm, and
if each cap has a height of z μm, we can formulate a static model for dV
= some function of volume V at time t (3)
the model of an E. coli cell (Figure 3) as dt
 y
2
 4 y y z
V ( x , y , z ) = xπ   + 2  π (1)
( )
= f V (t ) = f (V ) . (4)
 2  3 2 2 2 

cylinder two half-ellipsoildal
caps
This type of formulation, in which the derivative of a quantity
(dV/dt) is expressed as a function of the quantity (V) itself, constitutes
 x z an ordinary differential equation, which is also lovingly called a
=  +  πy 2 . (2)
diff. eq., o.d.e., or ODE. Very many models in systems biology are
 4 3
based on differential equations. The function f is not limited to a
Our static model allows us to pick any length, width, and cap height dependence on V; in fact, it most often contains other variables
and to compute the corresponding volume of the bacterium. as well. For example, f could depend on the temperature and the
Let’s now suppose the bacterium has just divided and is growing. substrate concentration in the medium, which will certainly affect
The size parameters x, y, and z become variables that depend on time the speed of growth, and on a lot of “internal” variables that describe
t, and we may write them as x(t), y(t), and z(t). It may even happen that the uptake mechanisms with which the bacterium internalizes the
these variables do not develop independently from each other, but substrate. Looking more closely, many biochemical and physiological
that the bacterium maintains some degree of proportionality in its processes are involved in the conversion of substrate into bacterial
shape, which would mathematically connect x(t), y(t), and z(t) through volume and growth, so that a very detailed dynamical model can
constraint relationships. The main consequence is that the volume quickly become quite complicated. Setting up, analyzing, and
is no longer static, but a function of time. The model has become interpreting these types of dynamical models is at the core of
a dynamic or dynamical model. computational systems biology.
Figure 2 To model the

Figure 1 The shape shape of a bacterium
of E. coli bacteria 2r1
(see Figure 1), we need
resembles cylinders 2r3 a cylinder and “caps”
with caps. The shape at both ends. The caps
may be modeled can be modeled with flat
approximately with the ellipsoids. Here, such an
geometrical shapes ellipsoid is shown with
shown in Figures 2 and 3. 2r2 two longer radii of the
same length r1 = r2 and
one shorter radius r3.
× 15,000 1 µm
Figure 3 Static model

of the volume of E. coli.
y The model is inspired
by the actual cell shape
(Figure 1) and composed
of a cylinder with caps at
z x both ends (Figure 2).
and dynamic models using a specific example. Static models may appear to be
insufficient, because they ignore change, but they play important roles in systems
biology, where they are, for instance, used to characterize the features of large
networks (see Chapter 3).
Within the realm of dynamic models, one needs to decide whether a discrete or
continuous timescale is better suited. An example of the discrete case is a digital
clock, which displays time only with a certain resolution. If the resolution is in
minutes, the clock does not give any indication whether the minute just started or is
almost over. A continuous, analog clock would, in principle, show the time exactly.
On one hand, the decision on a continuous or discrete timescale is very important,
because the former usually leads to differential equations and the latter to iterative
or recursive maps, which require different mathematical methods of analysis and
sometimes lead to different types of insights (see Chapter 4). On the other hand, if
we select very small time intervals, the behaviors of discrete models approach those
of continuous models, and differential equations, which are often easier to analyze,
may replace the difference equations.
The convergence of discrete and continuous models for increasingly smaller
time steps is not the only instance where one formulation is replaced with another.
More generally, we need to embrace the idea of approximations in other dimen-
sions as well (Figure 2.6). We are certainly interested in “true” descriptions of all
processes, but these do not really exist. Even the famous laws of physics are not
strictly “true.” For instance, Newton’s law of gravity is an approximation in the con-
text of Einstein’s general relativity theory. Not that there would be anything wrong
with using Newton’s law: apples still fall down and not up. But we should realize
very clearly that we are in truth dealing with approximations and that these are both
necessary and indeed very useful. The need for approximations is much more pro-
nounced in biology, where laws corresponding to those in physics are almost non-
existent. Chapter 4 discusses typical approximations used in systems biological
modeling and Chapter 15 discusses more philosophical issues of laws in biology.
Somewhat of an approximation is the intentional omission of spatial consider-
ations. In the real world, almost everything occurs in three spatial dimensions.
Nonetheless, if the spatial aspects are not of particular importance, ignoring them
makes life much easier. With diabetes, for instance, it may be important to study the
distribution of glucose and insulin throughout specific organs and the cardiovascu-
lar system of the body. But for modeling the administration of insulin in a given
situation, it may also be sufficient to consider merely the overall balance between
glucose and insulin. A model that does not consider spatial aspects is often called
(spatially) homogeneous. Models accounting for spatial aspects typically require
partial differential equations or entirely different modeling approaches (see Chap-
ter 15), which we will not discuss further at this point.
Finally, a systems model may be open or closed. In the former case, the system
receives input from the environment and/or releases material as output, while in
the latter case materials neither enter nor leave the system. F(x), A(x) A(x)
The model types we have discussed so far are called deterministic, which simply F(x)
means that all information is completely known at the beginning of the computational 2
experiment. Once the model and its numerical settings are defined, all properties
and future responses are fully determined. In particular, a deterministic model does
not allow any types of random features, such as environmental fluctuations or noise 1
affecting its progress. The opposite of a deterministic model is a probabilistic or
stochastic model. The latter term comes from the Greek word stochos meaning aim,
target, or guess, and implies that a stochastic process or stochastic model always x
2 4 6 8
involves an aspect of chance. Somewhere, somebody in the system is rolling dice
and the result of each roll is fed into the model. As a consequence, the variables Figure 2.6 Approximations are
become random variables and it is not possible to predict with any degree of cer- necessary and very useful for
tainty the outcome of the next experiment, or trial as it is often called. For instance, simplified representations of reality
in a mathematical model. In this
if we flip a coin, we cannot reliably predict whether the next flip will results in heads
illustration, A(x) (dashed blue line) is a linear
or tails. Nevertheless, even if we do not know exactly what will happen in future tri- approximation of a more complicated
als, we can compute the probability of each possible event (0.5 for each coin toss), nonlinear function F(x) (red line). If only the
and the more often we repeat the same random trial, the closer the collective out- range between 2 and 5 for the variable x is
come approaches the computed probability, even though each new trial again and biologically relevant, the approximation is
again exhibits the same unpredictability. sufficiently accurate and easier to analyze.
It is no secret that biological data always contain variability, uncertainties, mea-

surement inaccuracies, and other types of noise. Thus, as a final point to ponder
when choosing a model: should we—must we—always use a model that somehow
captures this randomness? Not necessarily. If we are primarily interested in average
model responses, a deterministic model will generally do just fine. However, if we
need to explore all possible outcomes, the worst and best cases, or the likelihood of
specific outcomes from a system with a lot of uncertainty, we may be best advised to
choose a stochastic model.
Because this is an important and subtle issue, let’s study similarities and differ-
ences between deterministic and stochastic models in more detail. As we dis-
cussed, each trial in a stochastic system is unpredictable, but we can make very
accurate predictions about average or long-term behaviors. By comparison, deter-
ministic models present us with situations where everything is exactly and unam-
biguously defined. Yet, there are strange cases where the behavior of the system
over time is so complicated that we cannot forecast it. A quite simple, but very
instructive example is the blue sky catastrophe, which is discussed in Box 2.2. This
system is completely determined and does not contain any noise or stochastic
features. All information about the system is fixed in its definitions and settings,
and yet the system is able to generate chaotic time courses that are impossible to
anticipate. Chapter 4 discusses more cases of this nature.
As an example of a stochastic process, suppose we irradiate a culture of microor-
ganisms with ultraviolet light (see, for example, [4]) in order to introduce random
mutations. This method of random mutagenesis is often used to create strains that
out-compete the wild type in a specific situation of interest. For instance, we may
want to find a strain with improved tolerance to acid in its environment. We know
from genetics that most mutations are neutral or decrease rather than increase fit-
ness. Nonetheless, one mutation in a million or a billion may improve the microor-
ganism’s survival or growth rate, and this particular mutation will rise to the top and
multiply in a properly designed random mutagenesis experiment. For simplicity,
let’s suppose that the mutations occur randomly throughout the genome with a rate
of between three and four mutations per kilobase (1000 base pairs of DNA) [5].
A typical question in this scenario is: What is the probability of actually finding
three or four mutations in a given DNA segment that is one kilobase long? Our first
inclination might be to claim (wrongly) that the probability is 1, because isn’t
that what we just assumed? Not quite. We assumed that the average number of
mutations is between three and four, and need to realize that it could happen that
there are only one or two mutations, or five or six, owing to the randomness of the
mutation process. Models from probability theory allow us to assess situations of
this type. For instance, we can forecast the probability of encountering one, two,
or three mutations in a given piece of DNA. Or we can make statements like “with
p percent probability, we will find at least four mutations.” Would it be possible
to find 200 random mutations within 1000 base pairs? In a true stochastic pro-
cess, this outcome is indeed theoretically possible, although the probability is
extremely low.
Even in this clear-cut scenario with precise conditions and well-defined ques-
tions, different models are available. For instance, we could use the binomial model
or the Poisson model. Both are applicable in principle. The binomial model would
be better suited for small numbers (here, short segments of DNA), while it would be
very cumbersome for large numbers (long segments of DNA). In fact, a pocket cal-
culator would not be able to manage the numbers that would have to be computed
for 1000 kilobases. The Poisson model uses a clever approximation, which is not
very accurate for small numbers (short segments of DNA), but is incomparably eas-
ier to use and misses the correct results for large numbers (long segments of DNA)
by only very small errors. Boxes 2.3 and 2.4 discuss these two models as typical rep-
resentatives of simple stochastic models.
2.4 System Components

Biological systems models contain different classes of components, which we dis-
cuss in this section. The most prominent components are variables, which represent
biological entities of interest, such as genes, cells, or individuals. Because biological
BOX 2.2: AN UNPREDICTABLE DETERMINISTIC MODEL
Thompson and Stewart [2] describe a very interesting system that However, if we think we understand what the system is doing, we
can be formulated as a pair of ordinary differential equations. The are wrong. Continuing the same example until t = 500, without any
system, called the blue sky catastrophe, is completely deterministic, changes to the model or its parameters, presents big surprises: out
yet can produce responses that are utterly unpredictable, unless one of the blue, the system “crashes” and undergoes regular or irregular
actually solves the differential equations. The system has the form oscillations at a much lower level, before it eventually returns to
something like the original oscillation (Figure 1B). Continuing further
dx in time shows irregular switches between oscillations around 1 and
= y,
dt (1) oscillations around −1. Intriguingly, even minute changes in the initial
dy values of x and y or the parameter A totally change the appearance
= x − x 3 − 0.25 y + A sin t .
dt of the oscillation (Figure 1C, D). In fact, this chaotic system is so fickle
that even settings in the numerical solver affect the solution. The
If we set A = 0.2645 and initiate the system with x0 = 0.9, y0 = 0.4, plots in Figure 1 were computed in the software PLAS [3] with the
both variables quickly enter an oscillation that is quite regular; the standard local error tolerance of 10−6. If one reduces the tolerance
time course for x for the first 100 time units is shown in Figure 1A. to 10−9 or 10−12, the appearance of the figures is strikingly different.
(A) (B)
2 2
x 1 x 0
0 –2
0 25 50 75 100 0 125 250 375 500
time time
(C) (D)
2 2
x 0 x 0
–2 –2
0 125 250 375 500 0 125 250 375 500
time time
Figure 1 The intriguing model of a blue sky catastrophe. The system of two differential equations (with x0 = 0.9, y0 = 0.4, and A = 0.2645) appears
to oscillate quite regularly about a baseline value of 1 (A). However, without any changes, the oscillations later become irregular (B: note the different
scales). Even very slight changes in numerical features change the appearance of the oscillations dramatically. (C) A = 0.265, x0 = 0.9, y0 = 0.4.
(D) A = 0.2645, x0 = 0.91, y0 = 0.4.
systems usually contain many variables, they are called multivariate. Variables may
represent single entities or collections or pools of entities, such as all insulin mole-
cules within a body, no matter where they are located. The second class of compo-
nents consists of processes and interactions, which in some fashion involve and
connect the variables. The third class contains parameters. These are numerical
characteristics of the system, such as pH or temperature, or the turnover rate of a
chemical
BOX 2.3:reaction, whichPROCESS:
THE BINOMIAL is the number of molecules
A TYPICAL theMODEL
STOCHASTIC reaction can process
In the text, we asked: What is the probability of finding a certain which is even smaller. It is not difficult to see that the probabilities
number of mutations in a DNA segment, if one typically observes become smaller and smaller for higher numbers of mutations. So,
three or four mutations per 1000 base pairs? Let’s do some math to what’s the most likely event? It is that there is no mutation at all. This
find out. We formulate the problem as a binomial process, for which probability is given by the binomial probability distribution (1) just
we need two ingredients, namely the average mutation rate, which like the others, namely,
we assume to be 3.5 (instead of “3 or 4”) mutations per kilobase, and
the number of nucleotides within the DNA segment of interest. From 10 !
P(0; 10, 0.0035) = 0.00350 (1− 0.0035)10 = 0.9655. (4)
this information, we can compute how likely it is to find 3, 4, 1, 20, or 0 !10 !
however many mutations within some stretch of DNA.
In the language of probability theory, a mutation in our In over 96% of all 10-nucleotide DNA segments, we expect to find no
experiment is called a success and its (average) probability is usually mutation at all!
termed p; in our case, p is 3.5 in 1000, that is, 0.0035. A failure means From these three results, we can estimate the chance of finding
that the base is not mutated; its probability is called q. Because more than two mutations. Because all probabilities taken together
nothing else can happen in our experiment (either a base is mutated must sum to 1, the probability of more than two mutations is the same
or it is not; let’s ignore deletions and insertions), we can immediately as the probability of not having zero, one, or two mutations. Thus, we
infer p + q = 1. Put into words: “for a given nucleotide, the probability obtain
of a mutation plus the probability of no mutation together are a sure (5)
P(k > 2) = 1 − 0.9655 − 0.0339 − 0.000537 ≈ 0.00002,
bet, because nothing else is allowed to happen.”
In order to deal with more manageable numbers, let’s first assume
that the DNA piece contains only n = 10 bases. Then the probability P which is 2 in 100,000!
to find k mutations is given as If we increase the length of the DNA piece, we should intuitively
expect the probabilities of finding mutations to increase. Using
n! the same formula, the probability of exactly one mutation in a
P ( k ; n, p ) = p k (1− p )n−k , (1) 1000-nucleotide segment can be written as
k !(n − k )!
according to the binomial formula of probability theory, which

1000 !
P(1; 1000, 0.0035) = 0.00351(1− 0.0035)999 (6)
we will not discuss here further. In this formulation, the term n! 1! 999 !
(pronounced “n factorial”) is shorthand for the product 1⋅2⋅3⋅…⋅n.
The semicolon within the function indicates that k is the This is a perfectly fine formulation, but our pocket calculator goes on
independent variable, while n and p are constant parameters, which strike. 1000! is a huge number. To ameliorate the situation, we can
may, however, be changed from one example to the next. So, what is play some tricks. For instance, we see from the definition of factorials
the probability of finding exactly one mutation in the 10-nucleotide (n! = 1⋅2⋅3⋅…⋅n) that the first 999 terms in 1000! and 999! are exactly
segment? The answer results from substituting values for the the same and cancel out from the fraction, leaving us simply with
parameters and for k, namely, 1000 for the first term. With that, we can compute the probability
of 1 mutation as 0.1054 and the probability of no mutations at all as
10 ! 0.0300 (0! is defined as 1). Indeed, in contrast to the scenario with 10
P(1; 10, 0.0035) = 0.00351(1− 0.0035)9 = 0.0339. (2)
1! 9 ! bases above, we are now about three times more likely to encounter
one mutation than no mutation at all. Is it imaginable that we would
That is not a very big probability, so the result is rather unlikely: find 10, 20, or 30 mutations? The probability of finding 10 mutations
roughly 3%. What about two mutations? The answer is is about 0.00226. Formulate the corresponding probabilities for 20
and 30. It is not hard to do, but even with the tricks above we run
10 ! into problems computing the numerical results from the formulae.
P(2; 10, 0.0035) = 0.00352 (1− 0.0035)8 = 0.000537, (3)
2!8! Box 2.4 offers a solution!
chemical reaction, which is the number of molecules the reaction can process
within a given amount of time. The particular values of the parameters depend on
the system and its environment. They are constant during a given computer experi-
ment, but may assume different values for the next experiment. Finally, there are signal
(universal) constants, such as π, e, and Avogadro’s number, which never change.
In order to design a specific formulation of a model, we begin with a diagram
and several lists, which are often developed in parallel and help us with our book-
keeping. The diagram contains all entities of the biological system that are of interest P0 P1 P2
and indicates their relationships. These entities are represented by variables and
drawn as nodes. Connections between nodes, called edges, represent the flow of
material from one node to another. For instance, a protein in a signaling cascade
may be unphosphorylated or phosphorylated in one or two positions. Within a Figure 2.7 Different states of a protein in
short timespan, the total amount of the protein is constant, but the distribution a signaling cascade. The protein may be
unphosphorylated (P0) or singly (P1) or doubly
among the three forms changes in response to some signal. Thus, a diagram of this (P2) phosphorylated. The total amount of the
small protein phosphorylation system might look like Figure 2.7. protein remains the same, but, during signal
In contrast to the flow of material, a diagram may also contain the flow of transduction, material flows among the
information. For instance, the end product of a pathway may signal—through pools, driven by kinases and phosphatases,
feedback inhibition—that no further substrate should be used. This signal is distinctly which are not shown here (see Chapter 9).
BOX 2.4: THE POISSON PROCESS: AN ALTERNATIVE STOCHASTIC MODEL
The binomial model runs into problems for large numbers k and n. Thus, the probability of 10 mutations (in a 1000-nucleotide segment) is
As a demonstration that modeling problems often permit different
solutions, we provide here an alternative to the binomial model. e −3.5 3.510
P(10; 3.5) = = 0.00230. (2)
The mathematician Siméon-Denis Poisson (1781–1840) developed 10 !
a formula, nowadays called the Poisson distribution, that lets us
The probabilities of 3 or 4 mutations are 0.216 and 0.189,
compute binomial-type probabilities for large numbers k and n. The
respectively. The two together make up about 40% of all
two ingredients we need are the mutation rate λ and the number k of
cases, which appears reasonable, because the overall mutation
mutations whose probability we want to compute. How can Monsieur
rate is 3.5 per 1000 nucleotides. The probability of no mutation
Poisson get away with only two pieces of information instead of three?
is 0.0302, which is very similar to the computation with the
The answer is that he no longer looks at the success or failure at each
binomial model. With the Poisson formula, we can compute all
nucleotide, but substitutes the step-by-step checking with a fixed rate
kinds of related features. For instance, we may ask for the
that is really only valid for a reasonably long stretches or DNA. The
probability of more mutations (anywhere between 5 and 1000
mutation rate with respect to 1000 nucleotides is 3.5, whereas the rate
mutations). Similar to the binomial case, this number is the
for 100 nucleotides is one-tenth of that, namely, 0.35. Using a fixed rate
same as the probability of not finding 0, 1, 2, 3, or 4 mutations.
makes life simple for large numbers of nucleotides, but does not make
The probabilities of 1 or 2 mutations are 0.106 and 0.185,
much sense for very short DNA pieces.
respectively; the others we have already computed. Thus, the
Poisson’s formula, which is again based on probability theory [6], is
probability of finding 5 or more mutations is approximately
1 − 0.0302 − 0.106 − 0.185 − 0.216 − 0.189 = 0.274, which
e−λ λ k corresponds to almost a third of all 1000-nucleotide strings.
P( k ; λ ) = . (1)
k!
different from the flow of material, because sending out the signal does not affect the
sender. It does not matter to a billboard how many people look at it. Because of this
difference, we use a different type of arrow. This arrow does not connect pools, but
points from a pool to an edge (flow arrow). As an illustration, the end product X4
in Figure 2.8 inhibits the process that leads to its own generation and activates an
alternative branch toward X5. These modulation processes are represented with dif-
ferently colored or other types of arrows to make their distinction from edges clear.
Of course, there are many situations where we do not even know all the pertinent
components or interactions in the system. Different strategies are possible to deal with
such situations. In opportune cases, it is possible to infer components, interactions, or
signals from experimental data. We will discuss such methods in later chapters. If such
methods are not feasible, we design the model with the best information we have and
explore to what degree it answers our questions and where it fails. While we under-
standably do not like failures, we can often learn much from them. Indeed, we typi-
cally learn more from failures than from successes, especially if the model diagnosis
points toward specific parts of the model that are most likely the cause of the failure.
Parallel to drawing out the diagram, it is useful to establish four lists that help us
organize the components of the model. The first list contains the key players of the
system, such as metabolites in a biochemical system, genes and transcription factors
in a gene regulatory network, or population sizes of different species in an ecological
study. We expect that these players will change over time in their concentrations,
numbers, or amounts, and therefore call them dependent variables, because their
fate depends on other components of the system. For instance, it is easy to imagine
that variable X2 in Figure 2.8 depends on X1 and other variables. Typically, depen-
dent variables can only be manipulated indirectly by the experimenter.
The system may also include independent variables, which we collect in the sec-
ond list. An independent variable has an effect on the system, but its value or concen- –
X1 X2 X3 X4
tration is not affected by the system. For example, independent variables are often
used in metabolic systems to model enzyme activities, constant substrate inputs, and +
co-factors. In many cases, independent variables do not change in quantity or X5 X6
amount over time, at least not within the time period of the mathematical experi-
ments we anticipate to execute. However, it may also happen that forces outside the Figure 2.8 Generic pathway system, in
system cause an independent variable to change over time. For instance, a metabolic which X4 inhibits its own production
process and activates the generation of
engineer might change the amount of nutrients flowing into a bioreactor. The vari-
an alternative pathway toward X5 and X6.
able is still considered independent, if its changes are not caused by the system. The flow of material is represented with blue
The distinctions between dependent and independent variables on the one arrows, while modulation processes are shown
hand and between independent variables and parameters on the other are not with red (inhibition) and green (activation)
always clear-cut. First, a dependent variable may indeed not change during an arrows that distinguish them from flow arrows.
experiment and could therefore be considered independent. This replacement

would make the model simpler, but might preclude analyses of other model settings
where the variable does change. It actually happens more frequently that independent
variables are replaced by dependent variables in later model extensions. A reason
could be that the extended system affects the dynamics of a formerly independent
variable. Second, if an independent variable is constant throughout an experiment,
we could replace it with a parameter. However, it is often beneficial to keep param-
eters and independent variables separate. First, variables and parameters have a dif-
ferent biological meaning, which may help us with the estimation of appropriate
values for them. And second, the computational effort of including an independent
variable or a parameter is the same.
The third list is actually closer to a table or spreadsheet. It shows which of the
(dependent or independent) variables have a direct effect on any of the processes in
the system. As in the example of Figure 2.8, an effect may be positive (enlarging or
activating) or negative (diminishing or inhibiting). The list accounts for all edges
(flow of material between pools), as well as all signals that affect any of the edges.
The final list contains parameters that contribute to the external or internal con-
ditions of the system, such as pH, temperature, and other physical and chemical
determinants, which will ultimately require numerical specification.
Throughout the model design process, we will add to these lists quantitative
information about pool sizes, magnitudes of fluxes, strengths of signals, and normal
values and ranges of parameters.
The lists are the basis for establishing a diagram of the system. For the design of
this diagram, it is beneficial to begin by representing each variable with one of the
core modules in Figure 2.9. In the simpler case (Figure. 2.9A), each variable is dis-
played as a box with one process entering and one exiting. In the end, there may be
more or fewer arrows, but using one influx and one efflux is a good default, since it
reminds us later to include these processes when we set up equations. Indeed, a
dependent variable tends to deplete without influx and to keep accumulating
without efflux. Of course there are situations where depletion or accumulation is the
desired behavior, and, if so, we can secondarily remove the corresponding process.
However, these situations are quite rare. We may also use the default in Figure
2.9(B), which reminds us that many processes are regulated by other variables and
that the variable in question may send out signals that could affect the influxes or
effluxes of other variables. Independent variables typically do not have an influx,
but they have an efflux if they feed material into the system. Once all variables are
defined, they need to be functionally connected: the efflux of one variable may
become the influx of another, and the signal sent by one variable may affect the
influxes or effluxes of other variables.
Let’s illustrate this construction of a diagram with the example of a population
that is experiencing the outbreak of an infection. The model is intended to answer
some of the questions mentioned earlier. Following some old ideas of Kermack and
McKendrick [7], we use a similar terminology and keep things as simple as possible;
one should, however, note that thousands of variations on this model have been
proposed since Kermack and McKendrick’s days. We begin by defining just three
dependent variables, namely, the number of individuals susceptible to the disease
(S), the number of infected individuals (I), and the number of individuals that are
“removed” (R) from the two pools, because they have acquired immunity. We sup-
pose that all individuals could be in contact with each other, at least in principle,
and assume that a certain percentage of the immune individuals (R) lose their
immunity and become susceptible again. We also allow for the possibility that indi-
viduals are born or immigrate and that individuals may die while infected. A dia-
gram summarizing this population dynamics is shown in Figure 2.10. Processes
(A) (B) Figure 2.9 Default core modules. When

designing a diagram of a system, it is
beneficial to augment the display of each
variable with an influx and efflux (A) and
variable Xi variable Xi possibly with modification signals that
affect the variable or are sent out by the
variable (B).
Figure 2.10 Diagram of a model for the

vbirth vdeath spread of an infectious disease within
+ a population. The model contains three
S I
vinfection
dependent variables: S, the number of
vimmunity individuals susceptible to the disease; I,
the number of infected individuals; and
vsusceptibility R, the number of immune (“removed” in
R
the original terminology) individuals. Each
process, in which people move from one pool
to another, is shown as an edge and coded
as v with a specific index. The process vbirth
represents birth or immigration, vinfection is
the infection process, which requires contact
such as infection, recovery, and death are denoted with indexed variables v. For
between a susceptible and an infected
instance, the infection process is coded as vinfection and involves the contact between individual. The process vimmunity models the
individuals of types S and I. The label v is often used in systems models, because it fact that infected individuals may acquire
characterizes the velocity of a process. Some arrows have no source (vbirth), which immunity. Infected people may also die
means that we do not model explicitly where the individuals (or material) come from the disease (vdeath). Finally, vsusceptibility
from. Similarly, arrows may not have a target (vdeath) because we do not account for represents the rate with which individuals
their fate; individuals or molecules simply leave the system. lose their immunity and become susceptible
again. An immediately visible simplification
is that S and R do not die, unless they are
2.5 Model Equations first infected. The model does not include
independent variables.
Once we have completed our model diagram and established the lists of compo-
nents and processes, we need to translate this information into equations that
reflect the chosen model structure. At first, these equations are symbolic, which
means that no numerical values are yet assigned to the processes that characterize
the system.
For the infectious disease problem, we choose a simple differential equation
model, called an SIR model, which by its nature is moderately explanatory, dynamic,
continuous, deterministic, and independent of spatial aspects. More complicated
choices could include random processes associated with the infection of suscepti-
ble individuals or age distributions within the population with different degrees of
susceptibility and recovery.
Like Kermack and McKendrick [7], we use as a model structure the simple and
intuitive process description of mass action kinetics. This structure entails that
the terms in each equation are set up either as constants (birth and immigration)
or as a rate constant multiplied with the dependent variables that are directly
involved in the process. This strategy reflects that the number of individuals mov-
ing from one pool to another depends on the size of the source pool. The more
infected individuals there are, for instance, the more people will acquire
immunity.
The specific construction is implemented one differential equation at a time and
requires us to introduce new quantities, namely, the rates with which the various
processes occur. The change in S is governed by two augmenting processes, namely
birth and immigration (vbirth) with a constant rate rB, and replenishment of the sus-
ceptible pool with individuals losing immunity (vsusceptibility) with rate rS. At the begin-
ning, no immune individuals may be around (R = 0), but the model accounts for the
possibility, which may become reality later. The pool S is diminished by the infec-
tion process (vinfection), which has a rate rI and depends on both S and I, because an
infection requires that a susceptible and an infected individual come into contact.
Thus, the first equation reads
dS .
= S = rB + rS R − rI SI , S (t 0 ) = S0 . (2.1)
dt
Here we have snuck in a new symbol, namely a variable with a dot, S . This notation
is shorthand for the derivative of this variable with respect to time. We have also
added a second equation, or rather an assignment, namely the value of S at the
beginning of the computational experiment, that is, at time t0. Because we don’t
want to commit to a numerical value yet, we assign it the symbol S0. This initial value
must eventually be specified, but we do not have to do it right now. The system does
not contain independent variables.
The equations for I and R are constructed in a similar fashion, and the entire
model therefore reads
.
S = rB + rS R − rI SI , S (t 0 ) = S 0 , (2.2)
.
I = rI SI − rR I − rD I , I (t 0 ) = I 0 , (2.3)
.
R = rR I − rS R , R (t 0 ) = R0 . (2.4)
Here, rR is the rate of acquiring immunity and rD is the rate of death. The set-up
implies that only infected individuals may die, which may be a matter of debate, or
of a later model extension.
We can see that several terms appear twice in the set of equations, once with a
positive and once with a negative sign. This is a very common occurrence, because
they describe in one case the number of individuals leaving a particular pool and in
the other case the same number of individuals entering another pool.
The equations thus constructed are symbolic, because we have not yet commit-
ted to specific values for the various rates. Therefore, this set of symbolic equations
really describes infinitely many models. Many of these will have similar characteris-
tics, but it may also happen that one set of parameter values leads to very different
responses than a different set.
2.6 Parameter Estimation

No matter which particular type of model we select, the model structure alone is
usually not sufficient for a comprehensive assessment. For most specific analyses,
we also need numerical values for the model parameters. In a statistical model, such
a parameter may be the arithmetic mean. In a stochastic system, it might be the
average rate of an event in a binomial trial. In a deterministic model, we usually
need sizes or quantities of variables, rates of processes, and characteristic features of
the model components. As an example, let’s consider a mathematical function that
is often used in biochemistry to describe the conversion of one type of molecule,
called the substrate, into a different type, called the product. This conversion is usu-
ally facilitated by an enzyme. We will discuss biochemical conversion processes,
along with their context and rationale, in Chapters 4, 7, and 8. Here it is sufficient to
state that such a conversion is frequently modeled either with a so-called Michaelis–
Menten function or with a Hill function. Both have the mathematical form
Vmax S n
v(S ) = . (2.5)
n
KM + Sn
The parameter Vmax describes the maximal speed (velocity) of the conversion, which is
attained as S becomes very large; in mathematical jargon, Vmax is the asymptote of the
function v(S) as S goes to infinity. The second parameter is the Michaelis constant KM.
This describes a property of the enzyme facilitating the conversion and represents the
value of S where the substrate-to-product conversion v(S) runs at half-maximal speed:
v(KM) = Vmax/2. The third parameter is the Hill coefficient n. In the case of a Michaelis–
Menten function, we have n = 1, while true Hill functions are typically characterized by
n = 2 or n = 4; however, the Hill coefficient may also take other positive values, includ-
ing non-integer values such as 1.25. Depending on the numerical values of these three
parameters, the function has a different appearance (Figure 2.11; Exercises 2.1–2.4).
Parameter values may be obtained in various ways, but their determination is
almost always a complicated process. Indeed, in very many cases, parameter esti-
mation for a systems model from actual data is the most challenging bottleneck in
the entire modeling process. Two extreme classes for obtaining parameter values
are from the bottom up and from the top down; in reality, the estimation is often a
mixture of the two.
In a bottom-up estimation, parameter values are collected for each system com-
ponent and process. For instance, experimental methods of enzyme kinetics could
MODEL ANALySIS AND DIAGNOSIS 37
v(S) Figure 2.11 The parameters of a function

determine its shape. In this case of a Hill
12 Vmax function with n = 4 (v(S) = VmaxS4/(KM4 + S4);
green line), the unknown parameters
Vmax and KM can be determined from
experimental data (red symbols) of v(S)
6 plotted against the substrate concentration
(S). They are given respectively as the
asymptote for large S and the value of S
where v(S) equals Vmax/2. In the example,
S Vmax = 12 and KM = 10 (see Chapters 4, 5,
10 20 30 40 and 8 for details).
KM
be used to determine the KM of an enzyme in the example above (see Chapter 8).
Similarly, in the SIR example, one could tally how many people die from the disease
within a given period of time and translate this measurement into the parameter rD
in (2.3). Thus, parameter values are measured or collected one by one, until the
model is fully specified.
Top-down estimation methods are very different. In this case, one needs experi-
mental measurements of all dependent variables under different conditions or at
successive time points. In the case of a Hill function, the top-down estimation
requires data such as the red dots in Figure 2.11, which measure v(S) for different
values of S. In the SIR example, one would need measurements of S, I, and R at many
time points, following the outbreak of the disease. In either top-down estimation,
the data and the symbolic equations are entered into an optimization algorithm that
simultaneously determines those values of all parameters for which the model
matches the data best.
Chapter 5 discusses methods of parameter estimation in detail. In our example
of an infectious disease outbreak, parameter values could be derived directly or
indirectly from the disease statistics established by a community health center in a
bottom-up or top-down fashion.
Essentially all parameters in our case are rates (which generically describe
the number of events per time), so that it is necessary to decide on a time
unit. For our SIR example, we use days and set the rates correspondingly. Babies
are born (or people immigrate) at a rate of 3 per day, 2% of the infected individuals
actually die per day, and individuals lose immunity at a rate of 1% of R. The other
rates are self-explanatory. The initial values say that, at the start of the model
period, 99% of the individuals in a population of 1000 are healthy, yet suscepti-
ble to the disease (S = 990), that 1% are infected (I = 10), for reasons we do
not know, and that nobody is initially immune (R = 0). These settings are entered
into the symbolic equations (2.1)–(2.4), and the resulting parameterized equations
are thus
.
S = 3 + 0.01R − 0.0005SI , S0 = 990,
.
I = 0.0005SI − 0.05I − 0.02 I , I0 = 10 , (2.6)
.
R = 0.05I − 0.01R , R0 = 0.
In general, parameter estimation is complicated, and Chapter 5 is dedicated to

the description of standard and ad hoc estimation methods. For the remainder of
this chapter, we simply assume that all parameter values are available.
MODEL ANALySIS AND DIAGNOSIS

A typical model analysis contains two phases. The first consists of model diagnos-
tics, which attempts to ensure that nothing is obviously wrong with the model and
assesses whether the model has a chance of being useful. For example, a model in
which a variable disappears altogether when some physiological feature is changed
by just a few percent is not very robust and the actual natural system would not
survive in the rough-and-tumble outside world for very long. After we have received 20
a green light from the diagnostics, we enter the second phase of exploring what the ex
model is able or unable to do. Can it oscillate? Can some variable of interest reach a
level of 100 units? What would it take to reduce a variable to 10% of its normal value?
How long does it take until a system recovers from a perturbation?
function of x
Because we are dealing with mathematics, we might expect that we could directly 10 4x
compute all properties of a model with calculus or algebra. However, this is not nec-
essarily so, even in apparently simple situations. As an illustration, suppose our task
is to determine one or more values of x that satisfy the equation e x – 4x = 0. Is there a
solution? We can draw e x and 4x on the same plot (Figure 2.12) and see immediately
that the two intersect twice. Hence, the equation is satisfied at these two intersec- 0 x
tion points (see Exercise 2.5). Interestingly, and probably surprisingly, while there 0 1.5 3.0
are two clear solutions, there is no algebraic method that would let us compute Figure 2.12 Computer simulation can be
these solutions exactly. Our only systematic alternative is a numerical algorithm a very useful tool for model evaluation.
that (quickly) finds approximate solutions. More generally, situations in modeling The equation ex – 4x = 0 obviously has two
where mathematical analysis must be supplemented with computer methods are solutions, where the red and blue lines
the rule rather than the exception. intersect, but we cannot compute them
directly with algebraic methods. A suitable
computer algorithm easily finds both
2.7 Consistency and Robustness solutions, although only approximately.
Before we can rely on a model, we need to do some diagnostics. This model check-
ing can be a lengthy process, especially if we find flaws or inconsistencies that must
be ameliorated by changing or refining the model structure. The targets of the diag-
nostics are partly biological and partly mathematical. As a biological example,
assume that we are analyzing a metabolic network of pools and channels through
which material flows. As a minimum requirement for consistency, we need to
ensure that all material is accounted for at every pool and every time point, that
effluxes from each pool actually reach one of the other pools, or correctly leave the
system, and that the biochemical moieties of interest are preserved.
Many of the mathematical aspects of diagnostics follow rather strict guidelines, but
the techniques are not always entirely trivial. For instance, we need to assess the robust-
ness of the model, which essentially means that the model should be able to tolerate
small to modest changes in its structure and environment. Techniques for these pur-
poses are discussed in Chapter 4. In lieu of—or in addition to—assessing the model
with purely mathematical techniques, we can learn a lot about the robustness and
some of the features of the model through computational exploration. This is accom-
plished by using software to solve (simulate) the model under normal and slightly
changed conditions. The order of items to diagnose (see Table 2.1) is not important.
For our illustration, let’s start by solving the equations using a computer simula-
tion with the baseline parameters that we set above. We may use for this purpose
software like Mathematica®, MATLAB®, or the freeware PLAS [3]. Figure 2.13 shows
the result, which consists of one time course for each dependent variable. In our
example, the changes within the population are quite dramatic. The subpopulation
of susceptible people almost disappears, because almost everyone gets sick, but the
number of immune individuals eventually exceeds half the population size. Yet, the
infection does not seem to disappear within the first 100 days since the outbreak.
Even after a whole year, some infectives are still present (Figure 2.13B); the situation
is reminiscent of the persistence of diseases like tuberculosis. In general, we should
always check whether the simulation results appear to make sense or are obviously
wrong, before we launch any comprehensive mathematical diagnostics. In our
example, we cannot judge whether the model yields the correct dynamics, but at
least there is nothing obviously wrong or worrisome.
Let’s do some computational exploration. The first question is whether the
model in (2.6) reaches a state where none of the variables changes anymore. The
answer is yes: we can show such a steady state by solving the equations for a suffi-
cient time range (note that the number of immune individuals still climbs after one
year). After a long time, the variables eventually reach the values S = 140, I = 150,
and R = 750, and this steady state is stable. How we assess stability is discussed in
Chapter 4. If we compute the total population size after several years, we note that
there are now more people (1040) than at the beginning (1000). Is that reason to
worry? No, it just means that the initial values were below the “normal” steady-state
(A) Figure 2.13 Spread of an infectious

1000 disease through a population over
100 days and 1 year. Simulations for
shorter time periods (A) can provide detailed
information on early transition behaviors,
while simulations for extended time periods
number of individuals
I R (B) can give hints regarding long-term and

steady-state behaviors.
500
0
0 50 100
(B) time
1000
S
R
500
0 182.5 365
time
size of the population, which is driven by the birth and death rates. What happens
when we start with S = 1500? Because the steady state is stable, the population will
shrink and in the end approach 1040 again (confirm this!).
Much of the remaining diagnostics characterizes the sensitivity of the model.
The generic question here is: If one of the parameters is changed a little bit, what is
the effect? Sensitivity analysis can be done mathematically in one big swoop (see
Chapter 4). Specifically, all changes in steady-state values, caused by a small (for
example, 1%) change in each of the parameters, are computed with methods of dif-
ferentiation and linear algebra. However, we can also explore important sensitivity
features by changing a parameter value manually and checking what happens. For
instance, we could lower the infection rate to 20% of the present value: rI = 0.0001.
The consequences are dramatic, as Figure 2.14 demonstrates. Two pieces of good
1000
S
500
Figure 2.14 A model can be sensitive to

changes in some parameters and robust
I
against changes in others. As an example,
the spread of the disease is rather different if
0
0 365 730 the infection rate is smaller (here, rI = 0.0001;
time compare with Figure 2.12).
news are that fewer people become infected and that the total population size
increases to 1650. The bad news is that more people remain susceptible.
If one has the luxury of many data, it is beneficial not to use them all for model
design and parameterization and instead to keep some data for external valida-
tion. This validation may be accomplished intuitively or statistically. In the former
(obviously simpler) approach, one compares the results of simulations with actual
data and judges whether the two are sufficiently close. Often, but not always, such
a simulation begins with the system at steady state and a stimulus that perturbs
this state. Sometimes, one might be satisfied with an agreement in qualitative
behavior, such as “If the input to the system is increased, these variables should go
up, while those should decrease in value.” In other cases, one may expect a model
to be semiquantitative: “If the input to the system is doubled, variable 3 should
increase to between 120% and 140%.” A much more stringent criterion is true
numerical agreement. In addition to this type of intuitive assessment, one may use
statistical methods to evaluate the quality of fit between a model prediction and
observed data.
One should note that a good fit is not the only criterion of quality. A particularly
prominent counterexample is overfitting, where the model contains more parame-
ters than can be estimated reliably from the data. While the model fit for the training
data may be excellent, the model tends to fail if new data are analyzed. Methods
have been developed to diagnose different types of residual errors and problems of
overfitting [8, 9].
2.8 Exploration and Validation of Dynamical Features

Assuming that the model has survived the diagnostics of the previous phase, it
is now time to explore its range of dynamical features, keeping in mind that
even a comprehensive analysis might not detect all possible flaws and failures.
In fact, since every model is at best a good approximation of reality, it is only a
matter of time before this approximation leads to discrepancies with actual
observations.
The exploration of the model consists of testing hypotheses regarding scenarios
that had not been used for the model design. This process is often called a simula-
tion. Some of the scenarios may correspond to situations that had already been
tested biologically. In this case, a successful simulation may be used for supporting
the validity of the model or for explanations of the internal operations that connect
a particular input to the observed (and simulated) output. Such an explanation may
lead to new hypotheses of the type “since process A is apparently more influential
than process B, inhibition of A should have a stronger effect than inhibition of B.”
The new hypothesis may be tested mathematically, biologically, or both. It is also
possible that the model makes predictions that are counterintuitive. In this case we
need to check both our intuition and the model. If the model is correct, it could lead
to new explanations that were not evident before.
Continuing beyond the diagnostic explorations, the scenarios of this phase
are typically a little bit more complicated, and many will require changes in the
model structure, that is, in the equations. These types of explorations can be used
to evaluate the importance of assumptions, simplifications, and omissions. For
instance, we assumed in the infectious disease model that only infected individ-
uals die. What happens if we add mortality terms to the equations for S and R?
What happens if immunity is permanent, as it is for most people in the case of
chickenpox? Is it of consequence if some individuals are infected, but do not
show symptoms?
As a specific example, let’s investigate what happens if we vaccinate before the
outbreak. The first task is to figure out where to enter vaccination into the equations.
Any effective vaccination before the outbreak moves susceptible individuals into
the R pool. In other words, the only change occurs in the initial values, where S0 is
lowered and R0 correspondingly elevated. The exact numbers depend on the effi-
cacy of the vaccination campaign. It is not difficult to imagine that complete and
perfect vaccination would render S0 = 0, R0 = 990. One could also model that the
vaccination would not necessarily prevent the disease altogether but only reduce
the infection rate, allow faster immunity, and reduce the number of individuals in
the R pool who become susceptible again.
Another scenario involves quarantine for infected individuals. Ideally, the entire
pool I would immediately be subjected to quarantine until they recover, thus chang-
ing the initial value to I0 = 0. It is easy to see in this extreme case that no infection can
take place, because the infection term rISI equals zero. Therefore, no susceptibles
move from S to I, and rISI remains zero. A more interesting question is: Is it really
necessary to quarantine every infected person? Or is it sufficient to quarantine only
a certain percentage? Let’s explore this question with simulations. Suppose we con-
tinuously quarantine some of the infected individuals. Of course, actually sending
infected individuals into quarantine is a discrete process, but we can model the
process approximately by subtracting infectives with a constant rate from (2.3),
which becomes
.
I = rI SI − rR I − rD I − rQ I , (2.7)
where rQ is the rate of quarantining (Box 2.5).

Simulating the system reveals that quarantining even at a rate of rQ = 0.1 has
noticeable, positive effects: the number of infected individuals is roughly
halved throughout the first 100 days of disease (Figure 2.15A). If we quarantine
with a higher rate of rQ = 0.4, a few individuals still become infected, but the
infection level is very low (Figure 2.15B). If we quarantine an even higher per-
centage, the number of infectives quickly goes to zero and, perhaps surprisingly
at first, the susceptible pool increases. Why is this? It is due to births and immi-
gration, and to the fact that nobody in the model dies, unless she or he is first
infected.
Case closed? Not quite. Something interesting happens if we let the disease pro-
cess continue for a longer period of time. Because the quarantine is not perfect,
some infectives remain in the population, and this can lead to repeated, smaller
outbreaks. For instance, for rQ = 0.1, a secondary outbreak occurs after about 165
days. Smaller outbreaks follow before the system finally settles in a state that is still
not disease-free, although the infected subpopulation is smaller than in the case
without quarantine (Figure 2.16).
BOX 2.5: THE RATE OF QUARANTINING IN THE SIR MODEL
The formulation in (2.7) is only an approximation of real-world, repeated I = −rQ I . (1)

quarantining, because the change in I is continuous, whereas the actual
quarantining process is discrete. For instance, one might imagine an
idealized situation where, every morning exactly at 8:00, one would Starting with I0 = 100 and rQ = 0.1, the number of infectives over
send 10% of all infected individuals into quarantine. The model does time is reduced roughly by 10% per time unit (see Table 1).
not represent this daily routine but, instead, subtracts infectives at a For larger values of rQ, the correspondence is less accurate. For
constant rate throughout the day. Nonetheless, the approximation is instance, for rQ = 0.5, the number of infectives decreases from 100
sufficient for our exploration, as we can see from the following argument only to 60.65 and 36.79 between times 0, 1, and 2, respectively,
with a simplified situation. Suppose we start with 100 infectives, nobody which roughly corresponds to 40% decreases per time step, rather
becomes infected, nobody dies or becomes immune, and the only than 50% decreases. The basis for the discrepancy is similar to
change in I is quarantining. In this simplified scenario, (2.7) becomes continuously compounding interest.
TABLE 1: DECREASE IN THE NUMBER OF INFECTIVES IN THE SYSTEM WHEN THE CONTINUOUS QUARANTINING RATE
IS rQ = 0.1; THE DECREASE ROUGHLY CORRESPONDS TO 10% PER TIME UNIT
t 0 1 2 3 4 5 6 7 8 9 10
I 100 90.48 81.87 74.08 67.03 60.65 54.88 49.66 44.93 40.66 36.79
(A) Figure 2.15 Quarantining infectives

1000 S has a positive effect on the population.
I (A) Results of continuous quarantining at a
R
rate of rQ = 0.1 (solid lines) in comparison
with no quarantine (dotted lines).
(B) The corresponding result for rQ = 0.4.
500
0
0 50 100
time
(B)
1000
S
I
R
500
0 50 100
time
Now that we have seen the positive effects of quarantine, we can explore ques-
tions such as: Is it better to quarantine or to reduce the infection rate? Reduction of
the infection rate by a certain percentage is easy to implement: we simple reduce
the value of the rate rI. Figure 2.17 shows two results: the outcomes of quarantining
and reducing the infection rate are quite different.
It is clear that the possibilities for new scenarios are endless. Thankfully, once
the model is set up, many slightly altered scenarios are very easy to implement, and
the process can even become addictive, if we try to find just the right parameter
combinations for a desired outcome.
1000 S
I
R
500
Figure 2.16 Quarantining may lead to

further, milder outbreaks. For rQ = 0.1,
the disease seems to be under control
0 initially, but a second outbreak occurs after
0 365 730 about 165 days, followed by later, smaller
time outbreaks.
MODEL USE AND APPLICATIONS 43
(A) Figure 2.17 Reduction in the infection

1000 S rate has a drastically different effect than
I quarantining. (A) The infection rate rI has
R been lowered to half the original value.
(B) rI has been lowered to 20% of the original
value.
500
0
0 182.5 365
time
(B)
1000 S
I
R
500
0 182.5 365
time
MODEL USE AND APPLICATIONS

The most common uses of a model (see Table 2.1) may be classified in two categories:
• Applications of the model as it is, including slight extension.
• Targeted manipulation and optimization of the model structure toward a spe-
cific goal.
2.9 Model Extensions and Refinements

As an example of a model extension, suppose that it becomes clear that many
infected individuals do not show any serious symptoms and often do not even
know that they carry the disease. In the jargon of SIR models, such individuals are
called asymptomatics. How can the model account for them? The extension is
really a mini-version of the model design and analysis phases. First, we define a
new dependent variable, A, and identify how it interacts with the other variables in
the system (2.2)–(2.4). As for the initial model design, it is a good idea to sketch
these interactions in a diagram. At first, we might treat A just like I, except that we
should use different names for the processes, because they are indeed different
(Figure 2.18A). However, should we also consider that the contact between S and
I could lead to A, and that the contact between S and A might lead to I (Figure
2.18B)? The answer cannot be given with mathematical arguments and has to be
based on biological or medical features of the disease process. For instance, if all A
result from a weaker strain of the disease vector, an infection of S by A might always
result in A and an infection by I might always lead to I; let’s call this situation Case
1. But if the difference between A and I is due to an individual’s general health
(A) vinfection Figure 2.18 Diagrams accounting for

vdeath
asymptomatics. (A) The inclusion of A
requires new processes (red) with their own
+ names. (B) Further interactions (blue and
I
vimmunity green) are needed if an infection of S by I can
vbirth result in A or vice versa.
S R
vinfection–A
vimmunity–A
A
+
vdeath–A
vsusceptibility
(B) vinfection
vdeath
+
I
+ vimmunity
vbirth vA–I
S R
vI–A
+ vimmunity–A
A
+
vdeath–A
vinfection–A
vsusceptibility
status and his or her age and genetic predisposition, we should probably consider
cases where an infection of S by A can indeed result in I, and vice versa; let’s call
this situation Case 2.
For Case 1, the equation for the asymptomatics is similar to that for the infec-
tives, (2.3), namely,
.
A = rASA − rR− A A − rD− A A , A (t 0 ) = A0 . (2.8)
While the structure of the equation is the same as in (2.3) for I, the parameters have
different names and potentially different values. For instance, the death rate rD–A
might be different, because the disease in A is apparently not as bad as for I. Similar
arguments might hold for rA and rR–A. Does that cover Case 1? No, for sure we also
need to adjust the equations for S and R. It could even be that R is different for
asymptomatics and for infectives (why?); but let’s assume that that is not the case.
The adjusted equations for S and R are
.
S = rB + rS R − rI SI − rASA , S (t 0 ) = S 0 , (2.9)
.
R = rR I + rR− A A − rS R , R (t 0 ) = R0 . (2.10)
For Case 2, we need to add more terms: one each to the equations for A and I and
two terms to the equation for S; the equation of R is not affected:
.
A = rASA + rI − ASI − rR− A A − rD− A A , A (t 0 ) = A0 , (2.11)
.
I = rI SI + rA−I SA − rR I − rD I , I (t 0 ) = I 0 , (2.12)
.
S = rB + rS R − rI SI − rASA − rI − ASI − rA−I SA, S (t 0 ) = S 0 . (2.13)
(A) Figure 2.19 Effect on asymptomatics

1000 S A on the disease progression. (A) If
I asymptomatics recover twice as fast as I,
A they quickly disappear from the population,
R
whether their death rate is zero (as shown)
or not. (B) If asymptomatics recover with the
same rate as I, and if they do not die,

the infectives essentially disappear, and
500 the population grows (note different scale).
However, the disease does not disappear;
it is present in a growing number of
asymptomatics.
0
0 182.5 365
time
(B)
1500 S
I
A
R
750
0
0 182.5 365
time
As an illustration, let’s simulate Case 1. For the numerical settings, we will retain
most values from our earlier example (2.6). Regarding A, we assume that initially
half of the infected individuals are asymptomatic, so that A(t0) = I(t0) = 5. Further-
more, we assume that the infection rate for A and I is the same (rA = rI = 0.0005), that
no A die from the disease (rD–A = 0), and that they recover more quickly than
I (rR–A = 2rR = 0.1). The result of these settings is that the asymptomatics disappear
quickly and the disease ultimately assumes the same profile as we saw in Figure 2.13
(Figure 2.19A). Interestingly, the fact that none of the asymptomatics die does not
have much effect (results not shown). By contrast, if no A die, but the recovery rate
for A and I is the same, the outcome is very different: I essentially disappears, while
the R and A subpopulations grow to much higher levels (Figure 2.19B) The explana-
tion is that essentially nobody in the model dies anymore. To make the model more
realistic, one would need to introduce death rates for S, R, and A.
2.10 Large-Scale Model Assessments

In reality, most parameters of biological systems are not entirely fixed but can vary
within certain ranges, for instance, due to inter-individual variability. If the biologi-
cally relevant ranges of all parameter values are known or can be estimated, the model
may be used to evaluate best- and worst-case scenarios, along with the most likely
situations. This analysis can be accomplished with targeted stochastic simulations in
which parameters are varied, as above, or with large-scale, automated simulation
studies, in which thousands of parameter combinations are screened. The latter
method is often referred to as Monte Carlo simulation. The principle is the following.
Suppose there are 10 parameters of interest and the range for each is more or less
known. For instance, there could be information that the infection rate in our example
is between 0 and 0.002 and that the most likely value is 0.0005, as we set it in our base-
line model. Suppose that this type of information is available for the other nine param-
eters as well. A Monte Carlo simulation consists of the following iteration:
1. Automatically draw a random value for each parameter, according to the avail-
able information about its range.
2. Solve the equations for the given set of values.
3. Repeat this procedure thousands of times.
4. Collect all outputs.
The overall result of such a Monte Carlo simulation is a visualization or listing of a

very large number of possible outcomes and their likelihoods. Furthermore, statisti-
cal analysis of the output conveys insights into which parameters or parameter
combinations are most or least influential.
As an example, let’s use our infection model (2.6) again and test the effects of
different infection and death rates. We use a very simple Monte Carlo simulation,
consisting of just 20 random choices for the parameter rI, which is selected ran-
domly from the interval [0, 0.001] and, simultaneously, 20 random choices for rD,
which is selected randomly from the interval [0, 0.04]. The lower bound of these
intervals (0 in both cases) means no infection (death) at all, while the upper bounds
(0.001 and 0.04) correspond to twice the corresponding rates in (2.6). The simula-
tion proceeds as follows:
1. Select a random number for rI from the interval [0, 0.001].

2. Select a random number for rD from the interval [0, 0.04].
3. Plug these two values into (2.6).
4. Simulate the model for a time period of 365 days.
5. Repeat Steps 1–4 many (here just 20) times.
6. Superimpose all plots and analyze results.
The 20 time courses of S are shown in Figure 2.20 for two Monte Carlo simulations.
Owing to the random nature of the 20 picks of rI and rD, these two simulation runs of
exactly the same model structure lead to different results. We can see that most results
lead to final values of S between 100 and 500, but that some combinations of rI and rD
lead to much higher values. A statistical analysis of this type of simulation would
reveal that very low death rates lead to high (and sometimes even increasing) values
of S; note one such case in Figure 2.20B. Other steady-state values, peak values, and
time courses may be analyzed in the same fashion. In a more comprehensive Monte
Carlo simulation, one would simultaneously select random values for several param-
eters and execute so many simulations that the results would no longer change much
from one run to the next. Uncounted variations on this theme are possible. As a nota-
ble example, a parameter value does not have to be chosen from its interval with equal
probability, and one could instead specify that values close to the original (0.0005 and
0.02 in our example) are chosen more frequently than extreme values like 0.0000001.
The model may also be used to screen for particularly desirable outcomes. For
instance, health officials will be interested in shortening the period during which
many individuals are sick, and in minimizing the chance of infection. To develop
ideas for combined interventions, a Monte Carlo simulation could be a first step. It is
very efficient to try out alternative strategies toward these goals with a model and,
possibly, to optimize them with numerical methods from the field of operations
research. Of course, predicted results always have to be considered with caution,
because the model is a vast simplification of reality. Nonetheless, a good model is able
to guide and correct our intuition, and the more reliable the model becomes through
testing and refinements, the more competent and important its predictions will be.
2.11 Questions of Design

An academically important use of a model is to explain design and operating prin-
ciples, which offer the closest resemblance to “laws” that biology can offer. The
(A) Figure 2.20 Two results of a Monte Carlo

1200 S
simulation. The set-up in the simulations in
(A) and (B) is exactly the same; namely, (2.6)
is simulated with randomly chosen values
for the infection and death rates. Owing
to the probabilistic nature of this process,
the results are quite different in some

respects, but similar in others. Notably, the
600 simulation in (B) includes the random choice
of a very low death rate, which allows S to
grow. However, most other patterns in time
courses are rather similar, although they
differ numerically. Note the different scales
on the vertical axes.
0
0 182.5 365
time
(B)
2000 S
1000
0
0 182.5 365
time
generic question to be asked is: Why is this system organized in this particular fash-
ion and not in an alternative fashion? For instance, to increase the concentration of
a metabolite, one could increase the amount of substrate that is used for its produc-
tion, increase the production rate, or decrease the degradation of the metabolite. By
analyzing these options comparatively with two models (for example, one showing
increased production and the other showing decreased degradation), it is possible
to identify advantages and disadvantages of one design over the other. For instance,
we might find that a model with one design responds faster to a sudden demand for
some product than the alternative. This interesting topic will be discussed in much
more detail in Chapters 14 and 15.
2.12 Simplicity versus Complexity

This chapter has given you a flavor of what modeling means and what we can do
with mathematical models in systems biology. Chapters 3 and 4 will describe stan-
dard techniques for setting up, diagnosing, analyzing, and using models of different
types. Before we dive into the nitty-gritty details of the mathematics of modeling, we
should realize that the complexity of the mathematical representations and tech-
niques is not necessarily correlated with the usefulness of a model. Certainly, some
models have to be complicated, but sometimes very simple models are “better”
than complicated counterparts. For instance, many growth models are rather sim-
ple, even if they describe very complicated biological systems, such as the human
body. There is no chance that we will come up with a complete physiological model
of the human body in the foreseeable future, yet the simple growth curves of chil-
dren show amazingly small variation from one child to the next (Figure 2.21). Pre-
dictions based on these curves are much more reliable than predictions from any
large physiological model that currently exists.
Figure 2.21 Growth curves of girls

from birth to 3 years. Considering how
complicated the human body is, the normal
growth trends of 50% of the girls are within
surprisingly narrow bounds. (From the
National Center for Health Statistics.)
An example at the other end of the spectrum is the mathematical model in Figure
2.22. It adorns a T-shirt and the artist has even provided the correct answer, upside
down below the formula, as age = 50. The display is certainly intimidating to mere
mortals, but does that mean that we should judge this “model for age” as good? The
answer is a resounding No, for the following (and many more) reasons. First, the
model does not explain or predict anything. Most importantly, does the T-shirt show
the age of the wearer? Only if he or she happens to be 50 at the time; wearing the shirt
next year, the implied age will be wrong. Age is obviously a function of time, but
the T-shirt does not account for time. In fact, there is no true variable, except for the
integration variable x, which is merely a placeholder and disappears as soon as the
integration has been executed. There are no true parameters either, because the only
candidates, α and n, disappear in the summation, and the answer is always 50. In
other words, the formula does not permit any adaptation to the wearer and his or her Figure 2.22 Awe-inspiring, useless model
on a T-shirt. While the formula looks
true age. Instead, the display is a mathematically unnecessary complication of the
intimidating, it is simply a complicated way
simple statement age = 50 and not much more. If we are interested in a better T-shirt of representing the number 50. The “model”
model of age, which correctly captures the wearer’s age, this model must certainly contains no real parameters and only shows
contain time as a variable. Barring magic or serendipity, the formula must also con- the correct age if the wearer happens to be
tain a place for explicit or implicit information about the wearer’s year of birth. This 50 years old.
EXERCISES 49
information may be given directly or in some convoluted fashion, but it must be

there as some (potentially explicit or hidden) parameter. We also need to decide on
the accuracy of the improved T-shirt model. If age in years is sufficient, the model will
certainly be simpler than if age has to be computed to the minute.
The message from these two examples is that the simplicity or complexity of a
model is not indicative of its quality. Of course, there are situations where compli-
cated mathematics is needed, but there are also many cases where simplicity will do
just fine, if not better. The key criterion of quality of a model is, instead, its useful-
ness. Mathematical models are tools, and tools are only good and helpful if they
allow us to solve problems. Sometimes, simple tools are our best choice, and at
other times the complexity of the model must match the complexity of the system
we want to analyze. Because usefulness is the most important criterion, every model
analysis should begin with a clear understanding of why we want to design a model
and what this model is supposed to accomplish. Time and effort spent on these two
questions are the most prudent investments in the realm of modeling.
EXERCISES
2.1. Explore the range of shapes that a Hill function (2.5) 2.6. Write equations for the small protein system in the
can attain. First, keep the settings Vmax = 12 and signaling cascade of Figure 2.7. Model each process
KM = 10 and use different values for n (n = 4, 6, 8; 3, as a rate constant times the concentration of the
2, 1, 0.5, 0.2). Then, set n = 4 and vary Vmax or KM variable involved. For instance, the phosphoryla-
individually or both simultaneously. Discuss tion of P0 toward P1 should be formulated as r01P0.
differences and similarities in the shapes of the Begin with P0 = P1 = P2 = 1 and set all rate
resulting graphs. constants to 0.5. Implement the system in a
software program and simulate it. Explore the
2.2. The function L(S) = 1/(1 + e10 −S) is a shifted logistic
dynamics of the system by changing the initial
function, whose graph is an S-shaped curve.
values of P0, P1, and P2. Discuss the results.
Through trial and error, determine parameter
Now change the rate constants one by one or in
values for the Hill function (2.5) that approximate
combination, until you develop a feel for the
L(S) as closely as possible.
system. Discuss the results and summarize your
2.3. Many processes in biology can be approximated insights into the system.
quite accurately with power-law functions of the For Exercises 2.7–2.19, use the infectious disease model
type P(X ) = γX f (one variable) or P( X 1 , X 2 , … , X n ) = γ X 1f1 in
X 2f2(2.2)–(2.4)
X nfn and (2.6) and be aware that several of
P( X 1 , X 2 , … , X n ) = γ X 1 X 2 X n (n variables; n = 2, 3, 4, . . .).
f1 f2 fn
these exercises are open-ended. Before you execute a
The positive parameter γ is a rate constant and the particular simulation, make a prediction of what the
real-valued parameters f are kinetic orders. model will do. After the simulation, check how good your
Explore the shapes of these power-law functions for predictions were. Consider a problem “done” when your
different values of γ and f. For γ use (at minimum) predictions are consistently correct.
the values 0.1, 1, 10, and 100. For the parameter f
use (as a minimum) the values −2, −1, −0.5, −0.1, 0, 2.7. Change each initial value in (2.6) (up and down)
+0.1, +0.5, +1, +2. Display your results as tables and study the consequences.
and/or graphs. Summarize the results of your 2.8. What would it mean biologically if rI = 0 or if rS = 0?
analysis in a short report. Test these scenarios with simulations.
2.4. Compare the Michaelis–Menten function 2.9. What happens if no babies are born and nobody
V (S ) = Vmax S (K M + S ), with Vmax = 1 and KM = 2, immigrates? Explore what happens if nobody in the
with the power-law function P(S ) = γ S f , where model dies.
γ = 0.3536 and f = 0.5. Discuss your findings. 2.10. Double the death rate in the model and compare
Choose different values for Vmax and KM and the results with the original simulation.
determine values of γ and f such that the power-law 2.11. Alter the model in (2.2)–(2.4) so that susceptible (S)
approximation matches the Michaelis–Menten and immune (R) individuals may die from causes
function well for some range of values of S. Write a other than the infection. Discuss whether it is
brief report about your results. reasonable to use the same rate constant for the
2.5. Compute the two solutions for the equation death of S and R individuals. Implement your
ex – 4x = 0 in Figure 2.12 by trying out different proposed alterations in (2.6) and execute
values. Write down your thought processes simulations demonstrating their effects.
when selecting the next value to be tried. 2.12. Change the infection rate, starting from the value of
Remember these thought processes when you 0.0005 in (2.6), by raising it to 0.00075, 0.001, 0.01,
study Chapter 5. and 0.1, or by lowering it stepwise toward 0. Discuss
trends in consequences on the dynamics of the 2.20. Set up symbolic equations for a two-stage cancer
system. model that consists of the following processes.
2.13. Change the infection rate or the rate of acquiring Normal cells divide; some of them die. In a rare
immunity to different values and solve the system case, a normal cell becomes “initiated,” which
until it reaches its steady state, where no variable means that it is a possible progenitor for tumor
changes anymore. For each scenario, check the formation. The initiated cells divide; some die. In a
number of deaths per unit time, which is given by rare case, an initiated cell becomes a tumor cell.
rD I. Record and discuss your findings. Tumor cells divide; some die.
2.14. Predict the consequences of partial vaccination. 2.21. Implement the system in Box 2.2 in computer
Implement a vaccination strategy in the model and software such as Mathematica®, MATLAB®, or
test your predictions with simulations. PLAS [3], and compute solutions with different
values for A, and with slightly changed initial values
2.15. Compare the effectiveness of quarantine and of x or y. Explore what happens if A is close to or
vaccination. equal to zero.
2.16. Implement scenarios where certain percentages 2.22. Use the binomial model in Box 2.3 to compute the
of asymptomatics become infectives. Test the probability that exactly 5 out of 10 flips of a coin are
consequences with simulations. heads and 5 are tails. Compute the probability that
2.17. How could the model in (2.2)–(2.4) be expanded to exactly all 10 flips of a coin are heads.
allow for different levels of disease susceptibility 2.23. Use the Poisson model in Box 2.4 to compute the
that depend on age and gender? Without actually probability that exactly 5 out of 10 flips of a coin are
implementing these changes, list what data would heads and 5 are tails. Compute the probability that
be needed to construct such a model. exactly 50 out of 100 flips of a coin are heads and
2.18. Describe in words what data would be required to 50 are tails.
develop an infectious disease model like that in 2.24. Use the Poisson model in Box 2.4 to compute the
(2.2)–(2.4) that accounts for the gradual spread of probability that at most 3 out of 20 flips of a coin are
the disease throughout a geographical area. heads.
2.19. Without actually implementing them in equations, 2.25. Implement the equation from Box 2.5 and compute
propose extensions that would make the SIR model the numbers of infectives over time for different
in (2.2)–(2.4) more realistic. For each extension, rates rQ. In parallel, design a model where a certain,
sketch how the equations would have to be fixed percentage of infectives is quarantined at t = 0,
changed. 1, 2, . . . , but not in between. Compare the results of
the two modeling strategies.
REFERENCES
[1] Qi Z, Miller GW & Voit EO. Computational systems analysis of [6] Mood AM, Graybill FA & Boes DC. Introduction to the Theory
dopamine metabolism. PLoS One 3 (2008) e2444. of Statistics, 3rd ed. McGraw-Hill, 1974.
[2] Thompson JMT & Stewart HB. Nonlinear Dynamics and Chaos, [7] Kermack WO & McKendrick AG. Contributions to the
2nd ed. Wiley, 2002. mathematical theory of epidemics. Proc. R. Soc. Lond. A
[3] PLAS: Power Law Analysis and Simulation. http://enzymology. 115 (1927) 700–721.
fc.ul.pt/software.htm. [8] Raue A, Kreutz C, Maiwald T, et al. Structural and practical
[4] Meng C, Shia X, Lina H, et al. UV induced mutations in identifiability analysis of partially observed dynamical models
Acidianus brierleyi growing in a continuous stirred tank reactor by exploiting the profile likelihood. Bioinformatics 25 (2009)
generated a strain with improved bioleaching capabilities. 1923–1929.
Enzyme Microb. Technol. 40 (2007) 1136–1140. [9] Voit EO. What if the fit is unfit? Criteria for biological systems
[5] Fujii R, Kitaoka M & Hayashi K. One-step random mutagenesis estimation beyond residual errors. In Applied Statistics
by error-prone rolling circle amplification. Nucleic Acids Res. for Biological Networks (M Dehmer, F Emmert-Streib &
32 (2004) e145. A Salvador, eds), pp 183–200. Wiley, 2011.
FURTHER READING
Adler FR. Modeling the Dynamics of Life: Calculus and Probability Haefner JW. Modeling Biological Systems: Principles and
for Life Scientists, 3rd ed. Brooks/Cole, 2013. Applications, 2nd ed. Springer, 2005.
Batschelet E. Introduction to Mathematics for Life Scientists, 3rd ed. Murray JD. Mathematical Biology I: Introduction, 3rd ed. Springer,
Springer, 1979. 2002.
Britton NF. Essential Mathematical Biology. Springer-Verlag, Murray JD. Mathematical Biology II: Spatial Models and Biomedical
2004. Applications, Springer, 2003.
Edelstein-Keshet L. Mathematical Models in Biology. McGraw-Hill, Yeagers EK, Shonkwiler RW & Herod JV. An Introduction to the
1988 (reprinted by SIAM, 2005). Mathematics of Biology. Birkhäuser, 1996.
Static Network Models
3
• Identify typical components of static networks
• Understand the basic concepts of graphs and how they are applied to
biological networks
• Describe different structures of networks and their features
• Discuss the concepts and challenges of network inference
• Describe typical examples of static networks in different fields of biology
• Characterize the relationships between static networks and dynamic systems
All molecules in biological systems are components of networks. They are linked to
other molecules through connections that may take a variety of forms, depending on
what the network represents. Molecules may associate loosely with each other, bind
to each other, interact with each other, be converted into each other, or directly or
indirectly affect each other’s state, dynamics, or function. Rather than focusing on
one particular molecular process at a time, the computational analysis of biological
networks has the goal of characterizing and interpreting comprehensive collections
of connections. The methods for these analyses can be automated and applied, using
computers, to very large networks. Typical results are characteristic features of the
connectivity patterns within the investigated networks, which in turn provide clues
regarding the functions of particular components or subnetworks. For instance, an
analysis may reveal which components of a network are particularly strongly con-
nected, and this insight may suggest important roles of these components in differ-
ent situations. The analysis may also show which variables act in concert with each
other, and whether the connectivity pattern within the network changes, for exam-
ple, during development and aging or in response to perturbations, such as disease.
Many concepts and methods of network analysis are actually quite simple and
intuitive when studied in small networks. This simplicity is crucial for the real power
of network analyses, namely, their scalability to very large assemblages of molecules
and their ability to reveal patterns that the unaided human mind would simply not
be able to recognize. Thus, computers can be trained to use the basic methods over
and over again, establish network models from experimental data, identify their
global and local features, and ultimately lead to hypotheses regarding their func-
tionality. In this chapter, we will study these types of analyses with small networks,
because they show most clearly what features to look for and how to characterize
them. Several web tools are available for computational analyses of large biological
networks. Prominent examples include the free open-source platform Cytoscape [1,
2], the commercial package IPA from Ingenuity Systems [3], and other software
packages, such as BiologicalNetworks [4] and Pajek [5], all of which offer a wide vari-
ety of tools for data integration, visualization, and analysis.
52 Chapter 3: Static Network Models
STRATEGIES OF ANALYSIS
The ancient Greeks had a saying that all is in flux. Heraclitus of Ephesus (c. 535–
475 BC), to whom this wisdom is attributed, considered change as a central force in
the universe and once allegedly asserted that one can never step into the same river
twice. Indeed, everything in the living world is moving, rather than static—if not
macroscopically, then at least at the molecular level. So we must ask ourselves
whether there is even a point in studying static networks, which by definition do not
move. The answer is a resounding Yes, for a number of reasons. First, even if biological
networks change over time, changes in interesting components are often relatively
slow. For example, even if an organism is growing, its size is almost constant during a
biochemical experiment that takes a few minutes. Second, much can be learned from
assuming that a network is at least temporarily static. A typical example may be the
activity profile of enzymes in the liver, which certainly changes throughout the life of
an organism and even within a 24-hour period, but which may be more or less con-
stant within a period of a few seconds or minutes. And third, even if systems are
dynamically changing, certain important features of these same systems may be
static. The most prominent examples are steady states, which we have already dis-
cussed in Chapter 2 (see also Chapter 4), where material is moving through a system
and control signals are active but, for instance, the concentrations of metabolite pools
are constant. Steady states are usually easier to characterize than dynamic changes
and, once established, can provide a solid starting point for looking into dynamic fea-
tures of a system [6–8].
A very practical reason for focusing on static networks or static aspects of
dynamic systems is that the mathematics needed for static analyses is incomparably
simpler than that required for dealing with temporally changing networks or fully
regulated dynamic systems. As a consequence, even large static systems with thou-
sands of components can be analyzed very efficiently, which is not really the case
for dynamic systems of the same size. This relative simplicity is due to two aspects of
static networks. First, because there is no change with time, the time derivatives in
the differential equations describing the network are all equal to 0. Consequently,
these equations become algebraic equations, which are easier to handle. Second,
the structures of very many static networks are (at least approximately) linear, and
linear algebra and computer science have blessed us with many elegant and power-
ful methods for analyzing linear phenomena. In some cases, these methods may
even be applied to aspects of dynamic systems, if their subsystems are connected in
a static fashion. A surprising example is a group of pacemaker cells in the sino-atrial
node of the heart that can oscillate independently but are synchronized by their
static interaction network, thereby enabling them to generate coordinated impulses
that are required exactly once during each cardiac cycle (see [9] and Chapter 12).
The analysis of static networks can be subdivided in various ways. First, networks
fall into distinct classes, which permit different methods of analysis. In some cases,
for instance, actual material flows through the network, and accounting for this
material permits accurate bookkeeping of all amounts or masses in different parts of
the networks at different points in time. Metabolic pathway systems are prime
examples in this category and therefore play a special role in this chapter. To illus-
trate, suppose there is a single reaction between a substrate and a product within a
metabolic network. We immediately know that the amount of product generated in
this reaction is directly related to the amount of substrate consumed; in fact, the two
must be the same. This equivalence is very beneficial for mathematical analyses,
because it eliminates a lot of uncertainty. By contrast, a signaling network trans-
duces information, and whether or not this information is actually received and
used is typically not relevant to the source of the information. Furthermore, the
main purpose of such a signaling network is often an all-or-nothing response to an
external signal, such as switching on the expression of a gene in response to some
stress. There is no accounting for masses or specific amounts, and even if the signal
were much stronger, the effect would still be the same. Similarly, simply the pres-
ence of some component in a regulatory network may affect the function of the net-
work somewhere else. An example is the inhibition of an enzyme, which may alter
the steady state and dynamics of a metabolic pathway, but which does not change
the amount of the inhibitor itself.
INTERAcTION GRAphS 53
A second distinguishing criterion in the treatment of static networks is the type

of research activity, which can broadly be divided into analysis and construction (or
reconstruction, as it is usually referred to). A typical analysis assumes that the struc-
ture of a network is known and investigates features such as the connectivity among
its components and their contributions to the functioning of the network. By con-
trast, (re-)construction studies attempt to infer the connectivity of an unknown or
ill-characterized network from experimental observations. At first glance, this dis-
tinction may appear to be splitting hairs, but the two aspects require distinctly dif-
ferent approaches, and the latter is incomparably more difficult than the former. It
might seem logical to study the construction of a network first, before launching
into its analysis, but it is easier to discuss and learn the two aspects in the opposite
order. To do so, we will initially assume that we know the structure of the network
and later worry about where our knowledge of that structure may have come from.
Finally, rigorous network analyses have led to the discovery of network motifs,
which are small, yet prominent connectivity patterns that are found time and again
in diverse biological systems and therefore are expected to be particularly effective
for solving certain tasks. Large sections of Chapter 14 are devoted to this fascinating
topic (see also [10]).
INTERAcTION GRAphS
When we are interested in the structure of a network, the specific interpretation of its
components becomes secondary, and it does not truly matter for the mathematical
or computational analysis whether the interacting components are molecules, spe-
cies, or subpopulations: they simply become nodes. Nodes are sometimes also
called vertices (singular: vertex), components, pools, or species, or they may be
termed or described in a variety of other specific or generic ways. The second set of
ingredients of a network consists of the connections between nodes, which are typi-
cally called edges, but may also be termed links, interactions, processes, reactions,
arrows, or line segments. Among the edges, we actually have a bit more flexibility
than with the nodes, because edges may be directed, undirected, or bidirectional. All
three cases are relevant in one biological situation or another, and they may even
appear within the same network (Figure 3.1). No features apart from nodes and
edges are usually admissible.
Taken together, the nodes and edges form graphs, which are the principal
representations of networks. Quite intuitively, the nodes of a network can only be
connected by edges, and two edges can only be connected with a node between
them. Two nodes connected by a single edge are called neighbors. Typical biologi-
cal examples of directed graphs are transcriptional regulatory networks, where
nodes represent genes and edges denote interactions between them. The choice of
a directed graph seems reasonable in this case, because gene A might regulate the
expression of gene B, but the opposite might not be true. Protein–protein interac-
tion (PPI) networks, by contrast, are usually formulated as undirected graphs.
One might be tempted to surmise that all networks and systems could be repre-
sented in this fashion. However, we must be careful with such a sweeping inference.
For instance, metabolic pathway systems do contain nodes (metabolite pools) and
edges (chemical or enzymatic reactions and transport processes). However, they also
(A) (B) (C) (D)
Figure 3.1 Different types of generic networks. Static networks may contain (A) directed, (B) undirected, or (C) bidirectional edges, as well as a
regulation (D), such as inhibition (red) and activation (blue) signals, which require a more complex representation.
Figure 3.2 Euler’s problem of the Seven

Königsberg Bridges of Königsberg. Euler answered the
question whether it is possible to find a path
crossing each bridge exactly once, and in so
doing initiated the field of graph theory.
River Pregel
exhibit regulatory features, with which metabolites may affect the strength, capacity,
or amount of material flow through the edges of a pathway. These regulatory effects
are genuinely different from metabolic conversions and therefore require other types
of representations, for instance, in the form of arrows pointing from a node to an edge,
rather than to another node (see Figure 3.1D) [11]. As an alternative representation, it
is possible to formulate such regulated networks with bipartite graphs, which contain
two sets of different types of nodes. In the case of metabolic networks, one set repre-
sents metabolites, while the other set represents the physical or virtual locations
where the enzymatic conversions between metabolites take place. A modeling
approach following this paradigm of bipartite graphs is Petri net analysis [12, 13].
The subdiscipline of mathematics and computer science addressing static net-
works is graph theory, which has a long and esteemed history that reaches back to
Leonard Euler in the eighteenth century. Euler’s first interest in graph theory was
piqued by a puzzle he posited for himself in his adopted home town of Königsberg
in Prussia. The River Pregel dissected this city in a peculiar fashion and at the time
was crossed by seven bridges (Figure 3.2). Euler wondered whether it was possible
to find a route that would cross each bridge exactly once. Of course, being a mathe-
matician, Euler not only solved this particular puzzle but also developed an entire
theoretical framework, the foundation of graph theory, to decide in which cases a
route can be found, and when this is impossible.
Generically speaking, graph theory is concerned with the properties of graphs,
such as their connectivity, cycles of nodes and edges, shortest paths between two
nodes, paths connecting all nodes, and the most parsimonious manner of cutting
edges in such a way that the graph is split into two unconnected subgraphs. Many
famous mathematical problems are associated with graphs. One example, arguably
the first to be proven using a computer, is the four-color theorem, which postulates
that it is always possible to color a (geographical) map with four colors in such a
fashion that neighboring states or countries never have the same color. Another
famous problem, which combines graphs with optimization, is the traveling sales-
man problem, which asks for the determination of the shortest path through a coun-
try such that every city is visited exactly once.
Most graph analyses in biology address questions of how a network is connected
and what impact this connectivity has on the functionality of the network. Thus, these
analyses characterize structural features, such as the numbers and densities of connec-
tions in different parts of the graph. These types of features can be subjected to rigorous
mathematical analyses, some of which lead to watertight proofs, while others are of a
more statistical nature. For our purposes, we can minimize the technical details, focus
on concepts and representative scenarios, and provide some key references.
3.1 properties of Graphs

Let us start with a graph G that describes a biological network and consists of nodes
and edges, and denote the edge connecting nodes Ni and Nj by e(Ni, Nj). A first
important and rather intuitive characteristic of a node is its degree, denoted by (A)
deg(Ni), which represents the number of associated edges. For an undirected graph, N1 N2
deg(Ni) is simply the total number of edges at Ni, and this number is the same as the
number of Ni’s neighbors. In a directed graph, it is useful to define the in-degree
degin(Ni) and out-degree degout(Ni), which refer to the numbers of edges terminat-
ing or starting at Ni, respectively (Figure 3.3).
For studies of the entire connectivity structure of a graph G with m nodes N1, …,
Nm, one defines the adjacency matrix A, which indicates with 1’s and 0’s whether N3 N4
two nodes are connected or not. This matrix has the form
1 if e ( N i , N j ) is an edge in G , N5
A = (aij ) =  (3.1)
0 otherwise.
(B)
Here, undirected edges are counted as two directed edges that represent the for- N1 N2
ward and reverse directions. Examples of A are given in Figure 3.4. The adjacency
matrix contains all pertinent information regarding a graph in a very compact rep-
resentation that permits numerous types of analyses with methods of linear algebra.
Box 3.1 discusses an interesting use of A.
In addition to investigating degrees of nodes, it is of interest to study how edges
are distributed within a graph G. The principal measure of this feature is the cluster- N3 N4
ing coefficient CN, which characterizes the density of edges associated with the
neighborhood of node N. Specifically, CN quantifies how close the neighbors of N
are to forming a clique, which is a fully connected graph. Let’s begin with an undi- N5
rected graph and suppose that some node N has k neighbors. Then, there can at
most be k(k − 1)/2 edges among them. Now, if e is the actual number of edges among
N’s neighbors, then CN is defined as the ratio
Figure 3.3 Visualization of the concept
of degrees associated with nodes.
e 2e (A) In this undirected graph, node N1 has
CN = = (if G is undirected). (3.2) neighbors N2, N3, and N5, and therefore its
k(k − 1)/2 k(k − 1)
degree is deg(N1) = 3. Node N4 has only one
neighbor, N2, and therefore deg(N4) = 1.
For directed graphs, the maximal number of edges among neighbors is twice as (B) In this directed graph, node N1 has
high, because each pair of neighbors could be connected with directed edges in an in-degree degin(N1) = 1 and an out-
degree of degout(N1) = 3, while node N4 has
both directions. Thus, the maximum number of edges, among k neighbors is k(k − 1),
degin(N4) = 1 and degout(N4) = 0.
and the clustering coefficient is defined as
e
CN = (if G is directed). (3.3)
k(k − 1)
Note that the minimum of CN in both cases is 0 and the maximum is 1.
(A) (B)
2 2
1 3 1 3
4 4
Figure 3.4 Two graphs and their
adjacency matrices. In (A), each edge is
considered bidirectional, which leads to
twice as many 1’s as edges. (B) shows a
directed graph and its adjacency matrix. As
0 1 0 1 0 1 0 1
an example, the red ellipse indicates that
1 0 1 1 0 0 1 1
A= A= directed edges point from node 2 to nodes
0 1 0 1 0 0 0 0
3 and 4. The third row of the matrix contains
1 1 1 0 0 0 1 0 only 0’s because no arrows leave node 3.
BOX 3.1: COMPUTATION OF THE NUMBER OF PATHS BETWEEN NODES
Interestingly, one can use adjacency matrices, as shown in (3.1), Executing the analogous operations for all Bij yields
to compute the number of paths between two nodes that consist
of a fixed number of steps. These numbers are given by powers 0 1 0 1 0 1 0 1  0 0 2 1
of the matrices. For instance, to compute all paths of length 2 in      
0 0 1 1 0 0 1 1  0 0 1 0
the matrix on the right-hand side of Figure 3.4, we multiply the A2 =  = = B. (2)
0 0 0 0 0 0 0 0  0 0 0 0
matrix by itself, which results in a new matrix B. According to the      
0 0 1 0  0 0 1 0  0 0 0 0 
rules for this operation, each element Bij in matrix B is given as
the sum of the products of the elements in row i of matrix A and
According to this result, there are two alternative two-step paths from
the elements in column j of matrix A. For instance, element B13 is
node 1 to node 3, while there is one two-step path from 1 to 4 and one
computed as
from 2 to 3. These three solutions exhaust all possible two-step paths.
In this simple case, the results are easily checked from the graph. The
B13 = A11A31 + A12A32 + A13A33 + A14A34 = 0 ⋅ 0 + 1 ⋅ 1 + 0 ⋅ 0 + 1 ⋅ 1 = 2. (1) reader should confirm these results.
An example is shown in Figure 3.5. The graph in (A) is undirected, and because
N1 has 8 neighbors and 12 edges among them, C N1 = 12 /(8 × 7/2) = 0.4286. The
directed graph in Panel B also has 8 neighbors, but 15 edges; in this case,
C N1 = 15/(8 × 7) = 0.2679. It has a lower clustering coefficient, because there could
theoretically be 56 edges among the neighbors of N1 instead of the actual 15.
It is useful to define an overall measure of the network’s clustering. This is accom-
plished with the graph’s average clustering coefficient CG, which is defined simply as
the arithmetic mean of the CN for all nodes [14]:
m
∑C
1
CG = N . (3.4)
m N =1
Intriguingly, biological systems typically have clustering coefficients that are higher
than those of randomly connected graphs. Furthermore, with an increasing size of
the network, the clustering coefficient tends to decrease. This implies that biological
networks contain many nodes with a low degree, but only relatively small numbers
of hubs, which are highly connected nodes that are surrounded by dense clusters of
nodes with a relatively low degree [15]. For example, in both the protein interaction
network and the transcription regulatory network in yeast, the hubs are significantly
more often connected to nodes of low degree, rather than to other hubs (Figure 3.6)
[16, 17]. We will return to this feature later in this chapter.
(A) (B)
N2 N2
N9 N3 N9 N3
N8 N1 N4 N8 N1 N4
N7 N5 N7 N5
N6 N6
Figure 3.5 Visualization of the clustering coefficient of a node. Node N1 has 8 neighbors. In the undirected graph in (A), the nodes are connected
among themselves through 12 (blue) edges. Thus, C N1 = (2 × 12)/(8 × 7) = 0.4286. In the directed graph in panel (B) the nodes are connected among
themselves through 15 directed (blue) edges. Therefore, C N1 = 15/(8 × 7) = 0.2679.
Figure 3.6 One large cluster within the

protein–protein interaction network of
the yeast Saccharomyces cerevisiae. The
entire network consists of 1870 proteins as
nodes, which are connected by 2240 direct
physical interactions. The cluster presented
here contains about 80% of this network. The
graph clearly shows a number of hubs, which
are highly connected, as well as the vast
majority of nodes with a low degree. Each
node is colored to signify the phenotypic
result if the corresponding protein is
removed (red, lethal; green, nonlethal;
orange, slow growth; yellow, unknown).
(From Jeong JP, Mason SP, Barabási A-L &
Oltvai ZN. Nature 411 [2001] 41–42. With
permission from Macmillan Publishers Ltd.)
A very important characteristic of a graph as a whole is its degree distribution

P(k), which measures the proportion of nodes within G that have degree k. If mk is
the number of nodes of degree k, and m is the total number of all nodes, then
P(k ) = mk /m. (3.5)

Note that P(k) is not a single number, but a set of numbers, with one degree value for
each k (k = 0, 1, 2, . . .). The division by m assures that the sum of P(k) over all values of
k is equal to 1. For the computation of P(k), the directions of edges are typically ignored.
Many characterizations of graphs, including degree distributions, are performed
by using comparisons with random graphs, which were first studied by Erdős and
Rényi in the late 1950s [18]. In such an artificially created graph, the edges are ran-
domly associated with nodes. One can show with mathematical rigor that the degree
distribution in this case is a binomial distribution, which resembles a skinny bell
curve with a small variance. In other words, most nodes have a degree that is close to
average.
The binomial degree distribution within random graphs is very different from
what is frequently observed in biological and other real-world systems, where a few
hubs are connected to disproportionally many other nodes, while most other nodes
are associated with much fewer edges than in a random graph (see Figure 3.6). Spe-
cifically, P(k) often follows a power-law distribution of the form
P(k ) ∝ k −γ , (3.6)
where the exponent γ has a value between 2 and 3 [19]. Plotting such a power-law
degree distribution in logarithmic coordinates results in a straight line with slope γ
(Figure 3.7). A network whose degree distribution more or less follows a power law
is called a scale-free network. This terminology was chosen because the power-law
property is independent of the size (or scale) of the network.
While a power-law distribution is often observed for most nodes in a large biologi-
cal network, the plot is seldom truly linear for very highly and very sparsely connected
nodes within such a network. Nonetheless, except for these extremes, the power-law
function often holds quite nicely, which demonstrates, unsurprisingly, that most bio-
logical systems are clearly not organized randomly (see, for example, [20]).
An interesting feature of scale-free networks is that the shortest paths between
two randomly chosen nodes are distinctly shorter than those in random graphs.
Here, a path is an alternating sequence of nodes and edges in a graph, leading from
a start node to an end node, such as
N 6 ; e(N 6 , N 1 ); N 1 ; e(N 1 , N 4 ); N 4 ; e(N 4 , N 5 ); N 5 ; e(N 5 , N 2 ); N 2 (3.7)

(A)
(B)
14/27
12/27
fraction of nodes mk /m
10/27
8/27
6/27
4/27
2/27
0/27
0 2 4 6 8 10 12 14
node degree k
(C)
1
fraction of nodes mk /m
0.1
0.01
1 3 6 10
node degree k
Figure 3.7 Example of a small scale-free network. (A) Among the 27 nodes, three (red) are the most highly connected hubs. (B) The relative number
of nodes with k edges, plotted against the number of nodes with a given number of edges, k, shows a decreasing power-law relationship. (C) If the
relationship in (B) is plotted in logarithmic coordinates, the result is approximately linear, with negative slope γ. (Adapted from Barabási A-L & Oltvai ZN.
Nat. Rev. Genet. 5 [2004] 101–113. With permission from Macmillan Publishers Limited, part of Springer Nature.)
in Figure 3.8. Of particular interest is the shortest path (or distance) between two
nodes, because it contains important clues about the structure of the network. In
Erdős and Rényi’s random graphs [18], the average shortest distance is proportional
to the logarithm of the number of nodes, log m. By contrast, the average shortest N1
distance in a scale-free network with a power-law degree distribution is shorter and e(N6,N1)
increases more slowly for larger networks [21]. Indeed, analyzing over 40 metabolic
networks with between 200 and 500 nodes, Jeong and collaborators found that a few N6 e(N1,N4) N2
hubs dominated the connectivity patterns and that, as a consequence, the average
shortest distance was essentially independent of the size of the network and always
had an amazingly low value of about three reaction steps [22]. In a different inter-
e(N5,N2)
pretation, one may also consider the average distance as an indicator of how quickly
information can be sent through the network. According to this criterion, biological
networks are indeed very efficient. N5 N3
Most authors define the diameter of a network as the minimum number of steps
that must be traversed to connect the two most distant nodes in the network, and we e(N4,N5)
will follow this convention. In other words, the diameter is the shortest path between N4
the farthest two nodes. However, one must be cautious, because other authors use the
same term for the average shortest distance between any two nodes.
Figure 3.8 A possible path from start node
N6 to end node N2. The path uses the edges
3.2 Small-World Networks e(N6, N1), e(N1, N4), e(N4, N5), and e(N5, N2).
The path is not the shortest possible path,
If a network is scale-free (that is, if it has a power-law degree distribution and there- which would use e(N6, N3) and e(N3, N2). It is
fore a small average shortest distance) and if it furthermore has a high clustering customary not to allow the same node or
coefficient, it is called a small-world network. This terminology reflects the special edge to appear twice in a path.
structure of these networks, which consists of several loosely connected clusters of

nodes, each surrounding a hub. In this organization, most nodes are not neighbors
of each other, but moving from any one node to any other node in the network does
not take many steps, because of the existence of the well-connected hubs. We have
already mentioned metabolic networks, which have average path lengths in the
range of 2–5 steps. Ecological studies have found the same length ranges in many
food webs [23]. For other biological examples and further discussions, see [24, 25].
Outside biology, the situation is similar in commercial flight routes, which use major
hubs combined with connecting flights to smaller airports. The small-world prop-
erty is furthermore reminiscent of John Guare’s famous play Six Degrees of Separa-
tion, whose premise is that essentially every human is connected to every other
human by a chain of six or fewer acquaintances.
Interestingly, the small-world property emerges automatically when a network
grows by the addition of new links preferentially to nodes that are already highly
connected. Specifically, Barabási and collaborators (see, for example, [26]), who
have devoted a lot of effort to these types of graph analyses, showed that a distri-
bution with γ = 3 is the result of a growth process where, for each addition of new
nodes, the edges are connected to an existing node N with a probability that is
directly proportional to deg(N) [27]. Variations on the mechanism of preferential
attachment lead to other γ values that can range anywhere between 2 and ∞. One
should note, however, that entirely different mechanisms for the evolution of
scale-free and small-world networks have been proposed. For instance, according
to the duplication–divergence (DD) model, a gene or protein network evolves
owing to the occasional copying of individual genes or proteins, which is followed
by mutations and pruning [28]. The DD model can lead to γ < 2. See also Box 3.2.
BOX 3.2: CONSTRUCTION OF REGULAR, LOW-DIAMETER, AND RANDOM NETWORKS
Interesting connectivity patterns in networks can be constructed in structure of the network remained unchanged. The other extreme
various ways. Watts and collaborators [14, 29] started with a regularly case, p = 1, was shown to result in a random network. For values of p
patterned ring of nodes, in which each node was associated with strictly between 0 and 1, other networks with much reduced diameters
four edges. Namely, the node was connected to its two immediate were obtained. Watts and collaborators called them small-world
neighbors and also to their immediate neighbors (Figure 1). They networks, but the terminology was different at the time. Today, we
then selected a constant probability p anywhere between 0 and 1, might call them short-distance or low-diameter networks, but they are
with which each edge could be disrupted and reconnected to a new, no longer considered small-world because they do not exhibit high
randomly chosen, node. Of course, for the extreme case of p = 0, the clustering coefficients.
Figure 1: Construction of different types of

networks, according to Newman and Watts
[29]. The construction begins with a regular
network, in which each node is connected to
its immediate neighbors and their immediate
neighbors (top of figure). Considering one node
after another, the edge to one of its nearest
neighbors is removed with a probability p
(0 < p < 1) and replaced with a new edge to
a randomly selected node; double edges are
not allowed. In a second round, the process is
repeated for second-nearest neighbors (bottom
of figure). For p = 0, no edges are changed and
the regular network is retained; for p between
0 and 1, a few edges (green) are replaced and
the result is typically a network with shorter
p=0 0<p<1 p=1 diameter; for p = 1, all edges are replaced and
regular lower diameter random the network exhibits random connectivity.
It is not surprising that the hubs in a gene or protein network are often among
the most important components and that their destruction or mutation is frequently
fatal. In fact, it seems that the degree of a protein within a PPI network may be
directly correlated with the fitness of the gene that codes for it [30]. This feature cor-
relates in an interesting fashion with the robustness of a small-world network, or
specifically its tolerance to attacks: attacks on randomly selected nodes are easily
tolerated in most cases, because the likelihood of hitting a hub is low; by contrast, a
single, targeted hit on a hub can break up the entire network [16, 22]. For these
reasons, the identification of hubs may be important, for instance, for drug target
identification. The interesting co-occurrence of robustness and fragility has been
observed in many complex networks and systems, including both biological and
engineered examples, and may be a characteristic design principle of such systems
[31–33]. Other design principles and network motifs will be discussed in Chapter 14.
While molecular networks have recently received the most attention, it should
be mentioned that static networks have played a role in systems ecology for a long
time. Beginning with the pioneering work of Joel Cohen [34], graph representations
have been used to study food webs, the flow of material through them, and the
robustness and reliability of their connectivity structures. It was shown, for instance,
that typical food webs have maximal path lengths between 3 and 5, that marine
paths tend to be longer than those in terrestrial systems, and that the reliability of a
food web decreases with increasing length [23]. Food webs often exhibit loops,
where A feeds on B, B on C, and C on A [35].
While biological systems are often scale-free, their subnetworks sometimes have
a different structure. For instance, sets of nodes within a small-world network are
highly connected around hubs and may even form cliques, within which each node
is connected to every other node. Clearly, the connectivity patterns of a clique and
the larger network are different, and caution is in order when only parts of larger
networks are investigated.
One should also keep in mind that most networks are in reality dynamic. For
instance, PPI networks change during development and in response to disease. They
sometimes even change periodically on a shorter timescale, as is the case during the
cell cycle in yeast [36]. Han and collaborators [37] analyzed the temporal properties of
PPI hubs in yeast during the cell cycle and in response to stress, and postulated two
types of hubs: party hubs, which interact simultaneously with most of their neighbors;
and date hubs, which connect to different neighbors at different times and possibly in
different cellular locations. The dynamic changes associated with hubs may indicate a
control structure of organized modularity, where party hubs reside inside defined,
semi-autonomous modules, which perform some biological functionality, while date
hubs connect these modules to each other and organize their concerted function.
However, the distinction between party and date hubs is not always clear-cut [38].
As a final example of graph-based network analysis, let us look at the galactose
utilization gene network in yeast [39], which serves as a documentation example in
the Cytoscape network analysis platform [2], where the following analyses can be
retraced. The network consists of 331 nodes, which have an average of 2.18 neigh-
bors. Once loaded into Cytoscape, it can be visualized in a number of ways, for
instance, in the so-called force-directed layout, whose details can be customized to
the user’s taste (Figure 3.9). Each node can be clicked on screen, which reveals the
gene identifier. A separate spreadsheet contains further information. For instance,
the hub (YDR395W) has a role in the transport of mRNA from the nucleus into the
cytosol, while its neighbors code for structural components of the ribosome; a pos-
sible exception could be YER056CA, whose role is unknown. Thus, the closeness
within the network is reflected by functional relationships. By contrast, the gene at
hub YNL216W, which is a few steps removed, codes for a repressor activator protein.
With a few clicks, Cytoscape reveals typical numerical features of the network, such
as the clustering coefficient (0.072) and the average (9.025) and the longest (27)
lengths of the shortest paths. It is furthermore easy to display the distribution of
shortest paths (Figure 3.10A) and the degree distribution (Figure 3.10B). The latter
distribution more or less follows a power law, with the exception of very low and
very high degrees, as we discussed before. Many other static features of the network
can be displayed. Beyond these typical network features, Cytoscape plug-ins can
even be used for transitions to dynamic analyses of small subnetworks [40].
YNL216W
YDR395W
YPR102C
YIL052C
YLR075W
YNL069C
YGR085C
YER056CA
YDL075W
Figure 3.9 Visualization of a network. Every node in Cytoscape’s network visualization of the galactose utilization gene network in yeast can be clicked
on screen to reveal its identity and features. Neighboring nodes are often functionally related. See the text for further details.
(A) (B)
6000 200
100
5000
4000
number of nodes
frequency
3000
10
2000
1000
0 1
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 1 10 20
path length degree
Figure 3.10 Typical output from Cytoscape. (A) Distribution of shortest paths. (B) Degree distribution with regression line computed in Cytoscape (red)
and regression line ignoring very low and very high degrees (blue); the latter leads to a steeper slope (−2.45) than the former (−2.08).
DEpENDENcIES AMONG NETWORK cOMpONENTS

3.3 causality Analysis
So far, we have discussed the connectivity among nodes in a network, as well as
paths leading from a source to a target. Especially in undirected networks, the
connections represent associations. However, they do not necessarily imply causa-
tion, which is much more difficult to establish, but yields greater insight into the
functionality of the network.
The search for causes and effects is probably as old as humankind. We know that
Aristotle stated that the cause has to precede an effect, while Galileo in his 1638
book Two New Sciences postulated that a true cause (vera causa) must not only pre-
cede the effect but also be both necessary and sufficient. Two hundred years later,
John Stuart Mill additionally required of a true cause that other explanations for a
cause–effect relationship must be eliminated, and, in the 1920s, the geneticist Sewall
Wright formulated systems of linear equations and analytical tools for addressing
indirect effects. Wright’s studies can be considered the foundation of modern cau-
sality analysis (cf. [41]). Nowadays, causes and effects are often studied with statis-
tical methods. A prominent approach is Granger causality, which requires that if A
causes B, then past values of both A and B should be more informative for predicting
B than past information on B alone. Such a situation can be evaluated with linear
regression methods applied to stochastic processes. The techniques have recently
been extended to nonlinear cases, but the corresponding computations become
quite complicated. While a warning is often voiced that correlation does not imply
causation, techniques of path analysis, structural equation modeling, and Bayesian
statistics permit to some degree the reliable inference of true causes of observed
effects [42].
Typical questions are whether effects are direct or indirect and how to address
the role of hidden variables, which cannot be measured. In many cases, the meth-
ods involve a cause–effect diagram, which consists of a directed graph without
cycles, which means that one can never return to the same node. The problem with
causality analysis in biology is that realistic systems almost always contain a con-
siderable degree of circularity, which makes most statistical approaches to analyz-
ing or inferring network structures and their causality troublesome. In systems
with many cycles, causes may be distributed throughout the whole system, and
multiple related or unrelated factors may contribute to the observed effect with dif-
ferent weights. As a consequence of these typical complications, causality analysis
has found only limited applications in biology (for a successful exception in a rela-
tively simple case, see [43, 44]), with the exception of Bayesian methods, which are
often used for the reconstruction of genetic and signaling networks, as we will see
in a later section.
3.4 Mutual Information

A simpler task than identifying true causality concerns the mutual dependence of
two variables. As an example, suppose we would like to quantify how much more
likely it is that some gene of interest is expressed if we know that certain other genes
are expressed. It is not difficult to see that such an analysis of the likely co-expression
of sets of genes could give important clues regarding the structure of pathways and
might shed light on diseases such as cancer that involve many changes in gene
expression [45]. Questions of this nature may be addressed with statistical methods
characterizing the mutual information between two or more quantities. These meth-
ods are derived from Claude Shannon’s and Warren Weaver’s ideas concerning the
reliability of communication through noisy channels [46], which led to what is now
known as information theory. In this field, one studies how likely it is that one vari-
able X is on (present), while another variable Y is either on (present) or off (absent). If
X and Y are perfectly coupled to each other, the two will always be both on or both off,
or it could be that one is always on if the other is off. However, in most cases, the cou-
pling is not that clear-cut and one often wonders whether two variables influence
each other at all. The mutual information between X and Y characterizes their depen-
dence in cases where the coupling is perfect, strong, loose, or absent. Expressed
BAYESIAN REcONSTRucTION OF INTERAcTION NETWORKS 63
differently, the mutual information is a measure of the extent to which knowledge H(X) H(Y)
regarding one of the two variables reduces uncertainties in the other. If X and Y are
totally independent of each other, the mutual information is 0, because, even if we
knew everything about X, we would be unable to make reliable predictions regarding
Y. By contrast, a strong dependence between X and Y corresponds to a high value of H(X | Y) I(X; Y) H(Y | X)
the mutual information criterion. Specific examples of the use of mutual information
in molecular systems biology include inferences regarding the functional relevance
of gene–gene interactions [47] and the coordination between gene expression and
the organization of transcription factors into regulatory motifs [48].
A key concept for quantifying mutual information is the entropy H(X), which
measures the uncertainty associated with a random variable X. Here, the term ran- H(X, Y)
dom variable means that X may take different values with certain probabilities; for
example, the roll of a die may result in any number between 1 and 6. Depending on Figure 3.11 Mutual information and
the situation, the uncertainty surrounding X may depend on how much we know entropy. The Venn diagram visualizes the
about another variable, Y. For instance, if two genes G1 and G2 are often co- relationships between mutual information
I(X; Y) and different types of entropy, namely
expressed, and if we observe that G2 is indeed expressed in a certain situation, we
the entropies H(X) and H(Y) associated
are more confident (although still not certain) that G1 will be expressed as well. The with the random variables X and Y, the
uncertainty of X, given that we know Y, is expressed as the conditional entropy joint entropy H(X, Y), and the conditional
H(X | Y). Furthermore, the uncertainty of a pair of random variables is quantified as entropies H(X | Y) and H(Y | X).
the joint entropy H(X, Y), which is defined as the sum of the entropy H(X) and the
conditional entropy H(Y | X). Since H(X, Y) is the same as H(Y, X), we obtain the
relationship
H ( X , Y ) = H ( X ) + H (Y | X ) = H (Y ) + H ( X | Y ).
The mutual information I(X; Y) is the sum of the individual entropies minus the
joint entropy:
I ( X ; Y ) =  H ( X ) + H (Y )  − H ( X , Y ) .
These relationships can be visualized in a Venn diagram of entropies (Figure 3.11).

Shannon, Weaver, and others worked out the mathematical formulations of the vari-
ous entropies, and these can directly be extended to uncertainties and the mutual
information among more than two variables. A good introduction to the topic is [45].
BAYESIAN REcONSTRucTION OF INTERAcTION

NETWORKS
While it is relatively easy to characterize graphs using measures such as their degree
distribution or clustering coefficients, it is much more difficult to find the connec-
tivity within an unknown graph. This section introduces the basic concepts needed
to execute reverse engineering techniques in ill-characterized graphs. The impor-
tant message of this section is the confirmation that it is indeed possible to infer the
structure of unknown static networks with some reliability, at least under favorable
conditions. The technical details, which are not always simple and are unfortu-
nately difficult to avoid, should not cloud this message.
It is easy to imagine that one can infer the structure of a network only if one has
suitable data—lots of them. It is also clear that such data in reality are essentially
always corrupted by noise, which may come from genuine randomness or from
inaccuracies during the experimental measurements. For instance, the components
of a signaling system that are active at a particular point in time may depend on all
kinds of factors, including environmental perturbations, inbuilt redundancies, and
the recent history of the system. Overlaid on these perturbations are the experimen-
tal difficulties of obtaining accurate characterizations of active and inactive states.
Given such enormous challenges, is it possible—and indeed valid—to infer an
active, causative role of two components A and B in the stimulation of C, if C is usu-
ally active when A and B are both turned on? It is not difficult to imagine that the
reliability of such an inference grows the more often we observe the same pattern of
C being on if both A and B are on, and C being off if A or B or both are off. In other
words, it is easy to imagine that network reconstruction tasks require many data that
permit the application of statistical tools.
Useful data can come in two forms. In the first case, one has many independent
observations of co-occurrences of component activities, as in the case of A, B, and C.
In the second case, one may have measurements of many or all components of the
network, which were obtained as a sequence of many time points, following some
stimulus. For example, one could expose a cell culture to an environmental stress
and study the expression profile of a set of genes over many minutes, hours, or days
afterwards. In this chapter, we discuss only the former case of many co-occurrences.
We will return to the latter case in Chapter 5, which addresses inverse modeling
tasks for dynamic systems.
The strategy for attempting to infer the structure of a network from co-occurrence
data is of a statistical nature and goes back almost 250 years to Thomas Bayes [49]; it
is still called Bayesian network inference. If the data are plentiful and reasonably
good, the strategy is rather robust and addresses direct and indirect causal effects.
However, it also has its limitations, as we will discuss later.
Before we can outline the principles of Bayesian network inference, it is neces-
sary to review the very basics of probabilities. We do this in a somewhat nonchalant,
conceptual fashion, but there are many books that treat these ideas in a rigorous
mathematical fashion (for example, [50]).
By convention, probabilities are real numbers between 0 and 1 (or 100%), where
0 signifies the impossibility of an event and 1 denotes the certainty of an event. Sup-
pose we have nine new $1 bills and one new $20 bill in our wallet and they are not
sorted in any particular way. Because the bills are new and of the same size, weight,
and material, we cannot feel differences between them, and, pulling a bill blindly
out of the wallet, the probability of a $1 bill is 90% or 0.9, whereas the probability of
the $20 bill is 10% or 0.1. The probability of pulling a $5, $100, or $7.50 bill is 0, and
the probability of pulling either a $1 or a $20 bill is 1. All this makes intuitive sense
and can be proven with mathematical rigor.
Imagine now that we are repeating the experiment many times, always returning
the bill to the wallet, and recording every time what dollar amount we pulled. In the
previous example, if we pull a bill often enough, we expect to pick the $20 bill roughly
every tenth time. Now suppose we do not know how many $1 or $20 bills are among
the 10 bills are in our wallet. Pulling only one bill does not really answer the ques-
tion, but if we repeat the experiment many times, the numbers of $1 or $20 pulls
begin to give us an increasingly clearer picture. The argument is as follows: If we
blindly pull a bill 50 times, and if there is only one $20 bill, we would expect a total
of five $20 pulls. It might be four, it might be six, but five would be our best estimate.
By contrast, if there were six $20 bills among the 10 bills, we would expect a much
higher return, namely something like 30 pulls of $20 bills, corresponding to the 60%
chance for each pull. Thus, just by observing the outcome of many experiments, we
can infer something about the unknown internal structure of the “system.” Bayesian
inference is a sophisticated formalization of such a strategy. We are not 100% sure
about our results, because they are probabilistic, but our confidence grows with the
number of observations.
Continuing with the example, suppose we are in a game show where we have
to choose one of three wallets and then blindly pull one bill out of it. The wallets
have different colors, and, for whatever reason, we have a slight preference for
gray. To be specific, let’s say the probability to select the gray wallet is 40% and the
other two probabilities are 30% each. Of course, we don’t know what each wallet
contains. Suppose the brown wallet contains two $100 bills and eight $1 bills, the
gray wallet has six $20 bills and four $1 bills, and the green wallet contains ten $10
bills (Figure 3.12).
Clearly, our personal outcome of the game will ultimately be $1, $10, $20, or
$100, but what is the likelihood of each? It is easy to see that the probability of pull-
ing a particular dollar amount out of a wallet depends on which wallet we select in
the first place. In statistical terms, the probability is conditional, because it depends
(is conditioned) on the selection of a particular wallet. Figure 3.12 illustrates the
situation.
initial choice
Figure 3.12 Conditional probabilities
can be explored intuitively with playing
0.3 0.4 0.3 cards, dice, and games. Shown here is the
schematic of a fictitious game show, where
the contestant selects a wallet and pulls a bill
out of it.
0.2 0.8 0.6 0.4 1
$100 $20 $10 $1
In general terms, the conditional probability P(B | A) (that is, the probability of B
given A) of an event B, given that event A is true (is occurring or has happened
before) is formally given as
P ( B and A )
P (B | A) = . (3.8)
P ( A)
Here P(B and A) denotes the (joint) probability that both A and B are simultaneously
true, and it is tacitly assumed that the probability of A, P(A), is not zero. Because P(B
and A) is the same as P(A and B), we can formally rearrange the terms and write a
sequence of equations as
P ( B | A ) P ( A) = P ( B and A ) = P ( A and B ) = P ( A | B ) P ( B ). (3.9)
The equivalence of the outer expressions here is known as Bayes’ theorem, Bayes’
formula, or the Bayes rule, and is typically written as
P ( A and B ) P ( B | A ) P ( A)
P (A | B) = = . (3.10)
P (B) P (B)
The formula says that we can make a statement of the (unknown) occurrence of
event A given an (observed) outcome B, assuming that P(B) is not zero.
It is not difficult to make various forward computations with the formulae for
conditional probabilities, for instance, for our game show example. Because the
choice of the wallet and the pulling of a bill from it are independent events, their
probabilities simply multiply. Thus, the probability of choosing the brown wallet
and securing the grand prize of $100 is 0.3 × 0.2 = 0.06, which corresponds to a mea-
sly 6% probability. The probability of choosing the gray wallet and ending up with
$100 is nil. In this fashion, we can exhaust all possibilities and, for instance, com-
pute what the probability is of ending up with $20. Namely, the probability of $20
from the brown or red wallets is zero. The gray wallet is chosen with probability 0.4
and choosing a $20 bill among the bills is 0.6, so that the overall probability is 0.24.
To compute the probability of the meager outcome of $1, we need to consider the
brown and the gray wallets, and the probability is 0.3 × 0.8 + 0.4 × 0.4 = 0.4. Finally,
the probability of an overall outcome of $10 is 0.3 or 30%. As should be expected, the
probabilities for $1, $10, $20, and $100 sum up to 1.
Another question we can answer quite easily is the following: If we knew the
contents of the three wallets, would the choice of the gray wallet be optimal? The
answer lies in the expected value, which we may imagine as the average over very
many runs of the game. The easiest case for computing the expected values is the
red wallet, which, when chosen, will always reward us with $10. For the brown wal-
let, the expected value is $20.80, because in 100 runs of the game we would expect to
pull about twenty $100s and eighty $1s. The total dollar amount of these is $2080,
which corresponds to an average of $20.80 per run (even though we can never
actually get $20.80 in any given run!). For the gray wallet, the analogous computa-
tion yields (60 × $20 + 40 × $1)/100, which is only $12.40. Gray might be a nice color,
but it is a bad choice in this game!
All these types of computations are straightforward, because they follow well-
established probability theory and, in particular, the law of total probabilities.
This law says that the probability of an event B that is conditioned on distinct events
A1, A2, . . ., An can be computed as
n
P (B ) = ∑ P(B | A )P(A ).
i =1
i i (3.11)
We can see that this result is directly related to Bayes’ formula, if we consider all possible
and distinct events Ai and compute all conditional probabilities P(B | Ai). The prob-
ability of event B is then the sum of all these P(B | Ai), given that Ai actually happens,
which it does with probability P(Ai). A direct example is the preceding computation
of the least desirable outcome of $1 in the game show example. Each P(Ai) is called a
prior probability, and P(B | Ai) is again the conditional probability of B, given Ai.
3.5 Application to Signaling Networks

Many of the concepts of Bayesian analysis are applicable to genomic, proteomic, signal-
ing, and other networks. Here we use signaling as the prime example for illustration, but
it should be clear that gene and protein networks can be analyzed in a similar fashion.
Signal transduction is a key element for the proper function of cells and organ-
isms. Simplistically speaking, a cell receives a chemical, electrical, mechanical, or
other signal at the outside of its membrane. This signal is internalized in some fash-
ion, for instance with the help of G-protein-coupled receptors, and typically triggers
an amplification mechanism that is often referred to as a signaling cascade. Ulti-
mately, the received and amplified signal leads to some cellular response, such as
the expression of appropriate genes. Fundamental aspects of these fascinating sys-
tems are discussed in Chapter 9.
For an illustration here, let us apply our insights on conditional probabilities to
a signaling network like the one in Figure 3.13. Suppose two receptors, R and S, are
anchored in the cell membrane and respond independently to appropriate ligands.
The response to the binding of L consists of sending a signal to a signaling cascade
C, whose details are unimportant here, but which ultimately triggers some genomic
response G. For the cascade to be activated, both R and S should fire. Furthermore,
if a ligand binds only to S, the system triggers a different response Z. In reality, the
interactions are corrupted by noise, such as the binding of nonspecific ligands, so
that the cascade may sometimes be triggered even if only R or S is active, or it may
not be triggered, even if both are occupied by ligands. The terminology of noise is
commonly used, even though signaling cascades don’t make audible sounds.
Figure 3.13 A generic, simplified signaling

system responds to the binding of two
different ligands. The expected numbers of
correct or faulty responses can be computed
L
with conditional probabilities. See the text
for details.
P(R) = 0.2 R S P(S) = 0.4
P(C | R and S) = 0.9

P(C | only R) = 0.1
P(C | only S) = 0.2
P(Z | S) = 0.8
P(C spont) = 0.02 Z
C P(Z spont) = 0.05
G P(G | C) = 0.8
P(G spont) = 0.1
The connectivity of the system may be sketched out in a diagram, as shown in

Figure 3.13, which also shows probability tables of various events. For instance, the
probability that receptor R is active (occupied by a specific ligand L) is 0.2. The
probability that cascade C is triggered if both R and S are occupied by ligands, is
P(C | R and S) = 0.9, whereas spontaneous activation of the cascade has a very low
probability of P(C spont) = 0.02. The probability tables allow us to compute all
kinds of probabilities, such as the probability that a ligand binds to S and triggers Z.
Or we might compute the probability that (1) ligands bind to R and S; (2) the
cascade is activated; and (3) the result is a genomic response G. This latter scenario
is modeled as
P (G , C , R , S are “on”) = P (G | C ) P (C | R and S ) P ( R ) P ( S )

(3.12)
= 0.8 × 0.9 × 0.2 × 0.4 = 0.0576
The roughly 6% probability of the response may seem low, but it is mainly dictated
by the relatively low binding probabilities of ligands of R and S, which both must
happen for a normal response.
A typical situation for Bayesian inference is the following. If we observe that gene
G is expressed, can we make inferences on whether the cascade is activated and/or
whether R and/or S are present? Knowing the probability tables, we can infer with
high probability that C is active, because the probability that G is on, given that C is
active, is high (0.8), and spontaneous gene expression is not overly likely (0.1). It is
not difficult to imagine that the inferences become better and better as more obser-
vations on the on–off states of the nodes in the graph become available.
Bayesian inference really becomes interesting if we do not know the probability
tables, but must deduce them from observations on the states of the system. Not
surprisingly, the execution of such an inference is rather complicated and requires
sophisticated computer algorithms. Nonetheless, it is useful to understand the
underlying conceptual ideas. In general, a Bayesian network is formulated as a
directed acyclic graph (DAG), which means that its nodes Xi are connected by
arrows heading in only one direction and that there is no possibility of returning to
any of the nodes, once one leaves this node through any of its outgoing arrows; Fig-
ure 3.13 is an example. The parents of a node Xi, denoted by Parj(Xi), are defined as
those nodes that directly send a signal to Xi through a single connecting arrow. The
state of a parent may again be on or off, 1 or 0, or take other values, just as Xi. If the
graph consists of n nodes, then the probability P that the state of the system has
particular values (x1, x2, . . ., xn) can be computed, with a dependence on the state of
the parents, as
P( X 1 = x1 ∧ X 2 = x 2 ∧ ... ∧ X n = xn )
n
∏ P ( X = x | Par (X ) = x ).
(3.13)
= i i j i j
i =1
Here, the notation ∧ is the logical “and.” Thus, the expression on the left-hand side
formalizes the probability P that all the nodes Xi simultaneously have the particular
values xi. The notation 〈Parj(Xi) = xj〉 on the right-hand side refers to the set of all
parents Pj of Xi and their particular states xj. The probabilities in the product are
therefore probabilities that Xi has the value xi, conditioned on the states of all its
parents. The states of the parents can in turn be expressed in terms of their own
parents, and so on, as we exemplified above. This stepwise dependence is possible
because the graph is acyclic; if it had loops, we would be in trouble. With this
formalization, we can compute any conditional probability in the system. These
computations are often very tedious, but they are straightforward and can be
executed with customized software.
The computation of conditional probabilities requires that we know (a) the con-
nectivity of the network and (b) the probability tables at each node, as it was the case
in Figure 3.13. For many actual signal transduction networks in biology, we may
have a rough idea about their connectivity, but we usually do not know the probabil-
ities. Sometimes, we have information on whether the probabilities are high or
rather low, but we seldom know precise numerical values. That’s where Bayesian
inference comes in. Instead of being able to compute numerically forward, as we
did so far, one needs a large dataset of observations consisting of on or off states of
all or at least many of the nodes. The focus of the analysis is now on correlations of
on or off values among the nodes. For instance, two observations regarding the sig-
nal transduction system in Figure 3.13 could be
(R = off ; S = on ; C = off ; G = off ; Z = on)
and
(R = on ; S = on ; C = on ; G = on ; Z = off ).
If C and G are strongly correlated, as they are in the above example, many observa-
tions will have them either both on or both off. By contrast, such a correlation would
not be expected between R and Z.
The inference strategy begins with composing a network graph that contains all
known, alleged, or hypothesized connections, the direction of causality between
any two nodes, and, if possible, likely ranges of probabilities. All probabilities are
represented by symbolic parameters p1, p2, . . ., pm rather than numerical values, and
the goal is to determine the unknown values for these parameters in such a manner
that the parameterized model reflects all (or at least most) observed combinations
of states as accurately as possible. Needless to say, figuring out such a well-matching
configuration of many parameters constitutes an optimization task that is anything
but trivial, especially if there are many nodes. Several methods have been devel-
oped for this purpose in recent times, including Markov chain Monte Carlo (MCMC)
algorithms such as the Gibbs sampler and the Metropolis–Hastings method, and
they are all quite involved [8]. The basic concept of these optimization methods con-
sists of randomly choosing probabilities (first blindly or within known or assumed
ranges, and later within shrinking ranges that are determined by the algorithm) and
computing the corresponding joint distributions of all nodes (recall the Monte Carlo
simulations in Chapter 2). This forward computation with putative probabilities is
relatively simple and is executed thousands of times. A statistical comparison
between the computed joint distributions and the observed distributions is then
used to determine probability values that slowly improve during the algorithmic
iterations and, if successful, ultimately yield a constellation of probability values
with which the model matches the observations best. For further details, the
STATIc METABOLIc NETWORKS AND ThEIR ANALYSIS 69
interested reader may consult [8] and [51]. The same type of analysis is sometimes
used to decide which set of network connections among several candidates is most
likely, given observations on input and output nodes.
One theoretical and practical disadvantage of the Bayesian inference strategy is
that it does not permit graphs that contain any types of loops. For instance, chains of
arrows starting and ending at the same node, as well as feedback signals, are not
permitted. Furthermore, graphs or probability tables that change over time make
inferences much more complicated. Current research efforts are focusing on these
challenges.
It may seem that Bayesian inferences of actual signal transduction networks
might run into insurmountable problems. However, several success stories have
been documented. A premier example is the identification of a signaling network in
CD4+ T cells of the human immune system [52]. The data consisted of thousands of
single-cell flow-cytometric measurements of multiply phosphorylated protein and
phospholipid components in individual cells that followed external molecular
perturbations. Specifically, the data consisted of knowledge of the molecular stimu-
lus (inhibition or stimulation of a specific component) and the corresponding
state or condition of each of 11 phosphorylated molecules within a given cell,
15 minutes after the stimulus. The Bayesian inference successfully confirmed
known relationships between some of the signaling components and revealed
additional, previously unknown, causal relationships. Not surprisingly, the analysis
missed three causal relationships that corresponded to known loops in the signal
transduction network.
3.6 Applications to Other Biological Networks

Outside of signaling systems, graph methods and Bayesian approaches have been
used to reconstruct protein–protein interaction networks, where the main purpose
was to identify which proteins are interacting with each other, and which of these
interactions are functionally meaningful [53, 54]. As just one example, Rhodes and
collaborators [55] proposed a model of the human protein–protein interaction net-
work that integrated information from protein domain data with gene expression
and functional annotation data. Using Bayesian methods, they constructed and par-
tially validated a network consisting of nearly 40,000 protein–protein interactions.
Beyond the presence or absence of particular interactions, one is also often inter-
ested in the structure of these networks and the features of the underlying graphs,
such as types of hubs, cliques, and small-world or random structure, as we discussed
earlier in this chapter. Since large-scale protein–protein interaction networks often
contain hundreds of nodes, their visualization is a great challenge.
Bayesian methods have also been applied to gene regulatory networks and the
control of transcription [56, 57]. In genomic networks, one asks which genes are
functionally co-regulated and whether the expression of one set of genes might
depend on the expression of another set of genes. Answers can be found in micro-
array data from perturbation experiments, from the interactions between tran-
scription factors and promoters, or from time series data, in which gene expression
is measured multiple times following a stimulus (see, for example, [58, 59]). A
sophisticated example is the analysis of genetic variations among individuals and
their influence on the occurrence and severity of complex diseases such as obesity,
diabetes, and cancer [60].
Bayesian inference methods are seldom applied to metabolic networks. The
main reason is that the interactions in these networks are strictly dictated by bio-
chemistry, so that it is usually more or less clear how these networks are connected.
Because of this strong biochemical basis, metabolic networks allow a whole class of
other analyses, which we will discuss next.
STATIc METABOLIc NETWORKS AND ThEIR ANALYSIS

Metabolic networks will be described in greater detail in Chapter 8, but they are
intuitive enough to serve here as an illustration. Metabolic networks are uniquely
suited for analyses beyond clustering coefficients and degree distributions, because
they represent the flow of actual material. In contrast to proteins, whose interactions
are often seen in their association and co-location, metabolic pools receive and
release material, and all this material must be accounted for. Thus, if a certain
amount of material leaves one pool, the same amount will be received by other
pools. If glucose is converted into glucose 6-phosphate, we know that the amount of
glucose used-up will equal the newly added amount of glucose 6-phosphate. These
precursor–product relationships tremendously constrain what may happen in a
metabolic system and thereby allow a variety of interesting analyses.
For simplicity of discussion, we will focus on small systems, but it should be
noted that some analyses in the literature have addressed metabolic networks of
considerable size (see, for example, [15, 20, 22, 61–63]. Generically, metabolic sys-
tems describe how organic material is converted from one chemical form into
another, such as the conversion of glucose into glucose 6-phosphate in the first reac-
tion of glycolysis. As before, all compounds of interest are represented as nodes, and
enzymatic reactions and transport steps are represented as edges. As noted earlier,
an alternative to this type of representation are bipartite graphs, in which two differ-
ent types of nodes represent metabolites and enzymatic reactions, and which are
formulated and analyzed as Petri nets [12, 13]. In the spirit of this chapter, we study
the fluxes through metabolic systems at steady state, and this restriction allows us
to ignore many aspects, such as temporal, kinetic, and regulatory features, which we
will study in Chapter 8.
3.7 Stoichiometric Networks

In a generic network of metabolic reactions, organic compounds flow from node to
node. Under the assumption of a steady state, all fluxes entering a given node must
collectively be equivalent in amount to all fluxes leaving this node. With this restric-
tion to the steady state, metabolic systems become static networks and can be rep-
resented as graphs. For simplicity of discussion, let us assume that all edges in the
metabolic network are unidirectional; that is, we have a situation as in network (A)
of Figure 3.1, where material can only flow from left to right and up, but not back-
wards. The constraints imposed by the material flow lead directly to mathematical
relationships, which are collectively called stoichiometry and which we illustrate
representatively with node N3 in the generic pathway of Figure 3.14. Namely, we
obtain, quite obviously,
F13 + F23 = F34 . (3.14)
F20
N2
F12 F24
F01 F40
N1 N4
F23
F13 F34
Figure 3.14 Generic network with four
nodes and fluxes entering and leaving
N3
them. At any steady state of the network,
the sum of all fluxes entering a node is
equal in magnitude to the total efflux from
this node.
For a node such as N2, we dissect the bidirectional flux F20 into two unidirectional,
forward and reverse, fluxes F02 and F20, which allows us to write a simple linear bal-
ance equation like (3.14) again.
Can the simple linear relationship in (3.14) be consistent with a fully dynamic
metabolic systems model? The answer is “Yes”—if we restrict the system to the
steady state. To see this, let us suppose for a moment that the system is not at steady
state. As we learned in Chapter 2, we can set up differential equations that describe
the network dynamics by equating the change in the amounts of each node with the
balance of all fluxes entering and exiting this node. Thus, for node N3, we formulate
a differential equation as
dN 3
= F13 + F23 − F34 , (3.15)
dt
which directly expresses that F13 and F23 are adding material to the pool N3, while F34
is siphoning off some of the material from the pool. In this dynamic formulation, F13,
F23, and F34 are typically time-dependent functions of metabolites, enzymes, and
modulators. However, at steady state, each flux takes on a fixed numerical value and
all time derivatives in the system must equal zero, because there is no change in any
of the variables. The result is a purely algebraic system, which is much easier to ana-
lyze. In the case of node N3, we obtain
dN 3
= F13 + F23 − F34 ⇒ 0 = F13 + F23 − F34 , (3.16)
dt
which recoups the earlier result (3.14). What’s happening to regulation that might
affect the fluxes? The answer is that all regulators affecting F13, F23, and/or F34 are
also in steady state and therefore constant, so that each flux is truly constant and
reduces to a numerical value. As a consequence, even for large, regulated systems,
the steady state reduces to a system of linear algebraic equations, in which the fluxes
along the edges are the variables.
The relative simplicity of investigating fluxes at steady state has led to enormous
research activities on metabolic systems, which sometimes are quite large and
contain dozens, if not hundreds, of metabolites and reactions. One reason for this
strong interest has been a desire to manipulate the natural flux distributions toward
a desired goal, such as the enhanced production of an organic compound in micro-
organisms, and the field of metabolic engineering has fully embraced these
methods (see, for example, [64–66] and Chapter 14). In its basic form, this type of
investigation, called stoichiometric analysis, goes back several decades [67, 68].
At its core is a stoichiometric model, which describes the distribution of all mate-
rial fluxes in a metabolic network, just as (3.14) and (3.16) do for a single node. The
general idea is simple and intuitive. One displays all metabolites and the fluxes
between them in a graph and sets up a differential equation for each metabolite. All
metabolites can be collected in a vector S, and the related vector S = dS/dt contains
the derivatives of these metabolites. A stoichiometric matrix N describes the con-
nectivity within the network.
As an example, consider the small network in Figure 3.15, which contains four
substrate pools A, . . ., D and seven fluxes R1, . . ., R7. The corresponding stoichiomet-
ric matrix N is also shown in Figure 3.15, with entries of +1 showing material enter-
ing a pool and entries of −1 representing fluxes leaving the pool. If two molecules
were used to create one product molecule, the matrix would contain entries of 2.
R1 R2 R3 R4 R5 R6 R7
R3 R6
B C
1 –1 0 –1 0 0 0 A
R2
R5 0 1 –1 0 0 0 0 B
R1
A N=
0 0 1 0 –1 –1 0 C Figure 3.15 Generic flux network (left)
R4 D and its stoichiometric matrix N (right).
R7 0 0 0 1 1 0 –1 D The matrix N quantifies which network
components are connected to each other.
The system of linear differential equations describing the reaction network can
be written in matrix form as
.
S = NR, (3.17)
where the vector R contains the fluxes between pools. Thus, the dynamics of metab-
olites in the vector S is dissected into a matrix of stoichiometric coefficients, to
which a vector of fluxes is multiplied.
In most stoichiometric investigations, the dynamical solutions of these differen-
tial equations, which would reveal all metabolites Si as functions of time, are actu-
ally not of primary interest, and one focuses instead on the distribution of fluxes at
steady state. In this situation, all the time derivatives in the vector on the left-hand
side of (3.17) are zero, so that the ith differential equation becomes an algebraic lin-
ear equation of the type
0 = N i1R1 + N i 2 R2 + + N im Rm (3.18)
which has the same structure as (3.16) in the simple example above. Now the fluxes
Ri are considered variables, which have to satisfy (3.18). More precisely, they must
collectively satisfy one equation of this type for each metabolite pool.
In the ideal case, we would be able to measure many fluxes within a network,
which would allow us to compute the remaining fluxes. In most realistic cases, how-
ever, only a few influxes and effluxes can be measured, and they are typically not
sufficient to identify the entire network. But with the exclusive focus on the steady
state, we have entered the realm of linear algebra, which offers methods for a great
variety of tasks. For instance, it allows us to characterize the solution of systems of
equations like (3.18), whether or not we have sufficiently many measured fluxes for
a full identification.
As an example, Kayser and co-workers [69] analyzed the metabolic flux distri-
bution in Escherichia coli under different glucose-limited growth conditions
(Figure 3.16). It is easy to check that the incoming and outgoing fluxes balance
exactly for each metabolite pool and for both steady-state growth conditions
(left and right numbers in beige boxes). For instance, 100 units flow from glucose
into glucose 6-phosphate (G6P), and these are forwarded in three directions:
depending on the experimental conditions, 58 or 31 units flow toward fructose
6-phosphate (F6P), 41 or 67 units are used for the production of ribulose
GLC
100
1/2 G6P 41 / 67 R5P
58 / 31 22 / 39 19 / 28
9 / 17
RBP 6/6
Figure 3.16 Flux distributions in cultures
F6P X5P
of E. coli growing under different
E4P 13 / 22 conditions. The blue arrows represent
80 / 70
enzyme-catalyzed reactions, while the boxes
DHAP GAP S7P contain flux measurements for two different
growth conditions. All magnitudes of
fluxes are represented in relationship to the
169 / 157
glucose consumption step, which is set as
13 / 22 100. Abbreviations: DHAP, dihydroxyacetone
16 /17 G3P phosphate; E4P, erythrose 4-phosphate; F6P,
fructose 6-phosphate; GAP, glyceraldehyde
153 / 140 3-phosphate; GLC, glucose; G6P, glucose
6-phosphate; G3P, 3-phosphoglycerate; OAA,
PEP 53 / 62 oxaloacetate; PEP, phosphoenolpyruvate;
PYR, pyruvate; R5P, ribulose 5-phosphate;
23 / 26 123 / 104 RBP, ribose 5-phosphate; S7P, sedoheptulose
7-phosphate; X5P, xylulose 5-phosphate.
(Data from Kayser A, Weber J, Hecht V & Rinas
23 / 26 OAA PYR 123 / 104
U. Microbiology 151 [2005] 693–706.)
5-phosphate (R5P), and 1 or 2 units move toward unspecified uses, respectively.

Note the apparent doubling between F6P and glyceraldehyde 3-phosphate (GAP),
which is due to the fact that F6P contains 6 carbons and GAP only 3. The system also
produces carbon dioxide and uses other compounds, such as ATP, which are not
explicitly listed.
The balances at each metabolite pool can be formulated as a set of simple (stoi-
chiometric) equations. If not all fluxes can be measured, these equations can be
used to infer them. For instance, the efflux of 16 or 17 units of 3-phosphoglycerate
(G3P) in Figure 3.16 simply balances the known influx from GAP and the efflux
toward phosphoenolpyruvate (PEP).
3.8 Variants of Stoichiometric Analysis

If only a few fluxes can be measured, the solution of the stoichiometric equations
is unique only in rare cases. The reason is that essentially all metabolic networks
contain more reactions than metabolite pools. As a mathematical consequence,
the system of equations is underdetermined and permits many different solu-
tions, which can all be characterized with methods of linear algebra. Indeed, all
these solutions form a solution space of sometimes high dimension. To determine
which of the many possible solutions is most likely to exist in nature, it has become
customary to posit a suitable criterion of optimality. The most typical criterion is
that a microorganism strives to optimize its growth rate. With such an objective,
one uses optimization methods to identify the one solution that satisfies all stoi-
chiometric steady-state equations (that is, of type (3.18)) and at the same time
optimizes the growth rate. Different variants have been proposed to address
underdetermined stoichiometric systems [70]. The most widely used is flux bal-
ance analysis (FBA) [64, 71]. Besides the balance between incoming and outgo-
ing fluxes at each node and the use of an objective to be optimized, FBA accounts
for biologically motivated constraints on some or all fluxes, which the system
must satisfy. These constraints may be thermodynamic or of some other physico-
chemical nature, and the core method of analysis is constrained optimization.
Other variants include elementary mode analysis [72], which will be discussed in
Chapter 14, extreme pathway analysis [64, 73], and the minimization of metabolic
adjustment [74, 75].
It is worth noting that the flux distribution within a static metabolic network
depends strongly not only on the optimization objective, but also on the experimen-
tal conditions. For instance, the flux distributions in Figure 3.16 differ quite a bit for
different substrate availabilities. As an example, algae of strain Synechocystis sp.
PCC 6803 can either draw energy from substrates such as glucose (heterotrophic
growth) or produce energy through photosynthesis (autotrophic growth). While the
metabolic network is the same in both cases, the two modes of growth correspond
to distinctly different flux distributions [76].
No matter which variation of stoichiometric analysis is applied, flux analysis does
not tell us what the metabolite concentrations in the system are. The focus is entirely
on fluxes, and if information on metabolites is of interest, it must come from addi-
tional data. In the most straightforward case, all concentrations can be measured at
the steady state. In other cases, it is sometimes but not always possible to deduce the
steady-state concentration if features of the catalyzing enzymes are known.
FBA can to some degree be extended to situations where the system is not in a
steady state [77]. However, this type of extended analysis is more complicated.
3.9 Metabolic Network Reconstruction

In the cases above, we assumed that we knew the structure of the metabolic network.
What needs to be done if we don’t know it? The answer is that the scope of the prob-
lem depends on the availability and type of data. Traditionally, biochemists have
been using wet-lab experiments, trying to identify reactions, regulators, and kinetic
properties associated with a particular metabolite. These studies are very labor-
intensive and slow, but have collectively produced an enormous body of knowledge,
which was originally depicted in the Boehringer Map that adorned many hallways of
species X species Y Figure 3.17 Filling gaps in a metabolic

map through genome comparison.
Suppose that metabolic and proteomic
? analysis has so far not revealed an enzyme EX
EY that would be able to connect two separate
EX
parts of the metabolic network of species
X. Suppose that an enzyme EY catalyzing
the missing reaction is known in the related
? species Y, and that its gene GY has been
GX GY identified. Sequence comparisons may reveal
a homologous gene GX in species X and
identify it as a candidate for coding for the
unknown enzyme EX.
?
biochemistry departments. This map is now available electronically [78], along with
other databases such as KEGG [79] and MetaCyc [80].
In more recent times, high-throughput methods of molecular biology have
become available for metabolic network characterization. Many of these focus on
the actual existence of specific metabolites in the target network and use meth-
ods such as mass spectrometry and nuclear magnetic resonance (see, for exam-
ple, Chapter 8 and [81]). Since genome data are usually easier to obtain than
information on the connectivity of metabolic pathway systems, methods have
also been developed to infer pathways from the presence of genes, homologies
with other organisms, and the genomic organization into operons and other reg-
ulatory structures (see, for example, Chapter 6). For instance, consider the situa-
tion in which there is a gap in an incomplete metabolic network of species X,
where some metabolite apparently cannot be produced with any of the enzymes
known in X (Figure 3.17). Suppose that a related species Y possesses a gene GY
that codes for an enzyme EY catalyzing the needed reaction and that could poten-
tially bridge the gap in X. If so, one uses homology methods of bioinformatics in
an attempt to determine whether X’s genome contains a gene GX that has suffi-
cient similarity with GY. If such a gene can be found, one infers that GX might code
for an enzyme EX that is similar to EY and fills the gap [61]. Free software associ-
ated with BioCyc and MetaCyc, such as BioCyc’s PathoLogic and KEGG’s KAAS
[82], serve this specific purpose [83–85]. An intriguing case study using this
approach is [86].
A different situation occurs when the connectivity and regulation of a metabolic
map is known, but specific kinetic parameters are not. Many methods for estimating
such parameters have been developed in recent years (see, for example, [87]). None
of these methods are foolproof, however, and all require quite a bit of computing
experience. We will return to this question in Chapter 5.
3.10 Metabolic control Analysis

An interesting transition from static to dynamic models (Chapter 4) is metabolic
control analysis (MCA), which characterizes the effects of small perturbations in a
metabolic pathway that operates at a steady state [88–91]. MCA was originally con-
ceived to replace the formerly widespread notion that every pathway has one rate-
limiting step, which is a slow reaction that by itself is credited with determining the
magnitude of flux through the pathway. The rate-limiting step could be compared to
the neck of a funnel, which controls how much liquid can run through the funnel
per second, independent of how much liquid has accumulated in the funnel. In lin-
ear pathway sections without branches, the rate-limiting step was traditionally
thought to be positioned at the first reaction, where it was possibly inhibited through
feedback exerted by the end product. In MCA, this concept of a rate-limiting step
was supplanted with the concept of shared control, which posits that every step in a
pathway contributes, to some degree, to the control of the steady-state flux. MCA
formalizes this concept primarily with quantities called control coefficients and
elasticities, and with mathematical relationships that permit certain insights into
E1 E2 E5 E6 Figure 3.18 Generic metabolic pathway

I S1 S2 S5 O used to explain concepts of metabolic
v1 v2 v5 v6 control analysis (MCA). The linear pathway
consists of input I, output O, intermediate
substrates S1, . . ., S5, and reactions v1, . . ., v6
that are catalyzed by enzymes E1, . . ., E6.
the control structure of the pathway. While many variations of the original quanti-
ties in MCA have been defined over the years [88, 91], one elasticity coefficient and
two primary control coefficients stand out: one with respect to the flux through the
pathway and one with respect to concentrations.
It might be most intuitive to illustrate the basic concepts of MCA with the linear
pathway in Figure 3.18. This generic pathway consists of an input I, output O, inter-
mediate substrate concentrations S1, . . ., Sn, and reactions v1, . . ., vn+1 that are cata-
lyzed by enzymes E1, . . ., En+1; in the example, n = 5. The pathway is assumed to
operate at a steady state, and all perturbations are required to be small; strictly
speaking, they must be infinitesimally small. In the generic terminology of systems
analysis, the control coefficients are sensitivities, as we will encounter in Chapter 4.
They quantify the effects of small changes in parameters on features of the system as
a whole. Outside of the minute perturbations in enzymes or other quantities,
dynamical aspects are not addressed in the original MCA, although later studies
have moved into this direction (see, for example, [92, 93]). As we discussed in Chap-
ter 2, the steady state corresponds to a situation where Si = 0 for all substrates, as well
as I = 0 and O = 0. It is easy to see that in this situation all reaction rates must have
the same magnitude as the overall flux J, namely v1 = v2 = . . . = v6 = J; the reader
should confirm this conclusion.
The control coefficients measure the relative change in a flux or substrate con-
centration at a steady state that results from a relative or percent change in a key
parameter, such as an enzyme activity. In other words, one asks a question such
as: “If one alters the activity of one of the enzymes, Ei, by a small amount such as
2%, how much will the flux J through the pathway change?” It is assumed that
each vi is directly proportional to the corresponding Ei, so that the control coef-
ficients may be equivalently expressed in terms of vi or Ei. This assumption holds
in most cases, although one should be aware that there are exceptions [94]. Fur-
thermore, it is mathematically convenient to replace the relative change with a
change in the logarithm of the quantity in question, which is valid if the changes
are small (the reader should confirm this). The flux control coefficient is thus
defined as the derivative
∂J ∂vi vi ∂J ∂ ln J
CvJi = = = . (3.19)
J vi J ∂vi ∂ ln vi
Similarly, the concentration control coefficient quantifies the effect of a change in

enzyme Ei on the steady-state concentration of metabolite Sk and is defined as
vi ∂Sk ∂ ln Sk
CvSik = = . (3.20)
Sk ∂vi ∂ ln vi
In a slight variation to the usual sensitivities in systems analysis, the control coeffi-
cients are thus defined as relative sensitivities, which are expressed in terms of loga-
rithms. This use of relative (or logarithmic) quantities is generally accepted as more
suitable in metabolic studies, because it removes the dependence of the results on
volumes and concentrations. For instance, according to (3.20), the change in Sk,
written as ∂Sk, is divided by the actual amount of Sk, and it is assessed as a conse-
quence of a change in vi, written as ∂vi, that is divided by the actual magnitude of vi.
Thus, relative changes are being compared.
The control exerted by a given enzyme either on a flux or on a metabolite con-
centration can be distinctly different, and it is indeed possible that a concentration
is strongly affected but the pathway flux is not. Therefore, the distinction between
the two types of control coefficients is important. The distinction is also pertinent for
practical considerations, for instance in biotechnology, where gene and enzyme
manipulations are often targeted either toward an increased flux or toward an
increased concentration of a desirable compound [91] (see Chapter 14). Finally, one
should note that several variations on this concept of control coefficients have been
explored. For instance, response coefficients may be defined for the dependence of
fluxes, concentrations, or other pathway features on other, external or internal,
parameters [88, 91].
Each elasticity coefficient measures how a reaction rate vi changes in response to
a perturbation in a metabolite Sk or some other parameter. With respect to Sk, it is
defined as
Sk ∂vi ∂ ln vi
ε Svik = = . (3.21)
vi ∂Sk ∂ ln Sk
Because only one metabolite (or parameter) and one reaction are involved in this
definition, but not the entire pathway, each elasticity is a local property, which can
in principle be measured in vitro. Closely related to the elasticity with respect to a
variable is the elasticity with respect to the Michaelis constant Kk of an enzyme for
the substrate Sk (see Chapters 2, 4, and 8). It is simply the complement of the metab-
olite elasticity [95]:
ε Svik = −ε Kvik . (3.22)
The main insights provided by MCA are gained from relationships among the
control and elasticity coefficients. Returning to the example in Figure 3.18, we may
be interested in the overall change in the flux J, which is mathematically determined
by the sum of responses to possible changes in all six enzymes. It has been shown
that all effects, in the form of flux control coefficients, sum to 1, and that all concen-
tration control coefficients with respect to a given substrate S sum to 0:
n +1
∑C
i =1
J
vi = 1, (3.23)
n +1
∑C
i =1
S
vi =0 (3.24)
[89, 90]. These summation relationships hold for arbitrarily large pathways under
the most pertinent conditions (for exceptions, see [94, 96]).
The summation relationship with respect to flux control, (3.23), has two interest-
ing implications. First, one sees directly that the control of metabolic flux is shared
by all reactions in the system; this global aspect identifies the control coefficients as
systemic properties. And second, if a single reaction is altered and its contribution
to the control of the flux changes, the effect is compensated by changes in flux con-
trol by the remaining reactions. The summation relationship with respect to flux
control has often been used experimentally in the following fashion: one measures
in the laboratory the effects of changes in various reactions of a pathway, and if the
sum of flux control coefficients is below 1, then one knows that one or more contri-
butions to the control structure are missing.
A second type of insight comes from connectivity relationships, which establish
constraints between control coefficients and elasticities. These relationships have
been used to characterize the close connection between the kinetic features of indi-
vidual reactions and the overall responses of a pathway to perturbations. The most
important of these relationships is
∑C ε
i =1
J vi
vi Sk = 0. (3.25)
Figure 3.19 Simple linear pathway with

– feedback. If the goal is to increase the
S X2 X3 X4 P flux through the pathway, MCA provides a
v1 v2 v3 v4 tool for identifying the best manipulation
strategies.
A comprehensive example of the use of these relationships is the analysis of

mitochondrial oxidative phosphorylation, which clearly demonstrated that the flux
control is distributed among a number of enzymes, thereby refuting the formerly
expected existence of a rate-limiting step (for a review, see [91]). A much simpler yet
instructive example is the linear pathway with feedback in Figure 3.19, which has
been adapted from [95]. The following elasticities are assumed to be known, maybe
from in vitro experiments:
ε vX12 = −0.9, ε vX22 = 0.5, ε vX23 = −0.2, ε vX33 = 0.7, ε vX24 = −1, and ε vX44 = 0.9;
all other elasticities are equal to 0. For this unbranched pathway, the information on
elasticities, together with the connectivity theorem in (3.25), is actually sufficient to
compute the flux control coefficients. For instance, the connectivity relationship for
X2 is
CvJ1 ε vX12 + CvJ2 ε vX22 = 0.19 × (−0.9) + 0.34 × 0.5 = 0.
Solving the set of these equations yields
CvJ1 = 0.19, CvJ2 = 0.34, CvJ3 = 0.10, and CvJ4 = 0.38.
Suppose there is reason to believe that the reaction step v2 is a bottleneck and
that a goal of the analysis is to propose strategies for increasing the flux through the
pathway. A reasonable strategy might be to increase the amount of the enzyme in v2,
for instance, using modern tools of genetic engineering. Or would it be better to
modify the enzyme in such a manner that it is less affected by the inhibition exerted
by X4? Suppose we could experimentally alter the effect of X4 by changing the bind-
ing constant K 42 of the enzyme in v2 by p%. For relatively small values of p, this
change corresponds to a change in ln K 42, as we discussed before. Owing to the inhi-
bition signal, K 42 has an effect on v2, which is quantified by the elasticity in (3.21) and
(3.22). Furthermore, the effect on changes in v2 on the pathway flux J is given by the
flux control coefficient in (3.19). Multiplication yields
∂ ln J ∂ ln v2 ∂ ln J
CvJ2 ε Kvi 2 = − =− , (3.26)
4 ∂ ln v2 ∂ ln K 42 ∂ ln K 42
rearrangement of which leads to
∂ ln J = −CvJ2 ε Kvi 2 ∂ ln K 42 ≈ −CvJ2 ε Kvi 2 p%. (3.27)

4 4
Substituting numerical values yields a relative change in flux J of 0.34 × (−1) per per-
cent change in the binding constant K 42 of enzyme v2. The predicted change has a
negative value; thus, in order to increase the pathway flux, K 42 must be decreased.
We could also explore a strategy of changing the activity of the enzyme in v4,
arguing that a higher activity would convert more X4, thereby reduce the inhibition
of v2, and possibly lead to a higher overall flux. Equation (3.19) quantifies this
effect:
∂ ln J = CvJ4 ∂ ln v4 ≈ CvJ4 q% = 0.38. (3.28)

Indeed, the effect is about 10% stronger than in the previous strategy of decreasing
the binding constant K 42 . In this fashion, changes in kinetic constants or enzymes
can be computed in a comprehensive manner [95].
For this small pathway, the computations are quite simple. For larger systems,
matrix methods have been developed that permit a more structured assessment of
the control within pathways [97, 98]. It has also been shown that the fundamental
features of MCA are directly reflected and generalized in the steady-state properties
of fully dynamic models in biochemical systems theory (BST) [99, 100] and in the
lin–log formalism [92, 93]. Thus, MCA can be seen as a direct transition from static
networks to dynamic systems analyses.
EXERcISES
3.1. Try to figure out Euler’s problem of the Seven 3.4. Construct 10 directed graphs with five nodes
Bridges of Königsberg (see Figure 3.2): Is there a path that in each case form a single loop, such as
that uses every bridge exactly once? Does the N1 → N5 → N3 → N4 → N2 → N1, and contain
solution change if any one of the bridges is removed? no other edges. How many different graphs of
If so, does it matter which bridge is removed? this type are there? Study similarities in their adja-
3.2. Are there connected paths in the graphs in cency matrices and summarize them in a brief report.
Figure 3.20 that use every edge exactly once? 3.5. How are hubs reflected in the adjacency matrix of
Does it matter where one starts? an undirected network?
3.6. How are cliques reflected in the adjacency matrix of
an undirected network?
3.7. What happens to the adjacency matrix of a graph if
this graph consists of two subgraphs that are not
connected to each other? Can you make a general
statement? Does this statement cover both directed
and undirected graphs?
3.8. Using the method in Box 3.1, compute all three-step
paths of the matrix on the left of Figure 3.4.
3.9. Confirm that in an undirected graph the number of
edges between all neighbors of a node is at most
k(k − 1)/2.
3.10. Compute the clustering coefficients of the red
nodes in the undirected graph of Figure 3.21.
Figure 3.20 Simple graphs with different

connectivities. Is it possible to find paths
that use every edge exactly once?
3.3. Construct 10 undirected random graphs with m

nodes and a total of mk edges, using different
combinations of m and k. As a simple start, study a Figure 3.21 Graphs with four nodes.
graph with six nodes and determine random edges The connectivity patterns determine their
with two dice, which identify the nodes to be clustering coefficients.
connected. Record the degree distributions. As
a more efficient alternative, write a computer 3.11. What are the minimum and maximum values of CN
algorithm and compute the degree distribution in a network without multiple edges between two
for each graph. nodes?
EXERcISES 79
3.12. What is the average clustering coefficient of a 3.23. Compute the probability that there is a
clique? genomic response in the system in Figure 3.13
3.13. Provide rationale for the two different definitions of even though R and S are not active. Interpret
CN for directed and undirected graphs. P(G spont) as the probability that G is on,
although C is not active.
3.14. Construct five random graphs with 6 nodes and 12
directed edges. Write a computer program or use two 3.24. Search the Internet for an investigation of a protein–
dice to determine source and end nodes; do not allow protein interaction network. Write a short report
double edges. Compute for each case all shortest path about the goals, methods, and results of the study.
lengths and the average shortest distance. Pay particular attention to the visualization of the
network.
3.15. Search the Internet for actual biological networks
with power-law degree distributions. Study γ and 3.25. Search the Internet for an investigation of a
discuss what happens to the power-law relationship gene interaction network. Write a short
for very high and very low k. report about the goals, methods, and results of the
study.
3.16. Search the Internet for two studies that discuss party
hubs and date hubs. Write a short report about the 3.26. Search the Internet for an example of a
goals, methods, and results of each study. stoichiometric or flux balance analysis. Write
a short report about the goals, methods, and
3.17. Box 3.2 discusses the construction of different
results of the study.
connectivity pattern from an initial ring-like structure.
Execute this construction yourself with different 3.27. If you are familiar with linear algebra and the
probabilities p (0 ≤ p ≤ 1). Specifically, study a ring features of matrices, discuss under what
with 12 nodes, which is connected as shown in Box conditions (3.17) can be solved uniquely or
3.2, and perform the analysis with p = 1/6, 2/6, . . ., not at all.
5/6. Start with the first node and go around the ring in 3.28. Set up a stoichiometric model for the pentose
one direction. For each node, roll a die to determine pathway shown in Figure 3.22. This pathway
whether one of its edges should be replaced or not. exchanges carbohydrates with different numbers of
For instance, for p = 2/6, replace an edge only if you carbon atoms, as shown in Table 3.1. Thus, two
roll a 1 or 2. If an edge is to be replaced, use a coin to units of X1 are needed to produce one unit of X2 and
determine whether to remove the node’s connection X3, and each reaction V5,3 uses one unit of X5 to
to its immediate right or left neighbor. Use a coin and generate two units of X3. Note that X3 appears in two
a die to determine the node to which a new edge is to locations, but represents the same pool.
be connected (head: nodes between 1 and 6; tail:
nodes between 7 and 12). In cases where you identify
the source node itself or where the edge already Xyl (X7)
exists, roll again. After one lap, repeat the procedure,
this time using second-nearest neighbors. Once you CO2 V7,1
have finished the second lap, report the degree

distribution, the network’s clustering coefficient, and G6P (X6) V6,1 Ri5P; Xy5P; Ru5P (X1)
the network’s diameter. As a more efficient alterna-
tive, write an algorithm that implements the above
V1,2
procedure with m nodes and a probability
p (0 ≤ p ≤ 1).
3.18. Discuss the tolerance of random and small-world V5,6
S7P (X2) GAP (X3)
networks to random attacks and to targeted attacks,
V3,0
where one node is eliminated in each case.
V2,5
3.19. Load the galactose utilization network in Cyto-
scape. Explore different modes of visualization. F6P (X5)
3.20. Load a different sample network in Cytoscape E4P (X4)

V5,3
and determine its numerical features, such
V4,5
as the number of nodes, and clustering information.
GAP (X3)
3.21. Load the galactose utilization network in Cyto-
scape. Select two hubs and find out whether they
V3,0
and their neighbors are functionally related.
3.22. Compute the probability that the cascade in Figure
3.13 is active even though R and S are not active. Figure 3.22 Diagram of the pentose phosphate pathway. Given
Interpret P(C spont) as the probability that C is on, influxes and effluxes, internal fluxes can be computed. Table 3.1 shows
although neither R nor S is active. Compute the the number of carbon atoms in each metabolite. (From Wiechert W &
probabilities that R or S is not active as 1 − P(R) and de Graaf AA. Biotechnol. Bioeng. 55 [1997] 101–117. With permission
1 − P(S), respectively. from John Wiley & Sons.)
corresponding flux distributions. How much CO2 is

TABLE 3.1: NUMBER OF CARBON produced?
ATOMS IN EACH METABOLITE IN
THE PATHWAY IN FIGURE 3.22 3.30. If the influx in the pathway systems in Figures 3.16
or 3.22 were doubled, would doubling all other
compound Number of fluxes in the system lead to a valid solution? Argue
pool carbon Atoms from a mathematical and a biological point of view.
X1 5 3.31. If the rate of a reaction is given as v2 (S1 ) = kS1p, where
k and p are constant parameters, what are the
X2 7
concentration control coefficient CvS21 and the
X3 3 elasticity ε Sv12 ?
X4 4 3.32. If the rate v of a reaction is given as a Michaelis–
X5 6 Menten function of the type
X6 6
VS
X7 5 v(S ) = ,
K +S
3.29. For the same pathway as shown in Figure 3.22, where V and K are parameters, what is the
assume that the value for the input flux V7,1 is 20 corresponding elasticity? Does the elasticity depend
units and try different values for V3,0. Compute the directly on the substrate concentration?
REFERENcES
[1] Cytoscape. http://www.cytoscape.org/. [20] Wagner A & Fell DA. The small world inside large
[2] Shannon P, Markiel A, Ozier O, et al. Cytoscape: a software metabolic networks. Proc. R. Soc. Lond. B 268 (2001)
environment for integrated models of biomolecular 1803–1810.
interaction networks. Genome Res. 13 (2003) 2498–2504. [21] Chung F & Lu L. The average distance in random graphs with
[3] Ingenuity. http://www.ingenuity.com. given expected degrees. Proc. Natl Acad. Sci. USA 99 (2002)
[4] BiologicalNetworks. http://biologicalnetworks.net. 15879–15882.
[5] Pajek. http://pajek.imfm.si/doku.php?id=pajek. [22] Jeong H, Tombor B, Albert R, et al. The large-scale organization
[6] Purvis JE, Radhakrishnan R & Diamond SL. Steady-state kinetic of metabolic networks. Nature 407 (2000) 651–654.
modeling constrains cellular resting states and dynamic [23] Jordán F & Molnár I. Reliable flows and preferred patterns in
behavior. PLoS Comput. Biol. 5 (2009) e1000298. food webs. Evol. Ecol. Res. 1 (1999) 591–609.
[7] Lee Y & Voit EO. Mathematical modeling of monolignol [24] Lima-Mendez G & van Helden J. The powerful law of the
biosynthesis in Populus xylem. Math. Biosci. 228 (2010) 78–89. power law and other myths in network biology. Mol. Biosyst.
[8] Wilkinson DJ. Stochastic Modelling for Systems Biology, 5 (2009) 1482–1493.
2nd ed. Chapman & Hall/CRC Press, 2011. [25] Nacher JC & Akutsu T. Recent progress on the analysis
[9] Keener J & Sneyd J. Mathematical Physiology. II: Systems of power-law features in complex cellular networks. Cell
Physiology, 2nd ed. Springer, 2009. Biochem. Biophys. 49 (2007) 37–47.
[10] Alon U. An Introduction to Systems Biology: Design Principles [26] Barabási A-L. Linked: The New Science of Networks. Perseus,
of Biological Circuits. Chapman & Hall/CRC, 2006. 2002.
[11] Voit EO. Computational Analysis of Biochemical Systems: A [27] Albert R & Barabási AL. The statistical mechanics of complex
Practical Guide for Biochemists and Molecular Biologists. networks. Rev. Mod. Phys. 74 (2002) 47–97.
Cambridge University Press, 2000. [28] Vázquez A, Flammini A, Maritan A & Vespignani A. Modeling of
[12] Chaouiya C. Petri net modelling of biological networks. Brief protein interaction networks. ComPlexUs 1 (2003) 38–44.
Bioinform. 8 (2007) 210–219. [29] Newman ME & Watts DJ. Scaling and percolation
[13] Hardy S & Robillard PN. Modeling and simulation of molecular in the small-world network model. Phys. Rev.
biology systems using Petri nets: modeling goals of various E 60 (1999) 7332–7342.
approaches. J. Bioinform. Comput. Biol. 2 (2004) 619–637. [30] Fraser HB, Hirsh AE, Steinmetz LM, et al. Evolutionary rate in
[14] Watts DJ & Strogatz SH. Collective dynamics of “small-world” the protein interaction network. Science 296 (2002) 750–752.
networks. Nature 393 (1998) 440–442. [31] Carlson JM & Doyle J. Complexity and robustness. Proc. Natl
[15] Ravasz E, Somera AL, Mongru DA, et al. Hierarchical Acad. Sci. USA 99 (2002) 2538–2545.
organization of modularity in metabolic networks. Science [32] Csete ME & Doyle JC. Reverse engineering of biological
297 (2002) 1551–1555. complexity. Science 295 (2002) 1664–1669.
[16] Jeong JP, Mason SP, Barabási A-L & Oltvai ZN. Lethality and [33] Doyle JC, Alderson DL, Li L, et al. The “robust yet fragile”
centrality in protein networks. Nature 411 (2001) 41–42. nature of the Internet. Proc. Natl Acad. Sci. USA 102 (2005)
[17] Maslov S & Sneppen K. Specificity and stability in topology of 14497–14502.
protein networks. Science 296 (2002) 910–913. [34] Cohen JE. Graph theoretic models of food webs. Rocky
[18] Erdős P & Rényi A. On random graphs. Publ. Math. 6 (1959) Mountain J. Math. 9 (1979) 29–30.
290–297. [35] Pimm SL. Food Webs. University of Chicago Press, 2002.
[19] Barabási A-L & Oltvai ZN. Network biology: understanding [36] de Lichtenberg U, Jensen LJ, Brunak S & Bork P. Dynamic
the cell’s functional organization. Nat. Rev. Genet. 5 (2004) complex formation during the yeast cell cycle. Science
101–113. 307 (2005) 724–727.
REFERENcES 81
[37] Han JD, Bertin N, Hao T, et al. Evidence for dynamically [61] Reed JL, Vo TD, Schilling CH & Palsson BØ. An expanded
organized modularity in the yeast protein–protein interaction genome-scale model of Escherichia coli K-12 (iJR904 GSM/
network. Nature 430 (2004) 88–93. GPR). Genome Biol. 4 (2003) R54.
[38] Agarwal S, Deane CM, Porter MA & Jones NS. Revisiting [62] Wagner A. Evolutionary constraints permeate large metabolic
date and party hubs: novel approaches to role assignment networks. BMC Evol. Biol. 9 (2009) 231.
in protein interaction networks. PLoS Comput. Biol. 6 (2010) [63] Schryer DW, Vendelin M & Peterson P. Symbolic flux analysis
e1000817. for genome-scale metabolic networks. BMC Syst. Biol.
[39] Ideker T, Thorsson V, Ranish JA, et al. Integrated genomic and 5 (2011) 81.
proteomic analyses of a systematically perturbed metabolic [64] Palsson BØ. Systems Biology: Properties of Reconstructed
network. Science 292 (2001) 929–934. Networks. Cambridge University Press, 2006.
[40] Xia T, Van Hemert J & Dickerson JA. CytoModeler: a tool [65] Stephanopoulos GN, Aristidou AA & Nielsen J. Metabolic
for bridging large-scale network analysis and dynamic Engineering. Principles and Methodologies. Academic Press,
quantitative modeling. Bioinformatics 27 (2011) 1578–1580. 1998.
[41] Olobatuyi ME. A User’s Guide to Path Analysis. University Press [66] Torres NV & Voit EO. Pathway Analysis and Optimization in
of America, 2006. Metabolic Engineering. Cambridge University Press, 2002.
[42] Shipley B. Cause and Correlation in Biology: A User’s Guide [67] Clarke BL. Complete set of steady states for the general
to Path Analysis, Structural Equations and Causal Inference. stoichiometric dynamical system. J. Chem. Phys. 75 (1981)
Cambridge University Press, 2000. 4970–4979.
[43] Torralba AS, Yu K, Shen P, et al. Experimental test of a method [68] Gavalas GR. Nonlinear Differential Equations of Chemically
for determining causal connectivities of species in reactions. Reacting Systems. Springer, 1968.
Proc. Natl Acad. Sci. USA 100 (2003) 1494–1498. [69] Kayser A, Weber J, Hecht V & Rinas U. Metabolic flux analysis
[44] Vance W, Arkin AP & Ross J. Determination of causal of Escherichia coli in glucose-limited continuous culture. I.
connectivities of species in reaction networks. Proc. Natl Acad. Growth-rate-dependent metabolic efficiency at steady state.
Sci. USA 99 (2002) 5816–5821. Microbiology 151 (2005) 693–706.
[45] Anastassiou D. Computational analysis of the synergy among [70] Trinh CT, Wlaschin A & Srienc F. Elementary mode analysis:
multiple interacting genes. Mol. Syst. Biol. 3 (2007) 83. a useful metabolic pathway analysis tool for characterizing
[46] Shannon CE & Weaver W. The Mathematical Theory of cellular metabolism. Appl. Microbiol. Biotechnol. 81 (2009)
Communication. University of Illinois Press, 1949. 813–826.
[47] Butte AJ & Kohane IS. Mutual information relevance networks: [71] Varma A, Boesch BW & Palsson BØ. Metabolic flux balancing:
functional genomic clustering using pairwise entropy basic concepts, scientific and practical use. Nat. Biotechnol.
measurements. Pac. Symp. Biocomput. 5 (2000) 415–426. 12 (1994) 994–998.
[48] Komili S & Silver PA. Coupling and coordination in gene [72] Schuster S & Hilgetag S. On elementary flux modes in
expression processes: a systems biology view. Nat. Rev. Genet. biochemical reaction systems at steady state. J. Biol. Syst.
9 (2008) 38–48. 2 (1994) 165–182.
[49] Bayes T. An essay towards solving a problem in the doctrine of [73] Schilling CH, Letscher D & Palsson BØ. Theory for the systemic
chances. Philos. Trans. R. Soc. Lond. 53 (1763) 370–418. definition of metabolic pathways and their use in interpreting
[50] Pearl J. Causality, Models, Reasoning, and Inference. metabolic function from a pathway-oriented perspective.
Cambridge University Press, 2000. J. Theor. Biol. 203 (2000) 229–248.
[51] Mitra S, Datta S, Perkins T & Michailidis G. Introduction to [74] Lee Y, Chen F, Gallego-Giraldo L, et al. Integrative analysis of
Machine Learning and Bioinformatics. Chapman & Hall/CRC transgenic alfalfa (Medicago sativa L.) suggests new metabolic
Press, 2008. control mechanisms for monolignol biosynthesis. PLoS
[52] Sachs K, Perez O, Pe’er S, et al. Causal protein-signaling Comput. Biol. 7 (2011) e1002047.
networks derived from multiparameter single-cell data. [75] Segrè D, Vitkup D & Church GM. Analysis of optimality in
Science 308 (2005) 523–529. natural and perturbed metabolic networks. Proc. Natl Acad.
[53] Bork P, Jensen LJ, Mering C von, et al. Protein interaction Sci. USA 99 (2002) 15112–15117.
networks from yeast to human. Curr. Opin. Struct. Biol. [76] Navarro E, Montagud A, Fernández de Córdoba P &
14 (2004) 292–299. Urchueguía JF. Metabolic flux analysis of the hydrogen
[54] Zhang A. Protein Interaction Networks: Computational production potential in Synechocystis sp. PCC6803. Int.
Analysis. Cambridge University Press, 2009. J. Hydrogen Energy 34 (2009) 8828–8838.
[55] Rhodes DR, Tomlins SA, Varambally S, et al. Probabilistic model [77] Goel G, Chou I-C & Voit EO. System estimation from
of the human protein–protein interaction network. Nat. metabolic time series data. Bioinformatics 24 (2008)
Biotechnol. 23 (2005) 951–959. 2505–2511.
[56] Libby E, Perkins TJ & Swain PS. Noisy information processing [78] Expasy. http://web.expasy.org/pathways/.
through transcriptional regulation. Proc. Natl Acad. Sci. USA [79] KEGG: The Kyoto Encyclopedia of Genes and Genomes. http://
104 (2007) 7151–7156. www.genome.jp/kegg/pathway.html.
[57] Zhu J, Wiener MC, Zhang C, et al. Increasing the power to [80] MetaCyc. http://biocyc.org/metacyc/index.shtml.
detect causal associations by combining genotypic and [81] Harrigan GG & Goodacre R (eds). Metabolic Profiling: Its Role
expression data in segregating populations. PLoS Comput. in Biomarker Discovery and Gene Function Analysis. Kluwer,
Biol. 3 (2007) e69. 2003.
[58] Hecker M, Lambeck S, Toepfer S, et al. Gene regulatory [82] KAAS. http://www.genome.jp/tools/kaas/.
network inference: data integration in dynamic models—a [83] Pathologic. http://biocyc.org/intro.shtml#pathologic.
review. Biosystems 96 (2009) 86–103. [84] Karp PD, Paley SM, Krummenacker M, et al. Pathway Tools version
[59] Gasch AP, Spellman PT, Kao CM, et al. Genomic expression 13.0: integrated software for pathway/genome informatics and
programs in the response of yeast cells to environmental systems biology. Brief Bioinform. 11 (2010) 40–79.
changes. Mol. Biol. Cell 11 (2000) 4241–4257. [85] Paley SM & Karp PD. Evaluation of computational metabolic-
[60] Sieberts SK & Schadt EE. Moving toward a system genetics pathway predictions for Helicobacter pylori. Bioinformatics
view of disease. Mamm. Genome 18 (2007) 389–401. 18 (2002) 715–724.
[86] Yus E, Maier T, Michalodimitrakis K, et al. Impact of genome Rijksuniv. Gent Fak. Landbouwkd. Toegep. Biol. Wet.
reduction on bacterial metabolism and its regulation. Science 66 (2001) 11–30.
326 (2009) 1263–1268. [94] Savageau MA. Dominance according to metabolic control
[87] Chou I-C & Voit EO. Recent developments in analysis: major achievement or house of cards? J. Theor.
parameter estimation and structure identification of Biol. 154 (1992) 131–136.
biochemical and genomic systems. Math. Biosci. 219 (2009) [95] Kell DB & Westerhoff HV. Metabolic control theory: its role
57–83. in microbiology and biotechnology. FEBS Microbiol. Rev.
[88] Fell DA. Understanding the Control of Metabolism. Portland 39 (1986) 305–320.
Press, 1997. [96] Savageau MA. Biochemical systems theory: operational
[89] Heinrich R & Rapoport TA. A linear steady-state treatment differences among variant representations and their
of enzymatic chains. Critique of the crossover theorem and significance. J. Theor. Biol. 151 (1991) 509–530.
a general procedure to identify interaction sites with an [97] Reder C. Metabolic control theory: a structural approach.
effector. Eur. J. Biochem. 42 (1974) 97–105. J. Theor. Biol. 135 (1988) 175–201.
[90] Kacser H & Burns JA. The control of flux. Symp. Soc. Exp. Biol. [98] Visser D & Heijnen JJ. The mathematics of metabolic control
27 (1973) 65–104. analysis revisited. Metab. Eng. 4 (2002) 114–123.
[91] Moreno-Sánchez R, Saavedra E, Rodríguez-Enríquez S & Olín- [99] Savageau MA, Voit EO & Irvine DH. Biochemical systems
Sandoval V. Metabolic control analysis: a tool for designing theory and metabolic control theory. I. Fundamental
strategies to manipulate metabolic pathways. J. Biomed. similarities and differences. Math. Biosci. 86 (1987)
Biotechnol. 2008 (2008) 597913. 127–145.
[92] Hatzimanikatis V & Bailey JE. MCA has more to say. J. Theor. [100] Savageau MA, Voit EO & Irvine DH. Biochemical systems
Biol. 182 (1996) 233–242. theory and metabolic control theory. II. The role of summation
[93] Heijnen JJ. New experimental and theoretical tools and connectivity relationships. Math. Biosci. 86 (1987)
for metabolic engineering of microorganisms. Meded. 147–169.
FuRThER READING
Alon U. An Introduction to Systems Biology: Design Principles of Palsson BØ. Systems Biology: Properties of Reconstructed Networks.
Biological Circuits. Chapman & Hall/CRC, 2006. Cambridge University Press, 2006.
Barabási A-L. Linked: The New Science of Networks. Perseus, 2002. Stephanopoulos GN, Aristidou AA & Nielsen J. Metabolic
Davidson EH. The Regulatory Genome: Gene Regulatory Networks In Engineering. Principles and Methodologies. Academic Press,
Development and Evolution. Academic Press, 2006. 1998.
Fell DA. Understanding the Control of Metabolism. Portland Press, Zhang A. Protein Interaction Networks: Computational Analysis.
1997. Cambridge University Press, 2009.
The Mathematics of
Biological Systems 4
• Understand basic modeling approaches
• Discuss the need for approximations
• Explain concepts and features of approximations
• Linearize nonlinear systems
• Approximate nonlinear systems with power-law functions
• Describe the differences and relationships between linear and nonlinear
models
• Describe different types of attractors and repellers
• Design and analyze basic models with one or more variables in the following
formats:
○ Discrete linear
○ Continuous linear
○ Discrete nonlinear
○ Continuous nonlinear
• Execute the steps of a typical biological systems analysis
Chapter 2 introduced some general concepts for modeling biological systems in an

intuitive fashion, while skipping a lot of technical detail. Chapter 3 continued the
discussion by addressing static networks, which do not change over time. This chap-
ter presents mathematical and computational methods for analyzing dynamic
systems. The range of such methods is almost unlimited, and not even experts
master them all in detail. Nonetheless, there are various standard methods that are
sufficient for setting up and analyzing basic models in a variety of formats, even if
they may not have all the “bells and whistles” of the most modern mathematical
techniques and software packages. If you are satisfied with understanding, albeit
from a bird’s eye view, what mathematical modeling is all about, and do not neces-
sarily intend to engage in its minutiae, you may just browse through some sections
or skip the technical details in this chapter, as long as you have a good grasp of the
material in Chapters 2 and 3. However, if you love to “look under the hood,” design
your own models, and enjoy the thrills of exploring biology on your computer, this
chapter is for you. It will put you on the right track, and, even if it is not comprehen-
sive, it will provide you with a solid platform from which to dive into the rich litera-
ture of modeling, which spans a vast range from relatively easy introductions to very
84 Chapter 4: The Mathematics of Biological Systems
specialized texts and a growing number of journal articles. All analyses shown in
this chapter can be executed in Mathematica® and MATLAB®, and many can be
done with the easy-to-learn software PLAS [1] or even in Excel®.
As discussed in the introduction to modeling in Chapter 2, there are numerous
types and implementations of mathematical and computational models. Because of
their prominence in systems biology, this chapter is heavily biased toward dynamic,
deterministic systems models and discusses only a few stochastic models that
depend on random processes. The chapter requires, here and there, basic working
knowledge of ordinary and partial differentiation, vectors, and matrices, but most
material can be understood without deep knowledge of these topics. The main goal
of the chapter is to formalize concepts of model design, diagnostics, and analysis in
more depth than in Chapter 2 and to provide the skills to set up and analyze models
from scratch. In contrast to a typical math book, which uses the famous lemma–
theorem–proof–corollary–example style, this chapter introduces concepts and
definitions on the fly, uses examples as motivation, and stops before things get too
complicated. It is subdivided into discussions of discrete and continuous models,
as well as of linear and nonlinear models.
Linear systems models are of utmost importance for two reasons. First, they per-
mit more mathematical analyses than any other format, and, second, it is possible to
analyze many features of nonlinear systems with methods of linear analysis.
A very interesting feature germane to linear systems is the principle of superpo-
sition. This principle says the following. If we analyze a system by giving it an
input I1, to which it responds with output O1, and if we do a second experiment
with input I2 and obtain output O2, then we know that the output response to a com-
bined input I1 + I2 will be O1 + O2. Maybe this superposition sounds obvious, trivial,
or unassuming, but it has enormous consequences. Only if this principle holds is it
possible to analyze the behavior of a system by studying each component separately
and then validly summing the individual results. As a banal example, we could
establish in one experiment a relationship between the pressure in the tires of a car
and its performance, and in a separate experiment a relationship between
performance and different grades of gas. Under the assumption that tire pressure
and gas quality independently affect performance, the results are additive. The rea-
son why we have not chosen a biological example here is that most biological
systems are indeed not linear, thereby violating the principle of superposition. For
instance, the growth of a plant depends on genetics as well as environmental condi-
tions, including nutrients. However, the two are not independent of each other, and
separate experiments with different strains and different experimental conditions
may not allow us to make valid predictions of how a particular strain will grow under
untested conditions.
If we are able to forge a system representation into a linear format, an incredible
repertoire of mathematical and computational tools becomes available. In particu-
lar, the entire field of linear algebra addresses these systems and has developed con-
cepts such as vectors and matrices that have become absolutely fundamental
throughout mathematics and computer science. For instance, if the right-hand
sides of a system of differential equations are linear, the computation of the steady
state of this system becomes a matter of solving a system of linear algebraic equa-
tions, for which there are many methods.
Linear representations can come in discrete and continuous varieties. Linear
(matrix) models in biology can effectively describe discrete, deterministic aging
processes, as well as some stochastic processes (such as the Markov chain models
discussed later in this chapter) that capture the probabilities with which a system
develops over time. Discrete linear models are also at the heart of much of traditional
statistics, such as regression, analysis of variance, and other techniques, which are
efficiently formulated using matrix algebra. These types of statistical models are
very important, but are outside our scope here.
Continuous linear representations are crucially important for biology as well,
even though most systems in biology are nonlinear. The reason is that many analyses
depend on the fact that nonlinear systems can be linearized, which essentially
means that one may flatten out the nonlinear system close to a point of interest.
The justification for this strategy comes from an old theorem attributed to Hartman
and Grobman, which says that, under most pertinent conditions, important clues
DiScreTe Linear SySTeMS MoDeLS 85
for the behavior of a nonlinear system can be deduced from the linearized system
[2]. We will make use of this insight when we discuss stability and sensitivities.
DiScreTe Linear SySTeMS MoDeLS

4.1 recursive Deterministic Models
We begin our illustration of linear systems with discrete models that are constructed
recursively. This terminology means that time is only considered in discrete steps
and that time periods in between are ignored (as on the display of a digital clock).
Furthermore, the state of the system at time t is represented as a linear function of
the state of the system at time t − 1. In the simplest, single-variable, case, the state is
characterized by a single feature, whereas a multivariate state is characterized by
several, simultaneous features. For instance, the blood pressure “state” of a person
consists of two variables, namely, a systolic and a diastolic pressure reading. The
state of an ecosystem might be describable in terms of the numbers of organisms
within each species at a given time, but it could also include temperature, pH, and
other factors.
A typical single-variable example is the size Pt of a bacterial population at time
points t = 0, 1, 2, …. The time points may refer to minutes, hours, days, years, or any
other fixed time periods, such as 20-minute intervals. The index notation Pt is com-
monly used for discrete models, whereas the notation P(t) usually refers to a
continuous model, where any values of t are permitted.
For simplicity let us assume that the bacteria are synchronized and all divide
simultaneously every τ minutes. Without these assumptions, the situation quickly
becomes complicated [3, 4]. Then, we can formulate a simple recursive model of
the type
Pt +τ = 2 Pt , (4.1)
which simply says that at time point t + τ there are twice as many bacteria as at time
t, and this is true for any time point t. Note that, in this formulation, t may progress
from 1 to 2 to 3, or maybe from 60 to 120 to 180, while τ remains fixed at 1 or 60,
respectively, throughout the history of the population. Also note that the process
really describes exponential growth, although the equation is called a linear equa-
tion, because its right-hand side is linear.
Suppose the population starts with 10 cells at time t = 0 and the division time is
τ = 30 minutes. Then we can easily project the growth of this population, according
to Table 4.1: every 30 minutes, the population size doubles. Using (4.1) for the
second doubling step, we immediately see
P(t +τ )+τ = 2 Pt + τ = 2 ⋅ 2 ⋅ Pt = 22 ⋅ Pt , (4.2)
and this relationship is true no matter how far t has progressed in its discrete steps.
By extension, if we move forward n doubling steps of 30 minutes each, we find
Pt +nτ = 2n Pt . (4.3)
This formulation contains two components: the initial population size at some time
point t, which is Pt, and the actual doubling process, which is represented by the
term 2n. Because the power function 2n is defined as exp(n ln 2), this process is
called exponential growth. Much has been written about this process, and further
exploration is left for the Exercises. A feature to ponder is what happens (biologi-
cally and mathematically) at time points between t and t + τ or between t + 3τ and
TABLE 4.1: GROWTH OF A BACTERIAL POPULATION ACCORDING TO (4.1)

Time (min) 0 30 60 90 120 150 180
Population size 10 20 40 80 160 320 640
t + 4τ. Of course the bacterial population does not cease to exist, but the model does
not address these time points. The reader should think about how the model could
be altered to accommodate these missing time points.
A slightly more complicated recursive growth model, which has found wide
application in population studies, was proposed by Leslie [5]. The key difference
from the bacterial growth process in (4.1) is that the population in this model
consists of individuals with different ages and different proliferation rates. Thus, two
processes overlap: individuals age and also bear children, with a fecundity rate that
depends on their age. Furthermore, a certain portion of each age group dies. We will
discuss this model in Chapter 10.
Leah Edelstein-Keshet introduced a small, yet informative example of a recur-
sive model [6]. It describes, in a very simplified form, the dynamics of red blood
cells. These cells are constantly formed by the bone marrow and destroyed by the
spleen, but their number should remain more or less constant from day to day.
In the simple model, the loss is modeled as proportional to the current number of
cells in circulation, and the bone marrow produces a number of new cells that is
proportional to the number of cells lost on the previous day. Thus, there are two
critical numbers, namely,
Rn = number of red blood cells in circulation on day n
and
Mn = number of red blood cells produced by the bone marrow on day n.
These numbers depend recursively on cell numbers during the previous day.
Specifically, Rn is given as the number of red blood cells in circulation on the previ-
ous day, and this number is diminished by losses through degradation in the spleen
and increased by cells that were newly produced by the bone marrow on the
previous day. If the fraction of cells removed by the spleen is f, which typically is a (A)
1300
small number between 0 and 1, and if the production rate is denoted by g, which is
defined as the number of cells produced per number of cells lost, we can formulate
the following model equations: 1200
1100
Rn = (1 − f )Rn−1 + M n−1 , (4.4a) Rn
1000
M n = g f Rn−1 . (4.4b)
900
Similar to the time counter in the previous example, n is simply a counter represent-
ing the day under investigation. As a consequence, we can write these recursive 800
0 5 10 15 20
equations also with shifted indices as iteration number n
(B)
25
M n+1 = g f Rn , (4.5a)
20
Rn+1 = (1 − f )Rn + M n , (4.5b) 15

Mn
and the two formulations (4.4) and (4.5) are exactly equivalent. 10
We can now ask interesting questions. For instance, given starting values R0 and
M0, and f and g, how do the numbers Rn and Mn develop over time? The problem is 5
simple enough to allow an analysis in Excel®. As a specific example, suppose
R0 = 1000, M0 = 10, f = 0.01, and g = 2. Solving the equations from n = 0 to n = 1, n = 2, 0
and so forth shows that both Rn and Mn continue to increase (Figure 4.1). Resetting 0 5 10 15 20
iteration number n
g = 1, the plot is different: both data trends are flat. Setting f = 0.05, g = 1 causes an
initial jump, followed by flat trends at Rn ≈ 962 and Mn ≈ 48. Thus, the initial 1010
Figure 4.1 Trends in red blood cells over
cells become differently distributed (the reader should confirm these results).
time. (A) The number of cells in circulation,
The model allows us to compute under what conditions the system exhibits a Rn, and (B) the number of cells produced
steady state, which is defined by the situation where neither Rn nor Mn changes in by the bone marrow, Mn, versus number of
value from one iteration to the next. Specifically, the condition of a steady state iterations, n. The model settings were R0 =
requires Rn+1 = Rn and Mn+1 = Mn. The easiest way to assess these conditions is to call 1000, M0 = 10, f = 0.01, and g = 2.
up the equation for Mn, (4.4b), and plug the result into the equation for Rn+1, (4.5a).
The result is
Rn+1 = (1 − f )Rn + M n = (1 − f )Rn + g f Rn−1 , (4.6)
which contains R at three time points, but not M. For R and M to be in a steady state
on days n − 1, n, and n + 1, we must demand that Rn+1 = Rn = Rn−1, because if this
condition is given, then Rn+2, Rn+3 and all later values of R will remain the same. This
pair of equations has two solutions, namely, Rn+1 = Rn = Rn−1 = 0, which is not very
interesting, and
1 = (1 − f ) + g f , (4.7)
which, for any positive f, is satisfied exclusively for g = 1. What does that mean bio-
logically? The condition mandates that the bone marrow produces exactly as many
cells as are lost, which in retrospect makes a lot of sense. Interestingly, this is true no
matter the value of f.
So, what are the actual steady-state values for R and M? We know that g must be
equal to 1; so let us assume that that is true. Furthermore, let us pick some (yet
unknown) value for f and try to compute the steady-state values Rn and Mn from
(4.5) by again requiring Rn+1 = Rn and Mn+1 = Mn. The second equation yields
Mn = Mn+1 = f Rn; thus, the number of red blood cells produced by the bone marrow
on day n must equal the cells lost on day n. Plugging this result into the first equation
of the set just yields Rn = Rn+1 = (1 − f )Rn + f Rn = Rn, which means that the system is
at a steady state for any values of f and Rn, as long as g = 1 and Mn = f Rn. The reader
should confirm this result with numerical examples and ponder whether it makes
sense biologically.
It is beneficial for many analyses to rewrite these types of linear recursive
equations as a single matrix equation. In our case, this equation is
 R 1 − f 1  R 
 M  =  g f 0  M  n
. (4.8)
n +1
According to the rules for multiplying a vector by a matrix, we obtain the value for
 R
Rn+1 by multiplying the first row of the matrix with the vector   , which yields
 Mn
Rn+1 = (1 − f )Rn + Mn, and we obtain Mn+1 by multiplying the second row of the
matrix with the same vector, which yields Mn+1 = g f Rn. Thus, the matrix equation is
a compact equivalent of the recursive equations (4.5a, b). This representation is
particularly useful for large systems.
The matrix formulation makes it easy to compute future states of the system,
such as Rn+2 or Mn+4. First, we know that
 R 1 − f 1  R 
 M  =  g f 0  M  n+1
, (4.9)
n+2
 R
because n is only an index. Secondly, we can express the vector   according to
 M  n+1
the matrix equation (4.8). Combining these two aspects, the result is
 R 1 − f 1  R  1 − f 1  1 − f 1  R 
 M  =  g f =
0  M  n+1  g f 0  g f 0  M  n
n+2
2
1 − f 1  R 
= . (4.10)
 gf 0  M  n
According to the rules of matrix multiplication,
A B a b   Aa + Bc Ab + Bd 
 C = , (4.11)
D   c d   Ca + Dc Cb + Dd 
we obtain
2
1 − f 1  (1 − f )2 + g f 1− f 
= .
0  g f (1 − f ) g f 
 g f (4.12)
The matrix equation is not limited to two steps, and, indeed, any higher values of
Rn+k and Mn+k can be computed directly from Rn and Mn, when we use the corre-
sponding power of the matrix:
k
 R 1 − f 1  R 
 M  =  g f 0  M  n
. (4.13)
n+k
Matrix methods for recursive systems can also be used to analyze the stability of
steady states or to make statements regarding the long-term trends of the system [6].
We will discuss a popular matrix model of the growth of an age-structured popula-
tion in Chapter 10.
The red blood cell model is of course extremely simplified. Nonetheless, the
same structure of a recursive model can be used for realistic models of the dynamics
of red blood cells in health and disease [7, 8].
4.2 recursive Stochastic Models

The linear growth models discussed so far are entirely deterministic. Once they have
been defined and started, their fate is sealed. This is not so in a certain class of prob-
abilistic models, called Markov models or Markov chain models, which superfi-
cially have a format similar to a larger version of (4.8); a good introduction to these
models is [9]. The crucial difference is that they have stochastic features, which
means that they account for random events (see Chapter 2). Generically, a Markov
model describes a system that can assume m different states. At each (discrete) time
point, the system is in exactly one of these states, from where it transitions with
some probability into a different state or stays where it is.
An intuitive example is a pinball machine, where the states are the typical flip-
pers and bumpers. An adequate time step in this case is the time for the ball to move
from one state to the next. For simplicity let us assume that the travel times between
states are all the same and call the time points 0, 1, …, t, t + 1, …. At each time point,
the one and only ball is in one state, from where it transitions, with some probability,
to another state or stays in the same state. A Markov chain model is constructed by
looking at very many games with the same pinball machine. Imagine we could
record all movements of the ball from all games. If so, we would find balls and move-
ments all over the place. However, some movements would certainly be observed
more often than others, and, focusing on some given state i, each transition to some
other state k would be observed with a specific frequency. The idea behind a Markov
model is to use this frequency and interpret it as the probability of future transitions
from state i to state k. Thus, the Markov chain process is defined by a fixed set of
probabilities of transitions between all pairs of states: if a ball is in state i, the
probability of moving to state j is pij, and this is true no matter where the ball was
previously; the Markov model has no memory. For a process with m states, we can
put all probabilities into an m × m matrix, which is called the Markov matrix:
 p11 p12 p1,m−1 p1m 

 p21 p22 p2 ,m−1 p2 m 
 
M =  p31 p32 p3 ,m−1 p3 m  . (4.14)
 

 pm1 pm 2 pm ,m−1 pmm 
We can use M to compute the dynamics of the Markov chain. Let us assume that
at the beginning (t = 0) the ball is in state 1, which could be interpreted as the start
position. Thus, X1 = 1 and all other Xi = 0. At the next observation time point (t = 1)
in game 1, the ball has moved to another state (or stayed in its start position). In the
next game, the ball might have moved from the start position to a different state,
and, in a third game, it may have moved to yet another position. This process is ran-
dom for each game, but collectively governed by probabilities in M: p13 is the prob-
ability of moving from state 1 to state 3 in one time unit, and p16 is the probability of
moving to state 6. Thus, the first row of the matrix describes what percentages of
games are expected to move the ball from state 1 into any of the states i, namely p1i.
With the next move, the ball moves again, according to the probabilities in (4.14).
Tracking all these expectations and transitions amounts to a big bookkeeping prob-
lem, which the matrix equation (4.14) solves beautifully.
What do we know about the matrix elements pij? They describe the probabilities
of moving from state i to state j, where j must be one of the m states in the equation;
there is no other option. Of course, every probability is between 0 and 1. It is actually
possible that, for instance, p13 = 0, which would mean that it is impossible to move
from state 1 to state 3 in one step. The ball could get to 3 in several steps, but not in
one. If a probability pij is equal to 1, this means that the ball, if in state i, has no
choice but to move to state j.
Suppose a ball is in state 4. Because it must move to one of the other states
(or stay in state 4), all available probabilities, namely p4j (j = 1, …, m), summed
together must equal 1. Generally, we find, for each i,
∑p
j =1
ij = 1. (4.15)
It is possible that one of the probabilities for leaving state i is 1 and all others are
0. For instance, if pi2 = 1, then it is certain that every ball in state i in every pinball
game of this type moves to state 2 next. In particular, if pii = 1, the ball can never again
leave state i. Game over! Note that the summation over the first index (a column),
p1j + p2j + ⋯ + pmj, does not necessarily yield 1 (the reader should explain why).
Markov chain models can be found in diverse corners of systems biology.
For instance, they may describe the dynamics of networks, they can be used to com-
pute the probability of extinction of a population, and they appear in the form of
hidden Markov models in bioinformatics [10]. The adjective “hidden” in this case
refers to the situation that one cannot tell whether the process is in one hidden state
or another.
As an example, consider the transition diagram in Figure 4.2, which corresponds
to the matrix
 0.1 0.5 0.4

M =  0.2 0.3 0.5 . (4.16)
 
 0.8 0.2 0  0.1
Suppose the process starts at time t = 1 in state X2, which we denote as X1 = 0, X2 = 1, X1

X3 = 0. Linear algebra allows us to determine the probabilities of finding the process in
each of the three states at time t = 2. This computation is accomplished by entering the
initial values into a start vector and multiplying this vector by the transposed matrix
MT, whose rows are the columns of M and whose columns are the rows of M. Thus, 0.5 0.2 0.4 0.8
0.5
 X1   X1   0  0.1 0.2 0.8  0  0.2 X2 X3
 X 2  = M T  X 2  = M T  1 =  0.5 0.3 0.2  1 =  0.3 . (4.17) 0.3 0.2 0.0
           
 X 3   X 3   0  0.4 0.5 0   0  0.5
2 1
Figure 4.2 Diagram of a Markov process.
The process is in one state, X1, X2, or X3, at
Intuitively, the flipping of the matrix is necessary because M contains probabilities each time point and moves among these
of moving forward in time, whereas (4.17) in a sense describes the “backwards” states with the probabilities shown on the
computation of the state vector at time 2 from the state vector at time 1. figure.
After a second step, the probabilities of finding the process in each of the three
states are computed from the result in (4.17) as
 X1   X1   0.2  0.1 0.2 0.8  0.2  0.48

 X 2  = M T  X 2  = M T  0.3 =  0.5 0.3 0.2  0.3 =  0.29 . (4.18)
           
 X 3   X 3   0.5  0.4 0.5 0   0.5  0.23
3 2
In both results (4.17) and (4.18), the three probabilities add up to 1.

Similar to the discrete red blood cell model, the vector on the left-hand side of
(4.18) can also be computed in a single step from the vector at time 1, by using the
second power of the matrix MT:
 X1   X1 
 X 2  = (M T )2  X2 . (4.19)
   
 X 3   X 3 
3 1
In the same manner, the probabilities of finding the process in each of the three
states at any later time point can be computed directly from the probabilities of the
states at time 1 through the use of correspondingly higher powers of MT.
The square of the Markov matrix itself, M2, contains two-step probabilities: for
instance, the element (1, 2) expresses the probability of moving from state 1 to
state 2 in exactly two steps. The kth power of the matrix contains k-step
probabilities.
Under certain mathematical conditions, which are satisfied here [9], continuing
the same type of multiplication as in (4.19) eventually leads to something like a
steady state of this Markov chain process, namely, the so-called limiting distribution
among the states, which does not change with further multiplications. As this distri-
bution is approached, the columns of the powers of MT eventually become identical.
For instance, the 16th power of MT is
 0.35088 0.35088 0.35088

(M T )16 =  0.33918 0.33918 0.33918 , (4.20)
 
 0.30994 0.30994 0.30994
and higher powers are the same within the given numerical resolution.
Analogously, high powers of M no longer change, and we obtain for all high
powers k
 0.35088 0.33918 0.30994

M =  0.35088
k
0.33918 0.30994 . (4.21)
 
 0.35088 0.33918 0.30994
As an example, the element (1, 2) in this matrix is the probability to reach state 2
from state 1 in exactly k steps. Note that this element has the same value as (2, 2) and
(3, 2). This equivalence is no coincidence and can be interpreted as follows: after
sufficiently many steps, it no longer matters where we started from—in the world of
Markov models, history becomes less and less important for the present and future
the further it lies in the past.
Ultimately, each column of (MT)∞ and each row of M∞ consists of the compo-
nents of the limiting distribution, which is typically denoted as a vector of πi’s. In the
given case,
 π 1   0.35088
 π 2  =  0.33918 . (4.22)
   
 π 3   0.30994
DiScreTe nonLinear SySTeMS 91
This limiting distribution can also be computed as the non-negative solution of the
matrix equation
 π1   π1 
 π 2  = MT  π 2  , (4.23)
   
 π 3   π 3 
where the sum of the π ’s equals 1. The values πi may be interpreted as the probabili-
ties of finding the Markov process in state 1, 2, or 3, respectively, once the process
has had enough time to converge. They also reflect the average proportions of time
that the process spends in these states. Equation (4.23) expresses the fact that this
distribution does not change if the Markov matrix is applied to it. Nevertheless, the
Markov process remains stochastic, and the next state can only be predicted in a
probabilistic manner.
The Markov model falls into the huge field of stochastic processes, which repre-
sent phenomena that contain random components [9]. Other examples are the
binomial and Poisson processes in Boxes 2.3 and 2.4 in Chapter 2. The possibilities
of combining systems analysis with randomness are unlimited, and every time we
develop a deterministic model, we should ask whether we should account for sto-
chasticity. There are no general guidelines or answers to this question, and sound
judgment is the best advice. As discussed in Chapter 2, the goals and purposes of the
model, as well as the technical difficulties associated with each candidate method,
should govern this thought process. In favor of stochastic models is the observation
that experimental data always contain noise, which leads to some randomness, and
that some (many? or indeed all?) biological processes have features that make them
unpredictable. However, there is also much to be said in favor of deterministic
models. First and foremost, they are usually much simpler to analyze. Second, if
noise is superimposed on some otherwise rather smooth dynamical process,
subtraction of the noise may eliminate mathematical distraction and allow us to
home in on the truly important determinants of the process.
In the next section and toward the end of the chapter, we will discuss systems
that are chaotic and appear to be driven by randomness, although they are entirely
deterministic. In these cases, the immediate future is predictable, but details of the
long-term dynamics cannot be foreseen. The stochastic systems have the opposite
feature: their long-term behavior can be computed, but the next step cannot.
DiScreTe nonLinear SySTeMS

In contrast to linear systems, nonlinear models require us to make an additional,
rather challenging choice. Namely, we need to determine or select the most suitable
function(s) for the right-hand sides of the difference equations or ordinary
differential equations (ODEs) that characterize these models. Linear models
always have the same generic format, but in the case of nonlinear models, infinitely
many alternatives are available.
Nonlinear representations are often more realistic for biological systems than
their linear counterparts. For instance, the linear model in (4.1) may be a good
representation for the growth of small bacterial cultures, but it is clear that the model
eventually becomes unrealistic, because it predicts that the entire world would
ultimately be covered with a thick (and growing) layer of this particular bacterium.
In reality, all biological systems have mechanisms that slow down growth or
production, for instance, through feedback. These mechanisms are usually nonlin-
ear and lead to time trends that are distinctly different from the time courses we
observe in linear differential or difference equations.
To illustrate how quickly nonlinearities may become drastically different from
linear models, let us consider the logistic map. This recursive model is not all that
different from the bacterial growth equation (4.1) but contains an additional term
that makes the model nonlinear:
Pt +1 = rPt (1 − Pt ). (4.24)
The starting point P1 is somewhere between 0 and 1, and r is a parameter between 0

and 4. It is easy to compute Pt+1 from Pt and r, but it is very difficult to make general
predictions of what this little model will do if we let it run. We shall just look at a few
cases and leave the rest as an exercise. As the first illustration, suppose r = 1.5. Start-
ing with P1 = 0.5, the model quickly approaches its fixed point or steady state of ⅓.
Indeed, if we plug r = 1.5 and Pt = ⅓ into (4.24), we can easily check that Pt+1 also
equals ⅓ (Figure 4.3A), which implies that all further P’s are also ⅓; convince
(A) (B)
0.6 0.9
0.8
0.5
0.7
0.4 0.6
0.5
Pt 0.3 Pt
0.4
0.2 0.3
0.2
0.1
0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20 22 24
iteration t iteration t
(C) (D)
1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
Pt Pt
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 0 2 4 6 8 10 12 14 16 18 20 22 24
iteration t iteration t
(E)
1.0
0.9
0.8
0.7
0.6
Pt 0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
iteration t
Figure 4.3 Values of the logistic map (4.24) over time. Iterations of this map can lead to amazingly complicated behaviors. The transition from a simple
steady state to chaos leads through phases of period doubling: (A) r = 1.5; (B) r = 3.2 (period 2); (C) r = 3.5 (period 4); (D) r = 3. 56 (period 8); (E): r = 3.8
(chaos).
conTinuouS Linear SySTeMS 93
yourself that this is true. Turning this argument around, we can compute the fixed
point of a logistic model if we know r: namely, we require Pt+1 = Pt and numerically
solve for Pt. Thus, for r = 2, we obtain a fixed point of 0.5. Similarly, for r = 3.8, we
compute the fixed point as 0.7368 and confirm that we obtain Pt+1 = 0.7368 if
Pt = 0.7368, as well as Pt+2 = 0.7368, along with all future iterations. However, if we
start the recursion with r = 3.8 and a value of P1 that is not 0.7368, the model jumps
around erratically, and one can show that its fluctuations are so unpredictable that
in the field of dynamical systems they are called chaotic (Figure 4.3E). The transition
from an “orderly” approach of a steady state to chaos exhibits several identifiable
steps, where the periodicity of the oscillation changes. This so-called “period
doubling” happens for r ≈ 3.2 (period 2), r ≈ 3.5 (period 4), r ≈ 3.56 (period 8) (see
Figure 4.3B–D). Further doublings become more difficult to decipher. We will
discuss other chaotic processes toward the end of the chapter.
The logistic map has attracted a lot of attention in the literature, and features
such as the stability of its fixed point and the progression toward chaos have been
studied intensely (see, for example, [11]). Considering the surprising dynamics of
the simple logistic map, it is easy to imagine how it is possible for iterative maps with
many variables and more complicated nonlinearities to become extremely compli-
cated. Indeed, the cover of this book shows a nonlinear iterative map based on a
complex polynomial. Depending on the details of such a map, the results are called
fractals, Mandelbrot sets, or Julia sets. They can be quite diverse and are often
stunningly beautiful.
conTinuouS Linear SySTeMS

Even if a phenomenon is clearly discrete, it is often prudent to describe it with a
continuous dynamical model. For instance, if we want to study the growth of a bac-
terial culture, it does not make much sense to account for every bacterium. Instead,
we may want to focus on the mass, volume, or optical density of the entire culture,
and these features are most easily described with continuous variables. The switch
from discrete to continuous systems results in variables that no longer change in
steps but in a smooth manner. If we study the development of an embryo, the early
cell divisions of the fertilized egg are well described in a discrete fashion, and we
even talk about four- and eight-cell stages. Later, however, we measure the length or
weight of the embryo on a continuous scale, and growth is better described by a dif-
ferential equation that expresses the smooth change. This smooth change is given
by a derivative on the left-hand side and reflected by a continuous function on the
right-hand side.
As in the case of discrete models, continuous models may be linear or nonlinear.
And again, a note is in order to avoid confusion. Namely, even if the right-hand sides
of a set of differential equations are linear, the solutions to these differential equa-
tions are combinations of linear as well as exponential and trigonometric functions,
which are clearly nonlinear. The best-known example is that of sine–cosine oscilla-
tions. These oscillations can be represented by a pair of linear differential equations,
which allow us to study them in a better way than with the explicit sine and cosine
functions. This pair is
x′ = y, x (0) = 0,
(4.25)
y′ = −x, y (0) = 1.
First we need to convince ourselves that these two equations really do correspond to
a sine–cosine oscillator. Indeed, x and y are functions of time, namely, x(t) = sin t
and y(t) = cos t. The derivative of sin t is cos t, thus satisfying the first equation, while
the derivative of cos t is −sin t, thus satisfying the second equation. Furthermore,
for t = 0, the sine function is zero and the cosine is 1. Put the system in a numerical
solver and, voilà, sine and cosine oscillations emerge. We will return to this
system later.
In later sections, the right-hand sides of the differential and difference equations
that we consider are themselves nonlinear, and the solutions to these systems can
be much more complicated than exponential and trigonometric functions.
4.3 Linear Differential equations

The simplest example of this type is the linear differential equation
dX
= X = aX , (4.26)
dt
which describes exponential growth or decay. Expressed in words from left to right,
the equation says that change (growth), as expressed by the derivative on the left-
hand side, is directly proportional to the current size or amount of the quantity X,
with a proportionality constant a. If X and a are positive, the change is positive, and
X continues to grow. If a is negative, an initially positive X keeps decreasing. Ponder
what happens if both a and X are negative.
Note that we write X rather than X(t), even though X depends on time, and that
time t is no longer seen on the right-hand side. This omission is a convention of
simplicity, and it is implicitly understood that X is a function of time.
In the simple case of (4.26), we can actually compute the solution of the differen-
tial equation, which explicitly describes the dependence of X on time. It consists of
the exponential function X(t) = X0eat, because differentiation of X yields dX/dt =
aX0eat, which equals aX(t) and thus the right-hand side of (4.26). We see that X0 is
visible in the time-dependent solution, but not in the differential equation.
Nevertheless, X0 has a clearly defined meaning as the initial value of X at time t = 0.
This is easy to confirm, when we substitute 0 for t in X(t) = X0eat, which directly
returns the result X(0) = X0. In fact, this last statement is often added to the definition
of the differential equation in (4.26). As in the case of the recursive growth process,
the differential equation with a positive initial value goes either to infinity (positive
a) or to zero (negative a), unless a happens to be 0, which means that X remains at its
initial value X0 forever. If X0 is negative, the solution can go to minus infinity or zero,
depending on a. Because it takes infinitely long to reach the final value exactly, expo-
nential decay processes are often characterized by their half-life, which is the time it
takes for half of some amount of material to be lost. In addition to growth or decay
processes, equations like (4.26) are used to describe transport processes in bio-
chemical and physiological systems and for physical phenomena such as heating
and cooling. We will return to this important equation in Chapter 8.
More interesting than a single linear differential equation is a system of linear
differential equations. These appear to have the same general form as (4.26), namely,
dX
= X = AX , (4.27)
dt
but here X is now a vector and A is a matrix with coefficients Aij. We have already
encountered these models when we discussed stoichiometric models of metabolic
pathways (Chapter 3). They are also used to study physiological compartment
models, where nutrients, drugs, or toxic substances can move between the blood-
stream and the organs and tissues in the body (for details see, for example, [12] and
Chapter 13).
The main feature of systems like (4.27) is that they allow us to compute explicit
solutions, instead of requiring numerical simulation that are almost always the
only solution for nonlinear systems. If X has n elements, these solutions consist of
n time-dependent functions Xi(t) that are determined by the eigenvalues of A,
which are real or complex numbers that characterize a matrix. For small systems,
they can be computed algebraically, and for larger systems, there are numerous
computational methods and software programs. Assuming that the eigenvalues of
A are all different, the solution of (4.27) consists of a simple sum of exponential
functions whose rate constants equal the eigenvalues of the system matrix. How-
ever, because these exponential functions may have complex arguments, the solu-
tions also include trigonometric functions, such as sine and cosine. An example
was given in (4.25).
Many methods are available for studying linear differential equations, including
the so-called Laplace transform, which permits analyses without the need to
integrate the differential equations [13].
4.4 Linearized Models

Someone once said that the distinction between linear and nonlinear models is like
a classification of the animal kingdom into elephants and non-elephants. It is true:
there is essentially only one linear structure, but there are uncounted—and in fact
infinitely many—nonlinear structures. Nevertheless, linear models are of utmost
importance even for nonlinear systems, because some analyses of nonlinear sys-
tems can be reduced to a linearized analog. The strategy is to flatten the nonlinear
system in a small region around a point of particular interest, called the operating
point, OP. This point can be chosen arbitrarily, as long as it is possible to compute
derivatives of the nonlinear system there. In practical investigations, OP is often the
steady state.
The graph in Figure 4.4 shows the linearization of a curved surface at some
point P. The linearization is visualized as the tangent plane, which at P has exactly
the same value as the curved surface and also the same slopes. As a consequence,
all paths Ci on the surface that go through P have the same local directions Ti in the
tangent plane, and many analyses can validly be done in this tangent plane.
As an illustration, suppose a system has the form
dX
= f ( X ,Y ),
dt
(4.28)
dY
= g ( X ,Y ),
dt
where f and g are nonlinear functions. The first order of business is the selection of
an operating point, where we anchor the linearization. The main ingredient of the
linearization process is differentiation, and, because the system has two indepen-
dent variables, X and Y, the derivatives must be partial derivatives. These are
computed like regular derivatives under the assumption that the other variable is
kept constant. The linearization is based on a theorem of the English mathematician
Brook Taylor (1685–1731), who discovered that a function may be approximated
with a polynomial, which in the simplest case is a linear function (see Box 4.1).
For our demonstration, we approximate f and g in the vicinity of our operating
point of choice, (XOP, YOP), which, from now on, we always assume to be a steady-
state point. This assumption is of course not always true, but it reflects by far the
most important situation. We denote a solution in the vicinity of the steady-state
operating point by [XOP + X ′(t), YOP + Y ′(t)], so that X ′(t) and Y ′(t) are deviations in
the X- and Y-directions. Because we intend to stay fairly close to the operating point,
this solution is called a perturbation. The strategy is now to formulate this perturba-
tion as a system of differential equations. Simply plugging in the perturbed values
into (4.28) yields
d
( X OP + X ′) = f ( X OP + X ′ ,YOP + Y ′),
dt
(4.29)
d
(YOP + Y ′) = g ( X OP + X ′ ,YOP + Y ′),
dt
where the argument (t) is again not shown.
tangent plane
C1
T1 C2
T3
T2
P
C3
Figure 4.4 Illustration of the linearization

of a curved surface. At P, the tangent plane
and the curved surface touch and have the
same slopes.
BOX 4.1 APPROXIMATION
Very many approximations in mathematics are based on an old to see that the original function f and the Taylor approximation are
theorem by the English mathematician Brook Taylor (1685–1731), exactly the same at the operating point p: just substitute p for x.
who showed that any smooth function (one that has sufficiently An important special case is linearization, where n = 1. That
many continuous derivatives) can be represented as a power series. is, one keeps merely the constant term and the term with the first
If this series has infinitely many terms, it is an exact representation derivative and ignores everything else. This strategy sounds quite
of the original function within some region of convergence. radical, but it is extremely useful, because the linearity of the result
However, if the series is truncated, that is, if one ignores all terms permits many simplified analyses close to the operating point. The
beyond a certain number, the resulting Taylor polynomial is an result of the linearization is
approximation of the function. This approximation is still exact at
a chosen operating point and very similar to the original function L( x ) = f ( p ) + f (1) ( p )( x − p ), (2)
close to this operating point. However, for values distant from
which has the familiar form of the equation y = mx + b for a straight
the operating point, the approximation error may be large. The
line, with slope m and intercept b (Figure 1). Here, m corresponds to
coefficients of the Taylor polynomial are computed through
the derivative f (1)(p), and the vertical-axis intersect b is f(p) − p f(1) (p).
differentiation of the original function.
As an example, consider the shifted exponential function
Specifically, if a function f(x) has at least n continuous
derivatives, it is approximated at the operating point p by f (x) = ex +1 (3)
1 (2) and choose x0 = 0 as the operating point. The linearization
f ( x ) ≈ f ( p ) + f ′( p )( x − p ) +f ( p )( x − p )2
2! includes the value of f at x0, which is 2, and the first derivative,
(1)
1 1 evaluated at the operating point, which is
+ f ( 3) ( p )( x − p )3 + + f ( n ) ( p )( x − p )n .
3! n!
f (1) (0) = e x = 1. (4)
The expressions 2!, 3!, and n! are factorials, which we have already
Thus, the linearization of f at (x0) is
encountered in Box 2.3, and f (1), f (2), f (3), and f (n) denote the first,
second, third, and nth derivatives of f(x), respectively. It is not difficult L( x ) = 2 + 1⋅ ( x − 0) = 2 + x . (5)
10
f(x)
5
L(x)
Figure 1 Taylor approximation can be used to linearize a nonlinear

function. Here, L(x) is the linear approximation of the shifted
–2 0 2 exponential function f(x) in (3), developed at the operating point
x x0 = 0.
To analyze the right-hand sides of this equation, we use Taylor’s insight

(Boxes 4.1 and 4.2) and obtain the linear approximation
∂f ∂f
f ( X OP + X ′ ,YOP + Y ′) ≈ f ( X OP ,YOP ) + X′+ Y ′ , (4.30a)
∂X ( X OP ,YOP ) ∂Y ( X OP ,YOP )
∂g ∂g
g ( X OP + X ′ ,YOP + Y ′) ≈ g ( X OP ,YOP ) + X′+ Y ′. (4.30b)
The curly ∂ signifies partial differentiation and the vertical lines with subscripts indi-
cate that this differentiation is evaluated at the operating point. Because we chose
the operating point as a steady state, where dX/dt = dY/dt = 0 in (4.28), the terms
f(XOP, YOP) and g(XOP, YOP) are 0 and drop out.
We can gain some intuition about (4.30a, b) by dissecting them into recognizable
components (Figure 4.5). Expressed in words, the function f (or g) can be
BOX 4.2 LINEARIZATION OF NONLINEAR FUNCTIONS IN TWO VARIABLES
Intriguingly, an approximation analogous to the Taylor repre- At the operating point, L has the same value and the same
sentation in Box 4.1 can be performed in higher dimensions. For slopes as f.
the case of two dependent variables, one obtains the approxima- As an example, consider the function f(x, y) = yex and choose
tion the operating point (x0, y0) = (0, 0). The linearization includes the
value of f at (x0, y0), which is 0, and the first derivatives, evaluated
A( x , y ) = f ( x 0 , y 0 ) + ( x − x 0 )fx ( x 0 , y 0 ) + ( y − y 0 )fy ( x 0 , y 0 ) at the operating point. These are
1
+ [( x − x 0 )2 fxx ( x 0 , y 0 ) + 2( x − x 0 )( y − y 0 )fxy ( x 0 , y 0 ) (1) fx (0, 0) = ye x = 0 ⋅1 = 0 and f y ( 0 , 0 ) = e x = 1. (3)
2!
+ ( y − y 0 )2 fyy ( x 0 , y 0 )] + .
Thus, the linearization of f at (x0, y0) is
The first term on the right-hand side is the value of f at the chosen L( x , y ) = 0 + ( x − 0) ⋅ 0 + ( y − 0) ⋅1 = y . (4)
operating point (x0, y0), the expressions fx and fy are the first
derivatives of f with respect to x and y, respectively, and fxx, fxy, The linearization is flat with respect to the x-direction and has a
and fyy are the corresponding second derivatives of f, which are all slope of 1 in the y-direction. Some values of f and L are given in
evaluated at the operating point. Table 1. The approximation for small x-values is good, even if y is
Ignoring the second and all higher derivatives, we obtain the large, because both f and L are linear with respect to y. By contrast,
linearization L(x,y). Specifically, we can write large x-values lead to much higher approximation errors, because
L does not even depend on x, owing to the choice of x0 = 0 for the
L( x , y ) = f ( x 0 , y 0 ) + ( x − x 0 )fx ( x 0 , y 0 ) + ( y − y 0 )fy ( x 0 , y 0 ). (2) operating point, while f increases exponentially with respect to x.
TABLE 1: SELECTED VALUES OF f(x, y) AND L(x, y) FOR DIFFERENT COMBINATIONS OF x AND y
(x, y) (0, 0) (0, 1) (0.1, 0.1) (0.1, 1) (0.2, 0.2) (0.1, 0.5) (0.5, 0.1) (0.5, 0.5) (1, 0.1)
f(x, y) 0 1 0.115 1.105 0.244 0.553 0.165 0.824 0.272
L(x, y) 0 1 0.1 1 0.2 0.5 0.1 0.5 0.1
approximated at some point by using the value at this point and adding to it devia-
tions in the directions of the dependent variables. Each deviation consists of the
slope in that direction (expressed by the partial derivative) and the distance of the
deviation along the corresponding axis (here X ′ and Y ′, respectively). This method
works for any number of dependent variables.
The left-hand sides of (4.30a, b) can be simplified as well. Namely, we are allowed
to split the differential of each sum into a sum of two differentials. Furthermore,
dXOP/dt = dYOP/dt = 0, because XOP and YOP are constant values. Thus, only dX ′/dt
and dY ′/dt are left. Taking the simplifications of the two sides together, the
statements in (4.29) and (4.30) become
d ∂f ∂f
X′ ≈ X′+ Y ′, (4.31a)
dt ∂X ( X OP ,YOP ) ∂Y ( X OP ,YOP )
A(X,Y)
Y
∂f
—
∂X
A(XOP+X ′ ,YOP+Y ′ ) Figure 4.5 Visualization of part of a plane
A(X,Y) (pink) approximating some function
∂f f(X,Y), which is not shown, at some
—
∂Y operating point (XOP and YOP) (green).
Y OP The plane is constructed according to (4.30a).
The blue lines have the slope of f at the
operating point in the X-direction, while the
red lines have the slope of f at the operating
point in the Y-direction. The green lines
Y′ indicate distances in the two directions from
X′
the operating point (X′ and Y′, respectively).
X OP The point A(XOP + X′, YOP + Y′) (red) lies on
the plane but is only approximately equal to
f(XOP + X′, YOP + Y′) (not shown). The two are
equivalent for X′ = Y′ = 0 and similar for small
X values of X′ and Y′.
d ∂g ∂g
Y′ ≈ X′+ Y ′. (4.31b)
dt ∂X ( X OP ,YOP ) ∂Y ( X OP ,YOP )
Note that this linearization could be a rather crude approximation if f and g are
strongly curved. But the approximation may also be very good for moderate
deviations from the operating point. Without further information about f and g, we
just don’t know, even for rather large perturbations. We do know that the lineariza-
tion is good if the perturbation is sufficiently small (whatever that means in a
specific situation). A second important result is that the right-hand sides are linear.
We can rename the partial derivatives in (4.31), evaluated at the operating
point, as
∂f ∂f
a11 = , a12 = ,
(4.32)
∂g ∂g
a21 = , a22 = ,
which identifies the linearization of (4.28) as
dX ′
= a11 X ′ + a12Y ′ ,
dt
(4.33)
dY ′
= a21 X ′ + a22Y ′.
dt
In general, we may write the linearization of a system with m variables in matrix

form as
 ∂f1 ∂ f1 ∂ f1 

 ∂X 1 ∂X 2 ∂X m 
 X 1′     X 1′ 
 X 2′   ∂f 2 ∂f 2 ∂f 2   
X 2′
d   / dt =  ∂X 1 ∂X 2 ∂X m    . (4.34)
     
X     
 m′   ∂f m  X m′ 
∂f m ∂f m 
 
 ∂X 1 ∂X 2 ∂X m  OP
The matrix on the right-hand side is so important that it has its own name: the
Jacobian of the system (at the operating point OP).
The linearization describes approximately the distance of the nonlinear system
from its steady-state operating point. Specifically in our case,
X ′(t ) ≈ X (t ) − X OP ,
(4.35)
Y ′(t ) ≈ Y (t ) − YOP .
Close to the operating point, the linear representation is very good, but on moving
away from this point, the approximation error typically becomes larger. Important
investigations into the stability and sensitivity of linear and nonlinear systems are
executed in a close vicinity of the steady state, so that the accuracy of the approxi-
mation becomes secondary for these purposes (see later in this chapter).
As a numerical demonstration of the linearization process, consider the system
X = f ( X ,Y ) = 2Y − 3 X 4 ,
(4.36)
Y = g ( X ,Y ) = 0.5e X Y − 2Y 2 .
The system has the steady state (Xss ≈ 0.776, Yss ≈ 0.543), which we use as our
operating point (OP). The values of the partial derivatives, evaluated at OP, are
∂f
a11 = = −12 X 3 = −5.608,
∂X ( X OP ,YOP )
( X OP ,YOP )
∂f
a12 = = 2,
∂Y ( X OP ,YOP )
(4.37)
∂g
a21 = = 0.5e Y
X
= 0.590,
∂X ( X OP ,YOP )
( X OP ,YOP )
∂g
a22 =
∂Y
(
= 0.5e X − 4Y ) ( X OP ,YOP )
= −1.086.
( X OP ,YOP )
The linearization is thus
dX ′
= −5.608 X ′ + 2Y ′ ,
dt
(4.38)
dY ′
= 0.590 X ′ − 1.086Y ′.
dt
The variables X ′ and Y ′ are functions of t that describe approximately how far X(t)
and Y(t) deviate from the steady state (XSS = 0.776, YSS = 0.543).
Suppose the original nonlinear system uses as initial values X0 = Xss + 0.75 and
Y0 = Yss − (−0.2). The linearized system (4.38) may start with any initial values, of
course, but for a direct comparison with the original nonlinear system we choose
0.75 and −0.2. We solve this ODE system and then shift the solution to the operat-
ing point: X (t) = X ′(t) + Xss, Y (t) = Y ′(t) + Yss. Figure 4.6 shows that X and Y
approximate X and Y very well close to the operating point, which equals the steady
state. At the beginning of the simulation, the approximation of X is not quite as
good.
As a second example, recall the SIR model (2.6) that we analyzed in Chapter 2.
It has the form
S = 3 + 0.01R − 0.0005SI , S0 = 170 = Sss + 30 ,

I = 0.0005SI − 0.05I − 0.02 I , I 0 = 120 = I ss − 30, (4.39)
R = 0.05I − 0.01R , R0 = 700 = Rss − 50.
Note that, for convenience, the initial values are expressed here as deviations from
the steady state, which is (Sss, Iss, Rss) = (140, 150, 750), and which we use as our
operating point (OP). Without going through the theoretical derivation, we can
immediately compute the Jacobian, which is
2.0
~
X X′ X
Y Y′ Y~
0.5
Figure 4.6 Linearization of the system

in (4.36). Solution points of the nonlinear
system are shown as dots. X′ and Y′ are the
solutions of the linearized system (4.38),
and X = X′ + Xss, Y = Y′(t) + Yss constitute the
–1.0 linearization of the nonlinear system at the
0 2 4 operating point, which here is the steady
time state.
(A) (B)
50 700
~
DS S S
–75 350 ~
DI I I
~
DR R R
–200 0
0 450 900 0 450 900
time time
Figure 4.7 Linearization of the SIR system in (4.39). (A) The solution to the linearization in (4.41). (B) The same solution, shifted to the operating point
OP. The solution points of the nonlinear system are shown as dots. S, I, and R constitute the linearization of the nonlinear system at the steady-state
operating point. The linearization is excellent, because the nonlinear system does not deviate very far from the steady state. See Exercise 4.21 for a less
accurate approximation.
 ∂ f1 ∂ f1 ∂ f1 
 ∂S ∂I ∂R 
   −0.075 −0.07 0.01 
∂f ∂f 2 ∂f 2 
J = 2 =  0.075 0 0 , (4.40)
 ∂S ∂I ∂R   
 0 0.05 −0.01
 ∂f 3 ∂f 3 ∂f 3 
 
 ∂S ∂I ∂R  OP
where f1, f2, and f3 are the right-hand sides of the system in (4.39), and the derivatives
are evaluated at the stead-state operating point OP. Therefore, the linearized
equations are given as
 dS ′ 
 dt 
   −0.075 −0.07 0.01   S′
 dI ′  =  0.075 0 0   I′ . (4.41)
 dt     
 dR ′   0 0.05 −0.01  R ′
 
 dt 
We solve these equations with initial values that correspond to the deviations of the
nonlinear system from the steady state, namely, (30, −30, −50); see (4.39) and
Figure 4.7. Once computed, the solution is shifted to the operating point OP, which
= S′(t) + Sss, I (t) = I′(t) + Iss, R(t)
corresponds to the steady state: S(t) = R′(t) + Rss.
Figure 4.7 shows the result: the linearization reflects the nonlinear system very well,
because the deviation from the steady state is relatively minor.
conTinuouS nonLinear SySTeMS

As in the context of linear systems, continuous representations account for entire
time intervals and not just for discrete time points. Furthermore, continuous
systems become the more convenient choice if the number of players in the system
becomes large. Who wants to count (or account for) every cell in a culture or organ-
ism? As in the linear case, the switch to continuous variables leads directly to the
formulation of differential equations, but in the nonlinear case these afford us with
incomparably more choices: How do we choose the right-hand side? Should we use
polynomials? Or exponentials? Or combinations of all kinds of functions? Clearly,
there are unlimited possibilities, and nature has not provided us with an instruction
manual. As a consequence, nobody will be able to tell what the best or most
conTinuouS nonLinear SySTeMS 101
appropriate functions are, except for a minute class of very special situations. This
fact poses a significant challenge and is the subject of extensive ongoing research. In
most cases, one of two strategies is used. Either one makes assumptions regarding
functions that appear to be reasonable, for instance, because they somehow model
the mechanisms that drive the phenomenon under investigation. These functions
are sometimes called ad hoc, indicating that they were chosen for this particular
situation and are not claimed to have wider application. Or one uses a generic
nonlinear approximation that does not require specific assumptions and offers suf-
ficient flexibility in the repertoire of responses it can represent. An approximation of
this type is called canonical.
4.5 ad hoc Models

Ad hoc models use functions that for some reason seem to suit the given task the
best. Sometimes they are based on the principles of physics and sometimes they
simply seem to have just the right shapes and other properties. As an example of the
latter, the logistic function
a
F (x ) = (4.42)
b + e − cx
is frequently used, sometimes even with a = b = c = 1, because it describes a

sigmoidal function as it is often encountered in biology (Figure 4.8). Altering a and/
or b changes the output range and horizontal shift of the function, and c affects its
steepness. The reader should try out a few combinations of a, b, and c.
In the field of population dynamics, the same logistic function is more widely
used in the equivalent format of a differential equation of the type
N = αN − βN 2 , (4.43)
where N is the size of a population and α and β are parameters. The first parameter,
α, is the growth rate, and α/β is the final size or carrying capacity of the population
(see Chapter 10).
At first glance, (4.42) and (4.43) don’t seem to have much in common. However,
if we rename F in (4.42) as N and x as t and differentiate N with respect to t, the result
can be manipulated with simple algebra to attain the form (4.43), where the param-
eters α and β and the initial value N0 are combinations of a, b, and c. We should also
note that (4.43), slightly rewritten as N = βN(γ − N), with γ = α/β, looks a lot like the
logistic map, thereby explaining the similarity in name. However, the two are quite
different, because the left-hand sides have a different meaning.
The logistic model in (4.43) is the launch pad for very many models of popula-
tion dynamics, where different species compete for resources or where one species
feeds on another. It is also the base case for canonical Lotka–Volterra systems
(see later), which can describe an incredible variety of nonlinearities.
F(x) 1.0
c=1
0.9
0.8
c = 0.25
0.7
0.6
0.5
0.4
0.3
0.2
0.1 Figure 4.8 Graphs of the logistic function
(4.42). Depending on its parameter values,
–10 –8 –6 –4 –2 0 2 4 6 8 10 the logistic function models different slopes,
x final values, and horizontal shifts.
k1 k2 Figure 4.9 Schematic diagram of an

E + S E S E + P enzyme-catalyzed reaction. Substrate
k–1 S and enzyme E form a complex, which
may revert to S and E or break down to
yield product P while releasing the enzyme.
See Chapter 8 for details.
There is hardly a limit to the possible choices for ad hoc functions, and the
selection of a particular function depends heavily on the context and knowledge of
the biological phenomenon. Possibilities include trigonometric functions, simple
exponentials, or more unusual functions such as spikes or square waves. Indeed,
there are catalogs of functions with various shapes [14].
A widely used example of an ad hoc function based on alleged mechanics is the
Michaelis–Menten function of enzyme kinetics, which has the form
Vmax S
vp = . (4.44)
KM + S
It describes the speed vP with which a metabolic substrate S is converted into a

product P with the help of an enzyme E (Figure 4.9). The function has two
parameters: Vmax is the maximum speed vP can reach, while KM is the substrate con-
centration for which vP runs with half the maximum speed (see Chapters 2 and 8).
4.6 canonical Models

A rather different strategy for model selection is the use of canonical models, of
which there are several types. The designation canonical implies that the processes
of setting up and analyzing a model follow strict rules. Forced adherence to rules
could be seen as a real limitation, but it has significant advantages, as we will see
later in this chapter and in some of the application chapters. One attraction of
canonical models is that they allow us to get started very quickly even if we are faced
with a lot of uncertainty and scarce information on the details of a system. This
advantage becomes most evident if we ask what would happen if we did not have
rules and guidelines. As an example scenario, imagine being tasked to come up with
a mathematical model describing how the introduction of a new species affects the
dynamics of an ecosystem. It is immediately clear that all kinds of players are
involved in the case, but it is very much unclear how exactly they interact. All
interactions must follow the laws of physics and chemistry, but accounting for every
physical and chemical detail in the system would quite obviously be impossible—
and is also often unnecessary. Canonical modeling provides a compromise between
general applicability and simplicity that circumvents some of these problems by
using mathematically rigorous and convenient approximations. As a rule of thumb,
the less one knows about the details of the system, the more one benefits from
canonical models.
The ultimate canonical model is again the linear model, and we would love to
use it. Alas, linearity is often too limiting for biological phenomena that deviate a bit
from their normal steady state and exhibit responses that are very clearly nonlinear.
For instance, a linear model cannot represent stable limit-cycle oscillations, as we
see them in circadian rhythms and our own heartbeat (see later in this chapter).
These limitations mean that we have to look for nonlinear canonical models.
The oldest example is the Lotka–Volterra (LV) model [15–19]. Interestingly, this
model was proposed around the same time as the infectious disease model that we
introduced in Chapter 2, namely, in the 1920s. Its independent authors were Alfred
Lotka, an American pioneer in mathematical biology, and Vito Volterra, an applied
mathematician from Italy. The rules for setting up an LV model are simple. Enumer-
ate all dependent variables of interest; independent variables are merged with
parameters. Assume that each variable may potentially interact with every other
variable in the system and formulate each interaction as the product of the two
variables and a rate constant. In other word, this mechanism is of mass-action type,
as we encountered it in Chapter 2. LV models also permit a linear term in each

equation, which could represent birth, death, degradation, or transport out of the
system. Thus, each differential equation of an LV model has exactly the same
format:
X i = ∑a X X
j =1
ij i j + bi X i , (4.45)
where some (or many) of the parameters may have values of 0. If the system has only
one variable, the LV model with negative a11 and positive b1 becomes the logistic
function (4.43).
The LV model bears good news and bad news. The main good news is that it is
easy and intuitive to set up the equations and to compute the steady state (see later).
Second, the model has found very many applications in ecological systems analy-
ses, where much of the dynamics, like competition for food or predation, is driven
by one-to-one interactions (see [17] and Chapter 10). A further piece of very surpris-
ing and intriguing news is that LV equations are extremely flexible: they are capable
of modeling any type of differentiable nonlinearities and complexities, including
all kinds of damped or stable oscillations and chaos, if one includes enough auxil-
iary variables and does some mathematical trickery with them [18–20]. The bad
news is that, without such trickery, many classes of biological systems are not really
compatible with the LV format. For instance, suppose we want to model a simple
pathway as shown in Figure 4.10. We define four variables X1, …, X4 and want to
formulate the equations. However, this is a problem for the LV format. For instance,
the production of X3 should depend only on X2, but we are not allowed to include a
term like kX2 in the equation for X3. Furthermore, the only choice to account for the
feedback by X4 in the equation for X2 would be a term like κ X1X4, but such a term is
not permitted in the equation. As soon as we try to ameliorate this issue, we increase
the applicability of the model but compromise the mathematical advantages of the
LV format. For instance, if we allow all possible linear and bilinear terms, we arrive
at a mass-action system, as we used for the infectious disease model in Chapter 2.
We could also permit multilinear terms of the type α X1X2X4. In either case, the
extension would compromise the analytical advantages of the LV model. In particu-
lar, the true LV format allows us to compute steady states with methods of linear
algebra (see later in this chapter), whereas the extensions would require the use of
nonlinear methods and search algorithms.
Greater flexibility than in LV and mass-action models is afforded by canonical
generalized mass action (GMA) systems within the modeling framework of
biochemical systems theory (BST) [21–25]. In this type of canonical model, each
process vi is formulated as a product of power-law functions of the type
vi = γ ik X 1fik 1 X 2fik 2 X nfikn . (4.46)
As in mass-action and LV models, this format includes a non-negative rate con-

stant γik, which describes the turnover rate of the process, and a product of variables.
In this case, the product may include all, some, or none of the systems variables.
Each variable has its own exponent fikj, which is called a kinetic order and may take
any real value. If a kinetic order is positive, it signifies a positive or augmenting
effect, and if it is negative, it signifies an inhibitory or diminishing effect. If a kinetic
order is zero, then the variable, raised to zero, becomes 1 and is therefore neutral in
the product. Typical kinetic orders are between −1 and +2. The variables in the term
include dependent and independent variables (see Chapter 2) in exactly the same
format.
– Figure 4.10 Linear pathway with feedback.

X1 X2 X3 X4
It is not directly possible to represent this
pathway as a Lotka–Volterra system.
With this type of formulation, each equation in a full GMA model with n
dependent and m independent variables has the form
Ti n +m
X i = ∑ ∏X
k =1
±γ ik
j =1
fikj
j , (4.47)
where Ti is the number of terms in the ith equation. Here, the large “pi” denotes a
product; for instance,
∏Xj=1
j = X1X 2 X 3 X 4 . (4.48)
The format of (4.47) looks complicated, but is actually quite intuitive once it becomes
more familiar. One simply enumerates all processes (terms) affecting a variable,
decides which variable has a direct effect on a given process, and includes this
variable in the term. The variable receives an exponent, and the whole term is
multiplied with a rate constant.
As a small example with three dependent variables, consider a branched
pathway with one feedback signal, input from the independent variable X0, and acti-
vation by an independent modulator X4, as shown in Figure 4.11. Each equation
contains power-law representations of all fluxes vi that are directly involved. Thus,
the GMA equations are
X 1 = γ 11 X 0f110 X 2f112 − γ 12 X 1f121 − γ 13 X 1f131 ,

X 2 = γ 21 X 1f211 − γ 22 X 2f222 X 4f224 , (4.49)
X 3 = γ 31 X 1f311 − γ 32 X 3f323 .
The indices of the rate constants reflect, first, the equation and, second, the term.
The kinetic orders have a further index, which refers to the variable with which it is
associated. This convention is not always followed, however.
Some of the terms are actually the same, even though they seem different at first
glance. For instance, the flux v2 out of X1 is the same as the flux into X2. This equality
translates into properties of power-law functions, which in the given case can only
be equivalent if γ12 = γ21 and f121 = f211. Similar equalities hold for the two descrip-
tions of the process between X1 and X3.
Very similar to the GMA format is the so-called S-system format [25, 26]. In this
format, all processes entering a variable are merged into a single power-law term,
and similarly for all processes leaving the variable. Biologically, this strategy corre-
sponds to a focus on pools (dependent variables) in S-systems rather than on fluxes
in GMA systems. Owing to the merging of influxes and effluxes in one net flux each,
S-systems have at most one positive and one negative term in each equation, which
is very significant for further analyses (see later). Therefore, their general form, with
its own names for parameters, is always
n +m n +m
X i = α i ∏ j =1
g
X j ij − βi ∏X
j =1
hij
j , i = 1, 2, ..., n. (4.50)
Figure 4.11 Generic pathway with

X4 one feedback signal and one external
activator. The dependent variable X1 is
produced from a precursor X0, which is
+ modeled as an independent variable, and
X2 is used as substrate for the production of
v2 v4
X2 and X3. X2 exerts feedback inhibition on
– the production of X1, and X4 activates the
X0 X1 degradation of X2. Each process is modeled
v1
by a function vi, which in a GMA model (4.49)
v3 v5 has the form of a product of power-law
X3
functions (4.46).
For the example in Figure 4.11, the production of X1 is exactly the same in GMA
and S-systems. Also, the production and degradation processes of X2 and X3 remain
unchanged. The only difference between the GMA and S-system models occurs for
the efflux from X1: in the GMA system, this efflux consists of two branches, while it
consists of one net efflux in the S-system. Thus, the example in Figure 4.11 is
described by the S-system model
X 1 = α 1 X 0g10 X 2g12 − β1 X 1h11 ,

X 2 = α 2 X 1g 21 − β 2 X 2h22 X 4h24 , (4.51)
X 3 = α 3 X g 31
1 − β3 X h33
3 .
Note the similarities and differences between this formulation and the GMA form in
(4.49). If the S-system and the GMA system are identical at some operating point,
such as a steady state, then the equation β1 X 1h11 = α 2 X 1g 21 + α 3 X 1g 31 holds at this point.
Generally, the two types of models differ only at converging or diverging branch
points and are otherwise identical.
As a more interesting example, which has the same equations in both the GMA
and S-system formats, consider a toggle switch model for the SOS pathway in
Escherichia coli, with which the bacterium responds to stress situations where DNA
is damaged by sending out SOS signals [27]; it is shown diagrammatically in
Figure 4.12A. The main components are two genes, lacI and λcI, which are
transcribed into the regulator proteins LacR and λCI, respectively. The most inter-
esting feature of the system is cross inhibition: the expression of lacI is high when
λcI has a low expression level, and vice versa. Specifically, the promoter PL of the lacI
gene is repressed by λCI and the promoter Ptrc of the λcI gene is repressed by LacR,
(A)
DNA damage
+
RecA
+
λ CI
PL
λ cl lacI
Ptrc
–
LacR
(B) X6
+
X3
Figure 4.12 Cross inhibition in the SOS

pathway of E. coli. (A) Diagrammatic
+ representation of the SOS system. (B) Model
X2 representation, where the dependent
– variables are X1 = LacR, X2 = λCI, X3 = RecA,
and the independent variables are the
promoters X4 = PL and X5 = Ptrc, as well as
+ + X6 = DNA damage. In the simplified model,
X5 X4
X1 and X2 inhibit each other’s production
directly, rather than through the promoters
(see the text for further discussion). ((A) from
– Tian T & Burrage K. Proc. Natl Acad. Sci. USA
X1 103 [2006] 8372–8377. With permission from
National Academy of Sciences, USA.)
subsequently leading to elevated expression of the gene lacI. DNA damage leads to
activation of the protein RecA, which increasingly cleaves the λCI repressor protein.
Reduced amounts of λCI relieve the inhibition of PL and thus cause higher expres-
sion of lacI.
As discussed in Chapter 2, we identify the main players and categorize them as
dependent and independent variables (see Figure 4.12B). Clear dependent vari-
ables, whose dynamics is driven by the system, are the two regulator proteins
X1 = LacR and X2 = λCI, as well as X3 = RecA. The genes, while obviously important
biologically, are not modeled, because they serve in the model only as precursors of
the proteins. A more difficult debate is whether the promoters should be dependent
or independent variables. This is one of those decisions that need to be made during
the model design phase that we discussed in Chapter 2. On one hand, the promoters
are involved in the dynamics of the system. On the other hand, we have no informa-
tion on how they become available or how they are removed from the system.
Weighing our options in favor of simplicity, we make the decision to consider the
promoters as independent variables X4 = PL and X5 = Ptrc with constant values. They
are included in the production terms of X1 and X2. Because the promoters do not
respond to changes in the system, we let X2 directly inhibit the production of X1 and
we let X1 directly inhibit the production of X2. In addition to the promoters, X6 = DNA
is modeled as an independent variable.
To set up the equation for X1, we identify all production and degradation pro-
cesses, as well as the variables that directly affect each. Indirect effects are not
included; they are realized through the dynamics of the system. The production is
affected by its gene’s promoter PL and the inhibition by λCI. Its natural degradation,
for instance, through proteases, only depends on its concentration. Thus, the first
equation is
X 1 = α 1 X 2g12 X 4g14 − β1 X 1h11 . (4.52)
We do not know the parameter values, but we know already that all are positive,
except for g12, which represents the inhibition by λCI.
The second equation is constructed in a similar fashion. A slight difference here
is that its degradation is hastened by X3:
X 2 = α 2 X 1g 21 X 5g 25 − β 2 X 2h22 X 3h23 . (4.53)
The third equation describes the dynamics of RecA. Let us begin with its
degradation, which depends on its concentration; thus, we formulate the power-law
term β 3 X 3h33 . For the production of RecA, we assume as a default that it is constant.
However, DNA damage activates this production process, so that X6 should be
included with a kinetic order. Thus, the equation reads
X 3 = α 3 X 6g 36 − β 3 X 3h33 . (4.54)
The formulation of the symbolic model (without numerical values) is now complete.
Imagine how difficult this step would have been without canonical modeling, that
is, if we had been forced to select nonlinear functions for all processes.
For further analysis and for simulations, we need to assign numerical values to
all parameters, based on experimental data. This parameterization step is generally
rather difficult (see Chapter 5), and we shall skip it for this illustration. Instead, we
choose values for all parameters that are typical in BST (see Chapter 5 of [23]) and
explore the functionality of the model with these values. The chosen parameter val-
ues are α1 = 10, α2 = 20, α3 = 3, β1 = β2 = β3 = 1, g12 = −0.4, g14 = 0.2, h11 = 0.5,
g21 = −0.4, g25 = 0.2, h22 = 0.5, h23 = 0.2, g36 = 0.4, h33 = 0.5, X1(0) = 7, X2(0) = 88,
X3(0) = 9, X4 = 10, X5 = 10, and X6 = 1 (no DNA damage) or X6 = 25 (DNA damage).
We begin the simulation under normal conditions. One hour into the simulation
experiment, we induce DNA damage by increasing X6. The system responds to the
damage immediately with a fast increase in X1 and a rapid drop in X2 (Figure 4.13).
We furthermore take into account that DNA damage is repaired after 10 hours.
In response, λCI and LacR expression return rapidly to their normal states. Thus, as
100 Figure 4.13 Simulation of switches

between λCI and LacR expression
X2 states. The simulation begins with normal
conditions, where the expression levels
of λCI and LacR are 88 and 7, respectively.
After 60 minutes, DNA is damaged. As
a consequence, X3 increases rapidly
50 (not shown) and cleaves λCI, whose
concentration drops in response. The
inhibition of PL is greatly reduced, and LacR
rises. After 10 hours, DNA damage in the
simulation is repaired, and λCI and LacR
X1 expression return to their normal states.
Parameter values are given in the text.
0
0 600 1200
time
these representative results demonstrate, the model generates high expression of

λCI under normal conditions and high expression of LacR under conditions of DNA
damage. Moreover, the model exhibits sharp switches from one steady state to
another, as has been observed in experiments. While the term “switch” has no firm
definition, it is used to distinguish a step-like change from a more gradual
transition.
The format for processes in GMA and S-system models was not developed arbi-
trarily. It is the result of a general approximation strategy that is very convenient in
biological systems analysis and at the same time is based on solid mathematical prin-
ciples, namely Taylor’s theorem (see Boxes 4.1 and 4.2). The approximation is very
similar to the linearization we discussed earlier in the chapter (Boxes 4.3 and 4.4).
However, the result is nonlinear and usually offers a wider range within which the
approximation is good. Moreover, the power-law format of these models is so flexible
that it admits models of essentially arbitrary complexity [20].
A more recent alternative for a canonical representation of metabolic processes
is the lin–log model [28]. The name stems from the fact that the model is linear in
logarithmically transformed variables. Specifically, all metabolites and modulators
Xj and all enzymes ei are expressed in relation to their normal reference values
X 0j and ei0 , respectively. Furthermore, the rate is expressed in relation to the refer-
ence steady-state flux ji0 through the pathway. Thus, the lin–log model has the form
vi ei   Xj 
n +m
0
= 0
J i ei 
1 + ∑ ε ij0 ln  0   ,
 Xj 
(4.55)
 j =1 
where ε ij0 is the reference elasticity, which plays the same role as kinetic orders
in BST.
The lin–log formulation has convenient features with respect to the steady state
of the system and models situations with high substrate concentrations better than
power-law models. Alas, nothing comes for free: power-law models are much more
accurate than lin–log models for small substrate concentrations [29–31]. The lin–log
model has been developed out of the framework of metabolic control analysis
(MCA) (see [32–34] and Chapter 3). The collective experience with lin–log systems
is at this point dwarfed by the experience gained for LV systems and models
within BST.
Another recent idea is the use of S-shaped modules, rather than power-law
functions, for the description of the role of variables within a system [35]. This
formulation permits greater flexibility but also requires more parameter values.
An evident question is why we need so many different model structures. The
answer is that each model structure has different advantages and disadvantages,
for instance, with respect to the range within which the approximation is superb,
good, or at least acceptable. The models also differ in the ease with which standard
analyses are executable. For instance, LV, lin–log, and S-systems permit the explicit
computation of the steady state, which is not the case for GMA systems. While the
BOX 4.3 POWER-LAW APPROXIMATION
The power-law approximation of a nonlinear function F results this expression at some operating point of our choosing:
in a representation that consists of a product of power-law
functions (4.46). This approximation is obtained in the following dF x ace − cx b + e − cx
fashion: f= = − cx 2
x . (3)
dx F OP (b + e ) a OP
1. Take the logarithm of all variables and also of the approximat-

Algebraic simplification yields
ed function F; it is assumed that they are all positive.
2. In the second step, compute the linearization of the result
(see Box 4.1). cxe − cx cx
f= = . (4)
3. Take the exponential of the linearization result. b + e − cx OP
1+ be cx OP
The result is a function vi of the form (4.46). This procedure may Finally, we pick an operating point, such as x = 5, and substitute
sound complicated, but there is a mathematically valid and correct numerical values for the parameters. The result is the desired
shortcut, which says: kinetic order f:
1. Each kinetic order (exponent) is equal to the partial derivative 0.3 ⋅ 5

of the approximated function F with respect to the correspond- f= = 1.036. (5)
1+ 0.1e 0.3⋅5
ing Xj, multiplied by Xj, and divided by F.
2. The rate constant is subsequently computed by equating the The rate constant is computed in a second step. The function
power-law term and the approximated function F at the oper- and the approximation must have the exact same value at the
ating point. operating point (x = 5), and this value is F(5) = 0.619, according to
(4.32) with the parameters chosen above. Thus, we equate F and
Let us illustrate the power-law approximation procedure γ X f for x = 5 and solve for γ:
with an example, where the approximated function F has one
argument (see Box 4.4 for multivariate cases). As an example, we F( x ) = γ x f , (6)
choose again the logistic function (4.42), OP
and so
a
F( x ) = , (1)
b + e − cx
γ = F( x )x −f = 0.619 ⋅ 5−1.036 = 0.1168. (7)
OP
and assume that its parameters are a = 0.2, b = 0.1, and c = 0.3. We
also suppose that x is some physical quantity and that its relevant If we choose a different operating point, the approximation is
range is therefore restricted to positive values. The task is to find different. For instance, for x = 15, we obtain
parameters γ and f (we omit indices for simplicity) so that γ X f is a
power-law approximation of F(x). 0.3 ⋅15
f= ⋅
= 0.4499, (8)
Using the shortcut, we have to compute the derivative of 1+ 0.1e 0.315
F(x), and, because there is only one dependent variable, this is an
ordinary derivative: and so
γ = 1.8 ⋅15−0.4499 = 0.5323. (9)

dF ace − cx
= . (2)
dx (b + e − cx )2 Figure 1 shows F(x) and the power-law approximations at the two
chosen operating points. The ranges of the two approximations
Now we multiply this derivative by x and divide by F, and evaluate where the approximation is acceptable are quite different.
3.0
0.1168x1.036
2.5
2.0
F(x), γ xf
1.5 OP2
0.5323x0.4499
1.0
Figure 1 Function F(x) (blue) in (1) with parameters a = 0.2, b = 0.1
and c = 0.3 and two power-law approximations at x = 5 (red)
0.5 and x = 15 (green). At the operating points, F(x) and the respective
OP1
approximations have exactly the same value and slope. Depending
0 on the purpose of the model and the required accuracy, the first
0 5 10 15 20 approximation might be sufficiently accurate within the interval
x 3 ≤ x ≤ 14 and the second within the interval 12 ≤ x ≤ 20.
BOX 4.4 POWER-LAW APPROXIMATIONS OF MULTIVARIATE FUNCTIONS
Power-law approximations of functions with more than one and

dependent variable are constructed in an analogous fashion,
except that we now use partial derivatives. We demonstrate the ∂v P E
g2 =
procedure for functions with two arguments; it is then easy to see ∂E v P OP
how functions with more arguments are to be approximated. kS(K M + S ) E
As an illustration, suppose we are interested in a Michaelis– = (4)
(K M + S )2 v P
Menten process (4.44) that includes an enzyme whose activity OP
changes as a consequence of altered gene expression and protein kS(K M + S ) E (K M + S )

= = 1.
degradation. The first task is to make the enzyme activity E explicit (K M + S )2 kES OP
in the Michaelis–Menten function. It is actually embedded in the
parameter Vmax, which equals the product of a turnover rate of The two results are different in several ways. Most importantly,
the process and the total enzyme concentration, kE, so that the g1 depends on the operating point, and its value changes
function with explicit E reads depending on where we want to anchor the approximation. For
substrate concentrations close to 0, g1 is close to 1; for S = KM,
kES g1 = 0.5; and for very large substrate concentrations, g1 approaches
v p (S , E ) = . (1)
KM + S 0. Thus, the kinetic order g1 in this application is always between
0 and 1. By contrast, g2 does not depend on the operating point,
The corresponding power-law representation will have the form and its value is always 1. The reason for this particular situation is
that E contributes to the Michaelis–Menten process (1) as a linear
V = αS g1E g2 , (2) function, which corresponds to a power-law function with kinetic
order 1. Thus, with respect to E, the two functions V and vP are the
and our task is to compute its parameter values α, g1, and g2 such same for all operating values.
that V and vP have the same value at some operating point of The rate constant is again computed following the rationale
our choice and such that the slopes in the directions of S and E that if V is to be a proper approximation of vP, then vP and V must
are also the same. Only then is V a proper approximation of vP. have exactly the same value at the chosen operating point. Thus,
According to the shortcut for kinetic orders, which is the same as equating the two at a chosen operating point, for instance, S = KM,
in Box 4.3, we: leads to
1. Compute partial derivatives. v p ( S , E ) = V ( S , E ) = α S g1E g2 ( for S = K M and some value of E ), (5)
2. Multiply them by the corresponding variable.
3. Divide by vP at the operating point.
and so
These algebraic operations yield the following:
α = v P S − g1E − g2
∂v P S kEK M −0.5 −1
g1 = = KM E
2K M (6)
∂S v P OP
k
kE (K M + S ) − kES S = K M−0.5 .
= 2
(K M + S )2 vP OP
(3)
kEK M S(K M + S ) The original rate law and its power-law approximation are shown
= in Figure 1 for parameter values k = 1.5, KM = 10, and E = 2. With
(K M + S )2 kES OP
respect to E, the power-law representation is exact. In other words,
KM if the substrate concentration S is set at some operating point, the
= ,
(K M + S ) OP original and the power law are identical for all values of E.
2.5 V
2.0 vP
1.5
1.0
0.5
KM Figure 1 Power-law approximation of a Michaelis–Menten rate
0 law (1) where E is also a dependent variable. At the operating point
0 5 10 15 20 25 S = KM, E = 2, the original function and the corresponding power-law
S approximation have the same value and slope.
mathematical structures differ, experience has shown that the choice of a particular
canonical model is often (but not always) less influential than the connectivity and
regulatory structure of the model. If this structure correctly captures the essence of
the biological system, several of the alternative models may perform equally well.
However, there is no guarantee, and the modeler is forced to decide on a model
structure, possibly based on specific biological insights as well as on matters of
mathematical and computational convenience, or to perform comparative analyses
with several models (see, for example, [36]).
4.7 More complicated Dynamical Systems Descriptions

Most biological systems are too complicated to be describable by explicit functions,
and even ordinary differential equation (ODE) models are not always sufficient.
One important limitation of ODEs is their failure to represent spatial components,
in addition to changes over time. The mathematician’s first answer is often the use
of partial differential equations (PDEs), which capture not just dynamic trends
but also allow us to analyze variables with respect to changes in their location (see,
for example, [6, 37]). For instance, it is rather evident that we need spatial features if
we want to describe the formation of plaques in arteries, which critically depend on
the details of blood flow and are especially often found at branches in arteries where
backflow can form complex eddies. PDEs are very well suited for capturing the spa-
tial–temporal dynamics of such a system. Unfortunately, their analysis requires
significant investment in math, and we do not pursue them further in this text.
An alternative is agent-based modeling (ABM) [38, 39], which we will describe in
Chapter 15. A second issue with ODEs is that they do not directly allow for delays in
the dynamics of a system. This situation is simpler, because mathematical tricks can
be used to permit the modeling of delays [40].
Finally, ODEs are entirely deterministic and do not permit any account of
stochasticity. Accounting for stochastic effects can be crucial, as beautiful experi-
ments have shown that exact copies of the same cell may develop very differently
over time and that the differences are caused by stochastic variability at the molec-
ular level (see, for example, [41]). Alas, capturing molecular processes in such
detail requires much more complicated approaches, such as stochastic differen-
tial equations or hybrid systems (see, for example, [42–44]), and it is again a ques-
tion of judgment and of the modeling purposes as to whether the effort is
worthwhile.
STanDarD anaLySeS oF BioLoGicaL SySTeMS

MoDeLS
The preceding sections have demonstrated that the choice of a good model for a
biological system is neither trivial nor unique. We have encountered alternative
models even for simple growth processes, and the choices obviously increase for
more complicated systems. Although the various representations differ quite a bit,
there are general strategies of analysis. Indeed, much of this analysis is to some
degree straightforward and follows guidelines that are more or less independent of
the particular system under investigation. Thus, it is time to discuss the most
pertinent standard methods for analyzing biological systems. They fall into two
broad classes. One pertains to static features and is based on analyses of the system
within the neighborhood of a steady state, while the other class contains dynamic
features, such as the exploration of possible model responses to perturbations or the
characterization of oscillations.
4.8 Steady-State analysis

A steady state is a condition of a system in which none of the variables change in
number, amount, or concentration. This lack of change does not imply that nothing
is happening in the system. It simply means that all influxes and effluxes at all pools
and variables are in balance. As an example, consider the concentrations of
STanDarD anaLySeS oF BioLoGicaL SySTeMS MoDeLS 111
chemical components in the blood. There is ongoing metabolism, and blood cells

take up and deliver oxygen. Nevertheless, many of the components of the serum
remain within very close ranges about their normal states. These states are of special
interest in the analysis of most biological systems.
Steady-state analyses related to dynamical models address three basic questions.
First, does the system have one, many, or no steady states, and can we compute
them? Second, is the system stable at a given steady state? And third, how sensitive
to perturbations is the system at a given steady state? In addition to these more
diagnostic topics, one might be interested in controlling, manipulating, or optimiz-
ing the steady state of the system. Interestingly, most steady-state analyses are exe-
cuted with the use of linear mathematics, motivated and supported by the insight of
Hartman and Grobman that many behaviors of a nonlinear system can be studied
with the corresponding linear system [2]. In fact, now that we have gained some
experience with approximations, we can say more exactly what that means. It refers
to the (univariate or multivariate) linearization of the nonlinear system, usually with
the steady state chosen as the operating point.
In a linear system, the steady state is relatively easy to assess, because the
derivatives are zero, by definition of a steady state (that is, no overall change in any
of the variables), so that the system of differential equations becomes a system of
linear algebraic equations. Recalling (4.27), but allowing for an additional constant
input u, and restricting the analysis to the steady state yields
dX
= X = AX + u = 0. (4.56)
dt
Linear algebra tells us that there are three possibilities. The system may have no
steady-state solution, or exactly one solution, or it may happen that whole lines,
planes, or higher-dimensional analogs of planes satisfy (4.56). As a simple two-
variable example, let us look at the following five systems S1–S5, which are quite
similar:
X 1 = 2 X 2 − 2 X 1 ,
S1:
X 2 = X 1 − 2 X 2 .
In this example, there is no input, and both variables are decreasing toward the
“trivial” steady-state solution X1 = X2 = 0. We can confirm this easily and directly,
either by solving the system numerically or by setting the derivatives in both equa-
tions equal to zero and adding them together (which we are allowed to do according
to linear algebra). The result of the latter is X1 − 2X1 = 0, which is only possible for
X1 = 0. Plugging this value into one of the equations yields X2 = 0. Thus, the solution
(X1 = X2 = 0) is unique and not particularly interesting.
X 1 = 2 X 2 − 0.5 X 1 ,
S2:
X 2 = X 1 − 2 X 2 .
This system is very similar, except that the coefficient of X1 in the first equation is
−0.5 instead of −2. The solution is again exponential, but this time both variables
grow without ceasing. The system again satisfies that trivial solution X1 = X2 = 0, but,
in this case, the dynamical solution does not approach this steady-state solution.
In fact, if we start a simulation anywhere but at (0, 0), the system leads to unbounded
growth.
X 1 = 2 X 2 − X 1 ,
S3:
X 2 = X 1 − 2 X 2 .
In this case, the coefficient of X1 in the first equation is −1. Setting the derivatives in
both equations equal to zero and multiplying one of them by −1 shows immediately
that they are equal. Both require that X1 be twice as big as X2, but there are no other
conditions. A bit of analysis (or computer simulation) confirms that any
(A) X1 (B)
2
X1
concentration
X2
1
X2
(C) X1 (D)
2
concentration
X2
1
X1
X2
0
0 2 4 0 2 4
time time
Figure 4.14 Four solutions of the system S3, which differ solely in the initial conditions. (A) X1 = 2, X2 = 1; (B) X1 = 1, X2 = 1; (C) X1 = 1, X2 = 2;
(D) X1 = 0.5, X2 = 0.5.
combination consisting of an arbitrary value of X1 and half that value for X2 satisfies
the steady state. The system has infinitely many solutions, including (0, 0)
(Figure 4.14); all these solutions lie on a straight line, defined by X1 = 2X2.
X 1 = 2 X 2 − 2 X 1 + 1,
S4:
X 2 = X 1 − 2 X 2 .
The only difference between S4 and S1 is the constant input with value 1. This
change drastically changes the dynamic and the steady-state solutions. The latter is
no longer trivial (X1 = X2 = 0), but has the unique values X1 = 1, X2 = 0.5. In contrast
to the decreasing processes toward 0 in S1, an input is feeding the system, and the
system quickly finds a balance at (1, 0.5).
X 1 = 2 X 2 − X 1 + 1,
S5:
X 2 = X 1 − 2 X 2 .
This system also has an input of 1, and the only difference with S4 is the coefficient
of X1 in the first equation. Setting the derivatives in both equations equal to zero and
adding them yields the requirement 0 = 1, which we know cannot be satisfied. The
system has no steady state—not even a trivial one.
The computation of steady states in nonlinear systems is in general much harder.
In fact, even if we know that the system has a steady-state solution, we may not be
able to compute it with simple analytical means and need to approximate the solu-
tion with some iterative algorithm. We already saw a harmless-looking example in
Chapter 2, namely the equation e x − 4x = 0. We could easily envision that this equa-
tion described the steady state of the differential equation dx/dt = e x − 4x. If we
cannot even solve one nonlinear equation, imagine how little is possible in a system
with dozens of nonlinear equations! You might ask why we don’t use linearization
and compute the steady state of the linearized system. The reason this seemingly
clever idea won’t work is that we need an operating point for the linearization. But if
we don’t know what the steady state is, we cannot use it as operating point, and
choosing a different point is not helpful.
We may obtain the steady state(s) of a nonlinear system with two numerical
strategies. First, we can do a simulation and check where the solution stabilizes. This
strategy is very simple and often useful. It has two drawbacks, however. First, if the
steady state is unstable, then the simulation will avoid it and diverge away from it
(see the system S2 above and later). Second, in contrast to linear systems, nonlinear
systems may possess (many) different, isolated steady states. Isolated means that
points in between are not steady states. As an example, consider the single nonlin-
ear differential equation
x = −0.1( x − 1)( x − 2)( x − 4). (4.57)
We can convince ourselves easily that 1, 2, and 4 are steady-state solutions, because
for each of these values for x the right-hand side becomes zero. However, values in
between, such as x = 3, are not steady-state solutions. The simulation strategy for
computing the steady states would yield x = 4 for all initial values above 2 (see the
red lines in Figure 4.15). For all initial values below 2, the simulation would yield
x = 1 as the steady state, and, unless we started exactly at 2 (or used insight and
imagination), the (unstable) steady state x = 2 would slip through the cracks. In the
case presented here, the failure would be easy to spot, but in realistic models of
larger size that is not always so, and numerical solution strategies can easily miss a
steady state.
The second strategy for finding steady-state solutions uses search algorithms,
of which there are many different types and software implementations. We will
discuss some of them in the context of parameter estimation (see Chapter 5).
The generic idea of many of these algorithms is that we start with a wild guess (the
literature calls it an educated guess or a guesstimate). In the example of (4.57),
shown in Figure 4.15, we might begin with the (wrong) guess that 3.4 could be a
steady state. The algorithm computes how good this guess is by evaluating all steady-
state equations of the system and computing how different they are from zero; in
our case, there is only one equation, and the residual error, that is, the difference
from the best possible solution, is −0.1(3.4 − 1)(3.4 − 2)(3.4 − 4) = 0.2016. Next, the
x 3
Figure 4.15 Some dynamical (red, blue)

and steady-state (green) solutions of
0 (4.57). The steady state at 2 is never reached
0 5 10 unless the simulation starts exactly at 2,
time because it is unstable.
algorithm checks solutions close to the initial guess (for instance, 3.3 and 3.5) and
determines the direction where the solution has the highest chance of improving. In
our case, x = 3.3 yields an increased residual error of 0.2093, while x = 3.5 corre-
sponds to a lower (= better) residual error of 0.1875. The algorithm now assumes
that, because the error at x = 3.5 is lower, the best chances of improving the solution
lie with even higher values of x. Thus, the algorithm updates the current guess with
a higher number, such as 3.6. The technical details of this type of gradient search
(such as the exact direction and distance for the next move) can be very compli-
cated, especially for systems with many equations. More details are presented in
Chapter 5.
While almost no nonlinear models permit an easy computation of their steady
states, some canonical forms are shining exceptions. The first is the LV system that
we discussed before (see (4.45)). Setting the derivatives equal to zero yields
∑a X X
j =1
ij i j + bi X i = 0. (4.58)
It is not hard to see that Xi = 0 is a solution. Because all n equations have to be satis-
fied for a steady state, Xi = 0 must be true for all variables. We are again encountering
a trivial solution. This solution is not very interesting in most cases, but it is never-
theless useful: employing an old math trick, we record—but then intentionally
exclude—this trivial solution, which allows us to divide each equation by Xi, an
operation that was not permitted earlier owing to the possibility of Xi = 0.
For simplicity, let us assume that all variables should be different from 0 at
the steady state. Then we divide each equation by the corresponding Xi, and the
nontrivial steady-state equations of the LV system become
∑a X
j =1
ij j = −bi . (4.59)
This is a system of linear algebraic equations, and we can apply all the tools of linear
algebra. Of course, it is possible that some variables are 0 and others are not. Some
people, but not all, call these solutions trivial as well.
The fact that we can use linear algebra for a nonlinear system is indeed a rare
event, but LV systems are not the only cases. Something similar occurs with
S-systems (4.50) [45]. Suppose for simplicity that the system does not contain
independent variables and set the derivatives equal to zero, which yields
n n
αi ∏ j =1
g
X j ij − βi ∏Xj =1
hij
j = 0, i = 1, 2,..., n. (4.60)
This result does not look linear at all. However, once we move the β-term to the
right-hand side and insist that all variables are positive, we can take logarithms on
both sides:
 n   n 
ln α i + ln 
 ∏ j =1
g
X j ij  = ln βi + ln 
  ∏
j =1
h
X j ij  .
 (4.61)
Now remember that
 n  n
ln 
 ∏ j =1
Zj =
 ∑ ln Z
j =1
j
is true for any positive quantities Zj and that ln X p = ln e p ln X = p ln X is true for posi-
tive X and any real parameter p. Thus, we obtain
 n  n
ln 
 ∏j =1
g
X j ij  =
 ∑ g ln X
j =1
ij j (4.62)
and the corresponding equations for terms with hij. When we rename the variables
yi = ln Xi and rearrange the equation, we obtain the result
n n
∑j =1
g ij y j − ∑h y
j =1
ij j = ln βi − lnα i . (4.63)
Finally, we rename aij = gij – hij for all i and j and define bi = ln(βi/αi), which reveals
the intriguing truth that the steady-state equations of an S-system, which look very
nonlinear, are indeed linear:
a11 y1 + a12 y2 + a13 y3 + + a1n yn = b1 ,

a21 y1 + a22 y2 + a23 y3 + + a2n yn = b2 ,
a31 y1 + a32 y2 + a33 y3 + + a3n yn = b3 , (4.64)

an1 y1 + an 2 y2 + an 3 y3 + + ann yn = bn .
For later biological interpretations, we just must remember to translate back from
y’s to X’s with an exponential transformation, once the analysis is completed. The
same procedure works for systems with independent variables. The reader should
confirm this or see [23].
As a final exceptional case of a nonlinear system with linear steady-state
equations, consider the lin–log model (4.55) [28]. Setting the rates on the left-hand
side of (4.55) equal to zero yields
ei   Xj 
n +m
0

ei 
1+ ∑
ε ij0 ln  0   = 0,
 Xj 
(4.65)
 j =1 
and, justifiably assuming that the reference enzyme activities ei0 are not zero,
we obtain
n +m
 Xj 
∑ ε ln  X
j =1
0
ij 0
j
= −1. (4.66)
If we rename the variables as y j = ln( X j /X 0j ), we see that the steady-state equations

in y-variables are linear.
With the exception of these systems, very few classes of nonlinear models
have steady-state solutions that can be computed explicitly with methods of cal-
culus and linear algebra. Even mass-action and GMA systems do not allow us to
do this.
Once we have computed a steady state, either analytically or numerically,
we can use it as an operating point for linearization and analyze important features
such as stability and parameter sensitivities, as we will discuss next.
4.9 Stability analysis

Stability analysis assesses the degree to which a system can tolerate external
perturbations. In the simplest case of local stability analysis, we can ask whether the
system will return to a steady state after a small perturbation. An analogy often used
is a ball in a half-spherical bowl (Figure 4.16A). Left to its own devices, the ball will
roll to the bottom of the bowl, maybe roll up and down a few times, and finally stay
at the bottom. If we move the ball away from the bottom (that is, if we perturb it), it
(A) (B) (C)
stable marginally stable unstable
Figure 4.16 Illustration of different degrees of stability. (A) A ball placed in a bowl rolls to the bottom and returns to this position if it is slightly moved.
(B) A ball on a perfectly flat surface rolls to a new position when it is tapped. (C) A ball placed perfectly on top of an inverted bowl might stay there for a
while, but the slightest perturbation will cause it to roll to one of the sides.
will roll back to the stable steady state at the bottom of the bowl. Now suppose the
bowl is turned upside down (Figure 4.16C). With a very steady hand, we might be
able to place the ball on the top of the inverted bowl, but even the slightest perturba-
tion will cause it to roll down. The top of the bowl is a steady state, but it is unstable.
As an intermediate case, imagine a ball on a perfectly flat, horizontal surface
(Figure 4.16B). Now any perturbation will cause the ball to roll and, owing to fric-
tion, come to rest at a new location. All points of the flat surface are steady-state
points, but all are only marginally stable. Note that these considerations are true for
relatively small perturbations. If we were to really whack the ball in the example in
Figure 4.16A, it might fly out of the bowl and land on the floor. The same is true for
local stability analysis: it only addresses small perturbations. Before we discuss
technical details of stability, we should note that the terminology “the system is sta-
ble” is unfortunate, if not wrong. The reason is that a system often has several steady
states. Thus, the correct terminology is “the system is stable at the steady state …” or
“the steady state … of the system is stable.”
Whether the steady state of a dynamical system is stable or not is not a trivial
question. However, it can be answered computationally in two ways. First, we may
actually start a simulation with the system at the steady state and perturb it slightly.
If it returns to the same steady state, this state is locally stable. Second, we can study
the system matrix. In the case of a linear system, the matrix is given directly, while
nonlinear systems must first be linearized, so that the matrix of interest is the
Jacobian (4.34), which contains partial derivatives of the system. In either case,
the decision on stability rests with the eigenvalues of the matrix, and while these
(real- or complex-valued) eigenvalues are complicated features of matrices, the rule
for stability is simple. If any of the eigenvalues has a positive real part, even if it is
just one out of three dozen or so eigenvalues, the system is locally unstable. For
stability, all real parts have to be negative. Cases with real parts equal to zero are
complicated and require additional analysis or simulation studies. Fortunately,
software packages such as Mathematica®, MATLAB®, and PLAS [1] compute
eigenvalues automatically.
In the case of a two-variable linear system without input (4.27), the stability anal-
ysis is particularly instructive, because one can visually distinguish all different
behaviors of the system close to its steady state (0, 0). Specifically, the stability of (0,
0) is determined by three features of the matrix A, namely, its trace tr A = A11 + A22,
determinant det A = A11A22 − A12A21, and discriminant d(A) = (tr A)2 − 4 det A, which
are related to the two eigenvalues of A. Depending on the combination of these
three features of A, the behaviors of the system close to (0, 0) can be quite varied.
They may look like sinks, sources, attracting or repelling stars, inward or outward
spirals, saddles, or collections of lines, as sketched in Figure 4.17. To reveal these
images with a simulation, one solves the same system (4.27) with the same matrix A
from many different initial values and superimposes all plots (Figure 4.18). Similar
behaviors occur for higher-dimensional systems, where they are, however, more
difficult to visualize.
The top half of the plot in Figure 4.17 contains on the left the stable systems,
where both eigenvalues have negative real parts, and on the right the correspond-
ing unstable systems, where the eigenvalues have positive real parts. Special situ-
ations occur on the axes and on the parabola given by the discriminant d(A). For
example, on crossing the vertical axis, the behaviors are transitions from stability
d(A) det A d(A)
degenerate degenerate
stable unstable
node node
center
stable unstable
star star
stable unstable
node node
stable unstable (source)
(sink)
spiral spiral
tr A
stable lines unstable lines
saddle points
Figure 4.17 Stability patterns in two-dimensional linear differential equation systems without input—see (4.27). The elements of the system
matrix A determine its trace, tr A, determinant, det A, and discriminant, d(A). Different combinations of these features, in turn, lead to different system
behaviors close to the steady state (0, 0).
to instability. In the particular case of a center, which is the transition case between
stable and unstable spirals, all solutions collectively form an infinite set of con-
centric ellipses that cover the entire plane. The interesting case of a saddle point
lies in the center of the bottom half of the plot: coming exactly from one of two
opposite directions, the steady state appears to be stable; coming exactly from a 10
second set of two opposite directions, the steady state appears to be unstable;
coming from any other directions, the trajectories of the system seem to approach,
but then curve away from, the origin. A unique situation occurs for tr A = det
A = d(A) = 0. Here, the two eigenvalues are the same (so there is really only one
eigenvalue), with a value of zero. It could be in this case that A contains four zero
coefficients, so that both derivatives equal zero, and wherever the system starts, it X2 0
stays. It could also be that A is not filled with zeros, and the system contains one
line of steady states and all other trajectories are unstable. Further details can be
found in [46].
The eigenvalue analysis just described is very important, but it is also limited.
First, in the case of nonlinear systems, the deduction of stability is only guaranteed –10
to be correct for infinitesimally small perturbations. In reality, stable systems –10 0 10
tolerate larger perturbations, but there is no mathematical guarantee. Second, defi- X1
nite conclusions are difficult if one or more eigenvalues have real parts of 0. There
are other types of stability that cannot be addressed with eigenvalue analysis alone. Figure 4.18 Visualization of a stable
For instance, if one changes the parameter values in a system, it may happen that a spiral. The coefficients of A are
(A11, A12, A21, A22) = (−3, −6, 2, 1) and initial
stable steady state becomes unstable and that the system begins to oscillate around
values for the four superimposed trajectories
the unstable steady state. An obvious question is under what conditions something are red, (6, 6); purple, (6, 3); green, (−6, −3);
like that might happen. The answer to this question is complicated and constitutes and blue, (−6, −6). Other initial values would
an ongoing research topic in the field of dynamical systems analysis. We will show lead to similar trajectories in between,
an example in the section on systems dynamics. collectively covering the entire plane.
4.10 Parameter Sensitivity

Sensitivity analysis is concerned with the generic question of how much a system
is affected by small alterations in parameter values. In a biological context, such an
alteration could represent a mutation or some permanent change in an enzyme
activity that could be due to aging or disease. Questions of sensitivity are fundamen-
tally different from local stability analysis. In stability analysis, one or more of the
dependent variables are perturbed, the system responds to this perturbation, and
one studies the response, which is usually a return to the steady state. In sensitivity
and gain analysis, by contrast, parameters or independent variables are changed,
respectively. These changes are permanent in the sense that the system cannot
counteract them and they almost always cause the system to assume a new steady
state. Thus, the question is how different the new steady state is from the old. It may
happen that even large variations in parameters do not change the location of the
steady state much, but it can also happen that a minute change in some crucial
parameter throws the system off course. In a biomedical application, two individuals
may be exposed to the same environmental toxicant (perturbation), but respond
differently. One person may not be visibly affected, while the other may become
very sick. The difference could be due to different expression levels (parameter
values) of some key genes.
Sensitivity analysis is a crucial component of any systems analysis, because it can
quickly show if a model wrong. Good, robust models usually have low sensitivities,
which means that they are quite tolerant to small, persistent alterations, in which a
parameter remains altered. Thus, if some of the sensitivities in the model are very
high, we typically have cause for concern. However, there are exceptions, for instance
in signal transduction systems, where even small changes in signal intensity are
greatly amplified (see Chapter 9). Typical sensitivity analyses are again executed by
computing the derivative of a system feature with respect to a parameter. While
steady-state sensitivities are the most prevalent, it is also possible to compute sensi-
tivities of other features, such as trajectories or amplitudes of stable oscillations.
The easiest way to see how sensitivity analysis works is to study an explicit func-
tion. As a specific illustration, let us look again at the logistic function F(x) = a/
(b + e−cx), (4.42). We have already seen that changing the parameter c affects the steep-
ness of the sigmoidal function (see Figure 4.8). Let us now look at the sensitivity of the
function with respect to the parameter b. As with c, we can just change b to some
arbitrary different value, and maybe repeat this exploration a few times with different
numbers. This strategy gives us some idea in this simple case, but we are interested in
a more systematic way to compute the effect of b on F. So, we shall study the effect of
a generic, small change in b. To make our intent of testing the effects of b explicit, we
write the function as F(x; b). The result of a small change Δb in F(x; b) is given as F(x;
b + Δb). Using the Taylor approximation theorem (see Box 4.1), we can linearize this
expression by using the operating point F(x; b), the slope, and the distance of the devi-
ation (see (4.29) and (4.30)). The result in this case of a single-variable function is
∂F ( x ; b )
F ( x ; b + ∆b) ≈ F ( x ; b) + ∆b , (4.67)
∂b
where the expression with the ∂′s indicates the partial derivative of F with respect
to b. The change in F caused by a small change in b is
∆F ≈ F ( x ; b + ∆b) − F ( x ; b). (4.68)
Rearranging this equation and using (4.67) gives us the desired answer, namely,
a quantification of how much F changes if b is altered:
∂F ( x ; b )
∆F ≈ ∆b. (4.69)
∂b
To put this result in words: a small change in b is translated into a change in F with
magnitude ∂F(x; b)/∂b. If this expression is equal to 1, the change in F is equal to the
change in b; if it is larger than 1, the perturbation is amplified; and if it is smaller
than 1, the perturbation is attenuated. According to Taylor’s theorem, these answers
are absolutely accurate only for infinitesimally small changes in b, but they are
sufficiently good for modest—or sometimes even large—alterations in b. The take-
home message is that the effect of b on F is driven by the partial derivative with
respect to b, and this message is unchanged even for differential equations and
systems of differential equations.
As a numerical example with the logistic function, let us assume that a = 2 and
c = 0.25, and that the normal value of b is 1. The partial derivative of F with respect to
b is computed while keeping all other parameters and the variable x constant. Thus,
∂F ( x ; b ) −a
= . (4.70)
∂b (b + e − cx )2
Obviously, the right-hand side depends on x, which means that the sensitivity of F
with respect to b is different for different values of x. A typical example is the sensi-
tivity of the nontrivial steady state. In the case of F, the function approaches this
state for x going to +∞. Maybe we would prefer a finite number for our computation,
but we can still compute the sensitivity, because for x going to +∞, the exponential
function e−cx goes to zero. Therefore, the sensitivity is
∂F ( x ; b ) a
=− 2 for x → ∞. (4.71)
∂b b
Plugging this result into (4.69), along with the values a = 2 and b = 1, we obtain an
estimate for how much the nontrivial steady state of F will change if b is altered.
For instance, if Δb = 0.1, which would mean a 10% perturbation from its normal
value of 1, the steady state of F is predicted to change by −2 × 0.1 = −0.2. In fact, if we
numerically compute F(x; b = 1.1) for x going to +∞, the result is −2/1.1 = −1.81818,
which is close to the predicted 0.2 decrease.
It is not difficult to imagine that our function F(x; b) could be the steady-state
solution of a differential equation. Reinterpreting our procedure and results then
tells us how to do steady-state sensitivity analysis for a single differential equation,
and systems of equations follow the same principles. Software like PLAS [1]
automates this computation for S-systems and GMA systems.
The profile of sensitivities is an important metric for assessing biological
systems. If the sensitivities with respect to parameter p1 are relatively high in magni-
tude, while they are very close to zero for parameter p2, then it is important to obtain
the numerical value of p1 as accurately as possible, whereas a small error in p2 has
little effect on the system.
4.11 analysis of Systems Dynamics

Because the differential equations describing the dynamical system are typically
nonlinear, we seldom have the luxury of being able to compute explicit solutions, in
which each dependent variable can be formulated as a function of time. The vari-
ables are indeed functions of time, and we can grind them out on a computer with
so-called numerical integrators, but we cannot write them down, for instance, as
X(t) = X0eat, as we did in the case of linear differential equations. Nonetheless, there
is an entire branch of mathematics that studies dynamic phenomena, such as
damped and stable oscillations, bifurcations, and chaos with analytical and numer-
ical methods. Even introductions to these types of investigations tend to require
quite a bit of effort and mastery of a variety of mathematical techniques (see, for
example, [47]), but we will discuss some of these issues at a somewhat superficial
level toward the end of this chapter.
While there are some nonlinear ODEs that do have explicit solutions, they are
typically restricted to such special constellations of mathematical structures and
parameters that they must be considered extremely rare exceptions in the real world.
A consequence is that most dynamic analyses in systems biology are done by means
of computer simulations. As we have already discussed, rigorous mathematical
results are always to be preferred over a collection of numerical results, but in bio-
logical applications we can seldom get around computational simulation analyses.
20 Figure 4.19 Simulation of a bolus

experiment with a dependent variable. A
50% bolus of X1 is supplied at time t = 10. All
variables change in response, but the system
concentration quickly returns to its steady state.
X1
10
X2
X3
0
0 30 60
time
Typical analyses fall into four classes:

1. Bolus experiments
2. Persistent changes in structure or input
3. Comprehensive screening experiments
4. Analyses of critical points, where the system behavior changes qualitatively
Bolus experiments usually begin with the system in a stable steady state. Some
time period into the experiment, one of the variables is altered. If this variable is a
dependent variable, the simulation shows whether and how the system returns to
the steady state. The “how” here refers to the time it takes to return and the shape of
the transient, which is the collective name for the time course between the altera-
tion (stimulus) and the move back to the old steady state, to a new one, or to an
oscillatory behavior. Note that the time it takes to return to the steady state is, strictly
speaking, infinite in most dynamical models. For instance, if the solution is a
decreasing exponential function, the final value of 0 is mathematically only reached
when x reaches infinity. A solution to this predicament is to study how long it takes
the system to return to within a small percentage of the steady state, such as 1%.
As an example, consider again the simple generic pathway in Figure 4.11, where
time units are minutes. Using the GMA representation with parameter values γ 11 = 20,
f110 = 1, f112 = −0.75, γ12 = 0.4, f121 = 0.75, γ13 = 2, f131 = 0.2, γ22 = 0.5, f222 = 0.5, f224 = 1,
γ32 = 2.5, f323 = 0.2, X0 = 1, X4 = 2, the steady state is X1 = 10.59, X2 = 5.52, X3 = 3.47. As
a bolus experiment with a dependent variable, we raise X1 at time 10 minutes by 50%.
Immediately following, the other variables respond, showing slight overshoots, and
the system returns to within 1% of its steady state in under 20 minutes (Figure 4.19).
If the altered variable in the bolus experiment is an independent variable, such
as an input, then any alteration in its value will likely affect the steady state. There-
fore, the bolus is usually applied for only a few time units. Afterwards, the altered
variable is reset to its value at steady state. The simulation shows how the system
responds to this temporary change. As an example, let us increase X4 in the branched
pathway by 50% at time t = 10 minutes and reset X4 after at t = 45 minutes. The
responses are quite different (Figure 4.20). In particular, the system essentially
reaches a new steady state after about 35 minutes.
20
concentration
X1
10
X2
Figure 4.20 Simulation of a bolus
experiment with an independent variable.
X3
A 50% bolus of X4 is supplied at time t = 10
0 minutes and stopped at t = 45 minutes.
0 45 90 During the altered setting, the system
time assumes a different steady state.
Simulations targeting persistent changes in structure or input are implemented

with permanent alterations in parameter values or independent variables. A change
in a parameter value could model a mutation or disease, which for instance could
affect the rate in a growth model or an enzyme activity in a metabolic model.
An alteration in an independent variable could be interpreted as a change in envi-
ronmental conditions, which could affect the availability of a substrate or other
essential factor. In these types of simulations, one may study transients between the
change and the completion of the response, but the target is usually the effect of
the alteration on the steady state. An illustration is the previous example of a
change in X4. However, here one would not reset X4 at t = 45 minutes, and the sys-
tem would stay close to the elevated state we observe at t = 45 minutes. A different
example would be the change in the parameter γ32, which represents the rate of
degradation of X3.
Comprehensive screening experiments are usually launched if one is not entirely
sure what to look for. Many options are possible. In a grid search, every parameter is
allowed to take on a certain number of possible values; let us say 6. A complete
simulation now computes the system (transients and/or steady state) for every com-
bination of parameter settings, thereby producing a coarse, but in a way comprehen-
sive, picture of possible responses. The problem with this approach is that it easily
gets out of hand, because a system with n parameters requires 6n simulations, and, for
instance, 620 = 3.66 × 1015. Even for a computer executing 10 simulations per second,
it would take over 100 billion hours (more than 11½ million years) to complete the
grid search! The rapid increase in numbers of simulations in these types of studies
even has a name: combinatorial explosion. Various statistical techniques for over-
coming this problem have been developed. They prescribe how to sample the vast
space in a fashion that is manageable and at the same time statistically valid. Probably
the best known methods are the Latin square and Latin hypercube techniques [48].
An alternative strategy is Monte Carlo simulation, which we discussed in Chapter 2.
Analyses of critical points address situations where the system behavior changes
qualitatively and often abruptly. Most changes in parameter values have an effect
on the steady state and on transients between stimuli and responses, but the effects
are only gradual: the values go up or down a little bit, or the transient overshoots
more and takes longer to come close to the steady state. However, even some
extremely small alterations can throw a system very definitively off kilter. A qualita-
tive change of this type may mean that the system becomes unstable, that one of the
variables goes to zero or infinity, or that the system enters a stable oscillation that
continues with the same amplitude infinitely. While these are typical examples,
many weird things can happen in nonlinear systems [49], and some people have
actually referred to the collection as a zoo of nonlinear phenomena.
Points within the space of all parameters (meaning specific combinations of
parameter values) that have the characteristic of a threshold where the qualitative
behavior changes are called bifurcation points, because there is a fork in the param-
eter space: above the bifurcation value the system does one thing and below the
value the system does something else. Much of the field of nonlinear dynamics is
concerned with mathematical ways of characterizing such bifurcation points. As a
less elegant, but often easier, alternative, one can execute a large number of simula-
tions, maybe with Monte Carlo methods, trying to establish domains where all
parameter combinations lead to similar solutions, along with boundaries between
domains where the system changes qualitatively.
As an example, consider a small system that could describe the mutual activa-
tion pattern of two genes (Figure 4.21A) [50]. Corresponding power-law equations
with typical parameter values are shown in Figure 4.21B. Notice that one parameter
(α1) is left unspecified. It will be our critical parameter and could represent the
strength of activation by a transcription factor. Even though the system has only two
dependent variables, its responses can be quite complicated. For values of α1 below
1, the system has a stable steady state, to which it returns after perturbations (see,
for example, the exogenous increase in X1 at time 10 in Figure 4.21C). As soon as α1
is increased above 1, the steady state becomes unstable and the system enters an
oscillation that will continue until some external force stops it (Figure 4.21D).
The mathematical analysis of dynamical systems attempts to characterize these
bifurcation points at which the system dynamics changes drastically. This analysis is
(A) (B)
+
X1
X1 = α1X1 X2
• 0.4 –0.15 0.2
– – X1
•
X2 = X1 – X20.2
+
X2
(C) (D)
1.50 3.0
X2
X1
0.75 1.5
X2 X1
0 0
0 30 60 0 120 240
time time
α 1 = 0.9 α 1 = 1.02
Figure 4.21 Model of a simple gene circuit in which two genes affect each other’s expression level. (A) Diagram of the gene circuit. (B) Model
equations. Changing the strength α1 of the effect of a transcription factor on gene expression can lead to qualitatively different model responses (C, D).
typically quite involved, but happens to be simple for the specific form of two-
variable S-systems [51, 52].
oTHer aTTracTorS
As we discussed in Section 4.8, a stable steady state is a point to which trajectories of
a system converge. In the language of nonlinear dynamics, a stable steady state is
called an attractor, whereas an unstable steady state is called a repeller, because
close-by trajectories are seemingly pushed away from it. Of interest now is that these
more generic terms of attractors and repellers also include more complicated sets of
points that draw trajectories in or push them away. These more complicated attrac-
tors or repellers consist of lines or volumes in two, three, or more dimensions. They
may even be fractal, which means that their patterns are the same at different
magnifications, like the cover illustration of this book. Indeed, some attractors are as
difficult to comprehend as they are stunningly “attractive,” as a quick Google search
for images of “strange attractors” attests.
Two classes of attractors, beyond steady-state points, are particularly important
for systems biology. The first consists of limit cycles, which are oscillations with a
constant amplitude that attract nearby trajectories. The second contains various
chaotic systems, which, however, are not “all over the place” but move unpredict-
ably although they never leave some constrained, often oddly shaped, domain. The
literature on both types is enormous, and the purely mathematical analysis of
chaotic oscillators is quite complicated [53]. However, the basic concepts are easy to
explore with simulations. A good introduction can be found in [11].
oTHer aTTracTorS 123
4.12 Limit cycles

We have several times used the term “trajectory,” referring to the solution of a
system between two points in time and, specifically, between an initial value and a
steady-state point. Trajectories may be linear or nonlinear, simply curved, oscillat-
ing, or very complicated. One particular type of trajectory is special in that it leaves
and returns to the same point again and again. A simple example is the sine–cosine
oscillation. As we discussed earlier—see (4.25)—it can be represented as a pair of
linear differential equations:
x = y , x (0) = 0,
(4.72)
y = − x , y (0) = 1.
Instead of displaying this oscillation as a function of time (Figure 4.22A), it is often

more instructive to show it as a phase-plane plot, where y (cosine) is plotted against
x (sine). In the case of a sine oscillator, this plot is a circle (Figure 4.22B). Note that
this representation does not directly show time, so one cannot see how fast the sys-
tem completes a cycle. Nor can we directly see in which direction the solution pro-
gresses or where it starts. However, if we look at the time-course plot (Figure 4.22A),
we see that the solution starts at (0, 1) and that x increases, while y decreases. Iden-
tifying these trends in the phase-plane plot shows that the solution starts at 12
o’clock and moves clockwise. The phase-plane representation explains the termi-
nology of a “periodic closed trajectory,” because, after one period, the trajectory
closes and continues on the same path time and again. Such a closed trajectory is
also called an orbit.
A key feature of this oscillation is that it is not stable. We can easily test this by
perturbing the solution. Specifically, we solve the differential equations, stop at
some point in time, alter the numerical value of either x or y or both by a bit, and
record what happens. A typical result is shown in Figure 4.22C, where the variable x
is artificially changed from its current value of about −0.959 to a value of −0.25 at
time t = 5. Owing to the coupling of x and y, y follows essentially instantaneously.
Continuing the simulation, x and y still show sine and cosine oscillations, but the
perturbation has done permanent damage: the amplitude has been altered and
does not recover. The corresponding phase-plane plot is shown in Figure 4.22D.
The solution starts again clockwise at 12 o’clock. At time t = 5, it has not yet completed
one cycle, which it would do at t = 2π, but it is artificially perturbed when x is set to
−0.25. This perturbation corresponds to the dashed horizontal line segment. From
(A) x = sin t y = cos t (B)

y
1
0.5
0 x
0 t
10 20 –0.5 0 0.5
–0.5
–1
(C) (D)
Figure 4.22 Sine–cosine oscillation. (A) The
y pair of linear ODEs in (4.72) corresponds to
1
a sine–cosine oscillation that repeats every
2π time units. (B) The same oscillation is
0.5
represented in the phase plane as a circle.
10 (C) When the oscillation is perturbed, the
0 t 0 x
20 –0.5 0 0.5 oscillations retain the same frequency,
–0.5 possibly with a shift, but assume a new
amplitude. (D) The corresponding phase-
plane trajectory jumps (dashed line) to a new
-1 circle, which here has a smaller diameter.
then on, the oscillation continues clockwise along the inner circle and never returns
to the original circle, unless it is forced to do so.
More interesting periodic closed trajectories are limit cycles, because they attract
(or repel) close-by trajectories. If they attract, they are called stable oscillations or sta-
ble limit cycles, whereas repellers in this case refer to unstable limit cycles. Limit cycles
are most easily discussed with examples. One famous case was proposed in the 1920s
by the Dutch physicist Balthasar van der Pol (1889–1959) as a very simple model of a
heartbeat [54, 55]. Since his early proposal, many variations of this van der Pol oscilla-
tor have been developed and applied to a wide range of oscillatory phenomena [11].
The prototype van der Pol oscillator is described by a second-order differential
equation,
v − k(1 − v 2 )v + v = 0, (4.73)
where k is a positive parameter. The variable v with two dots represents the second
derivative with respect to time, while v with one dot is the typical first derivative that
we have used many times before. It is easy to convert this second-order equation
into a pair of first-order differential equations, without losing any features of the
original. This is done by defining a new variable
w = v. (4.74)
Because the derivative of the derivative of a variable is the second derivative,

we immediately obtain
w = v. (4.75)
Rearranging (4.73) and substituting the results into (4.74) and (4.75) for v and v
results in a first-order system, provided the appropriate initial values are chosen, is
exactly equivalent to (4.73):
w = k(1 − v 2 )w − v , w(0) = v0 ,

(4.76)
v = w , v(0) = v0 .

To start the system, initial values must be specified at time 0 for v as well as v.
Figure 4.23A confirms that the oscillation in v repeats with the same pattern.
The secondary variable, w, oscillates with the same frequency, but quite a different
shape (Figure 4.23B). Instead of displaying the oscillations as functions of time, we
can again use the phase plane, where w is plotted against v. The result is shown in
Figure 4.23C. Starting at the point on the far right, the solution initially decreases
clockwise. Displaying the solution only for time points spaced by 0.05 shows that the
points are dense in some sections, but not in others (Figure 4.23D). In the dense sec-
tions, the solution is comparatively slow, while it is fast for points that are farther
apart.
Intriguingly, if the van der Pol oscillation is perturbed, it recovers and returns to
the same oscillatory pattern. This feature is the key characteristic of a stable limit
cycle. For instance, suppose the system is perturbed at t = 18.6. At this point, the
system without perturbation has the values v ≈ 0.26 and w ≈ 10. Artificially resetting
v to 0 yields the plots shown in Figures 4.24 and 4.25. It is evident that the original
oscillation is regained, although with some shift in the direction of time.
Limit cycles may have very different shapes. A flexible way to create such shapes
is to start with a limit-cycle model in the form of a two-variable S-system, for which
Lewis [51] computed mathematical conditions. An example, adapted from [52], is
X 1 = 1.01( X 28 − X 1−4 ), X 1 = 1.7,

(4.77)
X 2 = ( X 1−3 − X 24 ), X 2 = 0.9.
Figure 4.26 shows the oscillations of this system in both time-course and phase-
plane format. If one selects initial values inside the limit cycle, but not exactly at the
unstable steady state (1, 1), all trajectories are attracted to the limit cycle. Similarly,
(A) (B) Figure 4.23 Visualization of the van

v w der Pol oscillator (4.76). (A) The original
4 20
variable v represents a limit-cycle oscillation
that was designed to mimic a heartbeat.
(B) The plot of the auxiliary variable w has
the same frequency, but quite a different
shape. (C) Phase-plane plot of w versus v.
0 0 (D) Plotting the solution shown in (C) at
spaced-out time points gives an idea of the
speed of the oscillation in different phases
of the oscillation. Dense points correspond
to slow speed, while more widely separated
t t points indicate a higher speed.
–4 –20
0 50 100 0 50 100
(C) (D)
w w
20 20
0 0
–20 v –20 v
–4 0 4 –4 0 4
oscillations beginning close enough outside the limit cycle are attracted as well.
If one starts further away, one of the variables may become zero, and the equations v
are no longer defined. 4
By multiplying both equations of the limit-cycle system by the same positive

functions of X1 and/or X2, the appearance of the limit-cycle oscillation can be
changed quite dramatically (Figure 4.27), although the phase-plane plot remains
unchanged (Exercise 4.61). In all cases, the limit cycle remains stable and attracts 0
close-by trajectories.
With the same methods as for systems without limit cycles, we can compute the
steady states of the system in (4.77). Since a negative power is not really defined for
X = 0, we will not discuss a possibly trivial steady state. The nontrivial steady state is
(1, 1). Checking the stability of this state, we find a pair of complex-conjugate –4 t
0 50 100
eigenvalues with a real part close to 0, which means that the steady state is unstable.
This observation reflects the fact that inside a stable limit cycle there lies either at Figure 4.24 Perturbation of the van der
least one unstable steady state or an unstable limit cycle. We also note that the Pol oscillator. As is typical for a limit cycle,
imaginary parts are nonzero, which indicates the potential for oscillations. If the the oscillation recovers from a perturbation
rate constant 1.01 is changed to 1.00, the real parts become zero. At this bifurcation and resumes the original pattern, but the
point, the stable limit cycle is “born” when the rate constant is slowly increased from perturbation may lead to a shift in phase.
below to above 1. For values smaller than 1, the system is stable.
(A) (B) (C) (D)

w w w w
20 20 20 20
0 0 0 0
–20 v –20 v –20 v –20 v

–4 0 4 –4 0 4 –4 0 4 –4 0 4
Figure 4.25 Phase-plane visualization of the perturbation in Figure 4.24. The van der Pol oscillator is perturbed at t = 18.6, where v is artificially
changed from 0.26 to −3 (red arrow). (A), (B), (C), and (D) show, in the phase plane, the same trajectory, which is stopped at t = 20, 30, 40, and 50,
respectively. The arrowheads have been added for illustration.
(A) X1 X2 (B) Figure 4.26 Limit cycles can have

X2
2 2 very different shapes. (A) Time-course
representations of the limit cycle in
(4.77). (B) The corresponding phase-
plane plot. (C) Starting with initial values
(X1, X2) = (1.01, 0.99), which are close to the
unstable steady state (1, 1), the oscillation
1 1 spirals outward, eventually approaching the
limit cycle. (D) Approach to the limit cycle
from inside (green) and outside (blue) the
limit cycle.
0 t 0 X1
0 30 60 0 1 2
(C) X1 X2 (D) X
2
2 2
1 1
0 t 0 X1
0 100 200 0 1 2
4.13 chaotic attractors

In 1963, the American mathematician and meteorologist Edward Lorenz [56]
studied weather patterns and noticed that even minute changes in initial values
could lead to totally different outcomes if one waited for a while. It was Lorenz who
introduced the concept of the butterfly effect. Using a simplified model of air flow,
Lorenz was able to reproduce this effect with three differential equations and
thereby ushered in an intensive exploration of systems that had features similar to
what Lorenz called a strange attractor. Trajectories in this and other, similar sys-
tems are confined to some volume in three-dimensional space, but their specific
values cannot be predicted unless one solves the equations. The amazing aspect is
that it is very easy to simulate these equations, which in Lorenz’s notation read
x = P( y − x ),
y = Rx − y − xz , (4.78)
z = xy − Bz.
(A) X1 X2 (B) X1 X2
2 2
1 1
Figure 4.27 Variants of limit-cycle

oscillations in (4.77). (A) Both equations
were multiplied by X 16 X 2−2 . (B) Both equations
were multiplied by X 1X 2−4. Note the different
0 t 0 t shapes and scales for the time axes and thus
0 15 30 0 5 10 the different speeds of the oscillations.
(A) x y z (B) x y z Figure 4.28 Simulation results for the

50 50 strange attractor proposed by Lorenz
[56]. (A) Time-course solution for t ∈ [0, 50].
(B) The same time-course in the interval
t ∈ [40, 50] exhibits irregularities more
clearly. (C) and (D) Phase-plane plots
corresponding to projections of the three-
10 10 variable oscillator onto the x–z and y–z
planes, respectively. See also Figure 4.29.
–30 t –30 t
0 25 50 40 45 50
(C) (D)
z z
50 50
25 25
0 x 0 y
–20 0 20 –30 0 30
Lorenz used the parameter values P = 10, R = 28, and B = 8/3, along with the initial
values (1, 1, 1). Figure 4.28 shows some output that demonstrates the erratic,
chaotic trajectories of the system, which never exactly repeat. A common character-
istic of chaotic oscillators is the fact that they are extremely sensitive to changes in
initial values, and even a change to (1.001, 1, 1) results in a different temporal
pattern. With such high sensitivity, even the solver settings can lead to different
results. Interestingly, long-term simulations lead to an attractor in three dimensions
that is overall always very similar, but differs in details. Expressed differently, any
trajectory lies somewhere within the attractor, but one cannot predict where exactly
it will be without executing a numerical simulation (Exercise 4.66).
The Lorenz attractor actually has three steady states. They are easy to compute
and, as it turns out, they are all unstable (Exercise 4.69). Starting a simulation close
to one of the steady states fills one of the holes in the attractor with outward spirals
that eventually merge into the attractor (Figure 4.29).
50
10
Figure 4.29 Filling the holes of the Lorenz

attractor. Starting simulations close to
the nontrivial unstable steady-state points
–30 yields outward spirals (red and green,
respectively) that eventually merge into the
x attractor. The pseudo-three dimensional plot
0 demonstrates that the solutions are confined
to two “disks.” Where exactly the future
solution will be within this attractor cannot
–10
0 30 60 be predicted; it can only be determined
z through simulations.
Figure 4.30 The Thomas oscillator. The

deceivingly simple set of ODEs in (4.79) leads
to a complicated strange attractor.
5
4
3
2
1
y 0
–1
–2
–3
–4
–5
5
5
0
x 0
z
–5 –5
The Lorenz system is probably the best known strange attractor, but many others
have been documented. One strategy for creating chaotic oscillators is to start with
a limit cycle and add a small sine function to it. An example was shown with the blue
sky catastrophe in Chapter 2 (see Box 2.2). A second example is the addition of a
function like 0.00165 sin(1.5t) to the first example of the limit cycle in (4.77). This
situation is considered in Chapter 12. Such an addition sometimes leads to chaos,
but not always. It is possible that the oscillation will become regular after some time
or that one of the variables will goes to ±∞ or 0 (Exercise 4.62). Even systems of
differential equations that have a simple structure, such as BST and Lotka–Volterra
models, can exhibit chaos (see Chapter 10 and [57–59]).
A particularly stunning example, because of its apparent simplicity, is the
Thomas oscillator
x ′ = −ax + sin y ,
y ′ = −ay + sin z , (4.79)
z ′ = −az + sin x ,
whose appearance changes dramatically depending on the choice of the parameter

a [60]. The complicated phase plane plot for a = 0.18 and initial values (x, y, z) = (0,
4, 0.1) is shown in Figure 4.30 for t ∈ [0, 2000]. Several other examples of chaotic
systems, including the continuous Rössler band [61] and variations of the Thomas
oscillator, as well as the discrete Hénon map [11], are discussed in Exercises
4.63–4.72.
Seeing how complicated the trajectories of seemingly simple systems can
become, one may ask how overwhelmingly complicated a large system could be.
Indeed, the sky is the limit. At the same time, large systems often contain mecha-
nisms that keep their subsystems from “misbehaving,” and it can easily turn out that
the behavior of a large system is quite dull. However, one never knows until one
studies a system with rigorous mathematical techniques.
eXerciSeS
4.1. Describe in words, as well as mathematically, what synchronized? What happens if τ is slightly different
happens if the initial size Pt in the recursive model for each bacterium?
Pt+τ = 2Pt in (4.1) is different from 10. 4.4. Is anything special about the number 2 in (4.1)–
4.2. What happens (biologically and mathematically) in (4.3)? What does it mean if 2 is replaced by some
(4.1) at time points between t and t + τ or between other integer, such as 3 or 4? What about 1?
t = 3τ and t = 4τ ? 4.5. Is it mathematically and/or biologically meaningful
4.3. Is it important to distinguish whether or not the if the number 2 in (4.1)–(4.3) is replaced by
bacterial culture in the text (Table 4.1) is −2, −1, or 0?
eXerciSeS 129
4.6. What is the final size of the population in (4.1)–(4.3)? 4.17. Compute the limiting distribution for the example in
What is the final size if 2 is replaced with 0.5, 1, or −1? (4.16)–(4.19) with methods of linear algebra. This
What effect does the initial population size P1 have distribution is the solution of the matrix equation
on the final size in each case?
4.7. Construct a recursive model that approaches a  π1   π1 
constant final value (steady state) that is not 0, +∞ or  π 2  = MT  π 2  .
−∞. Is this possible with a linear model? If so, construct    
 π 3   π 3 
an example. If not, make the model nonlinear.
4.8. For the example in (4.4), compute Rn+1 from Rn and Does the limiting distribution contain the same
then Rn+2 from Rn+1 for different values of f, g, Rn, and amount of information as M?
Mn.
4.18. What happens with the logistic map (4.24) if r is
4.9. For the example in (4.4), compute Rn+2 directly from greater than 4? What happens if r is less than 0?
Rn, using the matrix formulation, and compare the
results with those in Exercise 4.8. Compute Mn+4. 4.19. Simulate the Gauss map model
4.10. Compute the steady state(s) of the following two 2
recursive systems: Pt +1 = e aPt + b.
X n+1 = (1 − p) X n + Yn , Start with the values a = −6.2, b = −1, and P0 = 0.2.

A: Change P0 and record what happens. Increase b in
Yn+1 = qX n + 1;
small steps (for example, 0.1) toward 1. Does the system
X n+1 = (1 − p) X n + Yn , have a steady state? Can it be computed algebraically?
B:
Yn+1 = qX n . 4.20. Linearize a system of the form (4.28), with
4.11. Discuss the steady state(s) of the so-called Fibonacci
c 2YX n
series, which is given by an+2 = an+1 + an and is usually f ( X ,Y ) = c1 − ,
started with a1 = 1 and a2 = 1. Also discuss the steady c 3n + X n
states of the associated series of ratios bn+1 = an+1/an. g ( X ,Y ) = c 4 − c5Y ,
4.12. Discuss the implications of the assumption that the
travel times between states in the pinball example where ci are positive parameters and n = 1, 2, or 4.
are all the same. Define initial values. Give a possible interpretation
for these equations in terms of a metabolic or
4.13. What is the behavior of a Markov chain model in cellular system. Compute the functions and their
which all matrix elements are the same? What is the approximations for different sets of parameter
behavior if all matrix elements on the diagonal are 1? values. Consider different operating points. Write a
4.14. Describe the dynamics of a Markov model with the brief report about your findings.
matrix 4.21. Linearize the system in (4.39) at the steady state.
Select initial values of 50, 10, 100 for S, I, and R and
0 1 0 0 0 assess the quality of the representation.
1 0 0 0 0 4.22. Compute the Jacobian of a generic two-variable
 
M = 0 0 0 0 1 . linear system without input.
0 0 0 1 0 4.23. Show that y = ce x + e2x is the solution of the
 differential equation y′ − y = e2x.
 0 0 1 0 0
4.24. Show that the logistic function (4.42) is indeed the
solution of the logistic differential equation (4.43).
4.15. Do the matrix elements in each column of a Markov For this purpose, differentiate N(t) = a/(b + e−ct)
matrix necessarily sum to 1? Why or why not? with respect to t and express this derivative as a
 X1  function of N(t) itself.
4.16. Compute  X 2  for the example in (4.16)–(4.19) in 4.25. Explore the effects of changes in α, β, and the initial
  size N0 on the shape of the logistic growth
 X 3 
4 function (4.43).
 X1 
4.26. Does it make sense biologically and/or
two ways. First, multiply  X 2  by MT. Second, mathematically if N0 in (4.43) is greater than the
 
 X 3  carrying capacity? If feasible, simulate this case.
3
 X1  4.27. What happens mathematically in the logistic
equation (4.43) if both α and β are negative? Is this
compute the square of M and multiply  X 2  by it.
T
  situation biologically relevant?
 X 3 
2 4.28. Linearize the logistic function in (4.43) at its steady
Explain why both results are identical. state(s) and determine stability.
4.29. Study the roles of KM and Vmax in (4.44). How do KM X 1 = α 1 X 2g12 − β1 X 1h111 − β 2 X 1h112 ,
and Vmax relate to each other?
4.30. Revisit the SOS system (4.52)–(4.54) and consider PL X 2 = β1 X 1h111 − β 3 X 2h22 .
and Ptrc as dependent variables. Change the model
Is the diagram unique? Discuss your answer.
equations correspondingly. Redo the simulation in
Approximate the GMA system with an S-system.
Figure 4.13 with the expanded system. Write a brief
Discuss all conditions under which the S-system is
report on your findings.
equivalent to the GMA system for all values of X1
4.31. Prove the shortcut in Boxes 4.3 and 4.4 for computing and X2.
power-law approximations.
4.41. Compute the steady state of the S-system in (4.51) for
4.32. Compute power-law approximations for the Hill arbitrary kinetic orders and for rate constants that
function are all equal to 1; assume X0 = 1. Show all steps of
your computations. Discuss which features (steady
Vmax S n state, stability, dynamics, …) of the system change if
vH = n
KM + Sn all rate constants are set to 10.
for n = 1, 2, and 4 at different operating points. Select 4.42. Compute the steady state of the S-system
Vmax and KM as you like. Plot the approximations
together with the original Hill functions. X 1 = α( X 28 − X 1−4 ),
4.33. Compute power-law approximations for the function X 2 = X 1−3 − X 24
F ( X 1 , X 2 , X 3 ) = X 1e X 2 +2 X 3 , for α = 0.8, 1, and 1.01. Show all steps of your
computations. In each case, determine the stability
using different operating points. Assess the approxi- of the system. Simulate the system in PLAS.
mation accuracy for each operating point. Describe
your findings in a report. 4.43. Compute all steady states (trivial and nontrivial) of
the SIR model
4.34. Compute the power-law approximation of the
function
S = 0.01R − 0.0005SI ,
2 2Y +1
F ( X ,Y ) = 3.2 X e I = 0.0005SI − 0.05I ,
at the operating point X = 1, Y = −0.5. R = 0.05I − 0.01R.
4.35. Develop a power-law approximation of the function
4.44. Consider the following ethanol (E)-producing
3 −1 fermentation system, in which the growth of
f (x , y) = 4 x y . bacteria B is driven by the availability of substrate
(glucose, G):
at the operating point (x, y) = (1, 2).
4.36. Develop a power-law approximation of the function
B = μB − aB 2 ,
f ( x , y ) = 4 x 3 sin y V B
E = max − cE .
KM + B
at the operating point (x, y) = (1, 2).
4.37. If a Hill function with a Hill coefficient of 2 is In this formulation,
approximated by a power-law function, in which
numerical ranges do the kinetic order and the rate G2
constant fall? μ = μmax ,
K I2 + G 2
4.38. Implement the branched pathway system in (4.49)
with the parameter values X0 = 2, γ11 = 2, γ12 = 0.5, and μmax KI, Vmax, and KM are positive parameters.
γ13 = 1.5, γ22 = 0.4, γ32 = 2.4, f110 = 1, f112 = −0.6, Explain the biological meaning of each term in the
f121 = 0.5, f131 = 0.9, f222 = 0.5, f224 = 1, and f323 = 0.8. two ODEs. Assuming constant glucose, as well as G,
Start with X1 = X2 = X3 = X4 = 1, but explore different E, B ≠ 0, compute the power-law approximation of
values between 0.1 and 4 for X4. the fermentation system at its steady state. Execute
4.39. Determine the S-system that corresponds to the comparative simulations with the original system
GMA system in (4.49) with the parameter values in and its power-law approximation.
Exercise 4.38. The two systems should be equivalent 4.45. Compute the trivial and nontrivial steady state(s) of
at the nontrivial steady state and as similar as the Lotka–Volterra system
possible otherwise. Perform comparative
simulations.
N 1 = r1N 1K 1 (K 1 − N 1 − aN 2 ),
4.40. Construct a diagram that could correspond to the
following GMA system: N 2 = r2 N 2 K 2 (K 2 − N 2 − bN 1 ),
eXerciSeS 131
where r1 = 0.15, K1 = 50, a = 0.2, r2 = 0.3, K2 = 60,

b = 0.6, N1(0) = 1, and N2(0) = 1. Assess the stability
of the steady state. –
X3 X1 X2
4.46. Enumerate the trivial and nontrivial steady state(s) of
+
a Lotka–Volterra system with three dependent
variables.
4.47. Is it possible to construct a model that is at once a Figure 4.31 Generic pathway with two
Lotka–Volterra, GMA, and S-system? If your answer modulation signals. Even a seemingly simple
is no, say why. If your answer is yes, provide an pathway can exhibit quite complex responses.
example.
next step, vary X1(0) and X2(0) up and down and
4.48. Is every Lotka–Volterra system also a GMA system? Is
study the system responses. Finally, vary the
every GMA system also a Lotka–Volterra system? Is
constant value of X3 up and down and study
every S-system a GMA system? Is every lin–log
the system for different initial values X1(0)
system also a GMA system?
and X2(0).
4.49. Determine matrices A for the different regions in
4.57. Suppose all terms in a three-variable ODE system are
Figure 4.17. Implement the systems and solve them,
approximated by power-law functions at a nontrivial
starting with different initial values, as illustrated in
steady state. Discuss the similarities and differences
Figure 4.18.
of the Jacobians of the two systems (the original ODE
4.50. Study the stability of the differential equation (4.27) system and the approximating power-law system) at
with (A11, A12, A21, A22) = (−1, 2.5, −1.5, 1). Compute this steady state.
trace, determinant, and discriminant and locate the
4.58. Why should one typically expect the model of a
expected behavior of the system close to (0, 0) on
biological system to be flawed if some of its
the plot of Figure 4.17. Use simulations to explore
sensitivities are in a range between 10 and 1000?
the consequences of slight changes in any of the Aij.
Is it possible to create different stability patterns by 4.59. What does it imply (1) mathematically or (2)
varying just one of the Aij up and down? for biological experiments associated with the
system if a parameter value has very low
4.51. Study the stability of the differential equation (4.27)
sensitivities?
with (A11, A12, A21, A22) = (−1, 2.5, −0.4, 1). Explore the
consequences of slight changes in any of the Aij. 4.60. Modify the limit-cycle system (4.77) by multiplying
Summarize your findings in a report. only one of the two equations by a power function of
X1 and/or X2. Record the effects on the oscillation
4.52. Calculate the symbolic steady-state equations of the
and the phase-plane plot.
SOS systems (4.52)–(4.54). Enter numerical values
and compare with simulation results. 4.61. Explain why multiplication of the limit-cycle system
(4.77) by a function of X1 and/or X2 changes the
4.53. Describe and explain the behavior of the following
appearance of a limit-cycle oscillation in the time
system at its nontrivial steady state:
domain, but leaves the phase-plane plot
unchanged.
x = f ( x , y ) = 1 − x 2 y ,
4.62. Add to the first equation of the limit-cycle system
y = g ( x , y ) = xy 2 + y . (4.77) the function c sin(bt). Start with c = 0.00165,
test different values for b, and record the shapes of
4.54. Use the numerical system from Exercise 4.52 and the oscillations in time-course and phase-plane
multiply all the right-hand sides by −1. How would representations. Subsequently, alter c as well and
you interpret this operation? Do simulations and record the effects.
interpret the results. 4.63. Explore the discrete strange attractor proposed by
4.55. Use the numerical system from Exercise 4.52 and Hénon, which is defined recursively by
vary each parameter individually by 5% up and
down. Which parameter is most influential with xn+1 = yn − 1.4 xn2 + 1,
respect to the system’s steady state?
yn+1 = 0.3 xn .
4.56. Analyze the system shown in Figure 4.31, using the
S-system model Typical starting values are (x0, y0) = (1, 1).
To see the true shape of the chaotic attractor,
X 1 = X 2−2 X 3 − X 10.5 X 2 , execute several hundred iterations, for instance,
in Excel®.
X 2 = X 10.5 X 2 − X 20.5 .
4.64. Explore the discrete attractor proposed by Hénon
The system contains the independent variable X3, and defined in Exercise 4.63 after replacing the factor
which is constant. Begin the analysis by simulating 1.4 with 1.25.
the system with X1(0) = X2(0) = X3 = 1. Compute all 4.65. Compute the steady state(s) of the Hénon map in
steady states and assess their stability. In the Exercises 4.63 and 4.64.
4.66. Simulate the so-called Rössler attractor, which is and study how the trajectories interact with the
defined by the differential equations attractor itself.
4.70. Predict what the phase-plane plot of the blue sky
x′ = −y − z, catastrophe in Chapter 2 (see Box 2.2) looks like.
Implement the system in PLAS and check your
y ′ = x + ay , prediction.
z ′ = bx − cz + xz , 4.71. Explore the Thomas attractor in (4.79) with different,
small positive values of a. Solve the system with
x(0) = 0, y(0) = 4, and z(0) = 0.5 and study the phase
with a = 0.36, b = 0.4, and c = 4.5. Choose different
plane of x and y and the pseudo-three-dimensional
initial values, beginning with x = 0, y = −3, and z = 0.
representation of x, y, and z.
Summarize your findings in a report.
4.72. Explore the Thomas attractor in (4.79) with
4.67. Compute the steady state(s) of the Rössler system in
a = 0. Start the system with x(0) = 0, y(0) = 4,
Exercise 4.66 and assess its/their stability.
and z(0) = 0.5 and solve the system first for
4.68. Explore the Lorenz oscillator for slightly changed 500 time units and then for much longer time
initial values and parameter values. periods. Study the phase plane of x and y and the
4.69. Compute the steady states of the Lorenz attractor. pseudo-three-dimensional representation of x, y,
Start simulations at or close to these steady states and z.
reFerenceS
[1] PLAS: Power Law Analysis and Simulation. http://enzymology. [17] May RM. Stability and Complexity in Model Ecosystems.
fc.ul.pt/software.htm. Princeton University Press, 1973 (reprinted, with a new
[2] Guckenheimer J & Holmes P. Nonlinear Oscillations, introduction by the author, 2001).
Dynamical Systems, and Bifurcations of Vector Fields. [18] Peschel M & Mende W. The Predator–Prey Model: Do We Live
Springer, 1983. in a Volterra World? Akademie-Verlag, 1986.
[3] Voit EO & Dick G. Growth of cell populations with arbitrarily [19] Voit EO & Savageau MA. Equivalence between S-systems and
distributed cycle durations. I. Basic model. Mathem. Biosci. Volterra-systems. Math. Biosci. 78 (1986) 47–55.
66 (1983) 229–246. [20] Savageau MA & Voit EO. Recasting nonlinear differential
[4] Voit EO & Dick G. Growth of cell populations with arbitrarily equations as S-systems: a canonical nonlinear form. Math.
distributed cycle durations. II. Extended model for correlated Biosci. 87 (1987) 83–115.
cycle durations of mother and daughter cells. Mathem. [21] Savageau MA. Biochemical systems analysis. I. Some
Biosci. 66 (1983) 247–262. mathematical properties of the rate law for the component
[5] Leslie PH. On the use of matrices in certain population enzymatic reactions. J. Theor. Biol. 25 (1969) 365–369.
mathematics. Biometrika 33 (1945) 183–212. [22] Savageau MA. Biochemical Systems Analysis: A Study of
[6] Edelstein-Keshet L. Mathematical Models in Biology. SIAM, Function and Design in Molecular Biology. Addison-Wesley,
2005. 1976.
[7] Fonseca L & Voit EO. Comparison of mathematical frameworks [23] Voit EO. Computational Analysis of Biochemical Systems:
for modeling erythropoiesis in the context of malaria A Practical Guide for Biochemists and Molecular Biologists.
infection. Math. Biosci. 270 (2015) 224–236. Cambridge University Press, 2000.
[8] Fonseca LL, Alezi HA, Moreno A, Barnwell JW, Galinski MR & [24] Torres NV & Voit EO. Pathway Analysis and Optimization in
Voit EO. Quantifying the removal of red blood cells in Macaca Metabolic Engineering. Cambridge University Press, 2002.
mulatta during a Plasmodium coatneyi infection. Malaria J. [25] Voit EO. Biochemical systems theory: a review. ISRN Biomath.
15 (2016) 410. 2013 (2013) Article ID 897658.
[9] Ross SM. Introduction to Probability Models, 11th ed. [26] Shiraishi F & Savageau MA. The tricarboxylic-acid cycle in
Academic Press, 2014. Dictyostelium discoideum. 1. Formulation of alternative kinetic
[10] Durbin R, Eddy S, Krogh A & Mitchison G. Biological Sequence representations. J. Biol. Chem. 267 (1992) 22912–22918.
Analysis: Probabilistic Models of Proteins and Nucleic Acids. [27] Tian T & Burrage K. Stochastic models for regulatory networks
Cambridge University Press, 1998. of the genetic toggle switch. Proc. Natl Acad. Sci. USA 103
[11] Thompson JMT & Stewart HB. Nonlinear Dynamics and Chaos, (2006) 8372–8377.
2nd ed. Wiley, 2002. [28] Visser D & Heijnen JJ. The mathematics of metabolic control
[12] Jacquez JA. Compartmental Analysis in Biology and Medicine, analysis revisited. Metab. Eng. 4 (2002) 114–123.
3rd ed. BioMedware, 1996. [29] Heijnen JJ. Approximative kinetic formats used in metabolic
[13] Chen C-T. Linear System Theory and Design, 4th ed. Oxford network. Biotechnol. Bioeng. 91 (2005) 534–545.
University Press, 2013. [30] Wang F-S, Ko C-L & Voit EO. Kinetic modeling using S-systems
[14] Von Seggern DH. CRC Handbook of Mathematical Curves and and lin–log approaches. Biochem. Eng. 33 (2007) 238–247.
Surfaces. CRC Press, 1990. [31] del Rosario RC, Mendoza E & Voit EO. Challenges in lin–log
[15] Lotka A. Elements of Physical Biology. Williams & Wilkins, 1924 modelling of glycolysis in Lactococcus lactis. IET Syst. Biol.
(reprinted as Elements of Mathematical Biology. Dover, 1956). 2 (2008) 136–149.
[16] Volterra V. Variazioni e fluttuazioni del numero d’individui [32] Heinrich R & Rapoport TA. A linear steady-state treatment of
in specie animali conviventi. Mem. R. Accad. dei Lincei enzymatic chains. General properties, control and effector
2 (1926) 31–113. strength. Eur. J. Biochem. 42 (1974) 89–95.
FurTHer reaDinG 133
[33] Kacser H & Burns JA. The control of flux. Symp. Soc. Exp. Biol. [47] Strogatz SH. Nonlinear Dynamics and Chaos, 2nd ed.
27 (1973) 65–104. Westview Press, 2014.
[34] Fell DA. Understanding the Control of Metabolism. Portland [48] Hinkelmann K & Kempthorne O. Design and Analysis of
Press, 1997. Experiments. Volume I: Introduction to Experimental Design,
[35] Sorribas A, Hernandez-Bermejo B, Vilaprinyo E & Alves R. 2nd ed. Wiley, 2008.
Cooperativity and saturation in biochemical networks: a [49] Epstein IR & Pojman JA. An Introduction to Nonlinear
saturable formalism using Taylor series approximations. Chemical Dynamics. Oscillations, Waves, Patterns, and Chaos.
Biotechnol. Bioeng. 97 (2007) 1259–1277. Oxford University Press, 1998.
[36] Curto R, Voit EO, Sorribas A & Cascante M. Mathematical [50] Voit EO. Modelling metabolic networks using power-laws and
models of purine metabolism in man. Math. Biosci. S-systems. Essays Biochem. 45 (2008) 29–40.
151 (1998) 1–49. [51] Lewis DC. A qualitative analysis of S-systems: Hopf
[37] Kreyszig E. Advanced Engineering Mathematics, 10th ed. bifurcations. In Canonical Nonlinear Modeling. S-System
Wiley, 2011. Approach to Understanding Complexity (EO Voit, ed.),
[38] Bonabeau E. Agent-based modeling: methods and techniques Chapter 16. Van Nostrand Reinhold, 1991.
for simulating human systems. Proc. Natl Acad. Sci. USA [52] Yin W & Voit EO. Construction and customization of stable
14 (2002) 7280–7287. oscillation models in biology. J. Biol. Syst. 16 (2008)
[39] Macal CM & North M. Tutorial on agent-based modeling and 463–478.
simulation. Part 2: How to model with agents. In Proceedings [53] Sprott JC & Linz SJ. Algebraically simple chaotic flows. Int. J.
of the Winter Simulation Conference, December 2006, Chaos Theory Appl. 5(2) (2000) 1–20.
Monterey, CA (LF Perrone, FP Wieland, J Liu, et al., eds), [54] van der Pol B & van der Mark J. Frequency demultiplication.
pp 73–83. IEEE, 2006. Nature 120 (1927) 363–364.
[40] Mocek WT, Rudnicki R & Voit EO. Approximation of delays in [55] van der Pol B. & van der Mark J. The heart beat considered
biochemical systems. Math. Biosci. 198 (2005) 190–216. as a relaxation oscillation. Philos. Mag. 6(Suppl) (1928)
[41] Elowitz MB & Leibler S. A synthetic oscillatory network of 763–775.
transcriptional regulators. Nature 403 (2000) 335–338. [56] Lorenz EN. Deterministic nonperiodic flow. J. Atmos. Sci.
[42] Wilkinson DJ. Stochastic Modelling for Systems Biology, 2nd 20 (1963) 130–141.
ed. CRC Press, 2011. [57] Vano JA, Wildenberg J C, Anderson MB, Noel JK & Sprott
[43] Gillespie DT. Stochastic simulation of chemical kinetics. Annu. JC. Chaos in low-dimensional Lotka–Volterra models of
Rev. Phys. Chem. 58 (2007) 35–55. competition. Nonlinearity 19 (2006) 2391–2404.
[44] Wu J & Voit E. Hybrid modeling in biochemical systems theory [58] Sprott JC, Vano JA, Wildenberg JC, Anderson MB & Noel JK.
by means of functional Petri nets. J. Bioinform. Comput. Biol. Coexistence and chaos in complex ecologies. Phys. Lett. A
7 (2009) 107–134. 335 (2005) 207–212.
[45] Savageau MA. Biochemical systems analysis. II. The steady- [59] Voit EO. S-system modeling of complex systems with chaotic
state solutions for an n-pool system using a power-law input. Environmetrics 4 (1993) 153–186.
approximation. J. Theor. Biol. 25 (1969) 370–379. [60] Thomas R. Deterministic chaos seen in terms of feedback
[46] Wiens EG. Egwald Mathematics: Nonlinear Dynamics: circuits: analysis, synthesis, “labyrinth chaos.” Int. J. Bifurcation
Two Dimensional Flows and Phase Diagrams. Chaos 9 (1999) 1889–1905.
http://www.egwald.ca/nonlineardynamics/ [61] Rössler OE. An equation for continuous chaos. Phys. Lett. A
twodimensionaldynamics.php. 57 (1976) 397–398.
FurTHer reaDinG
easy introductions Kremling, A. Systems Biology. Mathematical Modeling and Model
Analysis. Chapman & Hall/CRC Press, 2013.
Batschelet E. Introduction to Mathematics for Life Scientists, 3rd ed. Kreyszig E. Advanced Engineering Mathematics, 10th ed. Wiley, 2011.
Springer, 1979. Murray JD. Mathematical Biology I: Introduction, 3rd ed. Springer,
Edelstein-Keshet L. Mathematical Models in Biology. SIAM, 2005. 2002.
Murray JD. Mathematical Biology II: Spatial Models and Biomedical
Applications, Springer, 2003.
General Texts on Mathematical Biology Strang G. Introduction to Applied Mathematics. Wellesley-
Adler FR. Modeling the Dynamics of Life: Calculus and Probability Cambridge Press, 1986.
for Life Scientists, 3rd ed. Brooks/Cole, 2012. Strogatz SH. Nonlinear Dynamics and Chaos, 2nd ed. Westview
Beltrami E. Mathematics for Dynamic Modeling, 2nd ed. Academic Press, 1994.
Press, 1998. Yeargers EK, Shonkwiler RW & Herod JV. An Introduction to the
Britton NF. Essential Mathematical Biology. Springer, 2004. Mathematics of Biology. Birkhäuser, 1996.
Ellner SP & Guckenheimer. J. Dynamic Models in Biology. Princeton
University Press, 2006.
Haefner JW. Modeling Biological Systems: Principles and
Biochemical and cellular Systems
Applications, 2nd ed. Chapman & Hall, 2005. Fell DA. Understanding the Control of Metabolism. Portland
Keener J & Sneyd J. Mathematical Physiology. I: Cellular Physiology, Press, 1997.
2nd ed. Springer, 2009. Heinrich R & Schuster S. The Regulation of Cellular Systems.
Keener J & Sneyd J. Mathematical Physiology. II: Systems Physiology, Chapman & Hall, 1996.
2nd ed. Springer, 2009. Jacquez JA. Compartmental Analysis in Biology and Medicine, 3rd
Klipp E, Herwig R, Kowald A, Wierling C & Lehrach H. Systems ed. BioMedware, 1996.
Biology in Practice: Concepts, Implementation and Palsson BØ. Systems Biology: Properties of Reconstructed Networks.
Application. Wiley-VCH, 2005. Cambridge University Press, 2006.
Savageau MA. Biochemical Systems Analysis: A Study of Function and Stochastic Processes and Techniques
Design in Molecular Biology. Addison-Wesley, 1976.
Segel LA. Modeling Dynamic Phenomena in Molecular and Cellular Pinsky M & Karlin S. An Introduction to Stochastic Modeling,
Biology. Cambridge University Press, 1984. 4th ed. Academic Press, 2011.
Torres NV & Voit EO. Pathway Analysis and Optimization in Ross SM. Introduction to Probability Models, 11th ed.
Metabolic Engineering. Cambridge University Press, 2002. Academic Press, 2014.
Voit EO. Computational Analysis of Biochemical Systems: A Practical Wilkinson DJ. Stochastic Modelling for Systems Biology,
Guide for Biochemists and Molecular Biologists. Cambridge 2nd ed. CRC Press, 2011.
University Press, 2000.
Parameter Estimation
5
• Understand and explain the concepts of linear and nonlinear regression
• Describe the concepts of various search algorithms
• Discuss the challenges arising in parameter estimation
• Execute ordinary and multiple linear regressions
• Estimate parameters in linearized models
• Estimate parameters for explicit nonlinear functions
• Estimate parameters for small systems of differential equations
Several earlier chapters have discussed the construction and analysis of mathe-
matical and computational models. Almost absent from these discussions was a
crucial component, namely, the determination of values for all model parameters.
Clearly, these values are very important, because they connect an abstract struc-
ture, the symbolic model, with the reality of biological data. Even the best model
will not reproduce, explain, or predict biological functionality if its parameter val-
ues are wrong. Furthermore, very few analyses can be performed with a model
when its parameter values are unknown. The reason for dedicating an entire chap-
ter to parameter estimation is that this key step of modeling is complicated. There
are some situations that are easy to solve, but if the investigated systems become
even moderately large, parameter estimation often emerges as the most severe
bottleneck of the entire modeling effort. Moreover, in spite of enormous efforts,
there is still no silver bullet solution for finding the best parameters for a model in
a straightforward and efficacious manner. Thus, we will go through this chapter,
classifying and dissecting the estimation problem into manageable tasks, present-
ing solution strategies, and discussing some of the reasons why computer
algorithms tend to fail. Even if you have no intention to execute actual parameter
estimation tasks in your future work, you should at least skim through the chapter,
in order to gain an impression of the difficulties of this task, while skipping some of
the technical details.
Generically speaking, parameter estimation is an optimization task. The goal is
to determine the parameter vector (that is, a set of numerical values, one for each
parameter of a model) that minimizes the difference between experimental data
and the model. In order to avoid cancellation between positive and negative differ-
ences, the objective function to be minimized is usually the sum of the squared
differences between data and model. This function is commonly called SSE, which
stands for sum of squared errors, and methods searching for SSE are called least-
squares methods.
136 Chapter 5: Parameter Estimation
As in other modeling contexts, the greatest divide in parameter estimation is

between linearity and nonlinearity. If a model is linear, or can be transformed so
that it is (approximately) linear, then life is good. Many methods are available, most
of which are quite effective, and there are numerous statistical tools characterizing
the quality of the solution. By contrast, as soon as a few nonlinearities are intro-
duced, the vast majority of these effective methods become inapplicable. The two
dominant solution strategies in these cases are the use of brute computational force
and the development of methods that work with reasonable efficiency at least for
some classes of problems.
The field has a second divide. First, it is sometimes possible to estimate param-
eters for each process in the system (or model) and then to merge all these local
descriptions in the hope of achieving a well-functioning model that encompasses
the interplay of all processes in an adequate fashion. This bottom-up strategy has
dominated parameter estimation for biological systems in the past, and is still of
great importance. However, it often fails, for a variety of suspected or unknown rea-
sons. For instance, kinetic data may have come from in vitro experiments that were
carried out under different conditions, or may even have been obtained from differ-
ent organisms. The result of this unfortunately typical situation is that the resulting
model, which was constructed with the best available local information, clearly
does not generate reasonable results for an all-encompassing model. The discrep-
ancy requires revisiting all assumptions, as well as the information used to param-
eterize the model. Sensitivity analysis (Chapter 4) is a useful diagnostic tool for this
step, because it identifies parameters that affect the solution most and therefore
need to be estimated most precisely. Experience shows that any successes in apply-
ing the bottom-up strategy to moderately large systems with maybe 40 or 50 param-
eters are almost always the product of an excruciatingly slow and cumbersome
effort that sometimes takes months to complete.
A modern, very appealing alternative is becoming increasingly feasible. This
alternative depends on experimentally determined time series of genomic,
proteomic, or metabolomic data. A general setting for measuring these data is as
follows. Typically, the biological system is in its normal state. At the beginning of the
experiment, the system is perturbed, and one measures the responses of several
components at many time points afterwards. Hidden in these time profiles is very
valuable information about the structure, regulation, and dynamics of the system,
and the trend toward using time series data is to be welcomed, because these data
are obtained from the same organism under the same experimental conditions. The
catch with time-series-based estimation is that the extraction of information is a dif-
ficult task that requires not only powerful computers but also ingenuity and insight
into the biological system and into the intricacies of computational algorithms
designed to execute the extraction in an effective fashion.
PARAMETER ESTIMATION FOR LINEAR SYSTEMS

5.1 Linear Regression Involving a Single Variable
In the simplest case of a single variable y depending on only one other variable x, the
relationship between the two can be plotted on an x–y graph. This scatter plot gives
an immediate impression of whether the relationship might be linear. If so, the
method of linear regression quickly produces the straight line that optimally
describes this relationship (Figure 5.1). The reason that this works so well, even on a
moderately sophisticated pocket calculator, is that it is straightforward to express the
difference Δi between each data point (xi, yi) and an imagined straight line traversing
the data set and to compute the two characteristic features of the line, namely its
slope and intercept, such that this line simultaneously minimizes all differences.
Specifically, a linear regression algorithm solves the problem sketched in Figure 5.1,
which shows data points (symbols) forming a stretched-out cloud, as well as the opti-
mized straight line going through this cloud. This regression line is mathematically
defined as the uniquely defined straight line that minimizes the sum of all squared
residuals or errors Δi. The reason to square these quantities is simply that we do not
want a positive Δi to compensate a negative Δj, and squaring is mathematically more
PARAMETER ESTIMATION FOR LINEAR SYSTEMS 137
20 Figure 5.1 Simple linear regression.

A more or less linear cloud is transected by
Δi the regression line, which is optimized in a
15 Δj
sense that SSE, the sum of squared errors Δi,
is minimized.
y 10
0
0 2.5 5.0 7.5 10.0
x
convenient than using absolute values. SSE is also called an error function, and
parameter estimation tasks are therefore equivalent with tasks of minimizing error
functions. The line fitting the data in Figure 5.1 has the representation
y = 1.899248 x + 1.628571, (5.1)
which is easily obtained from the data, for instance with Microsoft Excel•.
Linear regression is a straightforward exercise, and its result allows us to predict
the expected value of y for some x that has never been analyzed. For instance, in
Figure 5.1, we can rather reliably predict the approximate y-value that is expected
for x = 6.75. Namely, plugging x = 6.75 into (5.1) yields y = 14.4485, which fits nicely
into the picture.
Two issues require caution. First, extrapolations or predictions of y-values beyond
the observed range of x are not reliable. While we can reliably compute the expected
y-value for x = 6.75, computing y for x = 250 suggests y = 476.4406, which is a correct
computation but may or may not be a good prediction; we simply don’t know. If the
example is biological, the solution may be utterly wrong, because biological phe-
nomena usually deviate from linearity for large x values and saturate instead.
The second issue to keep in mind is that the algorithm blindly yields linear
regression lines whether or not the relationship between x and y is really linear. As
an example, the relationship in Figure 5.2 clearly appears to be nonlinear, maybe
following a trend as indicated by the green line. Nonetheless, it is easy to perform a
linear regression, which quickly yields the relationship
y = 0.707068 x + 8.151429. (5.2)
It is evident that this model (the red line in Figure 5.2) does not make much sense.
Thus, vigilance is in order. Often, simple inspection as in Figure 5.2 is sufficient. But
one should also consider assessing a linear regression result with some mathemati-
cal diagnostics [1]. For instance, one might analyze the residual errors, which should
form a normal distribution if the data really follow a linear trend. One may also
study the lengths of “runs” of subsequent data falling below or above the regression
line. For instance, in Figure 5.2, all low-valued data fall below the regression line,
almost all data in the center lie above the line, and almost all high-valued data are
20
15
y 10
Figure 5.2 Misuse of linear regression.

5
Fed with the presented data (blue), any linear
regression algorithm will return the optimal
0 linear regression line (red), even if the true
0 2.5 5.0 7.5 10.0 relationship between x and y is presumably
x nonlinear, as indicated by the green line.
below the line, which would be statistically highly unlikely if the data more or less
followed the assumed linear relationship.
5.2 Linear Regression Involving Several Variables

Linear regression can also be executed quite easily in cases of more variables
(see, for example, [1]). As the most intuitive example, suppose that some quantity
Z depends on the two variables X and Y. For instance, the risk of disease may
increase with blood pressure and body mass index. Thus, for every pair (X, Y ),
there is a value of Z. If the dependence is linear, the triplets (X, Y, Z ) are located
close to a sloping plane in three-dimensional space (Figure 5.3), and one may,
for instance, use the regress function in MATLAB• for a multiple linear regres-
sion, which means linear regression where Z is a linear function of several vari-
ables. For more than two variables, the result is difficult to visualize, but the
regress function nevertheless computes the higher-dimensional analog of a
plane, which is called a hyperplane.
An example for two independent variables is given in Figure 5.3, which shows
the data from Table 5.1 in different perspectives. Red and blue dots in (A) and (C)
correspond to data above and below the regression plane, respectively. The regres-
sion plane, computed with regress, is characterized by the function
Z = −0.0423 + 0.4344X + 1.1300Y . (5.3)
(A)
15
Y 10
(C)
25
5
10 12 14 16 18 20 20
X
Z
(B)
15
10
25 15
20
18
20 10
16
Y
14
Z 12 X
5 10
15
10 Figure 5.3 Result of a multiple linear regression analysis with the

5 Matlab function regress and the data from Table 5.1. (A) The
10
data are distributed throughout the ranges of X and Y. (B) A different
10 X perspective shows that all data points are close to a sloping plane,
Y 15
which in this perspective almost looks like a line. (C) In this third view,
15 20 red dots lie above the regression plane (beige) and blue dots below.
PARAMETER ESTIMATION FOR LINEAR SYSTEMS 139
TABLE 5.1: DATA FOR MULTIPLE LINEAR REGRESSION AND

CORRESPONDING VALUES OF THE REGRESSION PLANE
X Y Z (data) Z (regression)
17.43 8.72 16.73 17.38
10.49 7.75 13.58 13.28
16.69 7.58 15.47 15.77
17.06 10.69 19.32 19.46
13.10 6.85 13.82 13.39
18.46 10.71 21.06 20.07
17.27 12.10 20.55 21.13
16.04 10.77 19.15 19.10
14.95 9.25 16.68 16.90
17.74 7.11 16.30 15.70
17.98 8.89 17.38 17.82
13.67 8.12 14.89 15.07
15.09 10.68 19.42 18.58
10.22 8.28 13.27 13.75
11.04 5.12 10.08 10.54
11.39 12.34 19.60 18.85
11.40 10.87 17.57 17.19
10.71 12.98 19.93 19.28
10.20 10.33 15.45 16.06
12.30 5.02 10.95 10.98
16.97 14.32 22.96 23.51
17.79 14.10 22.68 23.62
13.22 11.05 18.06 18.19
10.91 14.01 20.35 20.53
19.30 5.26 14.31 14.29
15.60 12.04 21.42 20.34
18.02 13.23 22.52 22.73
10.12 10.05 15.15 15.71
19.50 10.73 20.77 20.55
16.22 9.07 17.56 17.25
It may be surprising that linear regression is applicable to some nonlinear func-

tions. This option is available if the functions permit a mathematical transformation
that makes them linear. A well-known example is an exponential function, which
becomes linear upon logarithmic transformation. An interesting case of lineariza-
tion is the Michaelis–Menten rate law (MMRL) for enzyme-catalyzed reactions (see
Chapters 2, 4, and 8). Suppose that the system of interest consists of a very short
metabolic pathway with just two reactions, as shown in Figure 5.4. The pathway E
converts an initial substrate S into a metabolite M in a reaction that is catalyzed by
an enzyme E, whose activity does not change during our experiment. The metabo-
lite is subsequently used up or degraded, and the degradation products of the reac-
tion are of no particular interest and therefore not explicitly included in the model. S M
A typical mathematical representation for the first process is a Michaelis–Menten
function, which is characterized by three quantities: the substrate concentration S,
Figure 5.4 A simple model for the
the maximally possible rate or velocity of the reaction, Vmax, which depends on the dynamics of a metabolite M. M is the
(constant) enzyme concentration, and the Michaelis constant KM, which quantifies product of a reaction, which is catalyzed
the affinity between the enzyme and its substrate (Figure 5.5). The default descrip- by enzyme E and uses substrate S. M is
tion for the degradation of the metabolite is a so-called first-order process with rate degraded, but the product of this process is
constant c, which implies that the speed of degradation is proportional to the not specified.
current metabolite concentration M (Figure 5.6). Thus, the overall change in metab- (A)
is given by the difference between production and degrada-
olite concentration, M,
tion, and, with the typical functions discussed above, it reads 2
v
= Vmax S − cM .
M (5.4)
1
KM + S
0
0 40 80 120
The model contains three parameters, Vmax, KM, and c, as well as the concentration
S
of S, which may or may not be constant. Suppose someone had executed experiments (B)
with varying levels of S and measured the production of metabolite (without degrada-
tion), thereby generating data consisting of pairs (S, v), where v = Vmax S (K M + S ). For 6
KM
clarity of illustration, let us suppose these data are error-free (Table 5.2 and Figure ———
Vmax
5.5A). Our task is to estimate optimal values for the parameters Vmax and KM. Clearly, 4
1/v
the process is nonlinear, and linear regression seems to be out of the question.
2
However, an old trick for dealing with Michaelis–Menten functions is based on the
1
observation that the function becomes linear if we plot 1/v against 1/S. This inverse ———
Vmax
0
plot of the data is shown in Figure 5.5B. Since the data in this representation follow 0 1 2
a linear function, we can use linear regression to determine the best-fitting straight 1/S
line through the data. It turns out from the action of taking the reciprocal of the
Figure 5.5 Noise-free data of a Michaelis–
Michaelis–Menten function that the slope of the regression line corresponds to KM/
Menten process. The process is part of the
Vmax, whose value here is 2.91, and that the intercept is 1/Vmax with a value of 0.45. model (5.4) and was measured in a separate
The two values allow us to compute Vmax and KM as 2.2 and 6.4, respectively. experiment. (A) Raw data. (B) Inversely
Suppose the degradation process was estimated in a separate step. Artificial data plotted data and determination of slope and
are given in Table 5.3. These data exhibit exponential decay, which perfectly intercept.
matches the formulation of the degradation term in (5.4). Thus, a logarithmic trans-
formation of M into ln M yields a straight-line decay function (see Figure 5.6B)
whose parameter values can be estimated by linear regression. The result of this
procedure is the value of c that best fits the degradation data, namely, c = 0.88.
Taken together with the estimation of Vmax and KM, we have now estimated all
parameters of the system by bottom-up analysis and can substitute them into (5.4)
to perform numerical simulations or other analyses. Validation of the estimates and
the model requires additional data, for instance in the form of responses of M to
changes in S.
(A)
TABLE 5.2: MEASUREMENTS OF REACTION SPEED 8
V VERSUS SUBSTRATE CONCENTRATION S
S V
0.5 0.16 M 4
1 0.30
1.5 0.42
0
2 0.52 0 5 10
time
4 0.85
(B)
6 1.06 3
8 1.22
10 1.34
ln M –2
15 1.54
20 1.67
25 1.75
–7
30 1.81 0 5 10
time
40 1.90
50 1.95 Figure 5.6 Degradation of M in (5.4).
75 2.03 (A) In Cartesian coordinates, the function
is exponential. (B) It becomes linear for the
100 2.07
logarithm of M.
PARAMETER ESTIMATION FOR NONLINEAR SYSTEMS 141
TABLE 5.3: TIME MEASUREMENTS OF M AND ln M FOR

ESTIMATING PARAMETERS IN (5.4)
t M ln M
0 8.000 2.079
1 3.318 1.199
2 1.376 0.319
3 0.571 −0.561
4 0.237 −1.441
5 0.098 −2.321
6 0.041 −3.201
7 0.017 −4.081
8 0.007 −4.961
9 0.003 −5.841
10 0.001 −6.721
Now suppose that no kinetic data are available, but that we have instead time- 2
course data for the system in the form of M as a function of time (Figure 5.7). In this
case, we cannot use the bottom-up approach to estimate systems features from
local information (kinetic data, linear degradation), but instead need to use a top-
down estimation that works more or less in the opposite direction by using observa- M 1
tions on the entire system (namely, time course data M(t)) to estimate all unknown
kinetic features (Vmax, KM, and c) simultaneously. This top-down estimation requires
a (nonlinear) regression procedure that uses the time course data, together with
information on S, and all at once determines values for Vmax, KM, and c that capture 0
0 5 10
the data the best. In this simple case, a decent nonlinear regression algorithm time
quickly returns the optimal values. More methods for this purpose will be discussed
later in this chapter. Figure 5.7 Time-course data for the model
One of the great features of linear regression is that it also works for higher- (5.4). The data cannot be linearized in any
dimensional problems, as we discussed before, and multiple linear regression also obvious fashion and require methods of
applies directly to linearized systems. For instance, consider a process described by nonlinear parameter estimation.
the two-variable power-law function
V = α X g Y h. (5.5)
To linearize the function, we take logarithms of both sides, with the result
lnV = ln α + g ln X + h ln Y , (5.6)
which permits the application of multiple linear regression to the new variables ln
X, ln Y, and ln V.
When using this strategy, one needs to keep in mind that transformations of
whichever type also affect the noise in the data. As a consequence, residual errors
that are normally distributed around the true nonlinear function are no longer nor-
mal in the linearized form. In many practical applications, though, this problem is
considered secondary in comparison with the simplification afforded by lineariza-
tion. If an accurate account of the error structure is important, one may first estimate
parameters by linear regression and then use them as start values for a nonlinear
regression that does not distort the errors.
PARAMETER ESTIMATION FOR NONLINEAR SYSTEMS

Parameter estimation for nonlinear systems is incomparably more complicated
than linear regression. The main reason is that there is just one linear structure but
there are infinitely many different nonlinear functions, and it is not known per se
which nonlinear function(s) might model some data sufficiently well, let alone
optimally, especially if the data are noisy. Even if a specific function has been
selected, there are no simple methods for computing the optimal parameter values
as they exist for linear systems. Moreover, the solution of a nonlinear estimation task
may not be unique. It is possible that two different parameterizations yield exactly
the same residual error or that many solutions are found, but none of them is truly
good, let alone optimal.
Because the estimation of optimal parameter values for nonlinear models is
challenging, many algorithms for optimization, and thus for parameter estimation,
have been developed over many decades. All of them work well in some situations,
but fail in others, especially for larger systems where many parameters have to be
estimated. In contrast to linear regression, it is typically impossible to compute an
explicit, optimal solution for a nonlinear estimation task. So one may ask: if there is
no explicit mathematical solution, how is it possible that a computer algorithm can
find a solution? The answer is that optimization algorithms iteratively search for
ever better solutions, and, if they succeed, the solution is very close to the best pos-
sible solution, although it is usually not precisely the truly best solution. The trick of
developing good search algorithms is therefore to guide the search in an efficient
manner toward the optimal parameter set and to abandon or drastically alter search
strategies early when they enter parameter ranges that do not show much promise.
A crucial question regarding the success of nonlinear estimation algorithms is
whether the task calls for finding the parameters of a mathematical function or of a
system of differential equations. In the former case, many methods work quite well,
and we can even use software like the Solver option with the Data Analysis group of
Excel• while the latter case is much more complicated, although the basic concepts
for both tasks are the same. We will discuss different types of methods, exemplify
some of them with the estimation of parameters in explicit functions, and toward
the end discuss parameter estimation for dynamical systems.
The currently available search algorithms may be divided into several classes.
The first consists of attempts to exhaust all possible parameter combinations and to
select the best among them. A second class consists of gradient, steepest-descent, or
hill-climbing methods. The term “hill-climbing” implies finding the maximum of a
function, while parameter-estimation searches for the minimum use steepest-
descent methods. Mathematically, these two tasks are equivalent, because finding
the maximum of a function F is the same as finding the minimum of −F. The basic
idea is simple. Imagine finding yourself in a hilly terrain and that it is very foggy so
you can only see a few feet in each direction. You are thirsty, and you imagine that
water is most likely to be found down in the valley. So, you look around (although
you can’t see very far) and check which direction is heading down. You walk in this
direction until you come to a point where a different direction leads down even
more steeply. With time, you have a good chance of getting to the bottom of the val-
ley. Gradient methods imitate this procedure. As a warning, it is easy to realize that
even in the hiking example this strategy is by no means failsafe. In fact, gradient
search algorithms tend to have problems with rough terrains and, in particular,
when they find themselves in a valley that is not as deep as some of the neighboring
valleys. As a result, the algorithms often get trapped in a local minimum, which may
be better than other points close-by, but is not as good as the desired solution of the
true (global) minimum, which is located in a different valley (Figure 5.8).
A third class of parameter estimation methods consists of evolutionary algo-
rithms, which operate in a fashion that is entirely different from the previous
methods. The concepts here are gleaned from the natural processes of fitness-based
selection, which we believe has time and again promoted the best-adapted indi-
viduals and species from among their competitors. The best-known evolutionary
method, the genetic algorithm (GA), begins with a population of maybe 50–100
parameter vectors, which consist of default values or of values supplied by the user.
Some of these values may be good and others not so. The GA solves the model with
each parameter vector and computes the residual error associated with each case.
The vectors are then ranked by their fitness (how well the model with the parameter
values of the vector matches the data), and the fitter a vector, the higher is its prob-
ability to mate. In this mating process, one part of one vector (the mother) is merged
with the complementary part of another vector (the father) to produce a new vector
(the newborn baby vector). The vectors are furthermore subject to small mutations
Figure 5.8 Gradient searches may fail in

rough terrains. The function shown here
(a Laplace function from MATLAB®) has
three minima, of which the one in the center
front (X = 22, Y = 11, Z = −6.2) is the lowest.
A gradient search initiated at the white
location will likely find this global minimum.
However, gradient searches starting at
the green or blue locations will easily get
trapped in local minima with Z = 0 and
Z = −3, respectively.
10
40
Z 0
20 Y
–10
0 10
20
30 0
40
X
that permit the evolution of the population. Many mating processes within the par-
ent population lead to a new generation of vectors, the model is again solved for
each vector, residual errors are computed, the best vectors are selected, and the
process continues to the next generation until a suitable solution is found or the
algorithm runs out of steam. Gradient methods and genetic algorithms presently
dominate the field of search methods for parameter estimation, but there are other
alternatives, which we will briefly mention later in this chapter.
5.3 Comprehensive Grid Search

The idea of exhaustive searches is really simple. One determines the possible
numerical range for each parameter (for instance, from biological knowledge of the
process), evaluates the model with very many or even all combinations of parame-
ter values within these ranges, and selects the parameter set that produces the best
result. Because the parameter combinations are usually selected in regular inter-
vals, this type of estimation is called a grid search.
As an illustration, let us return to the example of the Michaelis–Menten rate law
(MMRL) v = VmaxS/(KM + S) and pretend that our earlier linearization was not pos-
sible. Suppose we have measurements of the rate v for several substrate concentra-
tions S. We know from our understanding of the MMRL that the two parameters
Vmax and KM must be positive. Let us suppose that we also know that Vmax ≤ 10 and
KM ≤ 8. Thus, we have admissible ranges, which we subdivide into even intervals of
1. For the 80 parameter combinations, it is now straightforward to compute the
value of v for each substrate concentration S, subtract from this value the corre-
sponding experimental measurement, square the differences, and divide the sum
by the number of data points used. Thus, for S = 0.5, the experimental v = 0.16 from
Table 5.2 and the pair (Vmax, KM) = (1, 1), the first quantity of the sum is computed as
2
 1 × 0.5 
 − 0.16 . (5.7)
1 + 0.5 
Figure 5.9A shows a plot of the sum of squared errors for the given ranges of Vmax
and KM. The plot indicates two important results. First, many combinations of Vmax
and KM are clearly not contenders for the optimum, because the associated errors
are very high. Thus, a considerable amount of computation time has been wasted.
(A) (B)
50
0.30
40 0.25
0.20
residual error
residual error
30
0.15
20
0.10
10
0.05
2.8
2
0 4 0 2.4
0 5.5 Vmax
2 6 KM 5.7
4 5.9
6 6.1
Vmax 8 8 KM 6.3 2
6.5
10
Figure 5.9 Residual errors (SSEs) in a grid search. The search targeted the MMRL parameter values KM and Vmax in (5.4), using the data in Table 5.2.
(A) Coarse search. (B) Refined search within smaller parameter ranges.
Second , it is difficult to discern where exactly the parameter combination with the
minimal error is located. It looks to be somewhere close to Vmax between 2 and 3 and
KM around 6. We can use this information to refine the ranges for Vmax and KM and
repeat the grid search with smaller intervals. Figure 5.9B indeed gives a more
focused view and identifies the accurate solution Vmax = 2.2 and KM = 6.4. This solu-
tion is accurate, because the true values happened to be exactly on the grid we used.
If the true values had been Vmax = 2.01234567, KM = 6.54321098, we would at best
have found an approximate, relatively close solution.
This simple example already exhibits the main problems with grid searches.
First, one has to know admissible ranges for all parameters; otherwise one might
miss the optimal solution. Second, the method is not likely to produce very precise
results, unless one iterates the search many times with smaller and smaller inter-
vals. Third, the method wastes a lot of computation time for parameter combina-
tions that are clearly irrelevant (in our case, high Vmax and low KM). Fourth, and
maybe most important, the number of combinations grows very rapidly when many
unknown parameters are involved. For instance, suppose the model has eight
parameters and we begin by trying 10 values for each parameter. Then the number
of combinations is already 108, which is 100 million. For more parameters and more
values to be tried, these numbers become astronomical. The field of experimental
design in statistics has developed possible means of taming this combinatorial
explosion [1, 2]. Arguably most popular among them is Latin hypercube sampling.
The adjective Latin here comes from the Latin square, which, for size 5 for instance,
contains the numbers 1, 2, 3, 4, and 5 exactly once in each row and each column of
a 5 × 5 grid. The term hypercube refers to the analog of a regular cube in more than
three dimensions. The key idea of Latin square sampling is that one analyzes only
enough points to cover each row and each column of a grid exactly once. In a higher-
dimensional space, one searches a hypercube in an analogous fashion.
While grid searches are seldom effective, the attraction of obtaining a global
optimum, at least in an approximate sense, has triggered research into improving
grid searches without being overwhelmed by combinatorial explosion. As an exam-
ple, Donahue and collaborators [3] developed methods that search a multidimen-
sional grid more intensively in the most promising domains of the parameter space
and only coarsely in domains that are unlikely to contain the optimal solution.
Branch-and-bound methods are significant improvements over grid searches,
because they ideally discard large numbers of inferior solutions in each step [4, 5].
Branch-and-bound methods do this with two tools. The first is a splitting or branch-
ing procedure that divides the set of candidate solutions into two non-overlapping
“partitions” A and B, so that all solutions are accounted for and each one is either in
A or B but not both. Each solution is characterized by a score for the quantity that is
being optimized, which in our case is SSE. The second tool is the estimation of upper
and lower bounds for SSE. A lower bound, say for partition A, is a number that is as
small as or smaller than the SSE of any candidate solution (fitness of a parameter vec-
tor) in A, and the analogous definition holds for upper bounds. The key concept of
the branch-and-bound method is that if the lower bound of partition A is greater
than the upper bound of partition B, then the entire partition A may be discarded,
because no solution in A can possibly have a lower SSE (and thus a higher fitness)
than any of the solutions in B. The trick is therefore to split the candidate solution sets
successively and effectively into partitions and to compute tight lower and upper
bounds. In the end, branch-and-bound methods find globally optimal solutions, but
their implementation is difficult and they require considerable computational effort.
5.4 Nonlinear Regression

Most nonlinear functions cannot be transformed into linear functions, and they are
too complicated for exhaustive explorations of their parameter space. Moreover, a
direct computation of optimal parameter values, which is the hallmark of linear
regression, is no longer possible, and entirely different approaches must be used.
The most prominent among these are iterative methods that search in a guided or
random fashion for good parameter vectors (see, for example, [6]). Because the
parameter space that encompasses all parameter vectors is usually of high dimen-
sion and is difficult to fathom, the algorithms initiate the search, assess the quality
of fit with the current parameter values, search again, assess the quality of fit again,
and thereby operate in an iterative mode that cycles between searching and assess-
ing thousands of times. The algorithm starts with a default or user-provided param-
eter vector (sometimes collectively called a guesstimate), which may be quite good
or rather inadequate, solves the model with this parameter vector, and computes
the residual error. In the next step, the algorithm evaluates the model for a small
number of other parameter vectors in a close neighborhood of the start vector—
remember the search for water in a hilly, foggy terrain (see Figure 5.8)? The algo-
rithm then uses this information to determine in which direction the error decreases
the most. Pointing into this direction, the algorithm selects a new trial parameter
vector at some distance from the original vector. This procedure is repeated thou-
sands of times until, ideally, the algorithm converges to a parameter vector that
yields a lower SSE than all vectors in its neighborhood as well as all earlier vectors.
Understanding these concepts is more important than technical details, because it
reveals why search algorithms sometimes do not succeed and because the technical
aspects have been implemented in many software packages.
The iterative search strategy is most easily demonstrated in the case of a single
parameter. Suppose the error function R has the shape shown in Figure 5.10. The
plot shows on the horizontal axis the value of the parameter p and on the vertical
axis the residual error R, which is a squared quantity and therefore not negative. Our
task is to find the position along the horizontal axis (that is, the optimal parameter
p) where R is closest to zero, because this position minimizes the error.
Suppose we are searching for the minimum inside the interval [0, +8] and our
initial guesstimate is p = 7. The algorithm enters p = 7 into the model and evaluates
the residual error R based on experimental data; in the figure, R = 7.2 (the circled 1).
The algorithm now evaluates the model with two values of p that are a little smaller or
a little greater than 7, respectively. From this information, the algorithm concludes
Figure 5.10 Fictitious error landscape

10
for the search for the optimal value(s)
z of the parameter p. The residual error
1 associated with each value of p is denoted
R 5
by R. Numbers in circles indicate the starting
2 points for different iterations, and arrows
4 show each search direction, with the length
3
of each arrow representing speed. If the
0 search is started at position z, the algorithm
0 2 4 6 8 may get stuck in a local minimum close to
p p = 1.9.
that lowering the guesstimate has a better chance of reducing R than increasing it.
It also estimates the slope of the error function and uses it to determine by how much
the original estimate should be lowered; how this determination is accomplished is
not so important here. In our example, the algorithm determines the next candidate
point as p = 6 and computes R at this point as about 2.6 (the circled 2). The slope at
this point still suggests lower values for p, and because the same direction is taken
again, the speed is increased and the next candidate is about p = 4.5 (the circled 3).
Now the slope points into the opposite direction, which suggests raising the value
of p. The algorithm does this and reaches the point indicated by the circled 4. The
process continues in this fashion until the algorithm detects that the slope of the
error function is essentially zero, which means that the error can no longer be low-
ered by moving p a little bit to the left or right. The algorithm terminates and the
result is the approximate minimum. It should be kept in mind that the algorithm
does not know where the true minimum is. Thus, it has to estimate how far to go at
every step and may overshoot or undershoot the target for quite some while.
Figure 5.10 also indicates a potential problem with this type of algorithm. Imag-
ine the initial guesstimate is p = z. Using the same rationale as above, it is easy to see
that the algorithm might converge to a value close to p = 1.9 with a residual error of
about R = 3.75. Clearly, this is not the optimal value, but it satisfies the criterion that
the slope is zero and that other parameter choices close to p = 1.9 are even worse.
Exactly the same principles apply to estimation tasks involving two or more
parameters. In higher dimensions, we humans have a hard time visualizing the situ-
ation, but iterative search algorithms function in exactly the same fashion as
discussed above. For two parameters, we can still conceptualize the error function
geographically: it is a surface that shows the residual error for each (x, y), where x and
y are the two parameters with unknown values (see Figure 5.9 for a simple example).
As in Figure 5.8, this error landscape potentially has many hills and valleys, as well as
many mountain ridges and craters, and it is easy to imagine how a search algorithm
might end up in a local minimum that is not, however, the true lowest point. This
situation of being trapped in a local minimum is unfortunately quite common. In
contrast to our two figures (Figures 5.8 and 5.10), where the mistake is easy to spot, a
local minimum is difficult to diagnose in more complicated cases, where for instance
six or eight parameters are to be optimized. One obvious, though not failsafe, solu-
tion is to start the algorithm many times with different guesstimates.
5.5 Genetic Algorithms

Using nonlinear regression for parameter estimation, one is often unsure whether
the solution is truly the optimum or whether the algorithm might have converged to
a local minimum (see Figures 5.8 and 5.10). Since this issue is directly tied to the
core principles of iterative search algorithms, it is unlikely that small changes in
implementation would solve the problem. Instead, an entirely different class of
global search heuristics has been developed over the years. Global refers to their
ability to avoid getting stuck in local minima, as nonlinear regression algorithms
tend to do. Heuristics means that it is often impossible to prove with mathematical
rigor that these algorithms will indeed succeed.
The best known global search heuristics are genetic algorithms (see, for exam-
ple, [7, 8]). They fall into the larger category of evolutionary algorithms, where the
attribute of evolution is due to their inspiration by natural selection of the fittest
members in a population [9]. Genetic algorithms often find good approximate solu-
tions, but they tend to have difficulties determining very precise solutions. Because
this situation is opposite to that in nonlinear regression, it is sometime useful to start
a parameter search with a genetic algorithm and to refine the solution with a subse-
quent nonlinear regression (we will discuss an example of this type later).
Just like in natural evolution, genetic algorithms are based on individuals with
their specific fitness, and this fitness corresponds to their ability to mate and pro-
duce offspring. Furthermore, the evolution process is subject to mutations. In the
specific application of genetic algorithms to parameter estimation, each individual
is a parameter vector. In a simplified description, such a vector might look like the
diagram in Figure 5.11: it lines up the values of all parameters that need to be
p1 p2 p3 p4 p5 p6 Figure 5.11 A typical individual in the

application of a genetic algorithm to
parameter estimation. The parameters are
0.24356 71.8232 2.02110 16.2211 0.00443 31.1433
lined up either as floating point variables,
as they are shown here, in binary code (see
Figure 5.12), or in some other mathematical
representation.
estimated. In most cases, the numbers are actually converted into binary code, but
other encodings, such as floating point representations, have also been explored.
The starting population contains many such individuals (maybe 20, 50, or 100),
which are defined through a random assignment of parameter values that collec-
tively represent the admissible parameter ranges or are “seeded” based on prior
knowledge of the problem. Each individual corresponds to a solution of the model,
because entering the parameter values into the symbolic model allows us to com-
pute any desired features of this model. In general, the feature of interest could be
anything that can be quantitatively characterized with the model, such as the steady
state, how fast the model reaches a point of interest, how much the transients of the
model deviate from the normal state, or how fast the system oscillates. In the case of
parameter estimation, the feature of prime interest is the fit to a set of experimental
data, and thus SSE. For each individual in a given population, one simulates the
model with its parameters and computes the residual error between the data and
the corresponding values of the model. This error is used as the measure of fitness:
the smaller the error, the fitter the individual. Now comes the time of reckoning. Out
of the 20, 50, or 100 individuals in the population, a small number of individuals
(maybe 20%) are permitted to mate. They are randomly selected, but the probability
of being chosen is proportional to each individual’s fitness and thus the SSE. Indi-
viduals that were not chosen are discarded. In this scheme, fit individuals have a
higher chance of mating, but even unfit individuals may get lucky. Interestingly, the
inclusion of a small number of less-fit individuals retains higher genetic diversity
and to some degree prevents early termination at a local minimum, which corre-
sponds to a suboptimal parameter vector.
The mating process occurs as follows. Two individuals (parameter vectors) are
lined up next to each other, and offspring is produced by using the first few parame-
ters (or parts of the parameter vector) from the first individual and the rest from the
second individual, as sketched in Figure 5.12. Any positions within a parent vector
may also be mutated; that is, from 1 to 0 or from 0 to 1 in the case of binary coding.
Experience has shown that mutations are very important, but that their rate has to be
rather low, lest the algorithm wander through the parameter space without ever con-
verging. The result of the mating and mutation process is again an individual in the
form of a new parameter vector. The parameter values from this vector are entered in
the model and the fitness is computed. The mating process is usually organized such
that the population size remains constant. Exceptionally fit individuals may mate sev-
eral times, and most algorithms automatically move the fittest two or three individu-
als of the parent generation unchanged into the offspring generation. The process
terminates for one of several reasons. The following are the most typical among these:
• A solution has been obtained that satisfies preset quality criteria. Success!
• Solutions with the highest fitness have reached a plateau where they do not
change for many iterations from one generation to the next. This may be good or
bad news.
• The number of generations has reached a preset limit. It could be started again
with the last population of vectors.
individual 1 01010011 10010010 11001110 00110011 01001111 11100010 Figure 5.12 Generation of offspring in
a generic genetic algorithm. Individuals
individual 2 00111000 01110011 11100110 10000111 01001110 00011100 (parameter vectors) are coded here as binary
strings. The new offspring contains a part of
individual 1 and the complementary part of
offspring 01010011 10010011 11100010 10000111 01001110 01011100 individual 2. The newly combined offspring
is furthermore subject to mutations before
the next iteration (arrows).
• The algorithm encounters some numerical problem, for instance, while integrat-
ing differential equations.
In many cases, genetic algorithms work quite well, and, because they are
extremely flexible, for instance in terms of fitness criteria, they have become very
popular and are implemented in many software packages. Nonetheless, they are
certainly no panacea. In some cases, they never converge to a stable population; in
other cases, they may not find an acceptable solution; and in almost all cases, these
algorithms are not particularly fast. If problems arise, the first attempt at resolution
is to reset the algorithmic parameters, such as the population size, the mating prob-
abilities of fit or less-fit individuals, the mutation rate, the number of fittest individu-
als that directly survive into the next generation, bounds on parameters, and criteria
for termination. A specific example will be presented later in this chapter.
5.6 Other Stochastic Algorithms

In addition to genetic algorithms, several other classes of methods are available for
global optimization and parameter estimation. They include evolutionary program-
ming, for instance implemented as ant colony and particle swarm optimization, as
well as simulated annealing and the branch-and-bound methods that we discussed
before. In this section, we will discuss the concepts of these methods without giving
a lot of detail.
As in genetic algorithms, evolutionary programming is a machine learning
technique based on evolving populations of solutions. The main mechanisms for
improving fitness are mutation and self-adaptation to adjust parameters, and varia-
tions may include operations such as combining information from more than two
parents. One intriguing method within the class of evolutionary programming is ant
colony optimization (ACO) (see, for example, [10]). ACO was inspired by the behav-
ior of ants seeking a food source while wandering about their colony. The original
exploration is more or less random, but once an ant finds food, it returns to the nest
in a more or less direct fashion. Along the way, the ant leaves a pheromone trail that
can be perceived by other ants. This type of chemical communication allows other
ants preferably to follow the established pheromone trails rather than to follow ran-
dom paths. If a path indeed leads to food, more and more ants follow it and release
pheromone, thereby reinforcing the successful paths. However, if the path does not
lead to success or if it is unduly long, it is less traveled, the pheromone evaporates,
and its attraction dissipates. Over time, the short, efficient paths incur more traffic
and thus enjoy higher concentrations of pheromone. ACO attempts to mimic this
dynamic, adaptive mechanism of optimizing the path to a desired destination. Thus,
parameters that show up time and again in good solutions are rewarded by increas-
ing their probability of being selected in future generations, while other infrequent
parameters become less and less prevalent in the population. Algorithmic issues
related to ACO are discussed in [11].
Similar ideas also form the basis for particle swarm optimization (PSO), which
was inspired by the flight patterns of birds (see, for example, [12]). Like genetic
algorithms and ACO, PSO is a population-based stochastic search procedure,
where each particle in the swarm represents a candidate solution to the optimiza-
tion or estimation problem. The flight of each particle is influenced by the best
positions of the particle and the swarm within a high-dimensional space, and the
fitness is measured according to each particle’s own experience, which is aug-
mented by communication with its neighboring particles. Through this communi-
cation within the swarm, all particles share some of the information regarding
high-fitness solutions. Specifically, as soon as a particle detects a promising solu-
tion, the swarm explores the vicinity of this solution further. As a result, the parti-
cles generally tend toward optimal positions, while searching a wide area around
them. Expressed differently, PSO is quite efficient because it combines local and
global search methods. PSO methods are reviewed in [13].
Simulated annealing (SA) is a more established global optimization technique
that grew out of a Monte Carlo randomization method from the 1950s [14, 15] for
generating sample states of a thermodynamic system. The inspiration for this
method came from the technique of annealing in metallurgy. This technique uses
heating and controlled cooling of a material to reduce defects. The heat provides
energy that allows atoms to leave their current positions, which are states of locally
minimal energy, and to reach states of higher energy. During cooling, the atoms
have a better chance than before of finding lower-energy states. Therefore, the sys-
tem eventually becomes more ordered and approaches a frozen state of minimal
energy in a semi-guided random process. The analogy of this heating–cooling pro-
cess in SA is that a population of solutions traverses the search space randomly,
thereby allowing the current solutions to change a bit. The random mutations of
each individual solution are tested (that is, the model is computed with the given
parameter values), and mutants that increase fitness always replace the original
solutions. Mutations with lower fitness are not immediately discarded but are
accepted probabilistically, based on the difference in fitness and a decreasing tem-
perature parameter. If this temperature is high, the current solutions may change
almost randomly. This feature prevents the method from getting stuck in local min-
ima. During cooling, the temperature gradually decreases and toward the end the
solutions can only vary in a close neighborhood and finally assume local minima.
Either the former solutions are regained in the process or new solutions are found
that have a higher fitness. Because many solutions settle in their individual min-
ima, it is hoped that the global minimum is among them. The key to a successful
implementation is the choice of adequate algorithmic parameters. For instance, if
the temperature of the heated system is too low at the beginning or if cooling occurs
too rapidly, the system may not have enough time to explore sufficiently many
solutions and may become stuck in a state of locally but not globally minimal
energy. In other words, SA only finds a local minimum.
It is possible to combine several of these methods. Indeed, it is often useful to
start a search with a genetic algorithm in order to obtain one or more coarse solu-
tions, which are subsequently refined with a nonlinear regression. It is also possible
to use SA within a standard genetic algorithm by starting with a relatively high muta-
tion rate, which slowly decreases over time. Similarly, Li et al. [16] combined SA with
ACO. The advantage of this combination was again that ACO is a global search
method, while SA has features characteristic of probabilistic hill climbing.
5.7 Typical Challenges

No matter which search algorithm is used to determine optimal parameter values
for a nonlinear model, some challenges come up time and time again [17]. One
ubiquitous challenge in computational systems biology is noise in the data.
Although data don’t usually come in decibels, the expression is used to describe
deviations from an expected (usually rather smooth) trend line, which may be due
to variability among beakers, Petri dishes, cells, or organisms, and/or to inaccura-
cies in measurements. It does not require much imagination to infer from plots such
as Figure 5.13 that noise may become a challenge for parameter estimation, espe-
cially if it is so large that it obscures or hides the true trend line.
Another challenge, which occurs more often than one might think and which is
actually quite hideous, is the fact that the estimation task may have many solutions
with the same SSE. These solutions may all be connected in the parameter space or
they may be distinctly separated. An intuitive example arises when two parameters
always appear in the model in the same formation or combination, such as p1p2.
Clearly, for every p1 that some algorithm might determine incorrectly, there is a p2
that perfectly makes up for the error (with the exception of p1 = 0). It is useful to
express the situation in a slightly different fashion. Suppose that, in the true solu-
tion, p1 = 3 and p1 = 4, so that p1p2 = 12. Then, any solution that satisfies p2 = 12/p1
has exactly the same SSE as the true solution.
In the case of p1p2, the culprit is easy to identify, but in other cases it might not be
so obvious. For instance, consider the power-law function
f = pX aY b
(5.8)
with three parameters. Suppose data are available for X, Y, and f, as shown in Table
5.4. For clarity of illustration, these data are assumed to be noise-free. Initiating a
(A) (B) (C)

10 10 10
size v
size v
size v
5 5 5
0 0 0
0 2.5 5.0 0 2.5 5.0 0 2.5 5.0
time time time
Figure 5.13 Too much noise may hide the true functional relationship and create problems with parameter estimation. (A) The true trend (blue) is
clearly perceivable in spite of moderate noise. (B) In the presence of stronger noise, the true trend is more ambiguous; in fact, a straight line could possibly
fit the data. (C) Even more noise makes it impossible to identify the true trend, and linear as well as drastically different nonlinear trends could be most
appropriate.
search algorithm with different guesstimates for p, a, and b may lead to distinctly
different solutions that are all optimal, which here means that the residual error is 0.
The differences in solutions are not the consequence of noise, because the data in
this constructed example are noise-free. So, what’s the problem? In this case, the
redundancy is actually due to the fact that the data for X and Y in Table 5.4 happen
to be related by a power-law function, that is,
Y =α Xγ (5.9)
TABLE 5.4: DATA TO BE MODELED WITH THE POWER-LAW

FUNCTION (5.8)
X Y f
1 1.75 2.07
2 3.05 4.03
3 4.21 5.95
4 5.31 7.84
5 6.34 9.71
6 7.34 11.57
7 8.30 13.41
8 9.24 15.25
9 10.15 17.07
10 11.04 18.89
11 11.92 20.70
12 12.78 22.50
13 13.62 24.30
14 14.45 26.09
15 15.27 27.88
16 16.08 29.66
17 16.88 31.44
18 17.67 33.21
19 18.45 34.98
20 19.22 36.75
with parameters α = 1.75 and γ = 0.8 (Figure 5.14). Because of this relationship, Yb 20
may be substituted equivalently with (α X γ)b, no matter what the value of b. As a
consequence, if the parameters of f in (5.8) are to be estimated and the algorithm
finds some incorrect value for b, the parameters p and a are able to compensate 15
perfectly. As a numerical example, suppose the true parameters of f in (5.8) are
p = 2.45, a = 1.2, b = −0.3. By applying the transformation Y = 1.75X 0.8, according to
(5.9), we obtain Y 10
f = 2.45 X 1.2Y −0.3 5

= 2.45 X 1.2YY −1.3
(5.10)
= 2.45 X 1.2 (1.75 X 0.8 )Y −1.3 0
0 5 10 15 20
= 4.2875 X 2Y −1.3 . X
Thus, we have two distinct solutions to the parameter estimation task, namely the Figure 5.14 Dependences between
parameter vectors (2.45, 1.2, −0.3) and (4.2875, 2, −1.3), which produce exactly the variables may lead to redundant
same values of f. Moreover, these are not the only solutions. Instead of replacing Y parameter estimates. The error-free (X, Y )
with 1.75X 0.8, we could replace any power of Y with the corresponding representa- data of Table 5.4 are related to each other by
the power-law function (5.9) with α = 1.75
tion of 1.75X 0.8, and the resulting f would be characterized differently, but generate and γ = 0.8.
exactly the same fit to the data.
In the cases discussed so far, all equivalent solutions taken together formed
straight or curved lines within the parameter space. A distinctly different situation
is possible where two or more solutions are equivalent, and yet the solutions in
between are not. The easiest demonstration of this situation is graphical: imagine
that the error function R in Figure 5.15 depends on only one parameter p. Clearly,
there are two solutions where the residual error is minimal (with R = 2), namely
for p ≈ ±3.16, and all other values of the parameter p give solutions with higher
errors. It is easy to imagine similar situations for error functions that depend on
several parameters.
It is possible that estimating parameters for various subsets of the same data may
lead to major differences in the optimized parameter values, an observation called
sloppiness. Furthermore, and probably surprising, just computing averages of the
values from different estimations does not necessarily provide a good overall fit. As
an example, consider the task of fitting the function
w(t ) = (p1 − p2e − p3t )p4 (5.11)
to experimental data. This function, which has been used to describe the growth of
animals [18], has four parameters that describe its sigmoid shape over time. Sup-
pose two datasets were measured, and we fit them separately. The results (Figure
5.16A, B) are both good, and the optimized parameter values are given in Table 5.5.
30
Next we average the parameter values obtained from the estimations for the two
datasets. Surprisingly, the fit is rather bad (Figure 5.16C). Indeed, much better
parameter values are obtained with the earlier fits for only one dataset or from
simultaneously fitting the two datasets (Figure 5.16D). How is that possible? At a
coarse level, the answer is again that the parameters in many models are not inde-
pendent of each other. Thus, if one parameter value is artificially changed, it is to R 15
some degree possible to compensate the resulting error by changing a second
parameter value, as we encountered with the extreme case of parameters that
always form a product p1 p2. In more general terms, the acceptable parameter vec-
tors in these cases are actually part of a curved (possibly high-dimensional) cloud
whose shape is governed by the dependences among the parameters, and all other
points within this cloud are similarly acceptable parameter vectors [19, 20]. Averag- 0
ing of results does not preserve the relationships within this cloud, and the short –5 0 5
p
take-home message is that it is often not a good idea to average parameter values
one by one from two or more fits (Figure 5.17). Figure 5.15 Two optimal solutions may be
Fitting two or more datasets separately does have some advantages. For instance, isolated. Here, the error function R assumes
it indicates how stable an optimized parameter vector is, or how much it is affected the same minimum value of 2 for p ≈ −3.16
by natural variability and experimental uncertainty. If the parameter values change and p ≈ +3.16. The error is higher for all other
drastically from one dataset to the next, none of the parameter vectors is likely to be values of p.
(A) (B) Figure 5.16 Data fits for the model (5.11)
3.0 3.0 with parameter values listed in Table 5.5.
(A) Fit v1 for dataset 1. (B) Fit v2 for dataset 2.
(C) Fit v3 with averaged parameter values.
(D) Simultaneous fit v4 to datasets 1 and 2.
size v1
size v2
1.5 1.5
0 0
0 2.5 5.0 0 2.5 5.0
(C) (D)
3.0 3.0
v3
v2 v2
v1 v1
size v
size v
1.5 1.5
v4
0 0
0 2.5 5.0 0 2.5 5.0
time time
reliable for predictions of untested scenarios. If many datasets are available, it may
be beneficial to use some of the data as a training set for estimating parameters,
while retaining the remaining, unused data for validation. In other words, one
obtains parameters with some data and checks with the validation data how reliable
the estimates from the training phase are.
If only one dataset is available, it is still to some degree possible to explore the
reliability of the estimates. One option is sensitivity analysis (Chapter 4), while the p2U
other is the construction of clouds of candidate solutions. The basic concept is a p2
resampling scheme such as the bootstrap or jackknife method [21–23]. In boot- p2A
strapping, one draws a random sample from the data hundreds of times, estimates
the parameters for each sample, and collects all results, which again form clouds of p2L
acceptable parameter vectors. In the jackknifing technique, one data point at a time
is randomly eliminated and the parameter estimation is executed with the remain- p1L p1A p1U
ing points. Repeating the estimation many times, while leaving out a different point p1
every time, leads to different solutions, which provide an indication of how much
Figure 5.17 Averaging suitable parameter
values from different estimations is
not necessarily valid. Suppose suitable
TABLE 5.5: PARAMETER VALUES FOR DATA FITS WITH THE MODEL (5.11) IN parameter combinations ( p1, p2) (with low
FIGURE 5.16 SSEs) are located within the pale orange
region, while the orange and red regions
Dataset/ Figure are acceptable and barely satisfactory,
p1 p2 p3 p4
parameter set part respectively. All solutions outside the red
1 A 1.2 0.8 1 3 region are unacceptable owing to high SSEs.
The optimal solution is indicated by the green
2 B 2.5 2.45 0.35 0.8 dot. Even if the solutions at the two blue
Average of 1 and 2 C 1.85 1.625 0.675 1.9 dots, ( p1L , p2L ) and ( p1U , p2U ), are both suitable,
averaging their parameter values ( p1A , p2A )
Simultaneous fit D 1.68 1.6 0.6 1.2
leads to an unacceptable solution (purple).
PARAMETER ESTIMATION FOR SYSTEMS OF DIFFERENTIAL EquATIONS 153
the acceptable parameter vectors may vary. The obvious caveat with these tech-
niques is that all information is literally based on a single dataset and therefore does
not account for natural variability that comes from different experiments.
PARAMETER ESTIMATION FOR SYSTEMS

OF DIFFERENTIAL EquATIONS
Essentially all principles and techniques of parameter estimation discussed so far
are applicable to explicit functions as well as to dynamical systems consisting of sets
of differential equations. However, there are two main reasons why the estimation of
dynamical systems is much harder. First, systems of differential equations usually
contain many more parameters than individual functions, and we have already had
a glimpse into the challenges that come with larger numbers of parameters and the
potential for a combinatorial explosion of possible sets of parameters. We saw this
issue most clearly for grid searches, but also for evolutionary methods like genetic
algorithms that attempt to cover an exhaustive representation of the parameter
space. The second issue making the estimation of dynamic systems difficult is that
the necessary experimental data now consist of sets of data measured at a series of
time points and these are to be compared with the solutions of the differential equa-
tions, which therefore require numerical integration of the equations during every
parameter updating step. Typical data of this type consist of the expression of genes
or the concentrations of metabolites, which are measured every so many seconds,
minutes, or hours, following some stimulus. A specific example is a population of
bacteria that have been starving for some while. At the beginning of the experiment (A)
(t = 0), a substrate such as glucose is added to the medium, and the measurements 100
N
consist of concentrations of glycolytic metabolites every 30 seconds [24].
To estimate parameters from time-series data, one starts again with some
parameter vector (or a population of vectors), but, before a comparison between
model and data is possible, one must solve the system of differential equations. It
does not sound like much effort to perform this solution step, but computation-time N,S 50
studies have shown that this step may use more than 95% of the entire time needed
to optimize a parameter vector [25]. To make matters worse, unreasonable param-
eter vectors, upon which a search algorithm may easily stumble once in a while,
sometimes cause the numerical integration of differential equations to become so S
slow that the algorithm eventually runs out of time before any decent parameter 0
0 15 30
vector has been obtained. These computational issues, added to the combinatorial time
explosion of large numbers of parameters to be estimated and other issues men-
tioned previously, render the estimation of dynamical systems a very challenging (B)
Si
topic that continues to be the object of much attention within the community of Ni+1
computational biologists [26–30]. Since no general method is presently capable of
solving all estimation issues related to dynamical systems, several diverse shortcuts Ni s+i
and means of simplification and amelioration have been proposed and are dis-
cussed in the following. N,S
s–i
A most effective strategy to counteract the computational challenges of estimat- Ni–1
ing parameters in dynamical systems is the avoidance of integrating the differential
equations. The rationale is the following. The derivative on the left-hand side of a
differential equation is really the slope of the time course of the variable at a given
point. Thus, if we can estimate this slope, we can circumvent integrating the differen-
0
tial equation. It may be best to demonstrate the approach with a simple example. 0 ti–1 ti ti+1
Suppose we are studying the growth of a bacterial population, which begins with time
2 units (maybe 2 million cells) and grows to 100 units. As a mathematical descrip-
tion, we assume the famous logistic growth function, which was proposed over 150 Figure 5.18 Estimation of parameters
in (5.12). (A) Growth curve of a bacterial
years ago [31] and has the mathematical formulation
population with size N. At each time point,
the slope of the growth curve is estimated
N = aN − bN 2 , (5.12) and plotted as S. (B) The slope estimation
may be performed by averaging the
quantities si− and si+, as explained in the text.
which contains two parameters, a and b (see also Chapters 4 and 10). For ease of dis- The result is the estimated slope Si of the
cussion, we pretend to have noise-free data N, which are shown in Figure 5.18A. green line at the point (ti, Ni) of the growth
Because the data are so clean, it is easy to estimate, at each time point, the slope S of curve.
the growth curve; the slopes are also shown in the figure. The slope estimation in this
case may be done in the following fashion, which is sometimes called the three-point
method (Figure 5.18B). For the two subsequent data points Ni−1 and Ni, which are
measured at time points ti−1 and ti, compute the difference Ni − Ni−1 and divide it by
ti − ti−1; denote the result by si−. Do the analogous computation for Ni and Ni+1, which
are measured at ti and ti+1, and call the result si+ . The average of si− and si+ is an estimate
of the slope Si at (ti, Ni). This method does not allow the computation of slopes at the
first and last data points, but one can often do with two fewer points. Much more
sophisticated and accurate methods are available for smoothing time courses and
estimating slopes even in cases where the data are noisy [32, 33].
The slope estimation leads to an augmented dataset, which consists of three vari-
ables, namely ti, Ni, and Si (Table 5.6). Substituting the estimated slopes Si for N (ti )
transforms the estimation task with one differential equation into as many decou-
pled algebraic equations as there are data points (possibly without the first and last
points). The original task is now computationally much easier, because the algorithm
no longer needs to solve the differential equation numerically. In the new system, the
ith algebraic equation looks like
Si = aN i − bN i2 (5.13)
in symbolic form. For the parameter estimation, the values for N and S are entered
into this set of equations. As an illustration, the first two equations of this set, using
S from the three-point method, are
1.80 = a 4.34 − b 4.342 ,

(5.14)
3.51 = a 9.18 − b9.182.
Using a gradient method, such as nlinfit in MATLAB•, we obtain parameter val-
ues very quickly as a = 0.3899 and b = 0.0039, which are quite close to the true val-
ues a = 0.4 and b = 0.004, which we know from constructing the example.
Importantly, the method of estimating slopes and converting systems of differ-
ential equations into algebraic equations also works for systems of more than one
TABLE 5.6: DATASET USED FOR ESTIMATING THE PARAMETERS IN (5.12)*

t N S (true) S (three-point)
0 2 0.78 †
2 4.34 1.66 1.80
4 9.18 3.33 3.51
6 18.36 6.00 6.05
8 33.36 8.89 8.58
10 52.70 9.97 9.47
12 71.26 8.19 7.99
14 84.66 5.19 5.30
16 92.47 2.78 2.95
18 96.47 1.36 1.48
20 98.38 0.64 0.70
22 99.27 0.29 0.32
24 99.67 0.13 0.15
26 99.85 0.06 0.07
28 99.93 0.03 0.03
30 99.97 0.01 †
* The data consist of the numbers of bacteria (in units of millions), the true slopes, which are usually
not known, and the slopes estimated with the three-point method.
† Slopes at times t = 0 and t = 30 cannot be obtained with this method.
variable. Let us illustrate the method for a dynamic system with two variables. For
a change, this time the example comes from ecology. Suppose two species N1 and
N2 are living in the same habitat. N1 is an herbivore that once in a while is eaten by
a predatory carnivore from population N2. One standard description is a Lotka–
Volterra model (see Chapters 4 and 10) of the form
N 1 = N 1 ( a1 − b1N 1 − b2 N 2 ) ,
(5.15)
N 2 = N 2 ( a2 N 1 + a3 − b3 N 2 ) .
Following standard techniques for setting up the model, N1 grows exponentially

with rate a1 and has a natural death rate of b1. It is typical to represent the death term
by the square of N1, which is sometimes interpreted as crowding within the popula-
tion. The population is also subject to decimation by population N2 with rate b2. This
process depends on the number of encounters between the two species and is
therefore represented by the product of N1 and N2. N2 feeds on N1 with a rate a2, and
this process augments its growth from other food sources, which is assumed to
occur with rate a3. The predators die with the rate b3. Note that a2 and b2 typically
have different values, because the growth benefit to the predator is seldom exactly
equivalent to the effect on the prey.
Suppose we have reason to believe that the model formulation in (5.15) is adequate,
but that we do not know appropriate parameter values. We do have data, which stem
from a series of measurements around some devastating event for the prey. Specifi-
cally, the populations are both at steady state, with about four times as many predators
as prey. At time t = 1, about 80% of the prey population succumb to some disease. Sub-
sequent measurements of N1 and N2 show how both populations respond to the cata-
strophic event: initially, N2 lacks food and decreases a little bit, while N1 begins to
recover. Subsequently, both N1 and N2 increase and decrease with a few damped oscil-
lations and finally return to their original steady-state values. Since this is just a demon-
stration, let us begin by assuming that we have comprehensive, error-free measure-
ments, which are given in Table 5.7. In addition to the population sizes, the table also
lists the slopes of N1 and N2, which are denoted by S1 and S2, respectively. At time t = 1,
the abrupt disease event happens, and the slopes at this point are not defined.
The estimation now proceeds by setting up two sets of algebraic equations: one
for N1 and one for N2. The first two equations of the first set are
S1 (1.1) = N 1(1.1)[a1 − b1N 1 (1.1) − b2 N 2 (1.1)] ,

S1 (1.2) = N 1 (1.2)[a1 − b1N 1 (1.2) − b2 N 2 (1.2)]. (5.16)
Substituting numerical values for S1, N1, and N2 in the first equation yields
798.41 = 63.35 ( a1 − b1 63.35 − b2 26.89) . (5.17)
Similar equations with numerical values listed in Table 5.7 are composed for time
points 1.2, . . ., 2.9. Again using nlinfit in MATLAB•, we obtain the exact param-
eters for the first equation in less than a second. Indeed, although the residual error
SSE for the initial guesstimate (2, 2, 2) is huge at about 5 × 1010, the algorithm con-
verges to the correct solution (a1 = 40, b1 = 0.008, b2 = 1.0) within only seven itera-
tions. In exactly the same fashion, the parameter values for the equation for N2 are
computed very quickly as (a2 = 0.05, a3 = 0.01, b3 = 0.18). The optimized solution
perfectly matches the noise-free data (Figure 5.19).
It should be mentioned that this specific example would have allowed a simpler
method of estimation. Namely, we could have divided all equations in (5.16) by N1,
which would have made the estimation linear [34]. This strategy is left as Exercise
5.17. It is very powerful, because linear regression provides an analytical solution and
therefore does not require search algorithms, as we discussed before. In fact, this
strategy can be used even for very large systems, such as bacterial metapopulations
[35] (see Chapter 15). However, the Lotka–Volterra system is a special case, and the
slope estimation method usually does not convert nonlinear into linear models.
TABLE 5.7: DATA FOR THE ESTIMATION OF PARAMETERS IN (5.15)*

t N1 N2 S1 S2
0 139.77 38.88 0 0
0.25 139.77 38.88 0 0
0.5 139.77 38.88 0 0
0.75 139.77 38.88 0 0
1 139.77 38.88 — —
__________________________________________________________________________________†
1 27.95 38.88 — —
1.1 63.35 26.89 798.41 −44.71
1.2 204.04 30.86 1531.64 143.72
1.3 182.09 45.96 −1350.61 38.67
1.4 107.57 40.61 −158.29 −78.03
1.5 123.83 35.97 376.76 −9.81
1.6 156.03 37.95 125.05 37.21
1.7 145.80 40.31 −214.54 1.83
1.8 132.38 39.13 −25.07 −16.21
1.9 137.55 38.25 89.86 0.12
2 143.28 38.80 7.54 7.37
2.1 140.51 39.17 −41.59 −0.61
2.2 138.19 38.90 −0.11 −3.18
2.3 139.55 38.75 18.30 0.47
2.4 140.50 38.88 −1.14 1.39
2.5 139.82 38.94 −8.07 −0.31
2.6 139.45 38.88 1.06 −0.59
2.7 139.77 38.86 3.51 0.18
2.8 139.92 38.89 −0.71 0.25
2.9 139.76 38.89 −1.51 −0.10
* The data consist of error-free measurements of the sizes of the two populations N1 and N2, as well
as slopes at each time point.
† At time t = 1, population N1 abruptly falls to about 20% of its steady-state value, for reasons not
represented in the model. At this point, slopes are not defined.
250
Of course, actual biological data contain noise, and this noise is usually amplified
when slopes are computed. Nonetheless, the slope method is often a good start,
especially if the noise is moderate. As an example, suppose noisy data are given, with
N1
values starting after the crash of population N1 (Table 5.8). Using the same tech-
number
niques of replacing the derivatives with slopes (which are also noisy now), we use 125
nlinfit to compute parameters. Again, it does not take long to estimate the para-
meters, and the results are (a1 = 24.0126, b1 = 0.0031, b2 = 0.6159) and (a2 = 0.0376,
a3 = −1.6032, b3 = 0.0933). The first observation is that these numbers are different N2
from the true values, which is to be expected because of the noise in the data and the
slopes. Some values are off by quite a bit, and it can be seen in particular that a3 is
0
actually negative rather than positive, which would change its interpretation from a 0 1.5 3.0
birth to a death process. However, it is difficult to judge how good these parameter time
values are just by looking at the numbers. The first test is therefore a plot of the slopes
as functions for N1 and N2. An example is shown in Figure 5.20, where S1 is plotted Figure 5.19 Error-free data (symbols)
corresponding to the model (5.15) and
against N1. At first, the spiral in this plot may be confusing. It is due to the fact that N1
solution with optimized parameters
increases and decreases, thereby moving back and forth along the horizontal axis. At (lines). At time t = 1, 80% of population N1
the same time, S1 moves up and down, thereby yielding the spiral pattern. The red succumb to an unexplained disease. By time
spiral shows the true dynamics (with correct parameter values), the blue line with t = 3, both populations have returned to
symbols represents the noisy data in Table 5.8, and the green line shows the their steady-state values.
TABLE 5.8: DATA FOR ASSESSING THE ROLE OF NOISE IN THE ESTIMATION
OF PARAMETERS IN (5.15)*
t N1 with noise N2 with noise S1 with noise S2 with noise
1 20 44 — —
1.1 50 25 690 −60
1.2 210 30 1310 132
1.3 195 48 −1520 30
1.4 97 37 −120 −52
1.5 120 33 510 −18
1.6 144 41 112 50
1.7 148 41 −255 0
1.8 130 44 −10 −22
1.9 136 37 70 0
2 148 35 10 10
2.1 130 36 −50 0
2.2 142 40 0 −2
2.3 140 41 27 4
2.4 136 34 0 2
2.5 137 36 −12 −1
2.6 135 38 4 0
2.7 144 35 4 0
2.8 142 41 0 0
2.9 138 38 0 0
* Noisy measurements of the sizes of the two populations N1 and N2, as well as the slopes measured
at each time point, starting at time t = 1, after population N1 had abruptly fallen to about 20% of its
steady-state value.
dynamics of the model with optimized parameter values. Visual inspection suggests
that the fit is not so bad. The real test is the use of the optimized parameter values in
the model equations (5.15). Three results are shown in Figure 5.21: in (A), noisy data
and slopes are used only for N1, while the parameter values for N2 are taken as the
true values; (B) shows the analogous situation for noisy N2 and correct N1 data; and
(C) shows the results of adding noise to both N1 and N2. The main observations are
that the overall dynamic patterns are retained, but that the timing of the curves is not
quite right. This deviation in time is a typical outcome of using the slope method. The
reason is that when derivatives are substituted with slopes, the regression does not
explicitly use time. If there is no noise, the fit is usually very good. However, the con- 2000
sequences of noise often appear in the form of some time warp. This is the cost of the
ease and speed of the slope substitution method. 1000
If biological considerations mandate that a parameter value should be positive,

S1 0
as we would suspect in the case of a3, the parameter estimation procedure can be set
up to enforce solutions to stay within admissible ranges.
–1000
As a control, we estimate parameter values from the noise-free and noisy data,
using the differential equations (5.15) directly. Again, the result is typical. Assuming
–2000
the algorithm converges to the correct solution, this solution is usually better than 0 100 200 300
the solution obtained with the slope substitution method. However, it often takes N1
much, much longer to obtain this solution, especially for larger systems, and some-
times the algorithms do not converge at all to an acceptable solution. Figure 5.20 Assessing the oscillatory
The website supporting the book (http://www.garlandscience.com/product/ behavior of the estimated system (5.15).
Shown here is the plot of the slope S1 versus
isbn/9780815345688) contains MATLAB• code for a parameter estimation pro-
the size of population N1. The red curve
gram that first uses a genetic algorithm to obtain a coarse global solution and then shows the true, noise-free relationship,
hands the results over to a gradient method to refine this solution. The important the blue-connected symbols represent
settings for the genetic algorithm are the population size (PopSize), the number of noisy data, and the green curve shows the
generations (GenSize), the number of best genes retained for the next generation relationship optimized for noisy data.
(A) (B) (C)

280
N 140 N N
0
1 2 3 1 2 3 1 2 3
time time time
Figure 5.21 Simulations of the population model with parameter values estimated from noisy data. (A) Only N1 data are noisy. (B) Only N2 data are
noisy. (C) Both N1 and N2 data are noisy. The plots start at the same perturbed value at time 1 as in Figure 5.19.
(EliteCount), a limit to the number of consecutive generations if there is no signifi-

cant improvement in fit (GenLimit), and sets of admissible lower and upper
bounds for the parameters. These settings allow a lot of flexibility, and, even in our
case, they can make quite a bit of difference. With the settings on the website, the
algorithm often actually converges to a good solution. However, because genetic
algorithms are stochastic, running the same program with the same settings and
data several times typically leads to different solutions. Five examples are given in
Table 5.9.
The initial SSE of the best solution automatically created by the GA is of the
order of 105. It decreases to somewhere between 103 and 104 at completion of the
GA, depending on the random numbers the algorithm uses during the stochastic
evolution. Under opportune conditions, the subsequent gradient method reduces
this error to between 0 and 10. Interestingly, even if the residual error at the end of
the GA phase is very similar in two runs, the gradient method sometimes con-
verges to the correct solution and sometimes to an unsatisfactory local minimum,
which corresponds to the simple shoulder curve for N1 in Figure 5.22. The com-
bined process may take several minutes to complete, which is much longer than
the time needed for the decoupled system with estimated slopes. The residual
error is almost never 0, even in cases of convergence, but a solution with SSE ≈ 7
is essentially perfect in this case (see Figure 5.22).
Depending on the upper bounds for the unknown parameters, randomly cho-
sen start values may be so far off that the integration of the differential equations
stops immediately with an error message. If the GA does get started, it often termi-
nates with a higher SSE than in the cases before. In many cases, the gradient method
terminates with a bad solution, but there are also cases where the iterative search is
TABLE 5.9: TRUE AND OPTIMIZED SOLUTIONS FOR NOISE-FREE POPULATION

DATA (SEE (5.15))
Run a1 b1 b2 a2 a3 b3 SSE
1 89.6965 0.6422 0.0000 0.4296 2.4108 1.6351 13,664.2
2 99.8746 0.0000 2.5988 0.4181 0.0000 1.5318 11,939.6
3 38.4554 0.0064 0.9601 0.0525 0.0048 0.1879 7.44405
4 38.5346 0.0065 0.9623 0.0524 0.0012 0.1874 6.99142
5 20.7977 0.0000 0.5359 1.2265 4.8767 4.6515 9162.04
True 40 0.008 1 0.05 0.01 0.18 0
220
200
N1 (SSE ≈ 7)
180
N1 (SSE ≈ 10,000)
160
140
number
120
100
80
N2 (SSE ≈ 7)
60 N2 (SSE ≈ 10,000)
40
20
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0
time
Figure 5.22 MATLAB output of two optimization runs. Noise-free data are represented as symbols.
A “correct” solution with SSE ≈ 7, connected with straight-line segments, essentially hits all points
perfectly. However, the same optimization settings may run into a local minimum with SSE ≈ 10,000,
which corresponds to an unsatisfactory solution, at least for N1.
successful. Finally, if the data are noisy, we can of course not expect the SSE to
decrease to single digits. If the algorithm converges to the right neighborhood, the
SSE in the example is between 900 and 1000, which looks like a big number but in
fact is not a bad solution (Figure 5.23).
In the previous examples, we have implicitly assumed that we know the true ini-
tial values for integrating the differential equations. This is a common assumption,
which may or may not be acceptable. Instead of making the assumption, the initial
values may be considered as parameters that are to be optimized along with the
parameters we discussed above.
250 true dynamics N noisy dynamics N2

true dynamics N2 data N1
noisy dynamics N1 data N2
200
150
100
50
0
1.0 1.5 2.0 2.5 3.0
time
Figure 5.23 Effect of noise in the data on parameter estimates. The figure shows the true dynamics
of the population system (dashed lines), noisy data (dots), and the dynamics estimated from these
noisy data (solid lines). Even with an SSE of about 1000, the optimized solutions are not bad. In fact, the
fit for N2 (red) is essentially indistinguishable from the true solution.
STRuCTuRE IDENTIFICATION
Related to the task of estimating parameter values is that of structure identification. In
this case, it is not even known what the structure of the model looks like. For instance,
one could have the same data as above, but the system would be represented as
N 1 = F1 ( N 1, N 2 ) ,
(5.18)
N 2 = F2 ( N 1, N 2 ) ,
with unknown functions F1 and F2. Two approaches may be pursued to attack this
complicated problem. In the most obvious, although cumbersome, approach, one
could explore a number of candidate models that appear to have a chance of succeed-
ing, based on experience or some biological rationale. For instance, in the context of
enzyme kinetics, one could use Michaelis–Menten rate laws, sigmoidal Hill functions,
or more general constructions [36]. It does not take much imagination to see that the
numbers of possible candidates and their combinations are limitless. In general, the
true structure may never be known in detail.
An alternative is the use of canonical models, as discussed in Chapter 4. In a
population setting, one might begin with Lotka–Volterra models. For instance, in
the case of (5.18), one could try composing different sums and differences of terms
like c1N1, c2N2, c 3 N 12, c 4 N 22, and c5N1N2, which in our situation would actually include
the structure that we assumed for the example. For enzyme kinetic models, one
would prefer generalized mass action (GMA) or S-systems. These canonical mod-
els are relatively general and simple, and they are often great default models for
starting an analysis. However, they are local approximations, which may or may not
be able to capture the full dynamics of the observed data.
There are no easy solutions to the structure identification problem, and much
more research will be required to develop promising solution strategies. However, it is
likely that dissecting the task into three steps might be beneficial. In the first, one
would merely try to establish which variables are important. This feature selection
step ideally reduces the number of variables to be included in the model. Second, one
tries to characterize which variable affects which other variables, and, possibly,
whether the effects are likely be augmenting or diminishing. Once the likely topology
of the system has been established, at least in a coarse manner, one might select
mechanistic or canonical models for developing a specific symbolic and numeric rep-
resentation [37].
Other very difficult issues arise in structure identification tasks. First, it may not
even be known whether the right number of variables is included in the model,
which can obviously lead to problems. Second, it might happen that the data are
simply insufficient in quantity and/or quality for identifying the true model struc-
ture, and many groups have started analyzing this topic of identifiability with differ-
ent approaches (for example, [19, 38–41]). An often effective approach toward this
issue is the computation of the Fisher information matrix [42]. Finally, it is possible
that the estimated model contains more parameters than can be validly estimated
from the data, so that predictions toward new data become very unreliable. Meth-
ods of cross-validation can help to some degree with this overfitting issue [43].
As we saw in one of the examples, a search algorithm may suggest a negative
parameter value where we would have expected a positive value. More generically,
a single solution computed with a search algorithm sometimes turns out to be unre-
liable. To address this issue, it has become popular to search not for the uniquely
optimal solution but rather for an entire ensemble of parameter settings that all,
more or less, have the same SSE [39, 40, 44]. Expressed differently, the argument of
unidentifiability and sloppiness is being turned around, and the optimal solution
now consists of an entire set of acceptable parameter settings.
Finally, if several comprehensive and representative time series datasets are avail-
able, it might even be possible to establish a model in a nonparametric manner, that
is, without specification of functions or parameter values. In this case, the data are
computationally converted into a library that contains information about how each
process in the study is affected by the various state variables [45]. This new method
has not been tested enough to allow any final judgment to be made, but it does seem
to have the potential to serve as a rather unbiased alternative to traditional modeling.
ExERCISES 161
ExERCISES
Naturally, many of the exercises in this chapter require computational efforts and access to software, which is available,
for instance, in Microsoft Excel, MATLAB and Mathematica. Specific techniques needed in some of the exercises
include solutions to ODEs, linear and nonlinear regression, genetic algorithms, and splines. The data tables are
available on the website supporting the book (http://www.garlandscience.com/product/isbn/9780815345688).
5.1. Perform linear regression with the bacterial growth Vmax S
data in Table 5.10. Plot the results. Repeat the vI = ,
 I 
K M 1 + +S
analysis after a logarithmic transformation of the  K I 
data. Compare and interpret the results.
where I is the inhibitor concentration and KI is the

TABLE 5.10 inhibition constant? Support your answer with
mathematical arguments.
t Size t Size
5.4. Use linear regression to confirm the results (Vmax =
0 2.2 1.1 22 2.2, KM = 6.4) in the text (see (5.4) and Table 5.2).
0.1 2.9 1.2 21 5.5. Apply a nonlinear regression algorithm directly to
0.2 3.2 1.3 30 the data in Tables 5.2 and 5.3. Compare the results
with the linear estimation upon transformation.
0.3 6 1.4 35
5.6. Apply a nonlinear regression algorithm to the
0.4 6 1.5 38 differential equation (5.4), using the data given
0.5 6.2 1.6 47 in Table 5.12 and plotted in Figure 5.6. Furthermore,
0.6 8 1.7 58
assume S = 1.2. Look at the code on the book’s
website for inspiration.
0.7 11 1.8 70
0.8 14 1.9 81
0.9 13 2 90 TABLE 5.12
1 18 t M
0 2.000
1 1.061
5.2. Perform linear regression with the bacterial
growth data in Table 5.11. Plot the results. 2 0.671
Repeat the analysis after a logarithmic 3 0.509
transformation of the data. Compare and inter- 4 0.442
pret the results.
5 0.414
TABLE 5.11 6 0.403

7 0.398
t Size t Size
8 0.396
0 3 1.1 18
9 0.395
0.1 3 1.2 26
0.2 2.8 1.3 31
0.3 3.7 1.4 36 5.7. Lineweaver and Burk [46] suggested plotting 1/v
against 1/S, thereby reducing the nonlinear
0.4 3.9 1.5 44
estimation problem for Michaelis–Menten rate laws
0.5 4.4 1.6 50 to one of simple linear regressions. Woolf [47] had
0.6 5.6 1.7 64 another idea. If one plots v/S against v, the
0.7 7.4 1.8 72 Michaelis–Menten function also becomes linear
(Confirm this!). Even though Woolf came up with
0.8 11 1.9 88
this trick, the resulting plot is usually called a
0.9 14 2 96 Scatchard plot [48]. While the nonlinear method, the
1 16 Lineweaver–Burk method, and the Woolf method,
should all give the same results, they do so only if the
data are error-free. As soon as there is noise in the
5.3. Confirm that the Michaelis–Menten function data, equivalence is not guaranteed; in fact, the
becomes linear if one plots 1/v against 1/S. Is the results can be quite different. To analyze this
same true for a Michaelis–Menten process with situation, use the data in Table 5.13. For all four
competitive inhibition, which is given by the cases, perform nonlinear regression and the two
function types of linear regression mentioned above.
Summarize your findings, along with graphs and

TABLE 5.15
regression results, in a brief report.
S v (I = 10) v (I = 20) S v (I = 10) v (I = 20)
TABLE 5.13 1 1.8 2 11 11.7 10.7

2 2.2 2.2 12 11.5 10.1
v1 (errors v2 (errors at v3 (errors at
S v0 (no error) 3 6.1 3.2 13 11.2 10.2
distributed) low end) high end)
0.2 0.5454552 0.5 1 0.5 4 5.5 4.8 14 13 10.2
0.4 1 1.2 1.2 1.2 5 7.1 6.6 15 12.2 12.3
0.6 1.384616 1.25 1.5 1.25 6 9.3 7.9 16 12.4 12.7
0.8 1.714287 1.8 1.35 1.8 7 9 7.5 17 12.6 11.4
1 2 2.2 2.2 2.2 8 9.4 8 18 14.8 12.6
2 3 2.75 2.75 2.75 9 11.2 8.2 19 14.2 12.3
3 3.6 3.55 3.55 3 10 10.6 9.4 20 16.5 15.5
4 4 4.2 4.2 3.8

5 4.285715 4.2 4.2 5
(b) Use a gradient method to perform the param-
10 5 4.85 4.85 5.6
eter estimation. Start with different guesstimates.
20 5.454546 5.6 5.6 4.9 Run the estimation for different numbers of
iterations.
5.8. Imagine an experiment where a process V is 5.10. Visualize the error landscape for the nonlinear
governed by a substrate S and an inhibitor I. In the regression with data v1 in Exercise 5.7.
system, S decreases and I increases over time. Use 5.11. Estimate the 10 parameters of the function
the data in Table 5.14 and multiple linear regression
to compute the parameters for process V, which has
F (t ) = sin(at + b)[ct + d (t − e )2 + f (t − g )3 ]e h+i(t − j ) ,
the form V = γ S a I b. Plot the results in a pseudo-
three-dimensional plot. Explain why it seems that
which is plotted in Figure 5.24 from the noise-free
the results lie on a slightly nonlinear curve rather
data in Table 5.16. Compare the requirements,
than forming a (linear) regression plane.
procedures and results for different methods.
TABLE 5.14 8.0
t S I V
0 10 1.2 6.3
1 7.5 1.7 4.5
2 5.5 2.3 3.2
F 2.5
3 4 2.9 2.4
4 2.5 3.5 1.8
5 1.5 4.0 1.4
6 1.0 4.5 1.1
7 0.7 4.5 0.8
–3
8 0.4 4.8 0.6 0 12.5 25.0
9 0.2 4.8 0.5 t
10 0.1 4.9 0.3 Figure 5.24 The function F(t) with 10 parameters. This strange
function is used for estimation purposes.
5.9. (a) The Michaelis–Menten rate law (MMRL) with

competitive inhibition has the formula shown in 5.12. Use a genetic algorithm to estimate parameters for
Exercise 5.3. In addition to the parameters of the the MMRL with inhibition from the data in Table
MMRL, it contains the inhibition parameter KI and 5.15 (see Exercise 5.3). Look at the website code for
the concentration of inhibitor, which we assume can inspiration on how to set up the genetic algorithm.
be controlled in the experiment. Suppose two datasets Run the algorithm with different bounds on
(Table 5.15) were measured with different inhibitor parameter values.
concentrations I. Using grid searches, determine the 5.13. Estimate the parameters of an MMRL with
parameters from the two datasets separately and from inhibition from the data in Table 5.17, assuming
both together. Discuss your findings. that I = 20. Redo the analysis with the
ExERCISES 163
TABLE 5.16
t F t F t F t F t F
0.0 0.000 5.000 0.001 10.000 −0.005 15.000 0.002 20.000 0.000
0.1 4.184 5.100 0.064 10.100 −0.119 15.100 0.031 20.100 −0.004
0.2 6.471 5.200 0.122 10.200 −0.207 15.200 0.052 20.200 −0.007
0.3 6.929 5.300 0.162 10.300 −0.253 15.300 0.062 20.300 −0.009
0.4 5.964 5.400 0.174 10.400 −0.250 15.400 0.060 20.400 −0.008
0.5 4.152 5.500 0.150 10.500 −0.200 15.500 0.047 20.500 −0.006
0.6 2.080 5.600 0.093 10.600 −0.114 15.600 0.026 20.600 −0.004
0.7 0.223 5.700 0.011 10.700 −0.011 15.700 0.002 20.700 0.000
0.8 −1.119 5.800 −0.082 10.800 0.091 15.800 −0.021 20.800 0.003
0.9 −1.836 5.900 −0.166 10.900 0.171 15.900 −0.038 20.900 0.005
1.0 −1.976 6.000 −0.224 11.000 0.216 16.000 −0.047 21.000 0.006
1.1 −1.690 6.100 −0.242 11.100 0.219 16.100 −0.047 21.100 0.006
1.2 −1.170 6.200 −0.211 11.200 0.180 16.200 −0.038 21.200 0.005
1.3 −0.599 6.300 −0.137 11.300 0.110 16.300 −0.022 21.300 0.003
1.4 −0.110 6.400 −0.031 11.400 0.022 16.400 −0.004 21.400 0.000
1.5 0.222 6.500 0.088 11.500 −0.066 16.500 0.014 21.500 −0.002
1.6 0.383 6.600 0.195 11.600 −0.137 16.600 0.028 21.600 −0.004
1.7 0.401 6.700 0.269 11.700 −0.180 16.700 0.036 21.700 −0.005
1.8 0.326 6.800 0.293 11.800 −0.186 16.800 0.036 21.800 −0.005
1.9 0.213 6.900 0.260 11.900 −0.158 16.900 0.030 21.900 −0.004
2.0 0.103 7.000 0.176 12.000 −0.101 17.000 0.019 22.000 −0.002
2.1 0.022 7.100 0.054 12.100 −0.029 17.100 0.005 22.100 −0.001
2.2 −0.020 7.200 −0.082 12.200 0.045 17.200 −0.009 22.200 0.001
2.3 −0.030 7.300 −0.204 12.300 0.107 17.300 −0.020 22.300 0.002
2.4 −0.020 7.400 −0.290 12.400 0.146 17.400 −0.027 22.400 0.003
2.5 −0.004 7.500 −0.321 12.500 0.155 17.500 −0.028 22.500 0.003
2.6 0.009 7.600 −0.291 12.600 0.135 17.600 −0.024 22.600 0.003
2.7 0.012 7.700 −0.204 12.700 0.091 17.700 −0.016 22.700 0.002
2.8 0.006 7.800 −0.077 12.800 0.032 17.800 −0.005 22.800 0.001
2.9 −0.006 7.900 0.066 12.900 −0.029 17.900 0.005 22.900 −0.001
3.0 −0.020 8.000 0.196 13.000 −0.081 18.000 0.014 23.000 −0.002
3.1 −0.031 8.100 0.289 13.100 −0.115 18.100 0.020 23.100 −0.002
3.2 −0.036 8.200 0.327 13.200 −0.126 18.200 0.021 23.200 −0.002
3.3 −0.035 8.300 0.302 13.300 −0.113 18.300 0.019 23.300 −0.002
3.4 −0.026 8.400 0.219 13.400 −0.079 18.400 0.013 23.400 −0.001
3.5 −0.012 8.500 0.095 13.500 −0.033 18.500 0.005 23.500 −0.001
3.6 0.006 8.600 −0.045 13.600 0.017 18.600 −0.003 23.600 0.000
3.7 0.024 8.700 −0.176 13.700 0.060 18.700 −0.010 23.700 0.001
3.8 0.041 8.800 −0.271 13.800 0.090 18.800 −0.014 23.800 0.002
3.9 0.052 8.900 −0.314 13.900 0.101 18.900 −0.016 23.900 0.002
4.0 0.054 9.000 −0.296 14.000 0.093 19.000 −0.014 24.000 0.002
4.1 0.045 9.100 −0.222 14.100 0.068 19.100 −0.010 24.100 0.001
4.2 0.025 9.200 −0.108 14.200 0.032 19.200 −0.005 24.200 0.001
4.3 −0.005 9.300 0.024 14.300 −0.008 19.300 0.001 24.300 0.000
4.4 −0.040 9.400 0.149 14.400 −0.043 19.400 0.007 24.400 −0.001
4.5 −0.073 9.500 0.242 14.500 −0.069 19.500 0.010 24.500 −0.001
4.6 −0.097 9.600 0.287 14.600 −0.080 19.600 0.012 24.600 −0.001
4.7 −0.104 9.700 0.277 14.700 −0.075 19.700 0.011 24.700 −0.001
4.8 −0.089 9.800 0.215 14.800 −0.057 19.800 0.008 24.800 −0.001
4.9 −0.053 9.900 0.114 14.900 −0.029 19.900 0.004 24.900 0.000
25.000 0.000
bootstrapping and jackknifing methods. Compare 5.19. Construct the diagram of a small pathway system
and interpret results. Also, average the parameter that would be described by the following
values obtained from (A) different bootstrapping equations:
runs and (B) different jackknifing runs and assess
the fits. X 1 = α 1 X 3g13 − β1 X 1h11 ,
X 2 = α 2 X 1g 21 − β 2 X 2h22 ,
TABLE 5.17
X 3 = α 3 X 2g 32 − β 3 X 3h33 X 4h34 ,
S v(I = 20)
X 4 = α 4 X 1g 41 − β 4 X 4h44 .
1 2
4 4.8 Is the diagram unique? Estimate parameter values
8 6 for the system, using the datasets in Table 5.18,
10 9.4
either one at a time or combined.
12 12
14 10.2
TABLE 5.18
16 13.3
18 12.6 t X1 X2 X3 X4
20 12.2
DATASET 1
5.14. (a) Use a nonlinear regression algorithm to estimate 0 1.400 2.700 1.200 0.400
the parameters of the function f in (5.9) from the 0.2 1.039 3.122 1.591 0.317
data in Table 5.4. Redo the estimation several times 0.4 0.698 3.176 1.980 0.250
with different start guesses. Study the quality of 0.6 0.482 2.992 2.298 0.195
solutions with averaged parameter values. 0.8 0.372 2.711 2.519 0.157
(b) Transform the function f in (5.9) so that it 1 0.322 2.431 2.643 0.135
becomes linear. Try to use linear regression to 1.2 0.302 2.198 2.682 0.125
estimate the parameter values. Report your findings.
1.4 0.300 2.025 2.657 0.120
(c) Generate more equivalent representations of the 1.6 0.307 1.911 2.591 0.120
function f in (5.9), similar to that in (5.10), using the
1.8 0.322 1.846 2.506 0.123
relationship Y = 1.75X 0.8. Formulate a functional
description that encompasses all possible represen- 2 0.340 1.822 2.416 0.126
tations of this type. 2.2 0.359 1.826 2.335 0.130
(d) Create a different situation of an estimation 2.4 0.377 1.851 2.269 0.135
problem with multiple solutions. 2.6 0.393 1.886 2.222 0.139
5.15. Construct a function that has at least two minima 2.8 0.404 1.924 2.192 0.142
with the same value (recall the situation in Figure 3 0.412 1.959 2.177 0.144
5.15). Is it possible to have infinitely many, 3.2 0.415 1.988 2.174 0.145
unconnected minima with the same error? If you 3.4 0.415 2.010 2.180 0.146
think yes, provide an example. If not, provide a proof,
3.6 0.413 2.024 2.189 0.146
counterexample, or at least compelling arguments.
3.8 0.410 2.030 2.201 0.145
5.16. Use a spline function in MATLAB• to connect the
4 0.406 2.031 2.212 0.145
data on population sizes Ni in Table 5.7, use this
function to compute slopes at all time points, and DATASET 2
estimate parameter values.
0 0.200 0.300 2.200 0.010
5.17. Estimate the parameter values in (5.16) upon dividing
0.2 0.439 0.833 1.789 0.107
by N1, which makes the problem linear. Compare the
results with those in the text. Discuss the advantages 0.4 0.597 1.347 1.569 0.156
and potential problems of each method. 0.6 0.685 1.791 1.540 0.185
5.18. (a) Use the data with and without noise (see Tables 0.8 0.686 2.125 1.634 0.197
5.7 and 5.8) and estimate parameter values with and 1 0.627 2.327 1.787 0.195
without slope substitution, starting with different 1.2 0.548 2.406 1.954 0.184
sets of initial guesses for the parameter values. 1.4 0.479 2.392 2.105 0.170
(b) Estimate parameter values with and without slope 1.6 0.428 2.324 2.223 0.157
substitution, using subsets of the data with and without 1.8 0.395 2.235 2.304 0.147
noise. Use different subsets with 10 and 15 time points. 2 0.376 2.146 2.349 0.141
Compare results. Test whether averaging of the
2.2 0.367 2.071 2.363 0.137
parameter values from different runs improves the fits.
ExERCISES 165
TABLE 5.18 (CONTINUED) TABLE 5.19

t X1 X2 X3 X4 t X1 X2 X3 X4
DATASET 2 DATASET 1
2.4 0.365 2.013 2.356 0.135 0 0.109 3.145 1.304 0.817
2.6 0.368 1.975 2.336 0.135 0.2 0.468 2.336 2.290 0.539
0.4 0.430 1.998 2.778 0.445
2.8 0.373 1.954 2.309 0.136
0.6 0.354 1.499 3.216 0.329
3 0.380 1.946 2.281 0.137
0.8 0.340 1.131 3.162 0.259
3.2 0.387 1.949 2.256 0.139
1 0.352 0.871 3.455 0.244
3.4 0.394 1.958 2.237 0.140 1.2 0.326 0.673 3.223 0.222
3.6 0.399 1.970 2.223 0.142 1.4 0.353 0.590 3.061 0.233
3.8 0.402 1.982 2.214 0.143 1.6 0.362 0.513 2.733 0.252
4 0.404 1.993 2.211 0.143 1.8 0.406 0.481 2.620 0.285
2 0.451 0.455 2.473 0.311
DATASET 3
2.2 0.460 0.473 2.262 0.336
0 4.000 1.000 3.000 4.000 2.4 0.508 0.508 1.977 0.383
0.2 1.864 2.667 2.076 1.909 2.6 0.525 0.552 1.951 0.466
0.4 0.952 3.151 1.952 0.872 2.8 0.559 0.595 1.800 0.492
0.6 0.598 3.111 2.094 0.413 3 0.576 0.676 1.760 0.547
0.8 0.447 2.886 2.296 0.232 3.2 0.607 0.707 1.744 0.557
3.4 0.613 0.788 1.674 0.587
1 0.371 2.622 2.468 0.165
3.6 0.616 0.782 1.627 0.646
1.2 0.332 2.376 2.572 0.139
3.8 0.609 0.794 1.624 0.630
1.4 0.316 2.175 2.607 0.128
4 0.611 0.848 1.736 0.618
1.6 0.313 2.026 2.588 0.124
DATASET 2
1.8 0.320 1.926 2.533 0.124
0 2.017 3.046 1.143 0.115
2 0.332 1.870 2.460 0.125
0.2 0.711 2.962 2.320 0.966
2.2 0.348 1.849 2.385 0.129
0.4 0.443 2.652 3.225 0.754
2.4 0.365 1.853 2.317 0.132 0.6 0.381 2.052 3.249 0.485
2.6 0.381 1.874 2.261 0.136 0.8 0.385 1.494 3.636 0.312
2.8 0.394 1.904 2.221 0.139 1 0.338 1.206 3.769 0.275
3 0.404 1.937 2.196 0.142 1.2 0.352 0.835 3.608 0.246
3.2 0.410 1.967 2.184 0.144 1.4 0.342 0.630 3.448 0.234
1.6 0.353 0.568 3.332 0.245
3.4 0.413 1.992 2.182 0.145
1.8 0.400 0.541 3.276 0.270
3.6 0.412 2.010 2.187 0.146
2 0.409 0.472 2.524 0.295
3.8 0.411 2.021 2.195 0.145 2.2 0.469 0.514 2.580 0.342
4 0.408 2.027 2.205 0.145 2.4 0.517 0.485 2.414 0.329
2.6 0.530 0.544 2.081 0.424
5.20. Assume again the model structure of Exercise 5.19, 2.8 0.614 0.560 1.848 0.451
but with new parameter values: 3 0.636 0.606 1.927 0.506
3.2 0.644 0.685 1.899 0.585
X 1 = α 1 X 3g13 − β1 X 1h11 , 3.4 0.687 0.724 1.668 0.575
X 2 = α 2 X 1g 21 − β 2 X 2h22 , 3.6 0.615 0.772 1.748 0.664
3.8 0.650 0.821 1.871 0.626
X 3 = α 3 X 2g 32 − β 3 X 3h33 X 4h34 ,
4 0.602 0.753 1.500 0.619
X 4 = α 4 X 1g 41 − β 4 X 4h44 .
Use a gradient-based method, such as

lsqcurvefit in MATLAB•, to estimate param- 5.21. For the two-population example in the text (see
eter values for the system, using the noisy datasets Table 5.9 and Figure 5.22), what is the absolute
in Table 5.19, either one at a time or combined. deviation of the optimized solution from the true
Use as bounds for the parameters 0 ≤ αi, βi ≤ 20; solution at each time point if SSE ≈ 7 or
and −2 ≤ gij, hij ≤ 2. SSE ≈ 10,000?
5.22. Find out from the literature what sloppiness means 5.24. Write a brief report on overfitting and
in the context of parameter estimation. Write a brief cross-validation.
report. 5.25. Discuss the advantages and drawbacks of canonical
5.23. Find out details regarding Fisher’s information models for structure identification.
matrix. Write a brief report.
REFERENCES
[1] Kutner MH, Nachtsheim CJ, Neter J & Li W. Applied Linear [20] Raue A, Kreutz C, Maiwald T, et al. Structural and practical
Statistical Models, 5th ed. McGraw-Hill/Irwin, 2004. identifiability analysis of partially observed dynamical models
[2] Montgomery DC. Design and Analysis of Experiments, 9th ed. by exploiting the profile likelihood. Bioinformatics 25 (2009)
Wiley, 2017. 1923–1929.
[3] Donahue MM, Buzzard GT & Rundell AE. Parameter [21] Efron B & Tibshirani RJ. An Introduction to the Bootstrap.
identification with adaptive sparse grid-based optimization for Chapman & Hall/CRC, 1993.
models of cellular processes. In Methods in Bioengineering: [22] Good P. Resampling Methods: A Practical Guide to Data
Systems Analysis of Biological Networks (A Jayaraman & J Analysis, 3rd ed. Birkhäuser, 2006.
Hahn, eds), pp 211–232. Artech House, 2009. [23] Manly BFJ. Randomization, Bootstrap and Monte Carlo
[4] Land AH & Doig AG. An automatic method of solving discrete Methods in Biology, 3rd ed. Chapman & Hall/CRC, 2006.
programming problems. Econometrica 28 (1960) 497–520. [24] Neves AR, Ramos A, Nunes MC, et al. In vivo nuclear magnetic
[5] Falk JE & Soland RM. An algorithm for separable nonconvex resonance studies of glycolytic kinetics in Lactococcus lactis.
programming problems. Manage. Sci. 15 (1969) 550–569. Biotechnol. Bioeng. 64 (1999) 200–212.
[6] Glantz SA, Slinker BK & Neilands TB. Primer of Applied [25] Voit EO & Almeida J. Decoupling dynamical systems for
Regression and Analysis of Variance, 3rd ed. McGraw-Hill, pathway identification from metabolic profiles. Bioinformatics
2016. 20 (2004) 1670–1681.
[7] Goldberg DE. Genetic Algorithms in Search, Optimization and [26] Chou I-C & Voit EO. Recent developments in parameter
Machine Learning. Addison-Wesley, 1989. estimation and structure identification of biochemical and
[8] Holland JH. Adaptation in Natural and Artificial Systems: An genomic systems. Math. Biosci. 219 (2009) 57–83.
Introductory Analysis with Applications to Biology, Control, [27] Gennemark P & Wedelin D. Benchmarks for identification
and Artificial Intelligence. University of Michigan Press, 1975. of ordinary differential equations from time series data.
[9] Fogel DB. Evolutionary Computation: Toward a New Bioinformatics 25 (2009) 780–786.
Philosophy of Machine Intelligence, 3rd ed. IEEE Press/Wiley- [28] Lillacci G & Khammash M. Parameter estimation and model
Interscience, 2006. selection in computational biology. PLoS Comput. Biol. 6(3)
[10] Dorigo M & Stützle T. Ant Colony Optimization. MIT Press, 2004. (2010) e1000696.
[11] Catanzaro D, Pesenti R & Milinkovitch MC. An ant colony [29] Gengjie J, Stephanopoulos GN, & Gunawan R. Parameter
optimization algorithm for phylogenetic estimation under the estimation of kinetic models from metabolic profiles: two-
minimum evolution principle. BMC Evol. Biol. 7 (2007) 228. phase dynamic decoupling method. Bioinformatics 27 (2011)
[12] Kennedy J & Eberhart RC. Particle swarm optimization. In 1964–1970.
Proceedings of IEEE International Conference on Neural [30] Sun J, Garibaldi JM & Hodgman C. Parameter estimation using
Networks, Perth, Australia, 27 November–1 December 1995. meta-heuristics in systems biology: a comprehensive review.
IEEE, 1995. IEEE/ACM Trans. Comput. Biol. Bioinform. 9 (2012) 185–202.
[13] Poli R. Analysis of the publications on the applications of [31] Verhulst PF. Notice sur la loi que la population suit dans son
particle swarm optimisation. J. Artif. Evol. Appl. (2008) Article ID accroissement. Corr. Math. Phys. 10 (1838) 113–121.
685175. [32] de Boor C. A Practical Guide to Splines, revised edition.
[14] Metropolis N, Rosenbluth AW, Rosenbluth MN, et al. Equations Springer, 2001.
of state calculations by fast computing machines. J. Chem. [33] Vilela M, Borges CC, Vinga S, Vasconcelos AT, et al. Automated
Phys. 21 (1953) 1087–1092. smoother for the numerical decoupling of dynamics models.
[15] Salamon P, Sibani P & Frost R. Facts, Conjectures, and BMC Bioinformatics 8 (2007) 305.
Improvements for Simulated Annealing. SIAM, 2002. [34] Voit EO & Chou I-C. Parameter estimation in canonical
[16] Li S, Liu Y & Yu H. Parameter estimation approach in biological systems models. Int. J. Syst. Synth. Biol. 1 (2010) 1–19.
groundwater hydrology using hybrid ant colony system. In [35] Dam P, Fonseca LL, Konstantinidis KT & Voit EO. Dynamic
Computational Intelligence and Bioinformatics: ICIC 2006 (D-S models of the complex microbial metapopulation of Lake
Huang, K Li & GW Irwin, eds), pp 182–191. Lecture Notes in Mendota. npj Syst. Biol. Appl. 2 (2016) 16007.
Computer Science, Volume 4115, Springer, 2006. [36] Schulz AR. Enzyme Kinetics: From Diastase to Multi-Enzyme
[17] Voit EO. What if the fit is unfit? Criteria for biological systems Systems. Cambridge University Press, 1994.
estimation beyond residual errors. In Applied Statistics for [37] Goel G, Chou I-C & Voit EO. System estimation from metabolic
Biological Networks (M Dehmer, F Emmert-Streib & A Salvador, time-series data. Bioinformatics 24 (2008) 2505–2511.
eds), pp 183–200. Wiley, 2011. [38] Vilela M, Vinga S, Maia MA, et al. Identification of neutral
[18] Bertalanffy L von. Principles and theory of growth. In biochemical network models from time series data. BMC Syst.
Fundamental Aspects of Normal and Malignant Growth (WW Biol. 3 (2009) 47.
Nowinski, ed.), pp 137–259. Elsevier, 1960. [39] Battogtokh D, Asch DK, Case ME, et al. An ensemble method
[19] Gutenkunst RN, Casey FP, Waterfall JJ, et al. Extracting for identifying regulatory circuits with special reference to the
falsifiable predictions from sloppy models. Ann. NY Acad. Sci. qa gene cluster of Neurospora crassa. Proc. Natl Acad. Sci. USA
1115 (2007) 203–211. 99 (2002) 16904–16909.
FuRThER READING 167
[40] Lee Y & Voit EO. Mathematical modeling of monolignol [45] Faraji M & Voit EO. Nonparametric dynamic modeling.
biosynthesis in Populus. Math. Biosci. 228 (2010) 78–89. Math. Biosci. (2016) S0025-5564(16)30113-4. doi: 10.1016/j.
[41] Gennemark P & Wedelin D. ODEion—a software module for mbs.2016.08.004. (Epub ahead of print.)
structural identification of ordinary differential equations J. [46] Lineweaver H & Burk D. The determination of enzyme
Bioinform. Comput. Biol. 12 (2014) 1350015. dissociation constants. J. Am. Chem. Soc. 56 (1934) 658–666.
[42] Srinath S & Gunawan R. Parameter identifiability of power-law [47] Woolf B. Quoted in Haldane JBS & Stern KG. Allgemeine
biochemical system models. J. Biotechnol. 149 (2010) 132–140. Chemie der Enzyme, p 19. Steinkopf, 1932.
[43] Almeida JS. Predictive non-linear modeling of complex data by [48] Scatchard G. The attractions of proteins for small molecules
artificial neural networks. Curr. Opin. Biotechnol. 13 (2002) 72–76. and ions. Ann. NY Acad. Sci. 51 (1949) 660–672.
[44] Tan Y & Liao JC. Metabolic ensemble modeling for strain
engineers. Biotechnol. J. 7 (2012) 343–353.
FuRThER READING
Chou IC & Voit EO. Recent developments in parameter estimation Montgomery DC. Design and Analysis of Experiments, 9th ed. Wiley,
and structure identification of biochemical and genomic 2017.
systems. Math. Biosci. 219 (2009) 57–83. Voit EO. What if the fit is unfit? Criteria for biological
Eiben AE & Smith JE. Introduction to Evolutionary Computing, 2nd systems estimation beyond residual errors. In Applied
ed. Springer, 2015. Statistics for Biological Networks (M Dehmer, F
Glantz SA, Slinker BK & Neilands TB. Primer of Applied Regression Emmert-Streib & A Salvador, eds), pp 183–200. Wiley,
and Analysis of Variance, 3rd ed. McGraw-Hill, 2016. 2011.
Kutner MH, Nachtsheim CJ, Neter J & Li W. Applied Linear Statistical
Models, 5th ed. McGraw-Hill/Irwin, 2004.
Gene Systems
6
• Discuss the Central Dogma of molecular biology, as well as modern
amendments and refinements
• Describe the key features of DNA and RNA
• Identify the roles of different types of RNAs
• Outline the principles of gene regulation
• Set up simple models of gene regulation
• Summarize current methods for assessing gene expression and its location,
along with their advantages and drawbacks
THE CENTRAL DOGMA

Genes are the main carriers of biological information from one generation to the
next. They make us unique, are responsible for many of our good features and the
bad features of others, contribute to a number of chronic diseases, and are the prime
targets of evolution. In the olden days, which in this context means about half a cen-
tury ago, the role of genes was simple, namely:
Genes consist of self-replicating deoxyribonucleic acid (DNA), DNA is transcribed

into matching ribonucleic acid (RNA), RNA is translated into proteins, and proteins
manage the processes of daily life.
This unidirectional Central Dogma, proposed by Nobel Laureate Francis Crick

[1, 2], became the centerpiece of modern biology.
The Central Dogma is still mostly true today, although we have learned that
everything surrounding it is much, much more complicated: some DNA does not
code for genes; some organisms inherit their information through RNA; molecular
modifications such as DNA methylation can control transcription; and alternative
splicing in eukaryotic cells may lead to many different RNAs and hence many differ-
ent proteins from the same DNA. Maybe most significant of all, and in stark contrast
to Crick’s original ideas, proteins have a huge impact on DNA through their roles as
transcription factors. In addition, most organisms employ various types of RNAs
that affect which gene is transcribed when, where, and how much, and when a gene
is to be silenced. Moreover, large and small metabolites can control gene expres-
sion. Large signaling metabolites include sphingolipids, which can “sense” environ-
mental stresses and control the up- or down-regulation of appropriate genes in
response [3]. An example of a small metabolite is glucose, which indirectly affects
170 Chapter 6: Gene Systems
the transcription of many genes in organisms from bacteria to humans and, for (A) (B)
DNA DNA
instance, represses a wide variety of genes in the β-cells of the pancreas, which can
be an especially serious problem in people with diabetes [4]. In bacteria, any excess
of the rare amino acid tryptophan in the environment shuts down the expression of
genes that are responsible for tryptophan production, because this production is
mRNA mRNA
metabolically expensive and would be wasteful. These strategies of avoiding unnec-
essary action are implemented through regulatory mechanisms that interact with
the responsible section of the chromosome. In the case of tryptophan, the amino
acid directly activates a repressor of the DNA segment that contains the genes that
protein protein
code for the enzymes of the tryptophan pathway [5]. Furthermore, there is evidence
suggesting that metabolites can affect translation (see, for example, [6]). To make
matters even more complicated, we have started to discover that about 98% of our
DNA does not code for proteins, that genes can overlap each other, that the same
metabolites metabolites
stretches of DNA can lead to several alternative transcripts, and that regulatory
elements are not necessarily located adjacent to the genes whose transcription they
control [7]. 1950s – 1970s today
Thus, it is still true that genes are responsible for information transfer between
generations, but they are clearly not the sole agents (Figure 6.1). In fact, genes Figure 6.1 The original Central Dogma
only determine which proteins can be made by a cell, but they contain no direct compared with modern understanding. In
the original concept of the Central Dogma,
information about the quantities and the timing with which proteins are made [8].
transcription, translation, and enzymatic
This information is obviously critical, because otherwise all cells of an organism catalysis were proposed to form a linear
would continuously produce the same proteins in the same quantities. Instead, chain of processes, although nobody
genes, together with proteins and metabolites, act as the components of a well- doubted the role of regulation. We know
organized system: genes provide the template for the synthesis of RNAs, which now that a complex feedback structure
lead to proteins and metabolites, which in turn feed back to affect and control at every level is crucial for appropriate
gene expression in multiple ways. This shared manner of control is a compromise functioning.
between the original, unidirectional Central Dogma, where genes were seen as the
true drivers of life, and the notion of genes as mere databases, or as “prisoners” of
the physiological control of the organism, where they have no autonomy or
independent control [8].
It is worth noting that genetic information is not exclusively transmitted through
the organism’s chromosomes. For instance, we have learned in recent years that
epigenetic modifications in the form of methyl or acetyl groups attached to specific
locations on the DNA can be inherited (see later in this chapter). Plants contain
additional DNA in their chloroplasts, and the mitochondria of eukaryotes contain
DNA that is transmitted from the mother to her offspring. Inheritance of male mito-
chondrial DNA has been reported in some species, although it is rare [9]. The contri-
butions of chloroplast and mitochondrial DNA are generally small, but significant
nevertheless. In humans, the mitochondrial DNA codes for 13 proteins, 22 transfer
RNAs, and two subunits of ribosomal RNA.
Other nonchromosomal genetic elements include plasmids and transposable
elements. Plasmids are typically circular, double-stranded DNA molecules that are
natural in bacteria, although they have also been found in Archaea and in eukary-
otes [10]. A bacterium may contain hundreds of plasmids at the same time and can
share them with other bacteria, even from different species, in a process called
horizontal gene transfer, which is crucial for the evolution of new species [11].
Plasmids can replicate independently of the main chromosome. They often contain
genes that permit bacteria to survive under hostile conditions, such as exposure to
antibiotics, which in hospitals can result in bacterial strains that are resistant to
multiple drugs and very difficult to treat [12]. Perhaps the best-known example is
methicillin-resistant Staphylococcus aureus (MRSA), which is a formidable culprit
in many human infections [13]. On the positive side, plasmids have become
extremely important research tools in biotechnology and synthetic biology,
where they are used as vectors for the transport of genes into foreign species (see
Chapter 14).
Transposable elements, or transposons, are stretches of DNA that can move
from one location in the genome to another [14]. In some instances, the transposon
is cut out of its original location and moved, whereas in other cases, the transposon
is first copied and only the copy is relocated. In bacteria, transposons can carry
several genes, which might be inserted into a plasmid and make the organism resis-
tant to environmental stresses. Transposons can lead to important mutations and
KEy PROPERTiES Of DNA AND RNA 171
also to the expansion of the genome. It has been estimated that 60% of the maize
genome consists of transposons. As one might expect, some transposons have been
associated with human diseases. The Alu sequence can be found over a million
times within the human genome and has been associated with diabetes, leukemia,
breast cancer, and a number of other diseases [15].
The body of information on genes and on their structure, function, regulation,
and evolution is overwhelmingly huge and growing at breakneck speed, primarily
owing to the rapid evolution of sequencing techniques. Some of this information is
of great and direct pertinence to systems biology, while other aspects are less so.
Arguably most important for systems analyses is an understanding of which genes
are expressed when, and how this complicated process is regulated by proteins,
RNAs, and even metabolites. To hone this understanding in a fashion that is as brief
as feasible, this chapter begins with a barebones discussion of the key features of
DNA and RNA and afterwards devotes more attention to issues of the control and
regulation of gene expression. The chapter ends with a brief description of standard
methods for measuring gene expression. This description is certainly not sufficient
for learning how to execute these methods, and the intent is rather to provide a feel
for the different types of data that are of the greatest value to systems biology and
for their qualitative and quantitative reliability. Thus—as with the following chap-
ters on proteins and metabolites—this chapter is bound to disappoint many biolo-
gists, because it will only scratch the surface of the fascinating world of genes,
genomes, and gene regulatory systems. In particular, this chapter keeps many
aspects of bioinformatics to a minimum, and methodological issues of sequence
analysis, gene annotation, evolution, and phylogeny, as well as experimental
methods of genome analysis and manipulation, will be discussed rather cursorily
or not at all. Excellent texts and Internet resources, including [16–20], cover these
topics very well.
KEy PROPERTiES Of DNA AND RNA

6.1 Chemical and Physical features
The basic building blocks of DNA are nucleotides, which form two long polymer
strands. These strands are oriented in opposite (anti-parallel) directions and are
held together by hydrogen bonds. Because of the physical and chemical features of
the nucleotides, the two strands curl slightly, and the result is the famous double
helix, which James Watson and Francis Crick discovered in 1953, based in part on
Rosalind Franklin’s X-ray photographs of DNA and her ideas of two nucleotide
strands, and for which Watson and Crick were awarded the 1962 Nobel Prize in
Physiology or Medicine [21, 22].
DNA makes use of four different nucleotides. Each consists of a nitrogen-
containing nucleo-base, a five-carbon sugar called 2-deoxyribose, and a phosphate
group. The nucleo-bases fall into the biochemical classes of purines and pyrimi-
dines and are called adenine, cytosine, guanine, and thymine (abbreviated as A, C,
G, and T). A combination of a nucleo-base and a sugar is called a nucleoside, and if
a phosphate group is attached to the sugar as well, the resulting molecule is called a
nucleotide. RNA uses very similar components, but the sugar is ribose rather than
2-deoxyribose and the nucleo-base thymine is replaced with uracil (U). The build-
ing blocks are visualized in Figure 6.2.
The chemical structure of the nucleo-bases has important consequences, since
it is hydrogen bonding between them that holds the DNA strands together.
Namely, guanine on one strand of DNA always pairs with cytosine on the anti-
parallel strand, and adenine always pairs with thymine (Figures 6.3 and 6.4).
Thus, one speaks of GC and AT base pairs, and this notion of a pair has become
the most widely used size unit of DNA. Because of the strict pairing, one strand is
in some sense the mirror image of the other, and this fact is of utmost importance
for two fundamental processes. The first is DNA replication, which precedes cell
division: double-stranded DNA opens up and both strands serve as templates for
newly created DNA strands that bind to them. As a result of this hybridization
process, the double-stranded DNA is now available in two copies, which
(A) (B)
HO HO
base OH OH
O O
P
sugar
OH OH OH H
ribose deoxyribose
(in RNA) (in DNA)
nucleoside
nucleotide sugars
(C) (D)
NH2 O NH2 O O
H H3C H H
N N N N N N
N
N N N N N N N O
NH2 O O
R R R R R
adenine guanine cytosine thymine uracil
purine bases pyrimidine bases
Figure 6.2 Building blocks of DNA and RNA. (A) Generic composition of nucleotides. (B) The sugars in RNA and DNA differ slightly, by one oxygen atom.
(C, D) Both DNA and RNA contain purine and pyrimidine bases. Adenine, guanine, and cytosine are common to DNA and RNA, but the fourth base in DNA
is thymine, while in RNA it is uracil.
eventually end up in the two daughter cells. The replication process is facilitated
by specific proteins (Figure 6.5). The second fundamental process is the tran-
scription of DNA into RNA, which follows the same principles, except that uracil
replaces thymine.
The phosphate–deoxyribose backbone of each DNA strand has a definite orien-
tation, with a 3-prime (3′) end and a 5-prime (5′) end, which are defined by the
location of the bonds on the final sugar of the strand. When the strands form a
double helix, they do so in an anti-parallel manner, with one strand in the 3′-to-5′
direction binding to the other in the 5′-to-3′ direction, as shown for example in
Figure 6.4. In Figure 6.3, the backbone molecules are located at positions indicated
by R. A sequence of three base pairs GC-AT-GC is illustrated in Figure 6.4. DNA
consists of millions of base pairs that are connected in this fashion to form two
complementary strands. The chemical and physical properties of these strands
H H
N O H N
N N H O H3C
N N H N Figure 6.3 Guanine always pairs with
N N H N cytosine on the anti-parallel strand,
R N N N N while adenine always pairs with thymine.
R Hydrogen bonds between the partners are
N H O R O R shown as dotted red lines. Note that the
GC pair contains three bonds, whereas the
H AT pair contains two. The GC bonding is
guanine cytosine adenine thymine therefore stronger.
(A) (B)
3′ end 5′ end 3′ end
– N
O H2 OH
O O
5′ end
P N N
N H
C –
O N
O O O
G O N N H2
N O O–
T H 3C P
A O O
O O
O H2
P N N
H N
C – O N
G O N O O
O N N
5′ end O O–
P
N O
O H2
3′ end O O
P O N N
O N H
O – N
N O O
O N H2
N
O O–
P
OH O
3′ end –
O
5′ end
Figure 6.4 Sequence of three base pairs, namely GC, AT, GC, as shown in Figure 6.3. The connection between two neighboring bases is made by
sugars and phosphates, which form one phosphate–deoxyribose backbone per strand. The backbones run in an anti-parallel fashion. (A) Diagram of
backbones (blue and red) and bases with hydrogen bonds. (B) Molecular details.
Figure 6.5 Two simian virus SV40 protein

domains interact with DNA just before
replication. The high-resolution crystal
structure shows the origin binding domains
(OBDs) of a large T antigen (T-Ag) complexed
with the DNA fragment that contains
the origin for replication. The core origin
contains four binding sites, each consisting
of five nucleotides, which are organized as
pairs of inverted repeats. In the co-structure,
two T-Ag OBDs are oriented in a head-to-
head fashion on the same face of the DNA,
and each T-Ag OBD engages the major
groove. Although the OBDs are very close to
each other when bound to their DNA targets,
they do not contact each other. (Courtesy
of Huiling Chen and Ying Xu, University of
Georgia.)
Figure 6.6 Modern methods of atomic

force microscopy (AFM) make it possible
to visualize supercoiled DNA. The images
show 5600 base pairs of plasmid DNA from
Escherichia coli. (From Lyubchenko YL. Micron
42 [2011] 196–206. With permission from
Elsevier.)
300 nm 100 nm
cause double-stranded DNA to form a double helix, and this double helix itself
curls into supercoils that can be visualized with methods of atomic force micros-
copy (Figure 6.6).
6.2 Size and Organization of DNA

Interestingly, the size of a genome has no correlation with the size or complexity
of an organism, and is not even correlated with the number of protein-encoding
genes [23]. Human DNA consists of about three billion base pairs, 2% of which
form between 20,000 and 25,000 protein-encoding genes. Some plants have twice
as many genes. The lowly ameba Polychaos dubium apparently contains 200
times as much DNA as humans, while the pufferfish Fugu rubripes, with roughly
400 million base pairs, has the shortest genome of any vertebrate species,
although it contains about 30,000 genes, which is more than the number of pro-
tein-encoding genes in humans [24]. A cogent explanation for these apparent
discrepancies cannot be given at present, although there are many partial expla-
nations, mixed with plenty of speculation. For instance, alternative splicing of
gene transcripts can ultimately result in a much larger number of different pro-
teins than the number of coding genes might suggest [25]. In this “Lego-block”
strategy, short protein sequences (peptides) can be linked together in many vari-
ants. The enormous amount of noncoding DNA (98% in humans) contains differ-
ent types of RNAs that are involved in transcription, as well as in many regulatory
mechanisms that guide the expression of appropriate genes and the concatena-
tion of appropriate peptides, at the right time and in the right place (see later in
this chapter).
Limiting our comparisons of genome sizes to bacteria, we can see some trends.
If a bacterial species lives in a stable environment with sufficient nutrition, such as
the gut of an animal, the number of genes averages about 2000. By contrast, aquatic
bacteria have about 3000 genes, and terrestrial bacteria average about 4500 [26].
An intuitive explanation is that larger genomes permit more—and more complex—
metabolic pathway systems, which are needed if the organism has to survive in
drastically varying ecological niches. It has been estimated that bacteria have
roughly the same numbers of genes and proteins, but that the number of different
metabolites is an order of magnitude lower [27].
Technological advances in the past few decades have made it possible to deter-
mine the order of nucleotides in a given DNA strand. This sequencing capability,
which is now fairly automated, has opened incredible insights into the manner
with which nature codes life processes [28]. At first, a DNA sequence looks like a
random arrangement of A, C, G, and T, but in truth it contains an enormous amount
of information. At the lowest level, triplets of nucleotides in a stretch of coding DNA
(called codons) correspond to specific amino acids in proteins; we will discuss this
genetic code in greater detail in Chapter 7. Codons can also correspond to start
and stop signals for transcription. In other words, the DNA is a blueprint for the
Figure 6.7 Barcodes of four bacterial

strains. The y-axis in each block represents
the location of measurement within the
DNA of an organism, while the x-axis shows
in shades of gray the frequencies of all
possible combinations of nucleotides within
a window of 1000 nucleotides, starting at
nucleotide y. Amazingly, the frequencies
are almost independent of the location
within the DNA. It seems that the occasional
horizontal “stripes” correspond to genes that
the organism obtained during evolution
from different species through horizontal
gene transfer. (Courtesy of Fengfeng Zhou
and Ying Xu, University of Georgia.)
E. coli E. coli Burkholderia Pyrococcus

K-12 O157 pseudomallei furiosus
features of the protein for which it codes. Because triplets of four nucleotides per-
mit 64 different combinations and most organisms use only 20 amino acids, there
is some degeneracy, and the same amino acid may correspond to several codons.
Again, the choice of one of these triplets is not random, and different organisms
prefer some triplets over others, a phenomenon called codon-bias or wobble. In
fact, the preferences of specific triplets for the same amino acids are so species-
specific that it is possible to “barcode” species by their sequences [29]. This barcod-
ing is accomplished in the following fashion. One computationally moves along the
DNA sequence, nucleotide by nucleotide, and reports at each nucleotide location
the frequencies of all different combinations of four nucleotides and their anti-
parallel complements, such as GTTC/CAAG, GCTC/CGAG, CTAC/GATG, within a
window of the next 1000 base pairs. Collecting these data, it turns out that these
frequencies are far from equal or random, and instead form a very specific barcode
pattern when they are represented as different shades of gray (Figure 6.7). It has
been shown that these barcodes correlate with the phylogenetic similarity of organ-
isms and that they can be used to classify DNA from prokaryotes, eukaryotes, plas-
tids, plasmids, and mitochondria. They can also be used to characterize the
genomes of metapopulations, which consist of many different cohabitating species
(see Chapter 15).
6.3 Genes and Noncoding DNA

It was initially assumed that the DNA of an organism consisted of a neat sequence of
all genes. However, DNA has turned out to be much more complicated in terms of
its encoded information and organization. Indeed, even the concept of a gene has
changed over time.
In his revolutionary work on the inheritance of certain features of peas, the
Austrian monk Gregor Mendel (1822–1884) noticed that biological traits are
inherited from one generation to the next as discrete units. The discrete nature of
this process allowed the separation of traits and explained the old observation
that siblings can have different features. Mendel’s findings were more or less
ignored for a long time, until DNA was identified as the carrier of inheritance in
the 1940s. At that point, genes were seen as the DNA code for traits such as eye
color and body height, and possibly for characteristics like personality and
intelligence.
Modern biology uses a more specific definition for a gene, namely, as a heredi-
tary unit or as a specifically localized region of DNA that codes for a protein. The
direct function of a coding gene is therefore the protein that results from its tran-
scription and translation. By contrast, the assignment of higher-order function is
often questionable, because such functions usually involve many processes that are
executed by different proteins. It is therefore too ambiguous and vague to talk about
an “eye-color gene,” an “intelligence gene,” or an “Alzheimer gene.” Indeed, the
Gene Ontology (GO) Consortium [30], which assigns functions to genes, states that
“oncogenesis is not a valid GO term because causing cancer is not the normal func-
tion of any gene.” Instead, the GO Consortium uses three categories that character-
ize different molecular properties of gene products:
• The cellular component: this refers to parts of a cell or its extracellular environ-
ment where the gene product is located.
• Molecular function: this covers fundamental activities of the gene product, such
as binding or a specific enzymatic activity.
• Biological process: this refers to well-definable operations or molecular events,
but not to higher functions such as renal blood purification.
It is rare that a single gene is solely responsible for some macroscopic feature
or phenotype. Instead, the inheritance of complex, multifactorial traits such as
body mass, skin color, or predisposition to diabetes depends on the contribution
of many genes passed from parents to their children. This polygenic inheritance
makes genetic studies challenging, because it involves the co-regulation of sev-
eral genes and the control of possibly complex interactions between genes and
the environment through epigenetic effects (see later). These issues are currently
being investigated with analyses of quantitative trait loci (QTL), which are isolated
stretches of DNA that are associated with a phenotypic trait and may be distrib-
uted over several chromosomes [31] (see also Chapter 13). Many successes of
QTL analysis have been reported in plant breeding and crop development [32,
33], and also with respect to human diseases (see, for example, [34, 35]). How-
ever, while originally hailed as a new horizon in human genetics [36], QTL pose
the severe problem that they often explain only a modest percentage of pheno-
typic variability. In an investigation of human body size, 20 loci explained only 3%
of the variability in height among about 30,000 individuals [37]. It seems that
many traits, including chronic diseases, involve a large number of genes, whose
individual contributions are apparently very modest and therefore statistically
elusive [38].
If a gene is a specifically localized region of DNA that codes for a protein, is there
DNA that is not a gene? The answer is a resounding “Yes.” It is currently assumed
that humans possess between 20,000 and 25,000 genes. That sounds like a lot, but
only comprises an estimated 2% of our DNA! Much of the remaining DNA has
unknown functions, although one should be very careful calling it “junk DNA,” as it
was termed in the 1970s [39], because it is likely that the “junk” turns out to be good
for something. Indeed, the more we learn about this noncoding DNA, the more we
find that it has intriguing and very substantial roles in living cells, including regula-
tory roles at the transcriptional and translational levels. That these long stretches of
DNA are not at all useless is also supported by the observation that some of them are
remarkably stable throughout evolution and are even conserved between diverse
species such as humans and pufferfish, which implies that they are under strong
selective pressure and therefore most probably rather important [7]. At the same
time, other stretches of DNA are not well conserved between species, which simply
implies that we do not really understand the hidden signals in some noncoding
DNA and their role in gene regulation. Current research therefore focuses on identi-
fying markers for locations of noncoding elements and on characterizing their spe-
cific functions. In some cases, we have a vague idea of their roles, as reflected in the
names of such DNA stretches—promoters, enhancers, repressors, silencers, insula-
tors, and locus control regions, for example—but we have yet to develop a complete
inventory.
Our current understanding of the organization of DNA is as follows [19]. The vast
majority of DNA is noncoding and only a few percent of DNA codes for the synthesis
of proteins in eukaryotes; this percentage increases in less advanced organisms. The
average gene size in humans is about 3000 base pairs, but varies widely; the largest
known human gene is dystrophin, with 2.4 million base pairs. Coding regions are
usually rich in pairs of guanine and cytosine (GC pairs), which are more stable than
AT pairs owing to their higher numbers of hydrogen bonds (see Figure 6.3). The pat-
tern of distribution of genes within DNA cannot so far be explained and appears to
be random in humans. Other species have more regulatory patterns of genes and
noncoding regions. Long stretches of DNA containing lots of G’s and C’s apparently
flank genes and form a barrier between coding and noncoding regions. The role of
noncoding DNA is so far insufficiently understood, but some of this DNA is sus-
pected to serve regulatory roles.
The prediction of genes from sequence data is still a challenge. One piece of
evidence for identifying a gene is the open reading frame (ORF), which is a stretch
of DNA that does not contain an in-frame stop codon. The presence of a start
codon and a following sufficiently long ORF are often used as an initial screening
tool for potential coding regions, but they do not prove that these regions will be
transcribed and translated into protein. The challenges of gene prediction and
annotation are greatly magnified in the new field of metagenomics, in which all
genomes of large metapopulations are analyzed simultaneously. Such metapopu-
lations may consist of hundreds if not thousands of microbial species, which
coexist in environments such as the human gut, the soil, and the oceans (see
Chapter 15).
Even within a eukaryotic gene, the DNA often contains regions that are not
translated into protein. These so-called introns are typically transcribed, but later
removed in a process called splicing, whereas the remaining regions of the gene,
the exons, are merged into one piece of message and ultimately translated into
protein. The splicing process, which is controlled by a variety of signaling mole-
cules, permits the translation from different combinations of RNA pieces and
thereby increases the possible number of resulting proteins (alternative splice vari-
ants) considerably (Figure 6.8). It is currently believed that some introns are tran-
scribed as RNA genes such as microRNAs, which can regulate protein-encoding
gene expression (see later).
exon intron exon intron exon

gene
transcription
pre-mRNA
splicing
Figure 6.8 Exons and introns. Many
eukaryotic genes contain introns, which
RNA
ultimately do not correspond to amino acid
sequences in the protein for which the gene
translation codes. The splicing mechanism at the mRNA
level permits alternative combinations and
protein greatly increases the possible number and
variability of proteins.
Over the past decades, molecular biology and bioinformatics have developed an
enormous repertoire of laboratory techniques and computational tools for deter-
mining DNA sequences for prokaryotic and eukaryotic species, and even for indi-
vidual humans [40], as well as for comparing DNAs from different species. These
sequencing capabilities have allowed us to establish gene inventories for many spe-
cies, along with their variations; notable, very comprehensive databases with this
information include [19, 30, 41–43]. Furthermore, the journal Nucleic Acids Research
annually publishes a special issue on biological databases. The relative ease of DNA
sequencing has also led to much more accurate phylogenetic trees and other evolu-
tionary insights than had been possible before. The methods for sequencing and
sequence comparisons are fascinating; they are detailed in many books (for exam-
ple, [16, 17, 28, 44]) and uncounted technical papers and review articles. An ongoing
major challenge is the annotation of genes, that is, the identification of the function
of a coding region. The Gene Oncology Consortium [30] collects methods and
results of such efforts.
6.4 Eukaryotic DNA Packing

In prokaryotes, most of the DNA is arranged in a single chromosome, but some
may be located separately, forming plasmids, as we discussed before. These plas-
mids can be shared among members of the same species or even other species,
which leads to a potentially wide distribution of DNA and is a key mechanism of
evolution. In higher organisms, DNA is organized in a species-specific number of
chromosomes, which consist of DNA together with proteins that help package
DNA efficiently and contribute to the control of gene transcription. Of particular
importance are histones, which are spherical proteins around which DNA is
wound (Figure 6.9), and protein scaffolds that are used for the DNA packaging
process (Figure 6.10) [45]. A basic unit of about 150 DNA base pairs, wound around
a histone core, is called a nucleosome (Figure 6.11; see also Figure 6.9). The
nucleosomes are connected by stretches of about 80 base pairs of free DNA and
coiled into higher-order structures, which eventually form chromatin, which
makes up the chromosomes. The component chrom in these terms is Greek for
color and reflects the fact that chromatin, when stained, becomes visible under a
light microscope. The repeating pattern of nucleosomes and free DNA, which can
be seen with an electron microscope, permits the astonishing packaging of about
2 m of eukaryotic DNA into a nucleus with a diameter of only 10 μ. At the same
time, this packaging allows selective access to the DNA for transcription, which is
controlled by chromatin remodeling motors that consist of proteins and use ATP
for energy [46].
6.5 Epigenetics
We are just at the beginning of understanding how the environment, diet, and
diseases can affect transcription through epigenetic processes that influence
where and when genetic information is to be used (see Figure 6.9) [47]. In these
processes, small molecules such as methyl groups can be attached to DNA as
adducts and lead to heritable modifications of the chromatin. This means that it is
possible that environmental factors affecting an individual can still be found in
children and even grandchildren. The DNA adducts can allow the opening DNA
for transcription or silence gene expression, and such processes have been associ-
ated with various diseases and all stages of tumor development, from initiation to
invasion and metastasis. Fortunately, it seems possible, at least in principle, to
reverse adverse epigenetic effects through clinical treatment and healthy living
(see Figure 6.9C) [48].
RNA
RNA and DNA differ only slightly in their basic biochemical properties: the sugars
in their backbone differ by just one oxygen atom, and one of the four bases is
RNA 179
(A) active
+ + + +
+
nucleosome
+ H3K4 methyl mark

+ + + H3K9/H3K27 methyl mark
+ + +
X
acetylation mark
(B) inactive unmethylated CpG site

DNMT DNMT
methylated CpG site
HDAC HDAC
MBD MBD MBD MBD
(C) epigenetic therapy and/or dietary factors
DNMT
HDAC MBD
MBD
+ + + +
+
+ +
+ +
Figure 6.9 Diagram illustrating DNA organization around histones and epigenetic effects influencing the expression of genes close to a
nucleosome (blue spheres). (A) Promoters of active genes are often associated with CG pairs that are not methylated (white circles), with acetylated
histones (green crosses), and with methylated histone H3 (yellow hexagons); this configuration is conducive to transcription. (B) In some diseases such
as cancer, CG sites in the promoter region of tumor suppressor genes are often methylated (red circles). Methyl-CG-binding domains (MBDs), histone
deacetylases (HDACs), and DNA methyltransferases (DNMTs) make chromatin inaccessible and lead to repression of transcription. (C) Epigenetic therapy
could potentially decrease MBDs, HDACs, and DNMTs and reverse their deleterious effects. (From Huang Y.-W, Kuo C.-T, Stoner K, et al. FEBS Lett. 585 [2011]
2129–2136. With permission from Elsevier.)
thymine in DNA and uracil in RNA (see Figure 6.2). However, in spite of their close
similarities, DNA and RNA have very distinct roles in biology. DNA is the typical
carrier of genetic information, whereas different types of RNA have diverse roles
as intermediate information transducers and as regulators. One key difference is
that, unlike DNA, most RNA molecules are single-stranded. Because these single
RNA strands are not matched up with partners, as in the case of the DNA double
helix, they are small in size, or short-lived, or exhibit complex two- and three-
dimensional structures, collectively called secondary structure, where some of
the nucleotides form bonds with other nucleotides of the same RNA strand
(Figures 6.12 and 6.13).
6.6 Messenger RNA (mRNA)

For most organisms, mRNA provides the means of converting the DNA blueprint
into proteins. In this transcription process, DNA opens up and serves as a template
Figure 6.10 The tight packaging of DNA

is accomplished with special proteins.
Crystal structure of a superhelical protein
scaffold that supports the packaging of
DNA in the gut bacterium Salmonella
typhimurium. (Protein Data Bank: PDB
3NR7 [43].)
for RNA, which is generated as a new strand in which one nucleotide at a time is
added to match the corresponding nucleotide of the DNA strand. This transcription
requires an enzyme, called RNA polymerase, which can be extensively regulated by
transcription factors, as we will discuss later.
In eukaryotes, transcription occurs within the nucleus. The new RNA is trans-
ported from the nucleus to the cytosol, where ribosomes convert it into a string of
amino acids, which form the primary structure of proteins (see Chapter 7). This
translation process requires two other types of RNA, namely transfer RNA (tRNA)
and ribosomal RNA (rRNA).
(A) (B)
Figure 6.11 Images of a nucleosome. DNA

(black) is wound around a protein complex,
consisting of histones H2A (red), H2B
(yellow), H3 (green), and H4 (blue). (A) Side
view. (B) Top view. (From Narlikar GJ. Curr.
Opin. Chem. Biol. 14 [2010] 660–665. With
permission from Elsevier.)
RNA 181
77
(A) G G (B) C A 36
C
G C A
G
A U C
C G A A
C G A U C C
U A G C G
G A A
G C U U G
G C U C G
U C G
C U G
U
2191 G C C A
2171 U
U A
C A G
G
A A A C A
U C
C C G
C G A
1 G
A G C 3 65
97
G C
C A 36
A G G U C
A G U
A G
U A U
G C
C C G G A
U G
U U C U
G U,C
U C G A G
C G
G G G C U C A A A
A G U A
U A C 3883 C
G A A
U C 3717 3737
G
U
G C C 3638
A C A A G
A A C G C A U A A
2211 U A G 3877 A,C A A A A
G A U G C U U U G A, C
G G A U A
A U U C G U G A
G A U C C C U U G U G U G A G U
U C U G U C C C C U A C C U
A C C C A G C U U U C G U C A
G C C A A G U A U C G C
A
G C
C
2151 C G G G G G G G A C C C G A
G C U A U U U C
G A A G A U
A C U U G A G G G 3757 U G
C U A A 3837 G U C A 7 A
3857 G C C A 77 A
G U G A A C U C A
C U G 3 A
A A C A C C C C
C GG G G G U U C C U A A A
C G 37 G
2231 C G U 97 U
A U C U
C C U
G U G C U G C U
U
G C G A U A
G C G U
A
G U G
G A
G C U A
A A
G A
A G G C U A
C,G C C
2150 A U G
38 U
17 U A
G C
G A U A
2131 U
C C
U,
C, G
Figure 6.12 Secondary RNA structures. Shown here are structures of (A) a putative promoter and (B) the untranslated region at the 3′ end of the
Pelargonium line pattern virus. (From Castaño A, Ruiza L, Elenaa SF & Hernández C. Virus Res. 155 [2011] 274–282. With permission from Elsevier.)
5′ 166
C
A
G C A
5′
G C 120 A
J2a/3 (loop 2)
P2b (stem 1)
95 G C C
C G A
U A A
G C A
U U A
100 U U A 175
U U A
U
J2b/3 (loop 1)
U C G
P3 (stem 2)
A U
U
110 G C 180 Figure 6.13 Single-stranded RNA often
C folds onto itself. The folding leads to
U A bonds that give the molecule the stable
105 U
C G two-dimensional and three-dimensional
C structure of a pseudoknot. Different colors
G C of nucleotides indicate their association with
stems (red and blue), and two loops (cyan
A 3′ and magenta). (From Yingling YG & Shapiro
BA. J. Mol. Graph. Model. 25 [2006] 261–274.
3′ With permission from Elsevier.)
polypeptide Figure 6.14 Schematic representation of

the translation of messenger RNA (mRNA)
into a polypeptide. A ribosome moves
along the mRNA (green). Corresponding to
empty tRNA with each codon on the mRNA, a transfer RNA
tRNA amino acid
(tRNA; yellow) with the appropriate amino
acid (blue octagons) is recruited. It enters
the ribosome and permits the binding of
its attached amino acid to the growing
polypeptide chain (blue chain).
mRNA
ribosome
6.7 Transfer RNA (tRNA)

This type of RNA is relatively short, consisting of only about 80–100 nucleotides. It is
synthesized from so-called RNA genes, which are usually located within stretches of
noncoding DNA. The role of each tRNA is to transfer a specific amino acid to the
growing peptide chain that is being constructed by a ribosome (Figure 6.14). Each
tRNA contains a triplet of nucleotides that corresponds to a sequence of three DNA
nucleotides, that is, a codon, which “codes” for the target amino acid (see Chapter 7
for a more detailed discussion of codons). This amino acid attaches to the 3′ end of
the tRNA and, once associated with the ribosome, is covalently bound to the already
formed polypeptide, a process that is catalyzed by the enzyme aminoacyl tRNA
synthetase.
Each tRNA is single-stranded, but forms bonds between its own nucleotides that
result in a secondary cloverleaf structure. This structure consists of three stems with
loops, a variable loop, and an acceptor arm that permits the amino acid to attach.
The center arm contains the anticodon region, whose three nucleotides match the
codon on the mRNA and ensure that the correct amino acid is selected. To fit into
the ribosome for translation, the tRNA must bend at its variable loop into an L-shape
(Figure 6.15).
6.8 Ribosomal RNA (rRNA)

Ribosomes facilitate the process of translating mRNAs into chains of amino acids.
They consist of proteins and rRNAs and are composed of two subunits (see Figure
6.14). The smaller one attaches to the mRNA and the larger one associates with the
tRNAs and the amino acids they carry. When reading of an mRNA is completed, the
two subunits separate from each other. Functionally, ribosomes are considered
(A) acceptor arm (B) T arm T stem acceptor stem (C)

5′ 3′
acceptor T arm
3′
stem T loop acceptor
D stem T stem 5′
arm
G15 D loop acceptor arm
G15
D arm
T loop G15
D stem D arm
T arm
variable loop
anticodon arm
D arm variable loop

anticodon stem anticodon stem
D loop
anticodon loop
anticodon loop
anticodon arm anticodon arm
Figure 6.15 Cloverleaf structure of a transfer RNA (tRNA). (A) This secondary structure consists of four arms. The acceptor arm allows the attachment
of a specific amino acid, whereas the opposing arm contains the matching anticodon, which mirrors the corresponding codon on the mRNA. (B) To serve
its role within the ribosome, the tRNA has to bend at its variable loop into an L-shape. (C) The tRNA’s three-dimensional structure. At position G15, this
particular tRNA has been modified. (From Ishitani R, Nureki O, Nameki N, et al. Cell 113 [2003] 383–394. With permission from Elsevier.)
RNA 183
Figure 6.16 Crystal structure of a rather

simple “hammerhead” ribozyme from
the satellite RNA of the tobacco ringspot
virus (sTRSV). This molecule was the first
hammerhead ribozyme discovered. It uses
a guanine residue as a general base during
RNA catalysis. The structure shows how
the single-stranded RNA folds and binds to
itself. Also seen are guanosine triphosphate
(balls and sticks) as a modified residue
and magnesium ions (spheres) as ligands.
(Protein Data Bank: PDB 2QUS [49].)
Figure 6.17 Structure of a ribozyme. The

atomic structure of the large ribosomal
subunit from the Dead Sea archaeon
Haloarcula marismortui shows that the
ribosome is a ribozyme, which catalyzes
the formation of peptide bonds during
translation. (Protein Data Bank: PDB 1FFZ
[50].)
enzymes, belonging to the class of ribozymes, because they are catalysts for the
linking of amino acids through peptide bonds (Figures 6.16 and 6.17).
6.9 Small RNAs

Several types of small, double-stranded RNAs have been discovered and character-
ized in recent years [51]. Because of their small size, they had been overlooked for a
long time, but it is now clear that they play important roles in numerous fundamen-
tal biological processes and diseases [52]. Small RNAs fall into several classes, the
most prominent of which are small-interfering RNA (siRNA) and micro-RNA
(miRNA). Interestingly, RNA interference (RNAi) pathways involving these types of
RNAs are highly conserved throughout evolution, which underscores their central OH
importance from a different angle. The first discovery of RNAi was actually made in
an effort to enhance the flower patterns of petunias [53], demonstrating once again P
that one can never foretell where some seemingly obscure scientific endeavor may
eventually lead.
Small interfering RNA goes by different names, including silencing RNA and
short interfering RNA, both of which luckily permit the same abbreviation siRNA.
These short, double-stranded RNA molecules consist of 20–30 nucleotides per
strand, of which usually 21–25 form base pairs, while the remaining two or three
stick out at either side and are not bound (Figure 6.18). The importance of siRNAs
derives from the fact that they can interfere with the expression of genes and even
silence it. This interference constitutes a potent regulatory mechanism that allows
the cell to slow down or shut down translation at appropriate times [54].
Once introduced into cells, siRNAs permit the silencing of precisely targeted
gene transcripts. The main mechanism seems to be the degradation of mRNA. This
RNAi process begins with long double-stranded RNA, which is processed by an
enzyme called dicer into small, sequence-specific siRNA. This siRNA triggers an
P
RNA-induced silencing complex (RISC), which is an enzyme that catalyzes the
cleavage of the target mRNA. Because siRNAs can be constructed artificially, the OH
discovery and experimental mastery of RNAi has led to an enormous repertoire of

Figure 6.18 Schematic representation of
tools for gene expression studies, including so-called loss-of-function techniques,
a small interfering (siRNA) molecule. A
and to a new class of potentially powerful therapeutics. typical siRNA consists of 21 base pairs and
Other small RNAs have been discovered in recent times. Similar to siRNAs are overhangs at both ends. P indicates the
micro-RNAs (miRNAs), which have also been called small temporal RNAs (stRNAs). phosphate group at the 5′ end, while OH
Like siRNAs, miRNAs are produced from double-stranded RNA by a dicer enzyme. represents the hydroxyl group at the 3′ end.
Their function is not quite clear, but it seems that they negatively regulate the expres-
sion of target transcripts. Apparently, miRNAs modulate protein expression through
(a) a slicer mechanism for high-homology mRNA targets, (b) slicer-independent
cleavage or degradation of mRNA, or (c) inhibition of translation without apprecia-
ble changes in mRNA levels [51]. It has been estimated that the human genome
might encode over 1000 miRNAs, which potentially target half of all genes. This
widespread influence is apparently possible because the homology between miR-
NAs and mRNAs does not have to be perfect. This imperfect matching is a significant
difference between siRNAs and miRNAs in animals [55].
The discovery of miRNAs has led to exciting new insights and options for manipu-
lating biological systems and possibly the treatment of cancer [56, 57]. However, one
must be careful not to jump to conclusions regarding potential drug treatments too
quickly, because the function of miRNAs in complex systems is not predictable from
silencing studies executed in simple experimental systems. The reason is that an
miRNA may silence some gene x whose normal function is down-regulation of
another gene y, so that the effect of the miRNA on gene y may be activating rather than
inhibiting. In fact, to understand the full impact of an miRNA it is necessary to con-
sider the entire system of direct and indirect influences of the miRNA on the genome.
Double-stranded RNAs can also activate gene expression, which has led to the
term small activating RNA (saRNA). Finally, so-called piwi-interacting RNAs (piR-
NAs) play an important role in germ-line development. It seems that these RNAs may
silence transcription by regulating DNA methylation [52]. The characterization of
small RNAs and their functional pathways is the subject of intense ongoing research.
Recently, Lai and collaborators reviewed numerous articles demonstrating how
our understanding of the roles of miRNAs for gene regulation can be enhanced
through mathematical modeling [58, 59].
6.10 RNA Viruses

As an exception to the Central Dogma, many viruses use single-stranded RNA
instead of DNA as their inheritable material; some even use double-stranded RNA
(Figure 6.19). The single strand is arranged either in a positive sense, which means
that it acts like a messenger RNA that had been transcribed from DNA, or in a nega-
tive sense, in which case the strand is first transcribed into a matching positive
strand. RNA viruses mutate very easily, because their polymerases do not possess
the proof-reading ability of most DNA polymerases. Thus, transcription errors go
GENE REGuLATiON 185
Figure 6.19 Electron micrograph of an

RNA-containing coronavirus. The virus
causes severe acute respiratory syndrome
(SARS), a disease that rapidly spread around
the world in 2003. Each virus is about 100 nm
in diameter. (From Murray CL, Oha TS & Rice
CM. Keeping track of viruses. In Microbial
Forensics [B Budowle, SE Schutzer, RG Breeze,
et al. eds], pp 137–153. Academic Press, 2011.
undetected, but are advantageous in a sense that some of them allow the virus to
develop resistance to environmental or pharmacological stressors. The positive
strand of RNA is directly translated into a single protein, which the host cell cuts and
modifies so that all proteins for virus replication are obtained.
GENE REGuLATiON
While essentially all cells of a multicellular organism contain the same genes, it is
obvious that not all cells execute the same functions. The key to this variability is the
coordinated expression of different sets of cell-type-specific genes at the appropri-
ate times. But even unicellular organisms do not continuously transcribe all genes,
and instead carefully regulate expression in order to satisfy the current demands for
metabolites and for the enzymes that are needed for their production. Because most
systems are simpler in microorganisms, very many investigations characterizing the
expression of genes have targeted model species such as the bacterium Escherichia
coli and the baker’s yeast Saccharomyces cerevisiae, and a vast body of information
has been assembled. And indeed, many of the fundamental concepts of gene regu-
lation can be learned best from simpler organisms, as long as one keeps in mind that
higher, multicellular organisms use additional means of control and regulation.
A major breakthrough in our understanding of the regulation of gene expression
was the discovery of operons in bacteria in the 1950s. According to the original
concept, each operon consisted of a defined stretch of DNA that contained at its
beginning at least one regulatory gene that allowed the bacterium to control the
expression of all other genes in the operon [60]. This view has been slightly modified
in recent years, because it turned out that an operon does not necessarily contain a
regulatory gene. Instead, the promoter region of an operon generally has
multiple cis-regulatory motifs that can bind with either repressors or inducers.
The regulation of an operon can occur through repressors or inducers.
Repressors are DNA-binding proteins that prevent the transcription of genes. These
repressors are themselves the products of genes, and this fact creates hierarchies of
regulation that we will discuss later. Repressors often have binding sites for specific
non-protein cofactors. If these are present, the repressor is effective in preventing
transcription; if they are absent, the repressor dissociates from the operator site, and
the genes in the operon are transcribed. Inducers are molecules that initiate
gene expression. An example is the sugar lactose, which functions as an inducer for
the so-called lac operon, which is responsible for lactose uptake and utilization.
Lactose inactivates the repressor and thereby induces transcription. Other exam-
ples include certain antibiotics, which can induce the expression of genes that make
the organism resistant [61, 62].
6.11 The lac Operon

It is illustrative to study the unit structure of an operon within the genome of an
organism in more detail. The first well-characterized example of this structure was
indeed the lac operon in E. coli [63]. Its discovery was a true breakthrough in genet-
ics and molecular biology, and François Jacob, André Michel Lwoff, and Jacques
Monod were awarded the 1965 Nobel Prize in Physiology or Medicine for their dis-
coveries associated with this and other operons.
The lac operon contains three structural genes that code for enzymes associated
with the uptake and utilization of the disaccharide (double-sugar) lactose, which
chemically is a β-galactoside and prevalent in mammalian milk. These enzymes are
β-galactoside permease, which is a transport protein responsible for the uptake of
lactose from the medium (gene 22acy), galactose transacetylase, which transfers an
acetyl group from acetyl coenzyme A (acetyl-CoA) to β-galactosides (gene lacA),
and β-galactosidase, an enzyme that converts lactose into the simpler sugars glu-
cose and galactose (gene lacZ). In addition to these structural genes, the lac operon
contains a promoter region P and an operator region O (Figure 6.20). Also of great
importance are a regulator gene lacI and a binding region for the catabolite activa-
tor protein (CAP). The names of these components are mainly historical.
The bacterium only transcribes the structural genes if their products are needed.
The cell manages the decision of whether to transcribe or not to transcribe with a
multicomponent regulatory mechanism. If lactose is available in the environment,
it serves as an inducer. It binds to the repressor, which is a protein that is coded for
by a regulatory gene. In the case of the lac operon, the regulatory gene is directly
adjacent to the operon, but in other cases it can be located quite distant from it. The
repressor is normally bound to the operator region. However, binding of the inducer
to the repressor causes the repressor to dissociate from the operator region. Without
the repressor bound to the operator, the enzyme RNA polymerase, which synthe-
sizes RNA, can in principle attach to the promoter region and transcribe the struc-
tural genes into mRNA. However, the promoter region often also contains response
elements in the form of short DNA sequences that permit the binding of transcrip-
tion factors under the appropriate conditions. Transcription factors recruit RNA
polymerase and are of utmost importance for the regulation of gene expression.
They will be discussed later in more detail.
In the case of the lac operon, the response elements react indirectly to high con-
centrations of glucose in the medium. Thus, glucose can repress the transcription of
glucose cAMP lactose
CAP RNAP R
lacI lacZ lacY lacA
DNA
CAP P O
BS
repressor β-galactosidase permease transacetylase gene product
R
lac operon
Figure 6.20 Diagram of the structure and function of the lac operon. The coding genes lead to the production of a repressor and three enzymes. Small
molecules such as glucose and lactose affect expression. See the text and Table 6.1 for details.
GENE REGuLATiON 187
TABLE 6.1: FUNCTION OF THE lac OPERON UNDER DIFFERENT

ENVIRONMENTAL CONDITIONS
Lactose Glucose Repressor CAP BS Promoter Operator Transcription
⇓ ⇓ U B F R off
⇑ ⇓ I B B F on
⇑ ⇑ I F F F off
⇓ ⇑ U F BF R off
Key: Arrows indicate high (⇑) and low (⇓) concentrations in the environment. I, induced; U,
uninduced; B, bound (in the case of the CAP-binding site, CAP BS, by CAP facilitated by cAMP, whose
concentration is inversely related to the glucose concentration in the medium; in the case of the
promoter P by RNA polymerase); F, free of binding partner; R, repressed. Transcription occurs only
under high-lactose and low-glucose conditions.
the lac genes. This repression is advantageous because the bacterium saves energy
by consuming the more easily digestible glucose rather than lactose. This repression
is accomplished in the following fashion. As a consequence of complex metabolic
and signaling mechanisms, the cell responds to environmental glucose levels by
regulating its internal level of cyclic adenosine monophosphate (cAMP). For low
glucose levels, cAMP is high, and for high glucose levels, cAMP is low. The signaling
molecule cAMP binds to CAP, which attaches to a specific binding site right next to
the promoter region (CAP BS in Figure 6.20). This binding permits the attachment of
RNA polymerase to the promoter and transcription of the structural lac genes, if
lactose is present. By contrast, if glucose in the medium is high, the bacterium does
not bother with lactose and waste energy synthesizing enzymes, and rather uses the
glucose. The high glucose concentration results in low cAMP, RNA polymerase does
not attach to the promoter region, and the organism does not transcribe the lac
genes. Taken together, the system expresses the structural genes only if lactose is
available but glucose is not (Table 6.1).
Several groups have developed detailed dynamic models of the lac operon (see,
for example, [64–67]). Each model focuses on a particular question, is correspond-
ingly structured, and has a different level of complexity, which again shows that bio-
logical questions and data determine the type, structure, and implementation of a
model (Chapter 2).
6.12 Modes of Regulation

In general, the regulation of an operon may occur through induction or repression,
and it may be positive or negative, so that there are four base forms of controlling the
initiation of gene transcription.
Positively inducible operons are controlled by activator proteins, which under
normal conditions do not bind to the regulatory DNA region. Transcription is initi-
ated only if a specific inducer binds to the activator. The inducer changes the con-
formation (three-dimensional shape; see Chapter 7) of the activator, which allows
its interaction with the operator and triggers transcription.
Positively repressible operons are also controlled by activator proteins, but these
are normally bound to the regulatory DNA region. When a co-repressor protein
binds to the activator, the activator no longer binds to the DNA and transcription is
stopped.
Negatively inducible operons are controlled by repressor proteins, which are
normally bound to the operator of the operon and thereby prevent transcription.
Specific inducers can bind to repressors, change their conformation so that binding
to DNA is no longer effective, and thus trigger transcription.
Negatively repressible operons are normally transcribed. They have binding
sites for repressor proteins, which, however, only prevent transcription if specific
co-repressors are present. Binding of the repressor and co-repressor leads to attach-
ment to the operator and prevents transcription.
The type of regulation found in each particular case appeared to be random at
first, but a later, careful analysis led to clear design principles dictating the mode of
control (see also Chapter 14). The discovery and interpretation of these design prin-
ciples resulted in the demand theory of gene regulation, according to which there
are strict rules for when a particular mode of regulation is beneficial and therefore
evolutionarily advantageous [61, 68]. As one example, a repressor mode is advanta-
geous if the regulated gene codes for a function that is in low demand within the
organism’s natural environment. By contrast, if the function is in high demand, the
corresponding gene is under activator control.
Much of what we know about operons and their functions was determined
through uncounted laboratory experiments, beginning even before the work of
Jacob and Monod [63]. By contrast, much of what we are now beginning to under-
stand about the higher-order organization of bacterial genomes has been deduced
with the help of sophisticated computational approaches. These approaches fall
mostly within the realm of bioinformatics, and we will only discuss some pertinent
results without describing the methods themselves in detail. The interested reader
is encouraged to consult [44].
In prokaryotes, all genes are located on a single chromosome, which consists of
a closed circle of double-stranded DNA, and on much smaller plasmids, which can
be transferred from one bacterium to another, as we discussed before. The chromo-
some usually contains a few thousand genes, while a typical plasmid contains
between a few and several hundred genes. In higher organisms, nonchromosomal
DNA is furthermore found in mitochondria and chloroplasts.
6.13 Transcription factors

The most prominent regulators of gene expression are transcription factors (TFs),
which include the repressors and activators we discussed before and which must
attach to a binding site in the vicinity of the promoter region to execute their func-
tion. TFs may up- or down-regulate the operon; some dual-regulator TFs can do
both, depending on the conditions. A TF is often activated through phosphorylation
by a kinase, which can “sense” specific changes in the intra- or extracellular
environment [69]. The kinase and the TF therefore constitute a two-component
signaling system that permits the cell to respond adequately to changes in the
environment. Chapter 9 discusses such systems in greater detail.
Typical bacterial genomes contain relatively few genes for TFs in comparison
with their number of operons, and while some genes can be expressed without TFs,
the mismatch between the numbers of TFs and operons has led to the inference
that at least some TFs must be responsible for several operons [69]. The collection
of all operons under the control of the same TF is called the regulon of the TF.
Because an operon may be regulated by several TFs, they may belong to different
regulons as well. Evidence suggests that bacteria tend to keep operons encoding
the same biological pathway, or even systems of related pathways, in nearby
genomic locations [70].
The prediction of TF-binding sites and regulons from sequence information is a dif-
ficult problem, but many algorithms have been developed for this purpose, and
numerous databases are available that contain information on known TF-binding sites
and regulons (see, for example, [69, 71]). In addition to regulons, bioinformaticists have
discussed über-operons (über is German for “above”), which are sets of genes or oper-
ons that are regulated together even after many rearrangements of the DNA of the
organism [72]. Other concepts include stimulons [73], which are sets of genes that
respond to the same stimulus, and modulons, which are regulons associated with sev-
eral pathways or functions that are under the control of the same regulator [74].
Responses to environmental stresses often require more than the up- or down-
regulation of a single operon. In fact, a complex response to a stress such as starva-
tion may consist of changes in the expression of dozens of genes. The coordination
of such responses therefore mandates means for co-regulating genes in different
locations of the chromosome. The premier mechanism for this coordination is a
hierarchy of TFs. Conceptually, such a hierarchy is not difficult to understand. Let us
suppose that six operons O1, …, O6 need to be called up for a proper response and
that they are under the control of one transcription factor each, which we just call
TF1, …, TF6 for simplicity. These TFs are proteins, which are coded for by their own
genes G1, …, G6. If these genes are under the control of a TF at the next level of the
GENE REGuLATiON 189
hierarchy, say TFA, then a single manipulation of TFA will result in expression
changes of all six target operons.
A complex example of this hierarchy is the control of processes converting sug-
ars into the amino acids lysine and glutamate in Corynebacterium glutamicum [75]
(Figure 6.21). Another impressive example is the gene regulatory network control-
ling metabolism in E. coli, which consists of about 100 TF genes that control about
40 individual regulatory modules. These modules can be triggered by various exter-
nal conditions and organize proper responses and control the expression of about
500 metabolic genes that code for enzymes [76, 77].
The hierarchical and context-specific mode of TF control is not restricted to bac-
teria, but can be found throughout evolution. An intriguing example is the analysis
of transcriptional networks that regulate transitions between physiological and
pathological states. For instance, the activation of genes in malignant brain tumors
was found to be governed by about 50 TFs, of which five regulate three-quarters of
the brain tumor genes as activators and one as repressor. Two TFs serve as master
regulators of the system [78]. Algorithms have been developed to study these types
of hierarchical TF networks [79].
Figure 6.21 Transcription is often

hierarchically controlled. Shown here is the
hierarchy of transcription regulators involved
in the conversion of glucose, fructose,
and sucrose into the amino acids lysine
and glutamate in C. glutamicum. Pertinent
metabolic pathways are indicated as black
arrows with corresponding gene names.
(From Brinkrolf K, Schröder J, Pühler A &
Tauch A. J. Biotechnol. 149 [2010] 173–182.
The responses of TFs to environmental stimuli are typically achieved through

small molecules that ultimately act as inducers for the TF genes. For example, a par-
ticular sphingolipid in yeast, which otherwise has no recognized function, was found
to regulate several dozen genes associated with the response to heat stress by influ-
encing 11 TFs positively and 2 negatively [3]. Moreover, groups of sphingolipids were
associated with the expression of clusters of genes with coherently related functions.
The alteration of TFs in response to environmental stimuli is not a one-time,
all-or-nothing event. Instead, TFs can be altered differentially with respect both
to magnitude and to time periods following a stimulus. For instance, in the
sphingolipid-induced change in TFs, clear time domains for different groups were
distinguishable, with some groups being turned on and others turned off about
10–15 minutes after the beginning of heat stress [3].
Numerous other mechanisms can regulate the activity of TFs; for a detailed
description, see [80]. Cells can sequester the regulator of some physiological function
if this function is not needed at the time. For instance, if a sugar like maltose is not
available in the medium, the regulator controlling the utilization of maltose is dis-
abled. Mechanisms such as phosphorylation can control the activity of TFs spatially
and temporally. Similarly, it is possible to fine-tune the cellular concentration of a TF
through its expression and degradation. Finally, a TF or regulator may be translocated
within the cell, for instance from the cytosol to the membrane in prokaryotes or
between the nucleus and the cytosol in eukaryotes, thereby changing its activity state.
6.14 Models of Gene Regulation

Gene regulation may be approached with distinct computational modeling tech-
niques. Different types of examples can be found in [79, 81–85] and excellent detailed
reviews are provided in [86–88]. Generically, the approaches fall into four categories.
First, one may look at the structure of a gene interaction network, which is
assessed as static, that is, the focus is on a fixed connectivity among genes under
defined conditions. We discussed graph methods and methods such as Bayesian
network inference in Chapter 3; see also the discussion of Boolean methods in
Chapter 9. In this type of approach, the roles of TFs and their interactions are only
implicit, although they are crucial, because genes do not interact directly, but by
means of their products, including proteins that serve as TFs or their modulators
and small signaling molecules, such as calcium, cAMP, and sphingolipids. Large-
scale gene interaction networks of thousands of genes and their interactions have
been reconstructed with experimental and computational means for model species
such as yeast and E. coli [89, 90].
A second approach is currently possible only with a much smaller scope. Here the
goal is to analyze the dynamics of gene regulatory networks. The typical techniques
involve various types of differential equations, including nonlinear, piecewise-
linear, or power-law-based ordinary differential equations (ODEs), as well as qual- (A)
itative differential equations, which are abstractions of ODEs, and delay differential
gene 1 gene 2
equations, which permit the modeling of processes that are not instantaneous. In the
forward mode, the differential equation models are typically designed with mRNAs
and proteins as variables. Sometimes, metabolites are represented as well, because
they may affect gene expression as co-inhibitors or co-activators, and these modulat-
ing effects are often formalized as Hill or power-law functions. It is quite evident that gene 3
the sky is the limit, and that one quickly arrives at complex systems as we discussed
them in Chapter 4. Specific examples of models with different levels of sophistication
have already been mentioned in the context of the lac operon. (B)
The biological literature on gene regulatory networks often represents genes as
nodes and gene–gene interactions as edges, as shown in Figure 6.22 (see also Chap-
ter 3). Although such representations might seem to depict clearly how one gene (or mRNA1 mRNA2
gene transcript) affects another, they do not provide a satisfactory basis for the
design of mathematical models, since, as we have seen, gene–gene interactions are
Figure 6.22 Schematic representations
mediated through transcription factors and other modulators, which are absent of interactions between (A) genes or (B)
from these diagrams. In particular, these representations do not indicate whether, transcripts. While intuitive, these types
for instance, a process reduces the transcription into mRNA or hastens the removal of diagrams are not well suited for the
of an mRNA. A clearer representation of the two-gene regulatory system in Figure design of mathematical models. See instead
6.22B uses the instructions from Chapter 2. It is shown in Figure 6.23 and Figure 6.23.
MEASuRiNG GENE ExPRESSiON 191
degradation transcription Figure 6.23 A representation of the

mRNA1 diagram in Figure 6.22B that is better
+ – suited for designing a mathematical
+ model. This representation appears to
degradation
protein1 be more complicated, but it makes the
translation processes involved explicit and immediately
prescribes the set-up of model equations
such as (6.1).
translation degradation
protein2
+
+
mRNA2
transcription degradation
immediately suggests how to design a corresponding model. For example, equa-

tions in the following power-law format can be obtained directly from this diagram,
and the theory behind these functions tells us that all parameter values are positive
except for g2:
i
(mRNA2 ) = α 1P1g1 P2g 2 − β1 (mRNA1 )h1 ,
i
(mRNA2 ) = α 2 P1g 3 − β 2 (mRNA2 )h2 ,
i (6.1)
P1 = α 3 (mRNA1 )g 4 − β 3 P1h3 ,
i
P2 = α 4 (mRNA2 )g 5 − β 4 P2h4 ,
In the reverse mode, ODE models have been used to infer gene interaction net-
works from expression data that had been recorded as time series (see, for example,
[91] and Chapter 5). Partial differential equation (PDE) models have been used to
account for reaction and diffusion processes within the cell [92].
The third class of models consists of stochastic formulations, which permit the
modeling of variability in gene expression among individuals and scenarios [93]
and can, for instance, model outcomes in systems where gene expression can toggle
on and off in an apparently random fashion [94].
Finally, it is possible to use rule-based or agent-based models, which permit
simulations of very diverse and complex scenarios, but are difficult to analyze math-
ematically [95]. We will discuss such approaches in Chapter 15.
MEASuRiNG GENE ExPRESSiON

When a gene is expressed, it is transcribed into mRNA, which may later be trans-
lated into protein. Thus, measuring gene expression means measuring the emer-
gence of mRNA or protein. If protein is actually produced, its amount can be
assessed with a variety of methods, including a Western blot. This name was chosen
as a play on an earlier DNA test, which had been developed by Edwin Southern. In
the Western blot technique, a cell culture or tissue sample is homogenized and pro-
teins are extracted. The protein mixture is separated through gel electrophoresis
based on molecular weight or other criteria. The most common variant, SDS-PAGE,
uses buffers loaded with the detergent sodium dodecyl sulfate (SDS) for a polyacryl-
amide (PA) gel electrophoresis (GE). The SDS is negatively charged and covers the
proteins, which therefore move through the gel toward a positively charged elec-
trode. Furthermore, SDS-PAGE prevents polypeptides and proteins from forming
three-dimensional structures (see Chapter 7), and therefore allows a separation of
the proteins by molecular weight, because smaller proteins move through the acryl-
amide gel faster. The proteins in each band of the gel are transferred (blotted) to a
membrane, where they are probed by the binding of specific antibodies.
Thousands of antibodies are available for this purpose. These antibodies can be
detected in various ways. The most common detection method uses antibodies conju-
gated to enzymes (often horseradish peroxidase, HRP), which convert a substrate to a
light-emitting product that can be detected and quantified. Technically, the primary
antibody recognizes the protein of interest and is produced by a known species (for
example, mouse). A secondary antibody, which is enzyme-linked, recognizes the con-
stant region of the mouse antibody; an example is the goat anti-mouse IgG-HRP, which
is a goat antibody that recognizes a mouse antibody and is linked to HRP. This con-
struct allows for amplification of signal and the requirement for relatively few of the
more expensive enzyme-linked antibodies. As an alternative to enzyme linkage, the
antibodies can carry a fluorescent or radioactive marker that makes them detectable.
The antibody methods are semiquantitative, but calibration experiments can
make them quantitative and therefore allow inferences regarding the expression of
the genes that code for the measured proteins. The implicit assumption with respect
to gene expression is that the target genes actually code for proteins and that the
amounts of protein produced are proportional to the degree of gene expression.
An alternative to protein-based methods is the measurement of the relative abun-
dance of the transcripts (that is, the mRNA molecules) that correspond to the target
gene. The assumption in this case is that the number of mRNA molecules is (linearly)
proportional to the degree of gene expression. For about a quarter century, the base
technique for this purpose has been the Northern blot (Figure 6.24A, B). Again, this
name was chosen in contrast to Southern’s DNA test. The Northern blot procedure
also uses gel electrophoresis, which in this case separates the mRNA molecules in a
tissue sample based on their sizes. Subsequently, complementary RNA or DNA is
applied. This RNA or DNA is labeled with a radioactive isotope such as 32P or with a
chemiluminescent marker. The probe is conjugated to an enzyme, which metabolizes
the inactive chemiluminescent substrate to a product that emits light. If the labeled
RNA or DNA finds a matching partner on the gel, it binds (hybridizes), thereby form-
ing a double-stranded RNA or an RNA–DNA pair. After washing off unbound RNA or
DNA, the chemiluminescence is detected. Alternatively, a radioactive label can be
detected with autoradiography by exposing an X-ray film, which turns black when it
receives radiation from the label. The degree of blackening is directly correlated with
the amount of RNA, which makes the method quantitative. A database contains hun-
dreds of blots for comparisons, searching, and identification [96].
A newer method of mRNA quantification is reverse transcription followed by a
quantitative polymerase chain reaction (RT-PCR) (Figure 6.24C). In the reverse-
transcription step, the mRNA is used to create a complementary DNA strand (cDNA)
with an enzyme called reverse transcriptase. This cDNA does not correspond to the
entire mRNA but consists only of a relatively short segment. The cDNA template is
now amplified with quantitative real-time PCR (which unfortunately is also
(A) (B)
Ov Gi Mb Ol Br Ov Gi Mb Ol Br
28S
3k
18S
Figure 6.24 Blotting techniques for
assessing gene expression. (A, B) Northern
blots of the gene expression (mRNA) of the
sox11a gene in ovary (Ov), gill filament (Gi),
midbrain (Mb), olfactory bulb (Ol), and brain
(Br) of the grouper: (A) separation on an
agarose gel; (B) result of hybridization with
labeled sox11a cDNA. (C) RT-PCR coupled
with Southern blot of the gene expression
(C) a b c d e f g h i j k l m n o p q r (mRNA) of the sox11a gene in different
tissues (a–q; r as a negative control) of the
sox11a
grouper. (From Zhang L, Zhu T, Lin D, et al.
Comput. Biochem. Physiol. B Biochem. Mol.
β-actin
Biol. 157 [2010] 415–422. With permission
from Elsevier.)
MEASuRiNG GENE ExPRESSiON 193
abbreviated RT-PCR, unless one includes “q” for “quantitative”: RT-qPCR). The ter-
minology “real-time” stems from the fact that one can use a fluorescent dye, which
attaches only to double-stranded DNA, to monitor in real time how much DNA has
been polymerized so far. qPCR is very sensitive and, once standards are established,
very accurate. The overall result of RT-qPCR is the number of mRNAs per cell or in
some small, defined volume. Alternative methods are described in [16].
Northern blots and RT-PCR are useful for accurate measurements of a few
mRNAs. For larger tasks, other methods have been developed. One is serial analysis
of gene expression (SAGE) [16, 97, 98]. This method permits rather exact measure-
ments of both known and unknown transcripts. It begins with the isolation of mRNA
from an input sample, such as some tissue. The mRNA is immobilized in a chroma-
tography column and converted into cDNA, which is cut with restriction enzymes
into short chains of 10–15 nucleotides (the so-called expressed sequence tag, EST).
The small cDNA chains are assembled to form long DNA molecules, called con-
catamers, which are replicated by insertion into bacteria and subsequently
sequenced. Computational analysis identifies which mRNAs were available in the
original sample and how abundant they were. In contrast to most other methods,
SAGE does not include an artificial hybridization step. The method is rather accu-
rate and does not require prior knowledge of gene sequences or their transcripts.
The second class of methods for large-scale gene expression analysis uses microar-
rays or DNA chips (Figure 6.25). These are cheaper than SAGE and permit larger num-
bers of genes to be tested. However, quantification of the results is more difficult and less
accurate than in SAGE. Microarrays and DNA chips both use hybridization and are
similar in concept, but differ significantly in the specific type of DNA they employ. In
both cases, thousands of tiny amounts of different DNA sequences, called probes, are
spotted onto defined positions on a glass, plastic, or silicon wafer or a nylon membrane,
where they are immobilized. The sample of mRNAs, characterizing the transcriptome
of the cells of interest, is converted into a mixture of cDNAs (called the targets), labeled
with a fluorescent dye, and applied to the chip or microarray. Wherever the cDNAs find
matching partners among the spotted DNA probes, they hybridize, forming double-
stranded DNA. After washing, which removes unbound or loosely bound cDNAs, the
chip or microarray is scanned with a laser that detects the fluorescence of the bound
cDNAs. The more cDNA is bound to a spot, the higher will be the intensity of fluores-
cence. The intensities are computationally converted into colors that reflect the up- and
down-regulation states of genes [17]. The truly attractive feature of microarrays and
DNA chips is that thousands of genes can be tested simultaneously.
In the case of a microarray, the DNA sequences are either synthetic short
stretches of DNA, called oligonucleotides, or other short DNAs, such as cDNAs or
products of a polymerase chain reaction. Because the amounts of DNA involved are
Figure 6.25 Different magnifications of

a cDNA microarray slide. The microarray
allows the profiling of changes in
gene expression, here associated with
chemotherapy of a human brain tumor.
Each spot is associated with a particular
gene. Different colors represent differences
in gene expression between healthy and
tumor cells. Green represents healthy
control DNA, red represents sample DNA
associated with disease, while yellow
represents a combination of the two. Black
areas correspond to spots where neither
the control nor the sample hybridized to
the target DNA. (From Bredel M, Bredel C &
Sikic BI. Lancet Oncol. 5 [2004] 89–100. With
very small, the spotting of DNA on the chip is done by a robot, the intensity of
fluorescence is measured with a laser scanner, and the results are interpreted with
computer software and with specialized statistics. Microarrays allow the character-
ization of RNA populations through very large numbers of spots.
A particularly intriguing option is the two-color microarray, which is used for the
comparison of DNA content in two samples, such as healthy cells and correspond-
ing cancer cells. For this method, cDNAs are prepared from the two cell types and
labeled with one of two variants of fluorescent dyes that emit light at different wave-
lengths. The two cDNA samples are mixed and hybridized to the same microarray.
Laser scanning measures the relative intensities of the two types of fluorescence,
and these intensities can be computationally converted into colors that indicate the
up- and down-regulation states of genes [17].
In the case of DNA chips, the probes are constructed artificially directly on the
chip by adding nucleotides, one by one, to growing chains of oligonucleotides in the
different spots of the chip. Using sophisticated methods of photolithography and
nucleotides that are light-activated, the resulting oligonucleotides in the different
spots have ultimately genuine, defined sequences. DNA chips permit an astonish-
ing density of up to 300,000 oligonucleotides per square centimeter, which allows
very detailed analyses, for instance, of single nucleotide changes in a DNA sequence.
Many variations on these types of methods have been developed; the interested
reader is referred to [16] and a rich body of literature. Each new method faces its
own technical challenges and has advantages and drawbacks. The technical chal-
lenges and complications sometimes correlate with the quality of the output data.
For instance, gene expression levels measured with the first microarrays only had an
accuracy of a few fold, and apparent changes in expression of 20% or 50% had to be
considered with caution. The accuracy of these methods has been improved quite a
bit and is continuing to do so. Large, rapidly growing collections of gene expression
data include the Gene Expression Omnibus [99], the Gene Expression Atlas [100],
and numerous organism- and tissue-specific databases. Collectively, these data-
bases contain information on several hundred thousand genes.
Newer alternatives for measuring gene-expression levels, which may eventually
replace microarrays, include emerging sequencing-based technique such as deep
sequencing, which records the number of times a particular base is read. This deep
sequencing of a transcriptome, also called RNA-Seq (pronounced RNA-seek),
reveals the sequence as well as the frequency with which particular RNAs are found.
RNA-Seq has several advantages over hybridization-based techniques, including
independence from the requirement of annotation, improved sensitivity, a wider
dynamic range, and an assessment of the frequencies of particular RNA segments
and of the percentage of the genome that was actually covered [101].
LOCALiZATiON Of GENE ExPRESSiON

In some systems, not only the degree of gene regulation, but also its localization
within an organism or even within a single cell is of interest. To determine localiza-
tion patterns of gene expression in space and time, one can again focus either on
mRNAs or on proteins that correspond to the target genes. Proteins can be detected
with labeled antibodies or by inserting the gene for the green fluorescent protein
(GFP) into the genome of the host organism in such a manner that it is under the
same control as the target gene. Not only that, but the DNA sequence of GFP can be
inserted in frame with another protein-coding region, resulting in a GFP-tagged
protein that renders it possible to follow the localization and dynamics of a protein
(see Chapter 7). Both genes are expressed together, and GFP permits visualization
of their location. mRNAs can also be visualized, because they contain specific local-
ization elements or zip codes that are recognized by corresponding RNA-binding
proteins, which can be tagged with a fluorescent marker such as tyramide.
As a particularly impressive example, different sets of genes are expressed in
specific locations and to specific degrees during development. Beautiful demon-
strations are presented in Figures 6.26 and 6.27, which show the localization of
different mRNAs and the collocation of mRNAs and proteins during the early
embryonic development of the fruit fly Drosophila [102, 103].
LOCALiZATiON Of GENE ExPRESSiON 195
Figure 6.26 Many mRNAs in Drosophila display striking subcellular localization patterns during development. In this high-resolution fluorescence
image, nuclei are shown in red and mRNAs in green. The head of the embryo will form on the left side of each image and the top will become its back.
(From Martin KC & Ephrussi A. Cell. Mol. Life Sci. 136 [2009] 719–730. With permission from Elsevier.)
(A)
crb Crb
(B)
BicD BicD
(C)
CG14438 CNN
(D)
anillin Anillin
Figure 6.27 Closely correlated distribution patterns (right column) of mRNAs (left column) and proteins (center column) during early
embryogenesis of the fruit fly Drosophila. The images were obtained using double-staining for selected mRNAs, proteins, or relevant markers.
(A) mRNAs/proteins located at the apex. (B, C) mRNAs/proteins associated with the cell division apparatus. (D) mRNAs/proteins residing at cell junctions.
Nuclei in the left and middle panels are shown in blue. Closer analysis demonstrated that mRNAs precede the appearance of the corresponding proteins.
(From Lécuyer E, Yoshida H, Parthasarathy N, et al. Cell. Mol. Life Sci. 131 [2007] 174–187. With permission from Elsevier.)
OuTLOOK
Genomes have been and will continue to be a very exciting area of biological
research. The developments in the field are often mind-boggling in their scope and
speed, and yet one might expect that we have so far only scratched the surface. In
the past, most gene studies have focused either on individual genes or on static gene
regulatory networks. The field has now begun to move toward the roles of small
RNAs, which comprise much more crucial and sophisticated regulatory systems
than anyone had expected, and toward other means of regulating and controlling
gene expression throughout the genome, so that the right gene is transcribed to the
right degree, at the right time, and in the right location. The regulation and dynamics
of these networks, and the localization and integration of their sub-networks, will
without doubt be a cornerstone of experimental and computational systems biology
throughout the foreseeable future.
ExERCiSES
6.1. Identify and describe examples of overlapping 6.14. Establish a list of diseases that are associated with
genes. RNA viruses.
6.2. Collect information on biological materials other 6.15. Determine which cells in the human body do not
than DNA that can be passed from one generation contain the same genes as most other cells. Are
to the next. there living human cells without genes?
6.3. Explore differences and similarities between DNA 6.16. Construct a Boolean model for the expression of the
transposons and retrotransposons. Summarize your structural genes of the lac operon. For this purpose,
findings in a report. look in the literature or on the Internet for the
basics of Boolean logic. Then assign to lactose and
6.4. GC base pairs contain three hydrogen bonds and
glucose each a value of 0 if they are absent or 1 if
AT pairs only two. Find out how much stronger GC
they are present in the environment and use a
base pairs are than AT base pairs.
combination of multiplication and addition to
6.5. Discuss in a two-page report what phylogenetic construct a formula for the “on” (1) and “off” (0)
trees are and how they are established. states of gene expression.
6.6. Gregor Mendel is sometimes called the “Father of 6.17. Determine the implications of mitochondrial DNA
Genetics.” Summarize his major accomplishments for studies of the evolution of the human race.
in a one-page report. Explain why studies of human evolution pay
6.7. Use Gene Ontology or some other web resource to particular attention to genes located on the
determine the function of gene Hs.376209. Y chromosome.
6.8. Explain why a search for “Parkinson’s disease” as a 6.18. Discuss the original Central Dogma of molecular
GO term fails, while searching for “Parkin” as a biology, as well as modern amendments and
“gene or protein” leads to genes associated with the refinements.
disease. 6.19. Search the literature for a study where the
6.9. Obtain information on several variations on QTL, connectivity of a gene network was inferred with
called eQTL (expression QTL) and pQTL (which statistical methods from observation (or artificial)
unfortunately can mean phenotypic, physiological, data. Write a one-page summary of methods,
or protein QTL). Describe their advantages and results, and shortcomings.
drawbacks over traditional QTL. 6.20. Search the literature for a study where the
6.10. How many DNA base pairs in humans are connectivity of a gene network was inferred
constituents of coding genes? from observation (or artificial) data with
methods based on differential equations. Write
6.11. Compare genetic and epigenetic inheritance. a one-page summary of methods, results, and
Summarize your findings in a report. shortcomings.
6.12. Find out how cells attempt to ensure that genes are 6.21. Kimura and collaborators described a method for
correctly translated into mRNA and proteins. Write inferring the structure of genetic networks. Like
a brief report. many other authors they created an artificial
6.13. Without searching on the Internet, speculate on genetic network (Figure 6.28), where blue and
how it is possible that only 1000 miRNAs can red arrows are supposed to indicate “positive”
specifically target half of all human genes. After- and “negative regulations.” What does
wards, check in the literature and the web how “regulation” in this case mean? What would be
close your speculations were to reality. its biological basis?
ExERCiSES 197
study the effects. Summarize your findings in a

report.
1
6.25. Implement a model for the artificial gene network in
Figure 6.30. Begin with the model
2 3 G1 = α 1 TF G2 − β1G1−0.5 ,
G 2 = α 2G1−1 − β 2G20.5 ,
and set all rate constants, as well as the initial values

and the value of the transcription factor TF, equal
to 1. Initiate the system with different initial values.
4 5
Second, study the effects of changes in the value of
TF. Finally, study how the systems’ dynamics
changes with alterations in parameter values,
Figure 6.28 Artificial genetic network, used for testing including the negative kinetic orders associated
an inference algorithm. (Adapted from Kimura S, Sonoda K, with G1.
Yamane S, et al. BMC Bioinformatics 9 [2008] 23. With permission
from Biomed Central.)
TF
6.22. Discuss different options for developing a
mathematical model for the gene network in Figure
6.28. Set up a mathematical model in symbolic form.
What kinds of data would you need to estimate + –
G1
parameter values for a mathematical model? What
+
type of algorithm would you need for the
estimation?
6.23. Read the review of models for gene regulation by –
Schlitt and Brazma [87] and write a brief report G2
about the use of difference models in the field.
6.24. Suppose two genes indirectly regulate each other’s
expression in the fashion shown in Figure 6.29, Figure 6.30 Artificial gene network. Although the system
consists of just two genes and one transcription factor, it can
where green arrows indicate direct or indirect
generate interesting dynamics.
activating effects and the red arrow indicates
repression. A possible model for this system has
the form
6.26. Set up an ODE model for the simple metabolic
pathway shown in Figure 6.31A, using power-law
X 1 = α 1 X 10.4 X 2−0.15 − X 10.2 , functions (see Chapter 4). Choose kinetic orders
X 2 = X 1 − X 20.2 . between 0.2 and 0.8 and adjust the rate constants to
your liking. Study the responses of the pathway to
changes in the input. In the second phase of the
Begin by setting α1 and the initial values equal to 1. exercise, use the same system, but augment it to
Study the responses of this system for different reflect the system in Figure 6.31B, where the end
values of the parameter α1. Argue whether α1 may product Z affects the production of a transcription
be positive or negative. Alter the kinetic orders and factor TF, which turns on a gene G, which codes for
an enzyme E, which catalyzes the conversion of X
into Y. Choose appropriate parameter values to your
liking. Before you simulate, make predictions of the
effects of the addition of TF, G, and E.
+
6.27. Many studies explicitly or implicitly assume that a
X1
– gene is transcribed into mRNA and that the mRNA
is translated into protein, so that there is a direct
linear correlation between gene expression and
+ protein prevalence. Search the literature and the
X2
Internet for information that supports or refutes the
biological validity of this assumption. Discuss the
implications of this assumption for the design of
Figure 6.29 Artificial system in which two genes indirectly models describing transcription–translation
regulate each other’s expression. processes.
(A)
input X Y Z
(B)
input X Y Z
E G TF
Figure 6.31 Pathways with and without genomic feedback. The responses of the system in (A) to changes in the input are easy to predict, but the same
is not true for the system in (B).
6.28. Explore the details of a Southern blot for DNA analysis 6.30. Explore which steps in a SAGE, microarray, or DNA
and summarize the highlights in a one-page report. chip experiment cause the greatest uncertainties in
6.29. Find out how one typically prevents an RNA from results. Compare the degrees of accuracy attainable
forming a secondary structure, which would hinder with the three methods.
hybridization.
REfERENCES
[1] Crick FHC. On protein synthesis. Symp. Soc. Exp. Biol. [16] Brown TA. Genomes, 3rd ed. Garland Science, 2006.
12 (1956) 139–163. [An early draft of an article describing [17] Zvelebil M & Baum JO. Understanding Bioinformatics.
the Central Dogma. Also available at http://profiles.nlm.nih. Garland Science, 2007.
gov/ps/access/SCBBFT.pdf.] [18] Alberts B, Johnson A, Lewis J, et al. Molecular Biology of the
[2] Crick FHC Central dogma of molecular biology. Nature Cell, 6th ed. Garland Science, 2015.
227 (1970) 561–563. [19] Human Genome Project (HGP). Insights Learned from
[3] Cowart LA, Shotwell M, Worley ML, et al. Revealing a the Human DNA Sequence. https://www.genome.
signaling role of phytosphingosine-1-phosphate in yeast. gov/12011238/an-overview-of-the-human-genome-
Mol. Syst. Biol. 6 (2010) 349. project/.
[4] Schuit F, Flamez D, De Vos A & Pipeleers D. Glucose- [20] Buehler LK & Rashidi HH (eds). Bioinformatics Basics:
regulated gene expression maintaining the glucose- Applications in Biological Science and Medicine, 2nd ed. CRC
responsive state of β-cells. Diabetes 51 (2002) S326–S332. Press, 2005.
[5] Merino E, Jensen RA & Yanofsky C. Evolution of bacterial trp [21] Maddox B. The double helix and the wronged heroine.
operons and their regulation. Curr. Opin. Microbiol. 11 (2008) Nature 421 (2003) 407–408.
78–86. [22] Watson JD & Crick FHC. A structure for deoxyribose nucleic
[6] Rock F, Gerrits N, Kortstee A, van Kampen M, Borrias M, acid. Nature 171 (1953) 737–738.
Weisbeek P & Smeekens S. Sucrose-specific signalling [23] Gregory TR & Hebert PD. The modulation of DNA content:
represses translation of the Arabidopsis ATB2 bZIP proximate causes and ultimate consequences. Genome Res.
transcription factor gene. Plant J. 15 (1988) 253–263. 9 (1999) 317–324.
[7] Elgar G & Vavouri T. Tuning in to the signals: noncoding [24] Fugu Genome Project. http://www.fugu-sg.org/.
sequence conservation in vertebrate genomes. Trends Genet. [25] Black DL. Mechanisms of alternative pre-messenger RNA
24 (2008) 344–352. splicing. Annu. Rev. Biochem. 72 (2003) 291–336.
[8] Noble D. Claude Bernard, the first systems biologist, and the [26] Wu H & Moore E. Association analysis of the general
future of physiology. Exp. Physiol. 93 (2008) 16–26. environmental conditions and prokaryotes’ gene
[9] Breton S, Beaupre HD, Stewart DT, et al. The unusual system distributions in various functional groups. Genomics
of doubly uniparental inheritance of mtDNA: Isn’t one 96 (2010) 27–38.
enough? Trends Genet. 23 (2007) 465–474. [27] Vaidyanathan S & Goodacre R. Metabolome and proteome
[10] Siefert JL. Defining the mobilome. Methods Mol. Biol. profiling for microbial characterization. In Metabolic Profiling:
532 (2009) 13–27. Its Role in Biomarker Discovery and Gene Function Analysis
[11] Keeling PJ. Functional and ecological impacts of horizontal (GG Harrigan & R Goodacre, eds), pp 9–38. Kluwer, 2003.
gene transfer in eukaryotes. Curr. Opin. Genet. Dev. 19 (2009) [28] Durban R, Eddy S, Krogh A & Mitchison G. Biological
613–619. Sequence Analysis: Probabilistic Models of Proteins and
[12] Nikaido H. Multidrug resistance in bacteria. Annu. Rev. Nucleic Acids. Cambridge University Press, 1998.
Biochem. 78 (2009) 119–146. [29] Zhou F, Olman V & Xu Y. Barcodes for genomes and
[13] Tang YW & Stratton CW. Staphylococcus aureus: an old applications. BMC Bioinformatics 9 (2008) 546.
pathogen with new weapons. Clin. Lab. Med. 30 (2010) [30] The Gene Ontology Project. http://www.geneontology.org/.
179–208. [31] Da Costa e Silva L & Zeng ZB. Current progress on statistical
[14] Sinzelle L, Izsvak Z & Ivics Z. Molecular domestication of methods for mapping quantitative trait loci from inbred line
transposable elements: from detrimental parasites to useful crosses. J. Biopharm. Stat. 20 (2010) 454–481.
host genes. Cell. Mol. Life Sci. 66 (2009) 1073–1093. [32] Mittler R & Blumwald E. Genetic engineering for modern
[15] Batzer MA & Deininger PL. Alu repeats and human genomic agriculture: challenges and perspectives. Annu. Rev. Plant
diversity. Nat. Rev. Genet. 3 (2002) 370–379. Biol. 61 (2010) 443–462.
REfERENCES 199
[33] van Eeuwijk FA, Bink MC, Chenu K & Chapman SC. Detection [57] Gregory PA, Bracken CP, Bert AG & Goodall GJ. MicroRNAs as
and use of QTL for complex traits in multiple environments. regulators of epithelial–mesenchymal transition. Cell Cycle
Curr. Opin. Plant Biol. 13 (2010) 193–205. 7 (2008) 3112–3118.
[34] Singh V, Tiwari RL, Dikshit M & Barthwal MK. Models to study [58] Lai X, Bhattacharya A, Schmitz U, Kunz M, Vera J. &
atherosclerosis: a mechanistic insight. Curr. Vasc. Pharmacol. Wolkenhauer O. A systems’ biology approach to study
7 (2009) 75–109. microRNA-mediated gene regulatory networks. Biomed. Res.
[35] Zhong H, Beaulaurier J, Lum PY, et al. Liver and adipose Int. 2013 (2013) Article ID 703849.
expression associated SNPs are enriched for association to [59] Lai X, Wolkenhauer O & Vera J. Understanding microRNA-
type 2 diabetes. PLoS Genet. 6 (2010) e1000932. mediated gene regulatory networks through mathematical
[36] Botstein D, White RL, Skolnick M & Davis RW. Construction modelling. Nucleic Acids Res. 44 (2016) 6019–6035.
of a genetic linkage map in man using restriction fragment [60] Mayer G. Bacteriology—Chapter Nine: Genetic regulatory
length polymorphisms. Am. J. Hum. Genet. 32 (1980) mechanism. In Microbiology and Immunology On-line.
314–331. University of South Carolina School of Medicine. http://
[37] Weedon MN, Lango H, Lindgren CM, et al. Genome-wide www.microbiologybook.org/mayer/geneticreg.htm.
association analysis identifies 20 loci that influence adult [61] Savageau MA. Design of molecular control mechanisms and
height. Nat. Genet. 40 (2008) 575–583. the demand for gene expression. Proc. Natl Acad. Sci. USA
[38] Weiss KM. Tilting at quixotic trait loci (QTL): an evolutionary 74 (1977) 5647–5651.
perspective on genetic causation. Genetics 179 (2008) [62] Adam M, Murali B, Glenn NO & Potter SS. Epigenetic
1741–1756. inheritance based evolution of antibiotic resistance in
[39] Ohno S. So much “junk” DNA in our genome. Brookhaven bacteria. BMC Evol. Biol. 8 (2008) 52.
Symp. Biol. 23 (1972) 366–370. [63] Jacob F, Perrin D, Sanchez C & Monod J. Operon: a group of
[40] Wheeler DA, Srinivasan M, Egholm M, et al. The complete genes with the expression coordinated by an operator. C. R.
genome of an individual by massively parallel DNA Hebd. Séances Acad. Sci. 250 (1960) 1727–1729.
sequencing. Nature 452 (2008) 872–876. [64] Wong P, Gladney S & Keasling J. Mathematical model of the
[41] DNA Data Bank of Japan (DDBJ). http://www.ddbj.nig.ac.jp/. lac operon: Inducer exclusion, catabolite repression, and
[42] National Center for Biotechnology Information (NCBI). diauxic growth on glucose and lactose. Biotechnol. Prog.
Databases. http://www.ncbi.nlm.nih.gov/guide/all/. 13 (1997) 132–143.
[43] NCBI. GenBank. http://www.ncbi.nlm.nih.gov/genbank/. [65] Vilar JMG, Guet CC & Leibler S. Modeling network dynamics:
[44] Xu Y & Gogarten JP (eds). Computational Methods for the lac operon, a case study. J. Cell. Biol. 161 (2003) 471–476.
Understanding Bacterial and Archaeal Genomes. Imperial [66] Yildrim N & Mackey MC. Feedback regulation in the lactose
College Press, 2008. operon: a mathematical modeling study and comparison
[45] Arold ST, Leonard PG, Parkinson GN & Ladbury JE. H-NS with experimental data. Biophys. J. 84 (2003) 2841–2851.
forms a superhelical protein scaffold for DNA condensation. [67] Tian T & Burrage K. A mathematical model for genetic
Proc. Natl Acad. Sci. USA 107 (2010) 15728–15732. regulation of the lactose operon. In Computational Science
[46] Narlikar GJ. A proposal for kinetic proof reading by ISWI and its Applications—ICCSA 2005. Lecture Notes in
family chromatin remodeling motors. Curr. Opin. Chem. Biol. Computer Science, Volume 3481 (O Gervasi, ML Gavrilova,
14 (2010) 660–665. V Kumar, et al., eds). pp 1245–1253. Springer-Verlag, 2005.
[47] Huang Y-W, Kuo C-T, Stoner K, et al. An overview of [68] Savageau MA. Demand theory of gene regulation. I.
epigenetics and chemoprevention. FEBS Lett. 585 (2011) Quantitative development of the theory. Genetics 149 (1998)
2129–2136. 1665–1676.
[48] Yoo CB & Jones PA. Epigenetic therapy of cancer: past, [69] Su Z, Li G & Xu Y. Prediction of regulons through
present and future. Nat. Rev. Drug Discov. 5 (2006) 37–50. comparative genome analyses. In Computational Methods
[49] Chi YI, Martick M, Lares M, et al. Capturing hammerhead for Understanding Bacterial and Archaeal Genomes (Y Xu
ribozyme structures in action by modulating general base & JP Gogarten, eds), pp 259–279. Imperial College Press,
catalysis. PLoS Biol. 6 (2008) e234. 2008.
[50] Nissen P, Hansen J, Ban N, et al. The structural basis of [70] Yin Y, Zhang H, Olman V & Xu Y. Genomic arrangement of
ribosome activity in peptide bond synthesis. Science bacterial operons is constrained by biological pathways
289 (2000) 920–930. encoded in the genome. Proc. Natl Acad. Sci. USA 107 (2010)
[51] Liu Q & Paroo Z. Biochemical principles of small RNA 6310–6315.
pathways. Annu. Rev. Biochem. 79 (2010) 295–319. [71] RegulonDB Database. http://regulondb.ccg.unam.mx/.
[52] Siomi H & Siomi MC. On the road to reading the RNA- [72] Che D, Li G, Mao F, et al. Detecting uber-operons in
interference code. Nature 457 (2009) 396–404. prokaryotic genomes. Nucleic Acids Res. 34 (2006) 2418–
[53] Napoli C, Lemieux C & Jorgensen R. Introduction of a 2427.
chimeric chalcone synthase gene into petunia results in [73] Godon C, Lagniel G, Lee J, et al. The H2O2 stimulon in
reversible co-suppression of homologous genes in trans. Saccharomyces cerevisiae. J. Biol. Chem. 273 (1998) 22480–
Plant Cell 2 (1990) 279–289. 22489.
[54] Deng W, Zhu X, Skogerbø G, et al. Organization of the [74] Igarashi K & Kashiwagi K. Polyamine modulon in Escherichia
Caenorhabditis elegans small noncoding transcriptome: coli: genes involved in the stimulation of cell growth by
genomic features, biogenesis, and expression. Genome Res. polyamines. J. Biochem. 139 (2006) 11–16.
16 (2006) 20–29. [75] Brinkrolf K, Schröder J, Pühler A & Tauch A. The
[55] Pillai RS, Bhattacharyya SN & Filipowicz W. Repression of transcriptional regulatory repertoire of Corynebacterium
protein synthesis by miRNAs: How many mechanisms? glutamicum: reconstruction of the network controlling
Trends Cell Biol. 17 (2007) 118–126. pathways involved in lysine and glutamate production.
[56] Chen J, Wang L, Matyunina LV, et al. Overexpression of miR- J. Biotechnol. 149 (2010) 173–182.
429 induces mesenchymal-to-epithelial transition (MET) in [76] Covert MW, Knight EM, Reed JL, et al. Integrating high-
metastatic ovarian cancer cells. Gynecol. Oncol. 121 (2011) throughput and computational data elucidates bacterial
200–205. networks. Nature 429 (2004) 92–96.
[77] Samal A & Jain S. The regulatory network of E. coli [90] Babu M, Musso G, Diaz-Mejia JJ, et al. Systems-level
metabolism as a Boolean dynamical system exhibits both approaches for identifying and analyzing genetic interaction
homeostasis and flexibility of response. BMC Syst. Biol. networks in Escherichia coli and extensions to other
2 (2008) 21. prokaryotes. Mol. Biosyst. 5 (2009) 1439–1455.
[78] Carro MS, Lim WK, Alvarez MJ, et al. The transcriptional [91] Gasch AP, Spellman PT, Kao CM, et al. Genomic expression
network for mesenchymal transformation of brain tumours. programs in the response of yeast cells to environmental
Nature 463 (2010) 318–325. changes. Mol. Biol. Cell 11 (2000) 4241–4257.
[79] Basso K, Margolin AA, Stolovitzky G, et al. Reverse [92] Reinitz J & Sharp DH. Gene circuits and their uses. In
engineering of regulatory networks in human B cells. Nat. Integrative Approaches to Molecular Biology (J Collado-
Genet. 37 (2005) 382–390. Vides, B Magasanik & TF Smith, eds), pp. 253–272. MIT Press,
[80] Martínez-Antonio A & Collado-Vides J. Comparative 1996.
mechanisms for transcription and regulatory signals in [93] McAdams HH & Arkin A. It’s a noisy business! Genetic
archaea and bacteria. In Computational Methods for regulation at the nanomolar scale. Trends Genet. 15 (1999)
Understanding Bacterial and Archaeal Genomes (Y Xu & JP 65–69.
Gogarten, eds), pp 185–208. Imperial College Press, 2008. [94] Gardner TS, Cantor CR & Collins JJ. Construction of a genetic
[81] Plahte E, Mestl T & Omholt SW. A methodological basis for toggle switch in Escherichia coli. Nature 403 (2000) 339–342.
description and analysis of systems with complex switch-like [95] Zhang L, Wang Z, Sagotsky JA & Deisboeck TS. Multiscale
interactions. J. Math. Biol. 36 (1998) 321–348. agent-based cancer modeling. J. Math. Biol. 58 (2009)
[82] Gardner TS, di Bernardo D, Lorenz D & Collins JJ. Inferring 545–559.
genetic networks and identifying compound mode of action [96] Medicalgenomics. BlotBase. http://www.medicalgenomics.
via expression profiling. Science 301 (2003) 102–105. org/blotbase/.
[83] Geier F, Timmer J & Fleck C. Reconstructing gene-regulatory [97] Sagenet. SAGE: Serial Analysis of Gene Expression. http://
networks from time series, knock-out data, and prior www.sagenet.org/.
knowledge. BMC Syst. Biol. 1 (2007) 11. [98] Velculescu VE, Zhang L, Vogelstein B & Kinzler KW. Serial
[84] Nakatsui M, Ueda T, Maki Y, et al. Method for inferring analysis of gene expression. Science 270 (1995) 484–487.
and extracting reliable genetic interactions from time- [99] NCBI. GEO: Gene Expression Omnibus. http://www.ncbi.nlm.
series profile of gene expression. Math. Biosci. 215 (2008) nih.gov/geo/.
105–114. [100] European Molecular Biology Laboratory–European
[85] Yeung MK, Tegner J & Collins JJ. Reverse engineering gene Bioinformatics Institute (EMBL–EBI). ATLAS: Gene Expression
networks using singular value decomposition and robust Atlas. http://www.ebi.ac.uk/gxa/.
regression. Proc. Natl Acad. Sci. USA 99 (2002) 6163–6168. [101] Croucher NJ & Thomson NR. Studying bacterial
[86] de Jong H. Modeling and simulation of genetic regulatory transcriptomes using RNA-seq. Curr. Opin. Microbiol.
systems: a literature review. J. Comput. Biol. 9 (2002) 67–103. 13 (2010) 619–624.
[87] Schlitt T & Brazma A. Current approaches to gene regulatory [102] Martin KC & Ephrussi A. mRNA localization: gene expression
network modelling. BMC Bioinformatics 8 (Suppl 6) (2007) S9. in the spatial dimension. Cell. Mol. Life Sci. 136 (2009)
[88] Le Novère N. Quantitative and logic modelling of molecular 719–730.
and gene networks. Nat. Rev. Genet. 16 (2015) 146–158. [103] Lécuyer E, Yoshida H, Parthasarathy N, et al. Global analysis
[89] Tong AH, Lesage G, Bader GD, et al. Global mapping of of mRNA localization reveals a prominent role in organizing
the yeast genetic interaction network. Science 303 (2004) cellular architecture and function. Cell. Mol. Life Sci.
808–813. 131 (2007) 174–187.
fuRTHER READiNG
Brown TA. Genomes, 3rd ed. Garland Science, 2006. Huang Y-W, Kuo C-T, Stoner K, et al. An overview of epigenetics and
Carro MS, Lim WK, Alvarez MJ, et al. The transcriptional network chemoprevention. FEBS Lett. 585 (2011) 2129–2136.
for mesenchymal transformation of brain tumours. Nature Noble D. The Music of Life; Biology Beyond Genes. Oxford University
463 (2010) 318–325. Press, 2006.
de Jong H. Modeling and simulation of genetic regulatory systems: Schlitt T & Brazma A. Current approaches to gene regulatory
a literature review. J. Comput. Biol. 9 (2002) 67–103. network modelling. BMC Bioinformatics 8(Suppl 6)
Durban R, Eddy S, Krogh A & Mitchison G. Biological Sequence (2007) S9.
Analysis: Probabilistic Models of Proteins and Nucleic Acids. Zvelebil M & Baum JO. Understanding Bioinformatics. Garland
Cambridge University Press, 1998. Science, 2007.
Protein Systems
7
• Discuss types of proteins and their roles
• Describe the basic chemical properties of proteins
• Explain the four hierarchical levels of protein structure
• Retrieve from databases information about the crystal structure of proteins
• Outline concepts of protein separation and proteomic techniques
• Discuss the basic concepts of protein structure prediction and protein
localization
• Describe interactions among proteins and between proteins and other
molecules
Biological systems are complicated machines, and proteins supply most of their
gears, gaskets, valves, sensors, and amplifiers. Indeed, proteins are the most versatile
and fascinating molecules in living cells. Although all proteins have generically
the same molecular composition, their sizes, functions, and roles could not be more
diverse, and their importance for life is hard to overestimate. Proteins are constantly
generated and destroyed, with some existing just for minutes or less, and others,
such as the proteins in the lens of the human eye, persisting for an entire human life.
Owing to their central importance and versatility, an astounding amount of
information has been accumulated over the past two centuries, since the Swedish
chemist Jöns Jacob Berzelius coined the term protein (from the Greek for “primary”)
in 1838, and uncounted databases for different aspects of proteins are now available
(see, for example, [1–4]). Proteins have been the subject of intense investigation in
biochemistry and protein chemistry; molecular, structural, and computational
biology; metabolic engineering; bioinformatics; proteomics; systems biology; and
synthetic biology. Yet, what we know today might just be the tip of the iceberg. Thus,
this chapter will not come close to portraying all aspects of proteins in a just manner,
and it will not even present much detail about relevant topics as far as they are
covered comprehensively in texts on computational biology and bioinformatics (for
example, [5]). Instead, it will focus on some of the main features of proteins and
discuss those roles that are of particular importance for the dynamic functioning of
biological systems. We begin with some chemical properties and the assembly of
proteins according to the building instructions provided in the genome. A good
starting point for exploring the known properties of a protein is the Human Protein
Atlas [6].
202 Chapter 7: Protein Systems
CHEMICAL AND PHYSICAL FEATURES OF PROTEINS

Proteins are polymers composed of amino acids, which are joined into a linear
string by peptide bonds (Figure 7.1). Each amino acid unit (or residue of the poly-
mer) has a particular molecular side chain, denoted by R in Figure 7.1. The chemical
properties of the side chains determine the features of the amino acids and collec-
tively of the protein. Importantly, the side chains can interact with each other
through different noncovalent mechanisms, such as hydrogen bonding and van der
Waals forces. Furthermore, the spatial arrangements of amino acid residues are
affected by hydrophobic effects, which tend to maximize the packing density of the
protein, as well as by covalent links and ionic bonds, called salt bridges. The interac-
tions between the side chains of distant amino acids of the same or a different pro-
tein lead to cross-linking and the formation of the tertiary (three-dimensional)
structure of the protein (see later). They are also responsible for the stability of a
protein. Of note is the most prevalent covalent link, namely, a disulfide bond
between two cysteine residues, which is so important that the resulting dimeric
(two-component) amino acid has its own name: cystine. Almost all of nature uses 20
l-amino acids (Table 7.1). Alas, as so often in nature, there are exceptions: a very
few organisms make use of two additional amino acids, namely, selenocysteine and
pyrrolysine [7, 8].
As we discussed in Chapter 6, proteins are generated through transcription,
translation, and possible subsequent modifications from stretches of DNA within the
organism’s genome. Specifically, each amino acid within a protein corresponds to a
DNA codon, consisting of a triplet of nucleotides within the genetic code, which is
transcribed into a matching RNA codon. Because DNA consists of four nucleotides,
4 × 4 × 4 = 64 different codons are possible. A few of these serve as start or stop sig-
nals for transcription, but more codons than necessary are left to code for the 20
amino acids, with the consequence that most amino acids are encoded by more than
one codon (Table 7.2). In fact, in cases such as alanine, the third nucleotide of
the codon is virtually, although not totally, immaterial. While different codons can
thus represent the same amino acid, a phenomenon called wobble, the reverse is not
true: every codon codes for one and only one amino acid, or, as a mathematician
might say, the transcription–translation system implements a “many-to-one map-
ping.” Nonetheless, the choice of a codon is apparently not random. For instance,
different species often prefer particular codons over alternatives, a phenomenon
R R
O O
H H
N C C N C C
H H
O O
H H
amino acid 1 H amino acid 2 H
R
carboxy
O terminus
H
H
R
N C C
O
O
H
H N C C
H
amino
terminus
peptide H
O
bond H water
Figure 7.1 Generic representation of a
peptide bond between two amino acids.
dipeptide R represents the specific side chain of each
H
amino acid.
CHEMICAL AND PHYSICAL FEATURES OF PROTEINS 203
TABLE 7.1: AMINO ACIDS IN ALPHABETICAL ORDER WITH THEIR ABBREVIATIONS

The generic backbone is shown in blue. Color coding of the amino acid’s name represents types of side chains: orange, hydrophobic; green, positively
charged; red, negatively charged; yellow, polar uncharged; turquoise, special cases; purple, nonstandard amino acids.
Generic Amino Acid Generic Amino Acid Alanine (A/Ala) Arginine (R/Arg) Asparagine (N/Asn) Aspartic Acid (D/Asp)
(simplified notation)
OH OH OH OH OH OH
O O O O O
O C
NH2 NH2 NH2 NH2 NH2
H C NH2
R
O O–
R
O
NH2
NH
H2 N
+ NH2
Cysteine (C/Cys) Glutamic Acid (E/Glu) Glutamine (Q/Gln) Glycine (G/Gly) Histidine (H/His) Isoleucine (I/Ile)
OH OH OH OH OH OH
O O O O O O
NH2 NH2 NH2 NH2 NH2 NH2
SH
O O N NH
–
O NH2
Leucine (L/Leu) Lysine (K/Lys) Methionine (M/Met) Phenylalanine (F/Phe) Proline (P/Pro) Pyrrolysine (O/Pyl)
OH OH OH OH OH OH
O O O O O O
NH2 NH2 NH2 NH2 NH NH2
+
NH3 NH
O
N
CH3
Selenocysteine (U/Sec) Serine (S/Ser) Threonine (T/Thr) Tryptophan (W/Trp) Tyrosine (Y/Tyr) Valine (V/Val)
OH OH OH OH OH OH
O O O O O O
NH2 NH2 NH2 NH2 NH2 NH2
HO
SeH OH
NH
OH
TABLE 7.2: AMINO ACIDS AND THEIR CORRESPONDING RNA CODONS

Ala GCA, GCC, GCG, GCU Leu CUA, CUG, CUC, CUU, UUA, UUG
Arg AGA, AGG, CGA, CGC, CGG, CGU Lys AAA, AAG
Asn AAC, AAU Met AUG
Asp GAC, GAU Phe UUC, UUU
Cys UGC, UGU Pro CCA, CCC, CCG, CCU
Gln CAA, CAG Ser UCA, UCC, UCG, UCU, AGC, AGU
Glu GAA, GAG Thr ACA, ACC, ACG, ACU
Gly GGA, GGC, GGG, GGU Trp UGG
His CAC, CAU Tyr UAC, UAU
Ile AUA, AUC, AUU Val GUA, GUC, GUG, GUU
START AUG STOP UAA, UAG, UGA
called codon usage bias (see Table 3 in [9]). In prokaryotes, this bias is so pronounced
that species can be barcoded based on codons [10] (see Chapter 6). Also intriguing,
research seems to suggest that alternative codons, while resulting in the same amino
acid, can nevertheless affect the function of a protein in some cases [11].
Each side chain makes an amino acid chemically unique. Because there are
20 standard amino acids, and because proteins contain many of them, the potential
variability among proteins is enormous. It might sound exaggerated to consider the
possible effects of single amino acid substitutions, but it is well known that even a
single change can indeed alter the structure of a protein drastically [12]. For instance,
sickle cell disease is caused by a single DNA mutation that results in one amino acid
substitution and has drastic physiological effects.
Proteins vary tremendously in size. At the low end of the size scale are pep-
tides, which consist of up to about 30 amino acids; this boundary between pep-
tides and proteins is not clearly defined. On the high end, the largest protein that
has so far been characterized is titin, which is a constituent of muscle tissue and
consists of about 27,000 amino acids. To gauge the degree of variability among
proteins, compute how many different proteins would be possible if they con-
sisted of exactly (or at most) 1000 or 27,000 amino acids. The size of a protein is
usually not given in terms of the number of amino acids though, but in atomic
mass units or daltons. Because of their large sizes, proteins are actually measured
in kilodaltons (kDa). As an example, titin has a molecular mass of close to
3000 kDa.
The physical and chemical characteristics of two neighboring amino acid side
chains slightly bend their peptide bond in a predictable fashion, and all these bends
collectively cause the protein to fold into a specific three-dimensional (3D) struc-
ture or conformation. This structure can contain coils, sheets, hairpins, barrels, and
other features, which we will discuss later. Rearranging the same amino acids in
another order would almost certainly result in a different 3D structure. One should
note in this context that some proteins contain unstructured peptide regions of vari-
ous lengths that allow different bending and folding, which in turn increases the
repertoire of roles a protein can assume [13].
Over 40 years ago, Nobel Laureate Christian Anfinsen postulated the so-called
thermodynamic hypothesis stating that, at least for the class of small globular pro-
teins, the structure of a protein is determined exclusively by the amino acid
sequence, so that under normal environmental conditions the protein structure is
unique, stable, and such that it minimizes the total Gibbs free energy [14]. This
assertion was based on in vitro experiments where he and his colleagues showed
that the protein ribonuclease, which had been unfolded with urea, spontaneously
refolded into the right conformation once the urea was removed. One might think
that disulfide bonds could be a driving force in this process, but it is our current
understanding that rather they are the result of correct folding and serve to stabilize
a protein, whether it is folded correctly or incorrectly.
The automatic folding posited by Anfinsen’s Dogma poses an interesting puzzle,
called Levinthal’s Paradox. Cyrus Levinthal computed that even if a protein
consisted of only 100 amino acids, it would allow so many conformations that cor-
rect folding would take 100 billion years [15]. Of course, that is out of the question,
and in reality the process is remarkably quick: it takes only a bit more than one min-
ute to fold a protein of 300 amino acids! While we do not completely understand
how this is possible, we do know that the folding of a newly created amino acid
chain into the proper 3D structure is facilitated by two means. First, a special class of
cytosolic proteins, called chaperonins, supports the folding task as the newly formed
chain of amino acids exits the ribosome. A prominent example is the GroEL/ES
protein complex. GroEL consists of 14 subunits that form a barrel, the inside of
which is hydrophobic, while its partner GroES resembles a lid. The unfolded amino
acid chain and adenosine triphosphate (ATP) bind to the interior rim of a cavity in
the barrel-shaped GroEL and trigger a change in conformation, as well as an asso-
ciation of the growing protein with GroEL. Binding between the two proteins causes
the subunits of GroEL to rotate and to eject the folded protein, and GroES is released.
Proteins often undergo several rounds of refolding until the appropriate conforma-
tion is reached. Modern methods of cryo-electron microscopy permit a glimpse into
the structure of proteins like GroEL/ES (see Figure 7.3 below). Chaperonins belong
to a larger class of molecules, called molecular chaperones. These not only facilitate
folding, they also protect a protein’s 3D structure in situations of stress, such as ele-
vated heat. In reference to this role, chaperones belong to the class of heat-shock
proteins (HSPs).
The second mechanism facilitating proper folding is accomplished through spe-
cial enzymes in the endoplasmic reticulum of eukaryotes. These enzymes, called
protein disulfide isomerases (PDIs), catalyze the formation and breakage of disul-
fide bonds between cysteine residues within the protein [16]. This mechanism helps
proteins find the correct arrangement of disulfide bonds. PDIs also serve as chaper-
ones that aid the refolding of wrongly arranged proteins. In the process of forming a
disulfide bond, the PDI is chemically reduced. The protein Ero1 subsequently oxi-
dizes PDI and thereby refreshes its active state.
While the chain of amino acids in a protein corresponds to the nucleotide
sequence in the corresponding mRNA and DNA, a protein may be chemically
altered once the transcription and translation processes are completed. Collec-
tively these alterations are called post-translational modifications (PTMs). They
consist of the attachment of biochemical groups, for example, phosphate, acetate,
or larger molecules such as carbohydrates or lipids. Furthermore, amino acids may
be removed from the amino end of the protein, enzymes may cut a peptide chain
into two, and disulfide bridges may form between two cysteine residues. It even
happens that the amino acid arginine is changed into a different amino acid, citrul-
line, which has no corresponding codon. Hundreds of PTMs are listed in a docu-
ment file within the UniProt Knowledgebase [3], and proteome-wide statistics on
PTMs are available at [17]. PTMs are of enormous importance, because they expand
the repertoire of protein functions manifold and permit fine-tuned, dynamic con-
trol of the specific roles of proteins. A prominent example of the functionality of a
PTM is phosphorylation or dephosphorylation in a specific location of a protein,
which in the case of an enzyme or signaling protein can lead to activation or deac-
tivation (see Chapters 8 and 9).
The turnover of proteins can be very fast or very slow, depending on the particu-
lar protein. If a protein is no longer needed, it is marked for degradation by the
attachment of arginine [11] or of multiple copies of a specific regulatory protein
called ubiquitin. This marking is recognized by a cellular protein recycling organ-
elle, called the proteasome, which disassembles the protein into peptides and
amino acids [18]. Aaron Ciechanover, Avram Hershko, and Irwin Rose received the
2004 Nobel Prize in Chemistry for the discovery of this process. Interestingly, ubiq-
uitin and ubiquitin-like proteins are directly involved in the assembly of ribosomes
[19]. This coupling of the nuclear ubiquitin–proteasome system with the assembly
of new ribosomal subunits ensures that these two central processes of protein pro-
duction and destruction are well coordinated.
Thus, the genetic code determines the amino acid sequence, the amino acid
residues and the protein’s environment determine the base structure of a protein,
and the structure and regulatory modifications determine its role. Special proteins
help other proteins fold into the right conformation, affect their activity, and mark
them for destruction in an ongoing dynamic process. During their short or long life-
times, the proteins in a functioning cell or organism have roles that are amazingly
diverse, as we will see throughout this and other chapters.
7.1 Experimental Protein Structure Determination and

Visualization
If a protein can be crystallized, its 3D structure can be determined with X-ray crys-
tallography [20]. In this very well-established technique, which has generated tens
of thousands of protein structures over the past decades, an X-ray beam is shot at
the protein crystal and diffracts (scatters) into many directions, which are deter-
mined by the atomic arrangements within the protein (Figure 7.2A). The angles and
intensities of the diffracted beams allow the reconstruction of a 3D picture of the
density of the electrons within the crystal, which in turn provides information on the
mean positions of the atoms and their chemical bonds. Another popular method for
3D reconstructions of proteins is NMR (nuclear magnetic resonance) spectroscopy
[21], which also has produced thousands of images of protein structures (Figure
7.2B).
A newer and very powerful alternative to X-ray crystallography is cryo-
electron microscopy (cryo-EM) (Figures 7.3 and 7.4). This method can be useful,
for instance, for the analysis of large proteins and protein complexes [26], as well as
of membrane proteins, which are often fragile and do not always yield sufficient
amounts of material for X-ray crystallographic methods [27, 28]. Cryo-EM is based
on freezing a thin suspension of molecules in vitreous ice on an electron micro-
scope grid and subsequently imaging the molecules.
The different structural techniques are complementary in providing information
on proteins. For example, cryo-EM may be used in conjunction with X-ray crystal-
lography for single-particle analysis. In this case, X-ray crystal structures of one or
more subunits are docked into the cryo-EM structure of a large complex that is not
amenable to 3D crystallization in its entirety. Some proteins are highly flexible,
which necessitates the study of each conformation in parallel.
(A) (B)
CH3
denatured HEWL
HN aromatic H
X2
native HEWL
HN H CH3
X1
10.0 8.0 6.0 4.0 2.0 0.0

ppm
Figure 7.2 X-ray crystallography and NMR spectroscopy have led to insights into thousands of protein structures. (A) Diffraction image of a
crystal of hen egg lysozyme. The protein structure is inferred from the angles and intensities of the diffracted X-ray beam. (B) One-dimensional 1H-NMR
spectra of denatured and native hen egg lysozyme. The presence (or absence) of protein structure can be assessed from the distribution of methyl,
Hα, and HN signals, where the latter identify the molecular positions of the hydrogen ions. In the native (structured) protein (bottom), the methyl signals
are shifted upfield and the Hα and HN signals downfield. In the denatured (unstructured) protein (top), these signals are absent from the regions of
interest. (From Redfield C. Proteins studied by NMR. In Encyclopedia of Spectroscopy and Spectrometry, 2nd ed. [J Lindon, GE Tranter & DW Koppenaal,
eds], pp. 2272–2279. Elsevier, 2010. With permission from Elsevier.)
(A)
(a) (b) (c)
(B)
(a) (b) (c) (d)
GroEL GroEL–ATP GroEL–GroES–ATP
Figure 7.3 Cryo-electron microscopy (cryo-EM) results for the molecular chaperone GroEL. (A) Electron micrograph (a) and the inferred ribbon
structure of the protein GroEL in end-view orientation (b) and in a side view (c) (Protein Data Bank: PDB 2C7E [22]). GroEL measures 140 Å in diameter
in the end-on orientation; the scale bar represents 100 nm. Courtesy of Inga Schmidt-Krey, Georgia Institute of Technology. (B) Cryo-EM-based,
surface-rendered views of 3D reconstructions of (a) GroEL, (b) GroEL–ATP, and (c) GroEL–GroES–ATP. The GroES ring is seen as a disk above the GroEL.
(d) Corresponding ribbon structure. It consists of a β-barrel structure with a β-hairpin forming the roof of the dome-like heptamer (Protein Data Bank:
PDB 2CGT [23]). ([a–c] from Roseman AM, Chen S, White H, et al. Cell 87 [1996] 241–251. With permission from Elsevier.)
(A) (B) (C)
Figure 7.4 Cryo-EM is a valuable tool for elucidating structure–function questions in membrane proteins. Here, two-dimensional crystallization
(A) and electron crystallography are used to study the human membrane protein leukotriene C4 synthase, which has a size of only 42 Å; the scale
bar represents 1 μm. (B, C) Inferred ribbon structures (Protein Data Bank: PDB 2PNO [24, 25]). (Courtesy of Inga Schmidt-Krey, Georgia Institute of
Technology.)
(A) (B) (C) (D)
Figure 7.5 Different PDB-Jmol representations of the crystal structure of human insulin, complexed with a copper ion (in the center). (A) Ribbon
cartoon representation. (B) Backbone representation. (C) Ball-and-stick representation. (D) Sphere representation. The last of these, also called the
space-filling or Corey–Pauling–Koltun (CPK) representation, uses a convention showing atoms in defined colors (for example, C, gray; O, red; N, blue; Cu,
light brown). (Protein Data Bank: PDB 3IR0 [29, 31].)
Because of the large number of amino acids in a protein, there is a huge reper-
toire of potential protein shapes and functions. Good software tools, such as the
open-source Java viewer Jmol in the Protein Data Bank (PDB) [1, 29], have been
developed to visualize the 3D shapes of protein crystal structures (Figure 7.5). One
should, however, keep in mind that a crystal structure cannot always be obtained.
For instance, some proteins contain unstructured peptide regions of various lengths,
which permit different folding, so that the global 3D structure and the functional
role of a protein are not always fixed but depend on its milieu [13]. Furthermore,
other molecules, such as sugars and phosphate groups, may be attached to proteins
at certain times and thereby change their conformation and activity state. Finally,
proteins have of course experienced changes in their structures throughout evolu-
tion, so that the “same” protein from different species may have different structures.
Interestingly, the most important features, such as the active sites of enzymes, have
been conserved much more strictly than other domains, leading to the terminology
of anchor sites and variable sites of protein structures [30].
AN INCOMPLETE SURVEY OF THE ROLES AND

FUNCTIONS OF PROTEINS
It is impossible to rank proteins by importance, because removal of any class or type
of proteins would be disastrous. Nonetheless, if one were to sort all proteins purely by
their abundance, the clear winner would be RuBisCO, a protein that most people have
never heard of. If you are surprised, consider that RuBisCO (ribulose 1,5-bisphos-
phate carboxylase oxygenase) is the first enzyme in the Calvin cycle, which is the met-
abolic pathway responsible for photosynthesis. It has been estimated that the total
photosynthetic biomass production on Earth is over 100 billion tons of carbon per
year [32], and RuBisCO is instrumental in all of it. Photosynthetic cyanobacteria (blue
algae) alone comprise an estimated biomass between 40 and 50 billion tons. This is an
enormous amount, considering that all humans together make up only about 250 mil-
lion tons. The most abundant protein in the human body is collagen, which accounts
for between a quarter and a third of the body’s total protein content [33]. Within
human cells, actin is the winner [34]. We can see that even the most abundant proteins
are distributed among different functional classes.
When discussing the function of a protein, one should be aware that proteins
often serve more than one role. For instance, the extracellular matrix consists of a
number of proteins and other substances that maintain a cell’s shape. The matrix
also aids the movement of cells and molecules, it is crucial for signal transduction,
and it acts as a storage compartment for growth factors, which can be released by
the action of proteases if they are needed for growth and tissue regeneration. The
same release processes are probably also involved in the dynamics of tumor forma-
tion and cancer metastasis. Specific examples of matrix proteins are fibronectins,
which attach to collagen and other proteins called integrins on the surfaces of cells.
Fibronectins allow cells to reorganize their cytoskeleton and to move through the
AN INCOMPLETE SURVEY OF THE ROLES AND FUNCTIONS OF PROTEINS 209
extracellular space. They are also critical in blood clotting, where they bind to plate-
lets and thereby assist in wound closure and healing.
The following will highlight the main function of some representative proteins,
but many of these have secondary, yet important, roles—for instance, by serving as
enzymes and as structural proteins that have a strong effect on signal transduction.
7.2 Enzymes
Enzymes are proteins that catalyze chemical reactions; that is, they convert
organic substrates into other organic products. For instance, the first enzyme of
glycolysis, hexokinase, takes the substrate glucose and converts it into the prod-
uct glucose 6-phosphate (see Chapter 8 for an illustration). This conversion
would not occur in any appreciable amount without the enzyme, because glu-
cose 6-phosphate has a higher energy state than glucose. The enzyme makes the
conversion possible by coupling the reaction to a second reaction. Namely, the
required energy cost is paid for by the energy-rich molecule adenosine triphos-
phate (ATP), which donates one of its phosphates to the product and is thereby
dephosphorylated to adenosine diphosphate (ADP), which contains less energy.
A key element of this coupling of processes is that the enzyme itself is unchanged
in the reaction and directly available for converting further substrate molecules.
The vast majority of all metabolic conversions in a living cell are catalyzed by
enzymes in this fashion. We will discuss details of such conversions in Chapter
8, which addresses metabolism.
With respect to the enzyme itself, it is of interest to study how a protein is capable
of facilitating a metabolic reaction. The details of such a process depend of course
on the specific enzyme and its substrate(s), but one can generically say that
the enzyme contains an active site to which the substrate molecule binds. This active
site is a physical 3D structure, which typically consists of a groove or pocket in the
surface of the enzyme molecule and whose shape very specifically matches the
Figure 7.6 Many pathogenic bacteria produce lipo-oligosaccharides, which are composed of
lipid and sugar components and resemble molecules on human cell surfaces. This molecular
mimicry can deceive membrane receptors in the host, which do not recognize the invaders as foreign.
As a consequence, the bacteria attach to the unsuspecting receptors and manage to evade the host’s
immune response. An example is the meningitis-causing bacterium Neisseria meningitidis, and a key
tool of its success is the enzyme galactosyltransferase LgtC, whose crystal structure is shown here.
The enzyme catalyzes an important step in the biosynthesis of a specific lipo-oligosaccharide; namely;
it transfers α-d-galactose from UDP-galactose to a terminal lactose. The enzyme (turquoise ribbons)
is shown in complex with sugar analogs (UDP 2-deoxy-2-fluoro-galactose and 4′-deoxylactose; stick
molecules), which respectively resemble the true donor and acceptor sugars encountered in nature.
Investigations of such complexes, combined with methods of directed evolution and kinetic analysis,
offer important insights into the mechanisms with which bacteria outsmart the mammalian immune
system and provide a starting point for targeted drug design. (Courtesy of Huiling Chen and Ying Xu,
University of Georgia.)
substrate (Figure 7.6). Because of this specificity, enzymes typically catalyze just
one substrate, or one class of substrates, with high efficiency. Initially, the active site
was assumed to be rather rigid, but the current understanding is that it is more flex-
ible, thereby allowing the transient formation of a complex between enzyme and
substrate. Chemical residues near to the active site act as donors of small molecular
groupings or protons, which are attached to the substrate during the reaction. Once
attached, the modified molecule falls off the enzyme as the product of the reaction.
Other classes of enzymes act on substrates by removing molecular groups, which
are then accepted by residues close to the active site. Once the product is formed,
the residues are returned to their original state.
The involvement of an enzyme in a metabolic reaction is of enormous impor-
tance for several reasons. First, very many reactions have to be executed against
a thermodynamic gradient. In other words, such a reaction would not be possible
without an enzyme. Second, even if the product has a lower energy state than the
substrate, the energy state of the substrate is still lower than a transition state that
needs to be overcome before the product can be formed. An enzyme is usually
needed to surmount this energy barrier. Specifically, the enzyme lowers the
energy requirement of the reaction by stabilizing the transition state. Third, the
cell can regulate the enzyme and thereby affect its activity. Through this mecha-
nism, the cell is able to steer substrates into alternative pathways and toward
products that are in highest demand. For instance, pyruvate may be used as the
starting substrate for the generation of oxaloacetate, which then enters the cit-
ric acid cycle, or it may be used to create acetyl-CoA, which can subsequently be
used in fatty acid biosynthesis or different amino acid pathways. Pyruvate can
also be converted into lactate and used by some organisms for the generation of
ethanol (Figure 7.7).
The “decision” of how much substrate should be channeled toward alternative
pathways and products is the result of complex regulation. The most immediate
controllers are the enzymes that catalyze the various possible conversions of the
substrate. Indirectly, the decision is much more distributed among factors that reg-
ulate the activities of these enzymes. This regulation is implemented in various, dis-
tinct ways. First, metabolites within or outside the target pathways can enhance or
inhibit the activity of an enzyme. This modulation occurs essentially immediately.
Figure 7.7 The metabolite pyruvate

is involved in numerous biochemical
reactions. These reactions are catalyzed
by enzymes (identified by their Enzyme
Commission (EC) numbers in boxes). Such
pathways are discussed further in Chapter 8.
(Courtesy of Kanehisa Laboratories.)
In some cases, the product of a pathway directly binds to the active site on the
enzyme, thereby competing with the substrate and slowing down the conversion of
substrate to product. An alternative is allosteric inhibition or activation, where the
product attaches to the enzyme in a location different from the active site. This phe-
nomenon leads to a physical deformation of the active site that alters the efficiency
of substrate binding and is called cooperativity. The second mode of regulation is
exerted by other proteins, which can quickly activate or deactivate an enzyme
through the attachment or removal of a key molecular grouping, such as a phos-
phate or glycosyl group. More permanently, the regulatory protein ubiquitin can
bind to the enzyme and mark it for degradation by the proteasome, as we discussed
before. Finally, the activity of enzymes can be changed over much longer time hori-
zons by means of gene regulation. Namely, the genes coding for the enzyme or its
modulators may be up- or down-regulated, which ultimately affects the amount of
enzyme present in the cell and therefore its throughput. As one might expect, the
different modes of regulation are not mutually exclusive, and situations such as
environmental stresses can easily evoke some or all of them simultaneously (see
Chapter 11).
7.3 Transporters and Carriers

In addition to enzymes, transport or carrier proteins are crucially important for
metabolism. Transport is needed between the inside and outside of a cell, within
a cell, and outside cells, for instance, in the bloodstream. Transport across the cell
membrane, in either direction, is often facilitated by transmembrane proteins,
which accomplish their task either through active transport or through facilitated
diffusion in the case of small molecules. A very important class of these proteins
carry ions across the membrane; this ion transport is central to a vast number of
processes, including the action and deactivation of nerve and muscle cells. We
will discuss this in greater detail in Chapter 12.
Transport of small molecules within cells often occurs via diffusion. However,
larger molecules are too big for this process to be efficient. Instead, they are often
packed into vesicles, where they can be stored and eventually expelled somewhere
else inside the cell or to the outside. Vesicles typically consist of lipid bilayers, but it
is the current understanding that proteins are often involved in their formation and
their docking to a membrane [35]. An important example of vesicle dynamics is the
movement of neurotransmitters, such as dopamine, from the nerve cell into the
synaptic cleft between cells, a process that is the basis for signal transduction in the
brain. Another example of vesicle transport is the uptake of lipids from the intes-
tines into the bloodstream via the lymphatic system [36].
A different, very interesting, type of movement within cells is facilitated by kine-
sin and dynein proteins, which use chemical energy from the hydrolysis of ATP to
move along microtubule cables [37]. This movement is involved in a number of pro-
cesses, including mitosis and the transport of molecules within the axons of nerve
cells. The microtubules themselves also consist of proteins.
Transport between neighboring cells can also occur through different mecha-
nisms. A particularly prevalent task is cell-to-cell communication, which is often
accomplished by sending ions or other small molecules through physical chan-
nels. These channels, called gap junctions, are formed by proteins that span the
intracellular space between neighboring cells. Specifically, these connectors con-
sist of two half-channels that each span the membrane of one of the neighboring
cells and are composed of four helices. This arrangement constitutes a positively
charged entrance within the cytoplasm, a funnel leading into a negatively charged
tunnel, and an extracellular cavity. The funnel itself is formed by six helices that
determine the maximal size of molecules entering the channel [19]. Gap junctions
can be opened or closed through rotating and sliding motion of their component
proteins, called connexins, which thereby create or close a center opening (Figure
7.8). Important examples of gap junctions are found in the heart, where they help
the cardiac myocytes contract in a coordinated fashion, making up the heartbeat
(see Chapter 12).
Transport outside the cell includes the delivery of oxygen from the arterial blood
to all tissues of the body. Responsible for this process is the iron-containing
(A) (B)
(C)
Figure 7.8 Gap junctions are composed of proteins called connexins. (A) Side view of a connexin complex. The red domains span the membranes
of two neighboring cells, while the yellow domains bridge the extracellular space. (B) Top view of the same connexin complex. (C) Symbolic representation
of the opening and closing of a gap junction by means of rotating connexin subunits. (Protein Data Bank: PDB 2ZW3 [38].)
metalloprotein hemoglobin, which accounts for about 97% of a red blood cell’s dry
mass (Figure 7.9). Like many enzymes, this protein is subject to potential competi-
tive and essentially irreversible inhibition of its binding of oxygen, for instance by
carbon monoxide, which is therefore very poisonous [39]. Another carrier protein is
serum albumin, which transports water-insoluble lipids through the bloodstream.
Albumin is the most abundant protein in blood plasma. It is a tightly packed globu-
lar protein with lipophilic groups on the inside and hydrophilic side groups on the
(A) (B) CH2 CH3
H
C
H3C CH2
N N
Figure 7.9 Heme and hemoglobin.

Fe CH (A) Human hemoglobin A is a globular
HC
metalloprotein (red) composed of four
subunits that anchor the planar heme
N N
groups (the ball-and-stick assembly) (Protein
H3C CH3 Data Bank: PDB 3MMM). (B) Each heme
C group consists of a complex porphyrin ring
H with four nitrogen atoms that holds an iron
(Fe2+) ion, to which oxygen can bind. (From
Brucker EA. Acta Crystallogr, Sect. D 560
HOOC COOH
[2000] 812–816.)
outside. The primary task of serum albumin is the regulation of osmotic pressure in
the plasma, which is needed for an adequate distribution of fluids between the
blood and extracellular spaces.
Another class of carrier proteins contains cytochromes, which operate in the
electron transport chain and thus in the respiration process. They receive hydro-
gen ions from the citric acid cycle and combine them with oxygen to form water.
This process releases energy, which is used to generate ATP from ADP. Cyto-
chromes can be inhibited by the poison cyanide, which therefore suppresses cel-
lular respiration.
Proteins in the extracellular matrix, such as fibronectins, are also involved with
the coordinated transport of materials between blood capillaries and cells. This
interstitial transport is important for the delivery of substrates. It is also the key
mechanism for exposing cancer cells to chemotherapy [40]. We will discuss fibronec-
tins later.
A particularly fascinating transport protein complex is the flagellar motor, which
allows many bacteria and archaea to move (Figure 7.10) [41]. The flagellum itself is
a hollow tube that is made from the protein flagellin and contains an internal cyto-
skeleton composed of microtubules and other proteins that render sliding and
rotating motions possible [37]. Its shaft runs through an arrangement of protein
bearings in the cell membrane. It is turned by a rotor that is located on the inside of
the cell membrane and consists of several distinct proteins. This flagellar motor
converts electrochemical energy into torque and is powered by a proton pump in
the cell membrane, which moves protons, or in some cases sodium, across the
membrane. The flow of protons is possible because of a metabolically generated
concentration gradient. The motor can achieve several hundred rotations of the fla-
gellum per minute, which translates into an amazing bacterial movement speed of
about 50 cell lengths per second. Furthermore, by slightly changing the positioning
of a specific type of motor protein, the organism can switch direction.
Flagella and similarly structured cilia are not only found in prokaryotes but are
also crucial in eukaryotes, where they often have a more complex microtubular
structure [42]. Prominent examples are sperm cells, cells of the fallopian tube, which
move the egg from the ovary to the uterus, and endothelial cells in the respiratory
tract, which use cilia to clear the airways of mucus and debris. Mammals often
develop immune responses to flagellar antigens such as flagellin, with which they
fight bacterial infections. Key components in these immune responses are proteins
flagellum
hook
rod
L-ring
outer membrane
P-ring
peptidoglycan
layer
Figure 7.10 Diagram of a flagellar motor
stator
from a bacterium. This complex protein
arrangement allows bacteria to swim in
cytoplasmic different directions. Similar assemblies are
membrane
found in sperm cells, in the fallopian tubes
MS-ring (where eggs need to be moved from the
motor switch
ovary to the uterus), and in the respiratory
Mot protein tract (where cilia clear the airways of mucus
C-ring and debris).
that recognize certain molecular patterns and are called Toll-like receptors (TLRs)
[43], which we will discuss in Section 7.5.
7.4 Signaling and Messenger Proteins

Cells in higher organisms communicate with other cells and with the outside on a
regular basis. They send and receive signals, which can be chemical, electrical, or
mechanical, and must respond to these signals in a coordinated manner. This pro-
cess, which is generically called signal transduction, is of obvious and utmost
importance, and the whole of Chapter 9 is devoted to it.
As just one example, let us consider a particularly important class of membrane-
spanning signaling proteins, called receptor tyrosine kinases (RTKs). Generically,
kinases are enzymes that transfer a phosphate group from a donor such as ATP to a
substrate. In the case of RTKs, this enzymatic action amounts to the initiation of sig-
nal transduction. Namely, when a ligand, such as an epidermal growth factor, cyto-
kine, or hormone, binds to the receptor on the outside of the cell, its protein domain
on the inside (that is, in the cytoplasm) initiates autophosphorylation. This process
leads to the recruitment of signaling proteins and thereby starts the intracellular sig-
nal transduction process. RTK-based processes are crucial in healthy cells and also
seem to be important in the development and progression of cancers, as well as neu-
rodegenerative diseases such as Alzheimer’s disease and multiple sclerosis. An
example of an epidermal growth factor receptor is shown in Figure 7.11.
In addition to receptors and signaling proteins, the signal transduction process
is supported by scaffold proteins, which interact with signaling proteins in a fashion
that is not entirely understood. It seems that scaffold proteins assist in the appropri-
ate localization of pathway components to the cell membrane, the nucleus, the
Golgi apparatus, and other organelles. Scaffold proteins furthermore assist the con-
densation of DNA and are involved with the recognition of surface structures of
viruses by antibodies [45].
It should be noted that some, although not all, hormones are proteins or pep-
tides, which act over long distances. Hormones are produced by animal cells in
glands, such as the pituitary, thyroid, adrenal glands, and the pancreas, are trans-
ported through the bloodstream, and bind to specific receptors in distant locations
of the body, where they signal the need for some action. Examples are the blood-
sugar-regulating protein insulin, growth hormone, the blood-vessel-control hor-
mones angiotensin and vasopressin, and the exhilaration-causing endorphins, as
well as the more recently identified “hunger hormones” orexin and ghrelin. Plants
also produce hormones.
Figure 7.11 Crystal structure of the

human epidermal growth factor receptor.
Two types of ligands are shown. The
stretched-out ligands at the top and center
left are nonaethylene glycol molecules. On
the right, one can see several N-acetyl-d-
glucosamine (GlcNAc) molecules, which are
components of a biopolymer in bacterial cell
walls, called peptidoglycan. Interestingly,
GlcNAc is also a neurotransmitter. (Protein
Data Bank: PDB 3NJP [44].)
Figure 7.12 A mouse antibody recognizes

a so-called Tn antigen (center top
and bottom). Individual spheres are
zinc molecules. Tn antigens have been
associated with many cancers and made
them targets for the development of cancer
vaccines. Tn stands for threonine/N-acetyl-
d-galactosamine α-O-serine. (Protein Data
Bank: PDB 3IET [46].)
7.5 Proteins of the Immune System

Proteins are crucial players at many levels of the vertebrate immune system. Anti-
bodies, or immunoglobulins, are globular plasma proteins that are produced by
white blood cells (Figure 7.12). They are either attached to the surface of immune
cells, called effector B cells, or circulate with the blood and other fluids. A typical B
cell carries tens of thousands of antibodies on its surface. Antibodies bind to neu-
tralize foreign molecules (antigens) as they are found on the surfaces of bacteria,
viruses, and cancer cells, as well as nonliving chemicals, such as drugs and toxins.
Owing to this binding, the invaders lose their ability to enter their target cells in the
host. The antibodies also initiate the destruction of pathogens by macrophages.
Antibodies typically consist of a complex of two heavy-chain and two light-chain
proteins, and while these always form the same generic structure, specific genetic
mechanisms have evolved to generate tips at the ends of these proteins that contain
extremely variable antigen-binding sites [47]. When B cells proliferate in response
to an insult, high mutation rates increase the variability of these sites. As a conse-
quence, the antigen-binding sites are capable of distinguishing millions of antigens
by their structural features, or epitopes. In autoimmune disease, the same antibody
responses are directed against the body’s own cells.
A second group of proteins that are crucial for immune responses is the large family
of cytokines. These signaling molecules are used for cell-to-cell communication and
drive the progression of inflammatory processes. Cytokines are produced by T helper
cells, which can respond to antigen peptides within a few seconds, yet require hours of
signal exposure to become fully activated [48, 49]. Cytokines help recruit further
immune cells and macrophages to the site of injury or a bacterial or viral infection.
Such an infection can trigger 1000-fold increases in cytokines, and slight imbalances in
the numbers and types of cytokines can lead to “cytokine storms,” sepsis, multi-organ
failure, and death. Well-known examples of cytokines include interleukins and inter-
ferons. We discuss some recent modeling approaches to understanding the exceed-
ingly complicated dynamic system of cytokines in Chapter 15. A final group of “defense
proteins” consists of enzymes in tears and the oils of the skin that guard against invad-
ers and thus contribute to the innate immunity of the body.
Prominent proteins at the intersection of signaling and immune response are the
Toll-like receptors (TLRs)—toll is German for “Wow!”—which function as sentinels
that recognize numerous epitopes, including peptidoglycans and lipopolysaccha-
rides, which are typical components of bacterial cell walls (Figure 7.13). Upon recog-
nizing a specific epitope, the TLR activates pro-inflammatory cytokines and signaling
pathways and thereby initiates a proper immune response. TLRs come in a number of
variants and form a very complex functional interaction network [52].
Figure 7.13 Toll-like receptors (TLRs)

sense and bind various ligands, a process
that initiates a signaling cascade and
leads to the production of inflammation
mediators. Shown here is the outer domain
of the human TLR3, which recognizes
double-stranded RNA, a molecular signature
of viruses. The horseshoe shape is due
to 23 repeated leucine-rich domains and
establishes a distinct receptor for pathogen
recognition, such as the cell surface
sugars fucose, mannose, and N-acetyl-d-
glucosamine (GlcNAc), which can be seen
inside and outside the receptor. (Protein
Data Bank: PDB 3CIG [50]; visualization with
methods from [51].)
7.6 Structure Proteins

Collagen is the most abundant protein in humans. It is found in many tissues that pro-
vide structural support to the body, including bone, cartilage, connective tissue, ten-
dons, ligaments, and the discs in the spine. It also constitutes between 2% and 5% of
muscle tissue and is a component of blood vessels and the cornea of the eye. Collagen
supports the structure of the extracellular matrix that provides support to cells and is
involved in cell adhesion and the transport of molecules within the space between cells.
Collagen consists of three polypeptide strands that contain a large number of
repeated patterns of specific amino acids, such as glycine–proline–AA or glycine–
AA–hydroxyproline, where “AA” represents some other amino acid. Owing to the
presence of these patterns, the amount of glycine is unusually high—the only other
protein with this feature seems to be a protein in silk fibers. The abundance of gly-
cine, combined with the relatively high content of proline and hydroxyproline rings,
causes collagen polypeptides to curl into left-handed helices. These helical strands
in turn are twisted into a right-handed triple helix, which is thereby a prime example
of a quaternary structure that is stabilized by numerous hydrogen bonds. Each triple
helix can form a super-coiled collagen microfibril, and many of these microfibrils
make up a strong yet flexible macro-structure. When hydrolyzed, collagen becomes
gelatin, which has multiple uses in the food industry.
Like other proteins, collagen is involved in control tasks as well, especially in the
context of cell growth and differentiation. For instance, it induces the receptor tyro-
sine kinase DDR2 (member 2 of the discoidin domain receptor family), which trig-
gers autophosphorylation of targets in the cytosol (Figure 7.14). To permit these
control tasks, some sections of the collagen polypeptide do not show the regular
curling patterns described above.
Complementing the load-bearing strength of collagen, elastins equip tissues
with shape flexibility, which is particularly important in muscle cells, blood vessels,
the lungs, and skin. Like collagen, elastins contain high amounts of glycine, but also
similar amounts of proline, valine, and alanine. Fibroblasts secrete the glycoprotein
fibrillin into the extracellular matrix, where it is incorporated into microfibrils that
are thought to provide a scaffold for the deposition of elastin.
Another structural protein is keratin, which is the base material for skin, hair,
nails, horns, feathers, beaks, and scales. Keratin contains many α-helices, stabi-
lized with disulfide bonds between cysteine residues. Similar to collagen, three
Figure 7.14 Human collagen (diagonal

strands) complexed with the discoidin
domain receptor DDR2 (to the right).
DDR2 is a common receptor tyrosine kinase
that is activated by triple-helical collagen.
The binding of collagen to DDR2 causes
structural changes in the surface loops
of the receptor and has been linked to
its activation. The receptors control cell
behavior and are incorrectly regulated in
several human diseases. (Protein Data Bank:
PDB 2WUH [53].)
α-helices are twisted around each other into a protofibril. Eleven of these form a
tough microfibril, in which nine protofibrils are wound around two center proto-
fibrils. Hundreds of microfibrils form a macrofibril. Interestingly, the toughness of
keratin is only matched by chitin, which is not a protein but a long-chain polymer
of N-acetylglucosamine that provides strength to the protective armor of crusta-
ceans like crabs and lobsters, and is also found on the surfaces of bacteria.
Actin is a globular protein in essentially all eukaryotic cells. It has two important
roles. First, it interacts in muscle cells with the protein myosin. Actin forms thin fila-
ments, while myosin forms thick filaments. Together they form the contractile unit
of the muscle cell (see Chapter 12). Second, actin interacts with cellular membranes.
In this role, actin and the proteins cadherin and vinculin form microfilaments.
These microfilaments are components of the cytoskeleton and participate in numer-
ous fundamental cellular functions, including cell division, cell motility, vesicle and
organelle movement, and the creation and maintenance of cell shape. They are also
directly involved in processes such as the phagocytosis of bacteria by
macrophages.
The lens of the mammalian eye is a complicated structure that contains fibers
composed of long transparent cells. More than a third of its weight is due to proteins.
These proteins, which have fascinated scientists for over 100 years, include crystal-
lins [54], collagen, and the membrane protein aquaporin [55], which maintains an
unusually tight packing between neighboring cells (Figure 7.15). Intriguingly, there
is very little turnover in the lens, and humans retain their embryonic lens proteins
throughout life. Age-related changes in the lens, such as cataracts, are therefore dif-
ficult to repair.
The versatility of proteins has resulted in uncounted roles associated with the
maintenance of shape and structure. In many cases, these roles are performed with
“mixed” molecules that consist of proteins and something else. For instance, apoli-
poproteins, together with several types of lipids, can form lipoproteins, which serve
as carriers for lipids in the blood, and also as enzymes and structural proteins. Well-
known examples are high-density and low-density lipoproteins (HDLs and LDLs),
which, respectively, carry cholesterol from the body’s tissues to the liver and vice
versa. They are sometimes referred to as “good” and “bad” cholesterols,
respectively.
Glycoproteins contain sugars (glycans) that are bound to the side chains of the
polypeptides. Glycans are often attached during or after translation of the mRNA
into the amino acid chain, in a process called glycosylation. Glycoproteins are often
integrated into membranes, where they are used for cell-to-cell communication.
Prominent examples are various hormones and mucins, which are contained in the
Figure 7.15 The vertebrate eye lens

protein aquaporin-0 (AQP0). AQP0 consists
of a complex of four subunits, which are
arranged in a three-dimensional lattice.
According to current understanding, AQP0
proteins mediate the connections among
the very tightly packed fiber cells of the lens.
(Protein Data Bank: PDB 2C32 [55].)
mucus of the respiratory and digestive tracts. Glycoproteins are important for two
aspects of immune responses. First, some of these proteins are mammalian immu-
noglobulins or surface proteins on T cells and platelets. Second, glycoproteins form
structures on the surfaces of bacteria such as streptococci, which are recognized by
the host’s immune system.
CURRENT CHALLENGES IN PROTEIN RESEARCH

Given the enormous distribution and versatility of proteins, it is not surprising that
intense research efforts are ongoing at all levels, from protein chemistry to the
dynamics of interacting protein systems. Outside pure chemistry and biochemistry,
current challenges fall into the categories of proteomics, structure and function pre-
diction, localization, and systems dynamics. Of course, there is not always a clear
boundary between these activities.
7.7 Proteomics
The proteome is the totality of proteins in a system, and proteomics attempts to
characterize it. This characterization typically includes an account of the inventory
and the abundance of many or all proteins at a given point in time. As is to be
expected, both inventory and abundance can significantly differ among the tissues
of an organism. Extensive research has addressed the spatial distribution and local-
ization of target proteins, as well as temporal changes due to physiological trends,
such as differentiation and aging, or in response to infection or disease. It is hoped
that early identification of these spatial and temporal changes might suggest diag-
nostic biomarkers that show up before a disease manifests.
Proteomics is indeed a grand challenge, because it is not known how many pro-
teins exist in model organisms. It is not even clear how to count them. For instance,
should the same protein with two different post-translational modifications count as
one or two proteins? Many signaling proteins and enzymes can be phosphorylated or
dephosphorylated, and these processes change the activity state of the protein. Should
both forms be counted separately? Methylation, glycosylation, arginylation, ubiquiti-
nation, oxidation, and other modifications raise similar questions. Finally, newly
formed transcripts can be spliced in different ways, thereby leading to different pro-
teins. These splice variants are presumably the reason that humans can manage with
only about 20,000–25,000 genes; some trees are thought to possess twice as many.
CURRENT CHALLENGES IN PROTEIN RESEARCH 219
A good start to identifying proteins, and to comparing and discerning two similar
tissues or protein systems, for instance in the characterization of cancer cells, is one-
or two-dimensional (2D) gel electrophoresis [56]. In the one-dimensional case,
proteins are first coated with a negatively charged molecule, called SDS (sodium
dodecyl sulfate). In this coating, the number of negative charges is proportional to
the size of the protein. Thus, the proteins can be separated according to their molecu-
lar size, which is reflected by their mobility through a gel that is placed into an elec-
tric field. In the typical 2D case, the proteins are first separated with respect to their
isoelectric point, which is the pH at which the protein has a neutral charge. In the
second dimension, they are separated with respect to size. In addition to these tradi-
tional features, other separation criteria have been developed (see, for example,
[57]). As the result of the separation in two coordinates, every protein migrates to a
specific spot on the gel. These spots can be visualized with various stains, such as a
silver stain, which binds to cysteine groups in the protein, or with a specific dye from
the wool industry, called Coomassie Brilliant Blue. The degree of staining indicates to
a reasonable approximation the amount of protein in a given spot. The 2D separation
is effective, because it seldom happens that two proteins have the same mass as well
as the same isoelectric point. If a protein is absent in one of the two systems being
compared, its spot remains empty. If it is phosphorylated or otherwise modified, it
migrates in a predictable fashion to a different spot. One important advantage of this
method is that one does not even have to know which protein has migrated to a par-
ticular spot. Thus, the method can be used as a discovery tool. Many software pack-
ages are available for supporting the analysis of 2D gel electrophoreses. A variation is
DIGE (difference gel electrophoresis), which shows the differences in proteomic pro-
files between comparable cells in different functional states (Figure 7.16) [58].
Numerous other methods for separating proteins have been developed in recent
years [59, 60]. These include MudPIT (multidimensional protein identification tech-
nology), which uses two-dimensional liquid chromatography, and the so-called
iTRAQ technique (isobaric tags for relative and absolute quantitation). iTRAQ is
used to compare protein mixtures from different sources in the same experiment.
The key is that the proteins in each sample are “tagged” with specific markers that
contain different isotopes of the same atom and therefore have slightly different
masses, which can be distinguished with mass spectrometry (see below).
Two-dimensional gel electrophoresis identifies spots where proteins are pres-
ent. In favorable cases, these same spots have been observed in other studies, and a
comparison with corresponding databases suggests what the so-far unknown pro-
tein might be [61]. If this comparison is unsuccessful, one can cut the protein spot
out of the gel or use the liquid from MudPIT and analyze it in order to deduce its
amino acid sequence. Once the sequence is known, comparisons with protein data-
bases provide clues about the nature and function of the target protein [2].
Figure 7.16 Two-dimensional difference

gel electrophoresis (DIGE) image of
lung alveolar type II cells from mouse.
The globin transcription factor GATA1 was
knocked down in these cells. Each spot
indicates the presence of a protein with a
specific charge and mass. Different colors
identify differences between knock-downs
and normal mice. (Courtesy of John Baatz,
Medical University of South Carolina.)
Figure 7.17 Tandem mass spectrum of
377.76 b6+2
the tryptic peptide RHPEYAVSVLLR from
427.27 b7+2
bovine serum albumin. The plot shows,
100
in the form of peaks, the fragments of the
587.27 y5
90 molecule as they were detected in the mass
spectrometer. The height of each peak
80 indicates the abundance of the observed
fragment, while the location indicates the
70 mass-to-charge ratio (m/z). Since peptides
tend to fragment at the linkages between
470.82 b8+2
511.35
relative abundance
60 specific amino acids, it is possible to deduce

the amino acid sequence of the peptide
401.25 y3
50
754.21 b6
from the ionized fragments. Designations
520.51 b4, b9+2
460.88 y8+2
like b4, b7, and y5 are codes within a

573.76 y10+2
40
500.29 y4
specific nomenclature for labeling peptide

342.22 b5+2
686.31 y6 fragments. (Courtesy of Dan Knapp and

294.10 b2
30
588.27
Jennifer Bethard, Mass Spectrometry Facility

288.18 y2
633.39 b11+2
260.68 b4+2
of the Medical University of South Carolina.)
853.22 b7
20
683.15 b5
755.21
757.30 y7
920.38 y8
1039.42 b9
1049.35 y9
940.22 b8
687.39
343.66
474.75
10
328.24
854.22
246.70
378.70
921.38
564.74
651.75
428.32
295.07
589.48
1038.76
235.08
1129.62
1153.42
688.35
560.20
739.23
941.32
603.57
902.38
198.99
855.39
546.67
825.34
781.98
814.53
972.20
994.70
0
200 300 400 500 600 700 800 900 1000 1100
m/z
The determination of the protein sequence from an isolated spot of a 2D electro-

phoresis or another sample of protein is typically accomplished with mass spec-
trometry (MS). One of the most widely used variants of this technique is MALDI–
TOF, which stands for matrix-assisted laser desorption/ionization–time of flight.
In this method, a sample of peptide solution is deposited on a substrate (the matrix).
A laser is beamed onto the matrix and desorbs (ejects) and ionizes (electrically
charges) the peptides. The ionized peptides are shot through an electric field toward
a detector, which measures for each peptide the time in flight. This time depends on
the mass of the peptide and the applied charge. Correspondingly, the output of a
MALDI–TOF experiment shows the amounts of material at the detector, which are
exhibited in the form of peaks, as a function of mass divided by charge (m/z). There is
ample literature on this and other separation techniques.
It is difficult to analyze large intact proteins by MS. A necessary preparatory step
is therefore the cutting of a protein into smaller peptides. This step is achieved with
enzymes such as trypsin that cut peptide bonds between predictable pairs of amino
acids. While this step solves the problem of large masses, it creates a new problem.
Namely, instead of one protein, one now has to address thousands of peptide frag-
ments that are mixed up. Thus, in a second round of a tandem MS, or MS/MS, analy-
sis, these peptides are typically forced to collide with an inert gas, which randomly
breaks one of the remaining peptide bonds in the fragment, thereby replacing it with
a positive charge. The result is a collection of “b” and “y” peaks, which correspond to
the left or right pieces of the broken-up peptide. The fragment y1 represents the first
amino acid in the peptide, the m/z difference between y1 and y2 represents the sec-
ond amino acid, and so on (Figure 7.17). Computer algorithms have been devel-
oped to reconstitute the peptides from these fragment mixtures. Current research is
focused on, among other things, miniaturization of the process, which would permit
proteomic analysis with smaller protein samples than is possible now [62], and the
development of methods that permit larger molecules to be analyzed without frag-
mentation [63].
7.8 Structure and Function Prediction

The amino acid sequence of a protein can be deduced from the nucleotide
sequence of its gene, if it is known, or determined with methods of gel electropho-
resis and mass spectrometry. Once the amino acid sequence is known, each resi-
due is known, and it may seem a small step to derive what the protein looks like
from the chemical and physical features of all the residues. Alas, this intuitive
deduction is utterly wrong, and the identification or prediction of the 3D structure
of a protein from its amino acids is in reality an extremely difficult challenge that
has attracted enormous experimental and computational attention. Nonetheless,
addressing this challenge is very important, because one rightly expects that the
3D structure of a protein is instrumental for its function. As a consequence, an
entire branch of biology, called structural biology, focuses on macromolecules
and, in particular, proteins.
In order to dissect the big challenge of protein structure and function predic-
tion into smaller steps, the field has defined four types of structures. The primary
structure of a protein refers to its amino acid sequence. The secondary structure is
a localized feature involving regular 3D shapes or motifs. The best understood of
these motifs are α-helices, β-sheets, and turns. Each α-helix is the consequence of
regularly spaced hydrogen bonds between backbone oxygen and amide hydrogen
atoms, while a β-sheet occurs when two strands are joined by hydrogen bonds and
each strand contains alternating residues. Turns and bends are due to three or four
specific sequences of amino acids that permit tight folds that are stabilized with
hydrogen bonds.
The local motifs are usually interrupted by less well-defined stretches within
the protein, and all defined and undefined portions combine to form the tertiary
structure, that is, the overall folding of the protein. The tertiary structure of a pro-
tein is maintained by a hydrophobic core, which prevents the easy dissolution of
the structure in the aqueous medium of the cell, and furthermore through chem-
ical bonds, such as hydrogen bonds, disulfide bonds, and salt bridges. One should
note that this tertiary structure is not ironclad, and post-translational modifica-
tions, changes in environmental conditions, and various cellular control mecha-
nisms of protein activation and deactivation can alter it significantly [11, 13]. This
flexibility in folding is crucial, because it is directly related to the function of the
protein. A premier example is the activation of an enzyme by a metabolic modu-
lator, which we discussed before. Finally, many proteins consist of subunits and
are functional only as complexes. The structure formed by two or more protein
subunits is called the quaternary structure of a protein. The functional rearrange-
ments, which may affect both the tertiary and quaternary structure, are often
called conformations, and transitions between such states are called conforma-
tional changes.
Two distinct approaches can be employed to identify the 3D structure of a pro-
tein. On the experimental side, crystallography, cryo-EM, and NMR are among the
most widely used methods. On the computational side, very large computers and
customized, fast algorithms are being used to predict the 3D structure of proteins
and their binding sites from their amino acid sequences, the chemical and physical
features of their amino acid residues, and possibly other available information [64–
67]. While some successes have been reported, this protein structure prediction task
is in general an unsolved problem. Reliable structure prediction is a very important
and ultimately rewarding problem, because it has immediate implications for an
understanding of disease and the development of specific drugs.
As a first step toward full predictability, powerful computational methods have
been developed to identify smaller, recurring 3D structures within proteins, such
as α-helices and β-strands. α-helices (Figure 7.18) cause the protein locally to
coil, while β-strands form stretched-out sheets, which are often connected with Figure 7.18 Helical bundles form a
U-turns or hairpins. In some channel proteins, these sheets are arranged into transmembrane potassium channel.
barrels (Figure 7.19). β-sheets are quite common and have been implicated in a The channel is shown in complex with
charybdotoxin. These types of ion channels
number of human diseases, such as Alzheimer’s disease. There is a rich literature play a critical role in signaling processes
on these and other structural motifs in proteins. and are attractive targets for treating
Prediction of the function of a protein from knowledge of its amino acid various diseases. The structure reveals how
sequence is very difficult. One must remember that while the nucleotide charybdotoxin binds to the closed form of
sequence provides the initial code for peptides and proteins, many modifications KcsA and thereby makes specific contact
can occur between the translation of mRNA and the availability of the fully func- with the extracellular surface of the ion
tional protein. These modifications can affect many physical and chemical prop- channel, resulting in pore blockage. This
erties of the protein and thereby alter its folding, stability, activity, and function. blockage yields direct structural information
about an ion channel complexed to a
Furthermore, peptides may be spliced together in different ways, thereby forming peptide antagonist. (Courtesy of Huiling
distinct proteins. Presently available methods for predicting the so-far Chen and Ying Xu, University of Georgia.)
Figure 7.19 Example of a β-barrel. The

outer membrane enzyme phospholipase
A of Escherichia coli participates in the
bacterium’s secretion of colicins, which
are toxins that are released into the
environment to reduce the competition
from other bacterial species. The protein
forms a β-barrel, whose activity is regulated
by reversible dimerization of the enzyme,
that is, by the formation of a complex of two
subunits (green and red). As a result of the
dimerization, the barrels become functional
channels for oxyanions. (Courtesy of Huiling
Chen and Ying Xu, University of Georgia.)
uncharacterized function of a protein include a variety of similarity, homology,

and structure comparisons with proteins of known sequence and function [61].
Also helpful in many cases are phylogenetic approaches and information from
the Gene Ontology Project [68] (see Chapter 6). All these methods fall within the
realm of bioinformatics and structural biology, and readers interested in the
topic are encouraged to consult texts such as [5]. An extension of these computa-
tional methods is the field of molecular dynamics, which attempts to describe
how molecules, for instance in proteins, move in order to achieve a specific
function.
7.9 Localization
To serve their specific functions, proteins must be in the correct location within a
cell or tissue. It is therefore of great interest to study which proteins are located
where. An intriguing method is the attachment of the small green fluorescent
protein (GFP) to a target protein [69, 70]. This attachment can be achieved
through genetic engineering by expressing in the cell of interest a fusion protein
consisting of the natural target protein with a GFP reporter linked to it (Figure
7.20). GFP was first isolated from a jellyfish, but many variants in different colors
have since been engineered. GFP consists of a β-barrel containing a chromo-
phore, which is a molecule that absorbs light of a certain wavelength and emits
another wavelength. GFP is not limited to locating proteins but is widely used in
molecular biology for reporting the presence of a variety of cellular components
and the expression of genes that are of particular interest. In fact, the discovery of
GFP has truly transformed cell biology, especially in conjunction with automated
real-time systems for fluorescence microscopy in living cells. For example, Sin-
gleton and colleagues used enhanced GFP-signaling sensors to investigate the
spatial and temporal processes of 30 signaling intermediates, including recep-
tors, kinases, and adaptor proteins, within individual T cells during their activa-
tion [71]. Because of the importance of protein localization, a dedicated website
has been established [72]. Martin Chalfie, Osamu Shimomura, and Roger Y. Tsien
were awarded the 2008 Nobel Prize in Chemistry for the discovery and develop-
ment of GFP.
A distinctly different methodology was developed by Richard Caprioli’s labora-
tory at Vanderbilt University. This methodology scans tissue slices and applies
MALDI individually to very many small areas in a tight grid. The result is a peptide
spectrum for every grid point, and visualization reveals a very detailed picture of
peptide abundances within the tissue section (Figure 7.21). This MALDI imaging
method even permits the creation of 3D pictures of peptide abundances in tissues
(Figure 7.22). For such an analysis, the tissue is thinly sliced and every slice is ana-
lyzed with 2D MALDI imaging. Computer software is used to stack up all these
images to create a 3D impression [73].
100
cells with free microtubules (MT)
mCherry-tubulin GFP-patronin mCherry-tubulin

GFP-tubulin
80
(% of total)
60
40
20
0
No Pat expression Pat expression
3 UTR RNAi 3 UTR RNAi
0 MT 1–5 MT > 5 MT
patronin merge with -tubulin patronin merge with -tubulin
interphase metaphase
prophase cytokinesis
Figure 7.20 Green fluorescent protein (GFP) and other tagging proteins have become powerful tools for localization studies. Shown here are the
specific intracellular locations of the proteins patronin and tubulin in different phases of the cell cycle. (From supplements to Goodwin SS & Vale RD. Cell
143 [2010] 263–274. With permission from Elsevier.)
2h 6h
(A) (A)
(B) (B)
(C) (C)
(D) (D)
Figure 7.21 Mass spectrometry can be used to image the spatial distribution of a drug and related metabolites in vivo. Shown are cross-sectional
images of a rat 2 and 6 h after dosing with the antipsychotic olanzapine. Organs are outlined in red. (A) Whole-body section across four gold-coated MALDI
target plates. (B) MS/MS ion image of olanzapine within the cross-section. (C) MS/MS ion image of the first-pass derivative N-desmethylolanzapine. (D) MS/
MS ion image of another first-pass derivative, 2-hydroxymethylolanzapine. The original drug, olanzapine, is visible throughout the body at 2 h, but shows
decreased intensity in the brain and spinal cord after 6 h. The first-pass derivatives accumulate in the liver and bladder, rather than the brain. (From Reyzer ML
& Caprioli RM. Curr. Opin. Chem. Biol. 11 [2006] 29–35. With permission from Elsevier.)
(A) (B) Figure 7.22 Reconstructed 3D localization

0.00 0.00 image of a rat brain tumor. The images
are composed of numerous 2D slices, each
evaluated with mass spectrometry. (A) and (B)
show the localization of different proteins.
X 1.05 X 1.05 (From Schwamborn K & Caprioli RM. Mol.
Oncol. 4 [2010] 529–538. With permission
from Elsevier.)
2.11 2.11
2.60 0.00 2.60 0.00
1.30 1.05 1.30 1.05
Z 0.002.11 Y Z 0.002.11 Y
7.10 Protein Activity and Dynamics

It should have become clear by now that proteins are drivers of dynamic activities at
many temporal and spatial scales. Some proteins interact with other molecules on a
timescale of seconds or even faster, as is the case with enzymes, whereas the dynamic
changes in some proteins may take many years, as we can see in the very slow bond-
ing of sugars to proteins in the lens of the eye, which clouds the lens in advanced age
and can lead to cataracts [74]. Correspondingly, studying proteins very often means
studying protein interactions. As is to be expected, these investigations can take very
different forms. At the low extreme of the time and space spectrum, the main topic of
interest is the docking of small ligands on proteins, which is the basis for enzyme
action and most drugs; these actions will be discussed in Chapter 8. Slightly further
up on the scale is the entire field of transcription factors and their crucial role in gene
regulation, which we discussed in Chapter 6. A further step up, we find protein–pro-
tein interactions, which, owing to their importance and prevalence, are often abbre-
viated as PPI, and can be found in numerous manifestations [75].
PPIs exhibit many aspects worth investigating. First, there is great interest in
complexes of a small number of peptides or proteins in quaternary structures. The
relevant methods of investigation require the techniques of theoretical chemistry
and algorithmic optimization, similar to those used for the analysis of ligand dock-
ing (Figure 7.23). Second, enormous effort in recent times has been devoted to PPI
networks of medium to very large sizes (Figure 7.24). The rationale is that decipher-
ing which proteins interact with each other, possibly in some specific location, will
provide insights into the functioning of a cell as a complex system.
Several approaches have been applied to these types of investigations. Arguably
the most prominent has been the yeast two-hybrid assay, which uses artificial fusion
proteins inside the cell nucleus of a yeast to infer the binding partners of a protein
[76, 77]. The assay is scalable, which permits the screening of interactions among
very many proteins. A significant drawback of this method is the often very high
number of false positives, which means that the test suggests many more interac-
tions than truly exist in a living cell. For relatively small interaction networks, it is
Figure 7.23 Interaction between

two proteins forming the cohesin–
dockerin complex of the cellulosome
in the anaerobic bacterium Clostridium
thermocellum. The cellulosome is a
multiprotein complex for the efficient
degradation of plant cell walls. The
β-sheet domain of cohesin (blue) interacts
predominantly with one of the helices of
the dockerin (red). The structure provides
an explanation for the lack of cross-species
recognition between cohesin–dockerin pairs
and thus provides a blueprint for the rational
design, construction, and exploitation of
these catalytic assemblies. (Courtesy of
Huiling Chen and Ying Xu, University of
Georgia.)
Figure 7.24 Network of 1785 unique

protein–protein interactions between
HIV and the human host. The graph was
assembled from a database of the National
gp41 Institute of Allergy and Infectious Diseases,
Nef
Division of AIDS. Red nodes represent HIV
capsid proteins and light orange nodes represent
p6
host factors. It is unclear at present
gp120 which interactions are direct and which
are functionally relevant. (From Jäger S,
Vif Gulbahce N, Cimermancic P, et al. Methods 53
RT [2011] 13–19. With permission from Elsevier.)
Tat
retropepsin
matrix
integrase Vpr
Rev
Vpu
nucleocapsid
possible to verify or refute these interactions with co-localization studies or through

immunoprecipitation with the help of specific antibodies [78]. It is also possible, at
least to some degree, to use computational validation methods [79]. Once PPIs have
been identified and validated, graph methods of static network analysis are used to
characterize the topological features of the interaction network, such as types of
nodes and connectivity patterns, which we discussed in Chapter 3. The effective
visualization of such networks is a challenging topic of ongoing research.
The analysis of PPIs becomes much more complicated when their dynamics is
of interest, or even if one just needs to know the directionality of the edges in the
network. Typical examples are signal transduction systems, where it is necessary
to identify which component is triggering changes in other components down-
stream (Figure 7.25). A different example of a short-term PPI is the transport of a
protein by another protein, for example between the nucleus and the cytoplasm.
Similarly, we have already mentioned the process of ubiquitination, which marks
a protein for degradation. PPIs with directionality and dynamics can be analyzed
with differential equation models. As a prominent class of examples, namely, sig-
naling systems, are discussed in Chapter 9.
Figure 7.25 Protein–protein interaction
network controlling the entry of cells into
unreplicated DNA
mitosis. Leland Hartwell, Tim Hunt, and
Paul Nurse received the 2001 Nobel Prize in
P
Physiology or Medicine for their discoveries
Cdc25C Cdc25C of key regulators of this process. While the
cyclin B
details are not critical here, the diagram
illustrates that there is clear directionality
P
in some signals. Unreplicated DNA blocks
cyclin synthesis Cdk1 Cdk1 Cdk1
the entry of cells into mitosis. The process
is governed by positive feedback between
the Cdk1–cyclin B complex and Cdc25 and
cyclin B cyclin B
by double-negative feedback between
Cdk1 and Wee1. This arrangement renders it
P possible for gradual mitotic cyclin synthesis
APC to lead to a switch-like increase in Cdk1
Wee1 Wee1
activation. Cdk1 is inactivated owing to a
negative feedback in which Cdk1 indirectly
targets cyclin B for degradation through APC.
unreplicated DNA (Adapted from Zwolak J, Adjerid N, Bagci EZ,
et al. J. Theor. Biol. 260 [2009] 110–120. With
EXERCISES
7.1. Search the literature to determine the relative 7.20. Determine from the literature the minimal
frequency of amino acids in different biological amounts of protein samples needed for different
systems. proteomics techniques.
7.2. Compute the theoretical number of different 7.21. List chemical and physical properties of amino
proteins consisting of 100, 1000, or 27,000 acids and their residues that are useful for protein
amino acids. structure prediction.
7.3. Make intuitive predictions of how much the total 7.22. Determine three reasons why protein structure
variability of proteins of a given length would prediction is challenging.
increase if pyrrolysine and selenocysteine were
7.23. Survey the literature and the Internet for different
found as often as standard amino acids. Check your
types of membrane proteins. Write a report
predictions with calculations.
describing their properties and functions.
7.4. Is the number of possible proteins with a fixed
number of amino acids affected by the relative 7.24. Explain in more detail the “positive feedback
frequencies with which different amino acids are between the Cdk1–cyclin B complex and Cdc25
found? Argue for or against a significant effect and/ and the double-negative feedback between Cdk1
or do some computations. and Wee1” in Figure 7.25.
7.5. Explore the visualization options of the PDB and 7.25. What would be the steps for setting up a dynamic
summarize the highlights in a report. model of the system in Figure 7.25? Once you have
determined a strategy, study the article by Zwolak
7.6. Search the literature and the Internet for important
and collaborators [80] and compare your strategy
physical, chemical, and biological properties of the
and theirs.
widely found proteins RuBisCO, collagen, and actin.
7.26. According to the Central Dogma of molecular
7.7. Explore whether plants have proteins, or even
biology (see Chapter 6), DNA is transcribed into
classes of proteins, that are not found in animals.
RNA, which is translated into protein, which in
7.8. Search the literature for a visualization of the this case is an enzyme that catalyzes a metabolic
allosteric inhibition site of an enzyme. reaction. Develop a simple cascaded model to
7.9. Retrieve from the literature a picture of an enzyme describe these processes, as shown in Figure 7.26,
as it binds to its substrate. where X1, X2, and X3 are mRNA, protein, and a
7.10. Discuss differences and similarities between metabolite, respectively, X4, X5, and X6 are precur-
prokaryotic flagella and eukaryotic cilia and their sors (explain what these should be), and X7 is a
molecular motors. transcription factor or activating metabolite. Set up
7.11. In a short report, explain differences between equations with power-law functions (Chapter 4) and
hormones and pheromones. Include a discussion select reasonable parameter values. Analyze what
of broad chemical categories that contain hor- happens if a gene is expressed. Study what happens
mones and pheromones. if X3 and X7 are the same metabolite.
7.12. Discuss in a one-page report how nature
accomplishes the enormous variety and adaptabil- X7
ity among antibodies. +

X4 X1
7.13. Discuss the terms “pro-inflammatory” and
“anti-inflammatory” in the context of cytokines. +
X5 X2
7.14. Discuss the concepts behind difference gel +
electrophoresis (DIGE). X6 X3
7.15. Why is it necessary to cut proteins into peptides
before using mass spectrometry? Find literature to Figure 7.26 Schematic of the Central Dogma.
back up your answers. X1 represents a transcribed mRNA, X2 a translated
protein, which here is an enzyme, and X3 a
7.16. Discuss the concepts on which the iTRAQ method metabolite produced in the reaction catalyzed by
is based. Weigh its advantages and disadvantages X2, X4, X5, and X6 represent sources or precursors
in comparison with MALDI–TOF. of each respective step. X7 is a transcription factor
7.17. Discuss the concepts behind the MudPIT method, or activating metabolite.
along with advantages and drawbacks.
7.27. How would you model post-translational
7.18. Discuss in a report advantages and disadvantages
modifications and alternative splice variants
of different separation techniques for proteins.
in the model in Exercise 7.26?
7.19. Explore the literature and the Internet and describe 7.28. Assume that metabolite X3 in Exercise 7.26 is the
the concepts of computer algorithms that substrate for a linear pathway beginning with Z1,
reconstitute peptides from digestion fragments. leading to Z2, Z3, Z4, and so on, until the final
EXERCISES 227
product Z10 is synthesized. Suppose Z10 strongly analysis with differential equations, set up a model
represses transcription at the top level of the for the cascade, select reasonable parameter
cascade. Compare simulation results with those in values, and perform sufficiently many simulations
Exercise 7.26. to write a confident report about the dynamical
7.29. Suppose a protein can be unphosphorylated, features of the cascade.
phosphorylated in some position A, phosphorylated
in some position B, or phosphorylated in positions
signal
A and B. Sketch a diagram of the different
phosphorylation states and transitions between
them. Design an ordinary differential equation
(ODE) model describing the transitions and states.
How do you assure that the total amount of protein Protein 1 Protein 1-P
does not change over time? Suppose an external
regulator inhibits one or two particular transitions.
What is its effect?
7.30. Consider the same phosphorylation states as in
Exercise 7.29. Design a Markov model describing Protein 2 Protein 2-P
the transitions and states. Compare features of the
model and of simulations with those in Exercise
7.29. Is it possible to account for external response
regulators?
7.31. Many signaling cascades are based on the Figure 7.27 Simplified signaling
phosphorylation and dephosphorylation of cascade. The protein at either level may be
proteins and have a structure as shown in simpli- phosphorylated (“-P”) or dephosphorylated.
fied form in Figure 7.27. Using methods of systems Only the phosphorylated form of Protein
1 triggers the second level of the cascade.
Figure 7.28 Protein–protein interaction

network of CRABP1. It is evident that different
nodes exhibit distinct connection patterns (see
Chapter 3). (From Akama K, Horikoshi T, Nakayama
T, et al. Biochim. Biophys. Acta 1814 [2011] 265–276.
7.32. Search the literature and the Internet for a binding protein CRABP1, which is involved
protein–protein interaction network whose in the development of the nervous system
structure was inferred with computational means. (Figure 7.28). What standard types of numerical
Write a one-page summary of the study. features that characterize this network can you
7.33. Akama et al. [81] describe the protein–protein extract?
interaction network of the cellular retinoic acid
REFERENCES
[1] Protein Data Bank (PDB). http://www.pdb.org/pdb/home/ [23] Clare DK, Bakkes PJ, van Heerikhuizen H, et al. An expanded
home.do. protein folding cage in the GroEL–gp31 complex. J. Mol.
[2] Swiss-Prot. http://ca.expasy.org/sprot/. Biol. 358 (2006) 905–911.
[3] UniProt. http://www.uniprot.org/. [24] Ago H, Kanaoka Y, Irikura D, et al. Crystal structure
[4] BioGRID. http://www.thebiogrid.org/. of a human membrane protein involved in cysteinyl
[5] Zvelebil M & Baum JO. Understanding Bioinformatics. leukotriene biosynthesis. Nature 448 (2007) 609–612.
Garland Science, 2008. [25] Zhao G, Johnson MC, Schnell JR, et al. Two-dimensional
[6] The Human Protein Atlas. http://www.proteinatlas.org. crystallization conditions of human leukotriene C4
[7] Stadtman TC. Selenocysteine. Annu. Rev. Biochem. 65 (1996) synthase requiring a particularly large combination of
83–100. specific parameters. J. Struct. Biol. 169 (2010) 450–454.
[8] Atkins JF & Gesteland R. The 22nd amino acid. Science 296 [26] Ranson NA, Clare DK, Farr GW, et al. Allosteric signaling of
(2002) 1409–1410. ATP hydrolysis in GroEL–GroES complexes. Nat. Struct. Mol.
[9] Brown TA. Genomes, 3rd ed. Garland Science, 2006. Biol. 13 (2006) 147–152.
[10] Zhou F, Olman V & Xu Y. Barcodes for genomes and [27] Schmidt-Krey I. Electron crystallography of membrane
applications. BMC Bioinformatics 9 (2008) 546. proteins: two-dimensional crystallization and screening by
[11] Zhang F, Saha S, Shabalina SA & Kashina A. Differential electron microscopy. Methods 41 (2007) 417–426.
arginylation of actin isoforms is regulated by coding [28] Schmidt-Krey I & Rubinstein JL. Electron cryomicroscopy
sequence-dependent degradation. Science 329 (2010) of membrane proteins: specimen preparation for two-
1534–1537. dimensional crystals and single particles. Micron 42 (2011)
[12] Seppälä S, Slusky JS, Lloris-Garcerá P, et al. Control of 107–116.
membrane protein topology by a single C-terminal residue. [29] Jmol: An Open-Source Java Viewer for Chemical Structures
Science 328 (2010) 1698–1700. in 3D. http://jmol.sourceforge.net/.
[13] Gsponer J, Futschik ME, Teichmann SA & Babu MM. Tight [30] Brylinski M & Skolnick J. FINDSITELHM: a threading-based
regulation of unstructured proteins: from transcript approach to ligand homology modeling. PLoS Comput.
synthesis to protein degradation. Science 322 (2008) Biol. 5 (2009) e1000405.
1365–1368. [31] Raghavendra SK, Nagampalli P, Vasantha S, Rajan S.
[14] Anfinsen CB. Principles that govern the folding of protein Metal induced conformational changes in human insulin:
chains. Science 181 (1973) 223–330. Crystal structures of Sr2+, Ni2+ and Cu2+ complexes of
[15] Levinthal C. How to fold graciously. In Mossbauer human insulin. Protein and Peptide Letters 18 (2011)
Spectroscopy in Biological Systems: Proceedings of a 457–466.
Meeting held at Allerton House, Monticello, Illinois (JTP [32] Field CB, Behrenfeld MJ, Randerson JT & Falkowski P.
DeBrunner & E Munck, eds), pp 22–24. University of Illinois Primary production of the biosphere: integrating terrestrial
Press, 1969. and oceanic components. Science 281 (1998) 237–240.
[16] Hatahet F & Ruddock LW. Protein disulfide isomerase: [33] Friedman L, Higgin JJ, Moulder G, et al. Prolyl 4-hydroxylase
a critical evaluation of its function in disulfide bond is required for viability and morphogenesis in
formation. Antioxid. Redox Signal. 11 (2009) 2807–2850. Caenorhabditis elegans. Proc. Natl Acad. Sci. USA 97 (2000)
[17] PTM Statistics Curator: Automated Curation and Population 4736–4741.
of PTM Statistics from the Swiss-Prot Knowledgebase. [34] Otterbein LR, Cosio C, Graceffa P & Dominguez R. Crystal
http://selene.princeton.edu/PTMCuration/. structures of the vitamin D-binding protein and its
[18] Vellai T & Takacs-Vellai K. Regulation of protein turnover complex with actin: structural basis of the actin-scavenger
by longevity pathways. Adv. Exp. Med. Biol. 694 (2010) system. Proc. Natl Acad. Sci. USA 99 (2002) 8003–8008.
69–80. [35] Ahnert-Hilger G, Holtje M, Pahner I, et al. Regulation of
[19] Shcherbik N & Pestov DG. Ubiquitin and ubiquitin-like vesicular neurotransmitter transporters. Rev. Physiol.
proteins in the nucleolus: multitasking tools for a ribosome Biochem. Pharmacol. 150 (2003) 140–160.
factory. Genes Cancer 1 (2010) 681–689. [36] Dixon JB. Lymphatic lipid transport: sewer or subway?
[20] Sherwood D & Cooper J. Crystals, X-rays and Proteins: Trends Endocrinol. Metab. 21 (2010) 480–487.
Comprehensive Protein Crystallography. Oxford University [37] Wade RH. On and around microtubules: an overview. Mol.
Press, 2011. Biotechnol. 43 (2009) 177–191.
[21] Jacobsen NE. NMR Spectroscopy Explained: Simplified [38] Maeda S, Nakagawa S, Suga M, et al. Structure of the
Theory, Applications and Examples for Organic Chemistry connexin 26 gap junction channel at 3.5 Å resolution.
and Structural Biology. Wiley, 2007. Nature 458 (2009) 597–602.
[22] Ranson NA, Farr GW, Roseman AM, et al. ATP-bound states [39] Birukou I, Soman J & Olson JS. Blocking the gate to ligand
of GroEL captured by cryo-electron microscopy. Cell entry in human hemoglobin. J. Biochem. 286 (2011)
107 (2001) 869–879. 10515–10529
REFERENCES 229
[40] Pietras K. Increasing tumor uptake of anticancer drugs with [62] Sen AK, Darabi J & Knapp DR. Design, fabrication and test
imatinib. Semin. Oncol. 31 (2004) 18–23. of a microfluidic nebulizer chip for desorption electrospray
[41] Kearns DB. A field guide to bacterial swarming motility. ionization mass spectrometry. Sens. Actuators B Chem.
Nat. Rev. Microbiol. 8 (2010) 634–644. 137 (2009) 789–796.
[42] Lindemann CB & Lesich KA. Flagellar and ciliary beating: [63] Grey AC, Chaurand P, Caprioli RM & Schey KL. MALDI
the proven and the possible. J. Cell Sci. 123 (2010) 519–528. imaging mass spectrometry of integral membrane proteins
[43] Bessa J & Bachmann MF. T cell-dependent and from ocular lens and retinal tissue. J. Proteome Res. 8 (2009)
-independent IgA responses: role of TLR signalling. 3278–3283.
Immunol. Invest. 39 (2010) 407–428. [64] Das R & Baker D. Macromolecular modeling with Rosetta.
[44] Lu C, Mi LZ, Grey MJ, et al. Structural evidence for loose Annu. Rev. Biochem. 77 (2008) 363–382.
linkage between ligand binding and kinase activation [65] Skolnick J. In quest of an empirical potential for protein
in the epidermal growth factor receptor. Mol. Cell. Biol. structure prediction. Curr. Opin. Struct. Biol. 16 (2006)
30 (2010) 5432–5443. 166–171.
[45] Holmes MA. Computational design of epitope-scaffolds [66] Skolnick J & Brylinski M. FINDSITE: a combined evolution/
allows induction of antibodies specific for a poorly structure-based approach to protein function prediction.
immunogenic HIV vaccine epitope. Structure 18 (2010) Brief Bioinform. 10 (2009) 378–391.
1116–1126. [67] Lee EH, Hsin J, Sotomayor M, et al. Discovery through
[46] Nemazee D. Receptor editing in lymphocyte the computational microscope. Structure 17 (2009)
development and central tolerance. Nat. Rev. Immunol. 1295–1306.
6 (2006) 728–740. [68] The Gene Ontology Project. http://www.geneontology.
[47] Brooks CL, Schietinger A, Borisova SN, et al. Antibody org/.
recognition of a unique tumor-specific glycopeptide [69] Phillips GJ. Green fluorescent protein—a bright idea for the
antigen. Proc. Natl Acad. Sci. USA 107 (2010) 10056–10061. study of bacterial protein localization. FEMS Microbiol. Lett.
[48] Weiss A, Shields R, Newton M, et al. Ligand–receptor 204 (2001) 9–18.
interactions required for commitment to the activation of [70] Stepanenko OV, Verkhusha VV, Kuznetsova IM, et al.
the interleukin 2 gene. J. Immunol. 138 (1987) 2169–2176. Fluorescent proteins as biomarkers and biosensors:
[49] Davis MM, Krogsgaard M, Huse M, et al. T cells as a self- throwing color lights on molecular and cellular processes.
referential, sensory organ. Annu. Rev. Immunol. 25 (2007) Curr. Protein Pept. Sci. 9 (2008) 338–369.
681–695. [71] Singleton KL, Roybal KT, Sun Y, et al. Spatiotemporal
[50] Bell JK, Botos I, Hall PR, et al. The molecular structure of the patterning during T cell activation is highly diverse. Science
Toll-like receptor 3 ligand-binding domain. Proc. Natl Acad. Signaling 2 (2009) ra15.
Sci. USA 102 (2005) 10976–10980. [72] PSLID—Protein Subcellular Location Image Database.
[51] Moreland JL, Gramada A, Buzko OV, et al. The molecular http://pslid.org/start.html.
biology toolkit (MBT): a modular platform for developing [73] Sinha TK, Khatib-Shahidi S, Yankeelov TE, et al. Integrating
molecular visualization applications. BMC Bioinformatics spatially resolved three-dimensional MALDI IMS with in
6 (2005) 21. vivo magnetic resonance imaging. Nat. Methods 5 (2008)
[52] Oda K & Kitano H. A comprehensive map of the Toll- 57–59.
like receptor signaling network. Mol. Syst. Biol. 2 (2006) [74] Ulrich P & Cerami A. Protein glycation, diabetes, and aging.
2006.0015. Recent Prog. Horm. Res. 56 (2001) 1–21.
[53] Carafoli F, Bihan D, Stathopoulos S, et al. Crystallographic [75] Nibbe RK, Chowdhury SA, Koyuturk M, et al. Protein–
insight into collagen recognition by discoidin domain protein interaction networks and subnetworks in the
receptor 2. Structure 17 (2009) 1573–1581. biology of disease. Wiley Interdiscip. Rev. Syst. Biol. Med.
[54] Bloemendal H. Lens proteins. CRC Crit. Rev. Biochem. 3 (2011) 357–367.
12 (1982) 1–38. [76] Chautard E, Thierry-Mieg N & Ricard-Blum S. Interaction
[55] Palanivelu DV, Kozono DE, Engel A, et al. Co-axial association networks: from protein functions to drug discovery.
of recombinant eye lens aquaporin-0 observed in loosely A review. Pathol. Biol. (Paris) 57 (2009) 324–333.
packed 3D crystals. J. Mol. Biol. 355 (2006) 605–611. [77] Fields S. Interactive learning: lessons from two hybrids over
[56] Rabilloud T. Variations on a theme: changes to two decades. Proteomics 9 (2009) 5209–5213.
electrophoretic separations that can make a difference. [78] Monti M, Cozzolino M, Cozzolino F, et al. Puzzle of protein
J. Proteomics 73 (2010) 1562–1572. complexes in vivo: a present and future challenge for
[57] Waller LN, Shores K & Knapp DR. Shotgun proteomic functional proteomics. Expert Rev. Proteomics 6 (2009)
analysis of cerebrospinal fluid using off-gel electrophoresis 159–169.
as the first-dimension separation. J. Proteome Res. 7 (2008) [79] Tong AH, Drees B, Nardelli G, et al. A combined
4577–4584. experimental and computational strategy to define protein
[58] Minden JS, Dowd SR, Meyer HE & Stühler K. Difference interaction networks for peptide recognition modules.
gel electrophoresis. Electrophoresis 30(Suppl 1) (2009) Science 295 (2002) 321–324.
S156–S161. [80] Zwolak J, Adjerid N, Bagci EZ, et al. A quantitative model of
[59] Latterich M, Abramovitz M & Leyland-Jones B. Proteomics: the effect of unreplicated DNA on cell cycle progression in
new technologies and clinical applications. Eur. J. Cancer frog egg extracts. J. Theor. Biol. 260 (2009) 110–120.
44 (2008) 2737–2741. [81] Akama K, Horikoshi T, Nakayama T, et al. Proteomic
[60] Yates JR, Ruse CI & Nakorchevsky A. Proteomics by mass identification of differentially expressed genes in neural
spectrometry: approaches, advances, and applications. stem cells and neurons differentiated from embryonic
Annu. Rev. Biomed. Eng. 11 (2009) 49–79. stem cells of cynomolgus monkey (Macaca fascicularis) in
[61] ExPASy. http://www.expasy.ch/. vitro. Biochim. Biophys. Acta 1814 (2011) 265–276.
FURTHER READING
Chautard E, Thierry-Mieg N & Ricard-Blum S. Interaction networks: Shenoy SR & Jayaram B. Proteins: sequence to structure and
from protein functions to drug discovery. A review. Pathol. function—current status. Curr. Protein Pept. Sci. 11 (2010)
Biol. (Paris) 57 (2009) 324–333. 498–514.
Latterich M, Abramovitz M & Leyland-Jones B. Proteomics: new Twyman R. Principles of Proteomics, 2nd ed. Garland Science,
technologies and clinical applications. Eur. J. Cancer 2014.
44 (2008) 2737–2741. Yates JR, Ruse CI & Nakorchevsky A. Proteomics by mass
Nibbe RK, Chowdhury SA, Koyuturk M, et al. Protein–protein spectrometry: approaches, advances, and applications.
interaction networks and subnetworks in the biology of Annu. Rev. Biomed. Eng. 11 (2009) 49–79.
disease. Wiley Interdiscip. Rev. Syst. Biol. Med. 3 (2011) Zvelebil M & Baum JO. Understanding Bioinformatics. Garland
357–367. Science, 2008.
Metabolic Systems
8
• Identify and characterize the components of metabolic systems
• Describe conceptually how the components of metabolic systems interact
• Understand the basics of mass action and enzyme kinetics
• Be aware of the various data sources supporting metabolic analyses
• Associate different types of metabolic data with different modeling tasks
• Explain different purposes of metabolic analyses
Our genome is sometimes called the blueprint of who we are. We have seen in previ-
ous chapters that this notion is somewhat simplified, because the macro- and
micro-environments have significant roles throughout the biological hierarchy,
including the expression of genes. Nonetheless, our genes do contain the code that
ultimately determines the machinery of life. The lion’s share of this machinery is
present in the form of proteins, and, like most machines, they take input and con-
vert it into output. In the case of living systems, the vast majority of both inputs and
outputs consist of metabolites. Most prominently, proteins convert metabolites,
including food or substrate, into energy and into a slew of other metabolites ranging
from sugars and amino acids to vitamins and neurotransmitters. Signal transduc-
tion is sometimes seen as distinct from metabolic action, but it is also often accom-
plished through metabolic changes, such as the phosphorylation or glycosylation of
proteins or the synthesis or degradation of signaling lipids. Frequently, it is at the
level of metabolites that a normal physiological process becomes pathological and
health turns into disease. A gene may be mutated but, on its own, this change does
not cause problems. However, if the mutation leads to an abnormal protein or regu-
latory mechanism and thereby affects the dynamics of a metabolic pathway, it may
ultimately cause an excess or dearth of specific metabolites, with consequences that
in some cases might be tolerable, but in other cases can compromise the affected
cell and possibly even kill it.
DNA, proteins, and metabolites are composed of the same “organic” atoms
(mainly C, O, N, and H), and there is really no essential biochemical difference
between them, except that DNA and proteins are large molecules with particular
chemical structures, while metabolites in the strict sense are much smaller by com-
parison. For instance, the paradigm metabolite glucose (C6H12O6) has a molecular
mass of (6 × 12 + 12 × 1 + 6 × 16) = 180, whereas hexokinase, the enzyme that con-
verts glucose into glucose 6-phosphate, has a molecular mass of roughly 100,000
daltons (Da) [1] (Figure 8.1). The average molecular mass of a nucleotide in DNA is
about 330 Da, which for the average-sized human gene of roughly 3000 bases cor-
responds to a molecular mass of about 1,000,000 Da. The largest known human
232 Chapter 8: Metabolic Systems
Figure 8.1 Typical metabolites are small

in comparison with proteins and DNA.
The enzyme hexokinase (blue) binds a
molecule of glucose (red). Note the huge
difference in sizes of the two molecules.
(Courtesy of Juan Cui and Ying Xu, University
of Georgia.)
gene codes for dystrophin, a key protein in muscle, which, if altered, is one of the
root causes of muscular dystrophy. This gene consists of 2.4 million bases and thus
has a molecular mass of almost 800 million daltons [2]. Indeed, a typical metabolite
is tiny in comparison with proteins and nucleic acids.
The conversion of metabolites into other metabolites has been the bread and
butter of traditional biochemistry. This conversion typically occurs in several steps
that form metabolic pathways. Many biochemistry books are organized according
to these pathways, containing chapters on central metabolism, energy metabolism,
amino acid metabolism, fatty acid metabolism, and so on, sometimes including less
prevalent pathways such as the synthesis of poisons in insects and spiders or path-
ways leading to secondary metabolites, such as color pigments, which may not be
absolutely required for survival. In reality, metabolism is not neatly arranged in
pathways, but pathways are merely the most obvious contributors to complex regu-
lated pathway systems, just as interstate highways are the most prominent features
of a much larger and highly connected network of roads, streets, and alleyways.
Traditionally, biochemistry has mostly focused on the individual steps a cell
employs to convert different metabolites into each other. Of particular importance
in these discussions are enzymes, which we discussed in Chapter 7. Each enzyme
facilitates one or maybe a few of the specific steps in a string of metabolic conver-
sions. In recent years, much of the emphasis on single reaction steps has shifted
toward high-throughput methodologies of metabolomics that try to characterize
the entire metabolic state of a living system (or subsystem) at a given point in time.
A good example is a mass spectrogram, which is capable of quantitatively assessing
the relative amounts of thousands of metabolites in a biological sample (see later).
This chapter is very clearly not an attempt to review biochemistry, and it will
mention specific pathways only in the context of illustrative examples. Rather, it
tries to describe some of the basic and general concepts of biochemistry and metab-
olomics that are particularly pertinent for analyses in computational systems biol-
ogy. It begins with the characterization of enzyme-catalyzed reactions, discusses
means of controlling the rates of these reactions, establishes the foundation for
mathematical methods that capture the dynamics of metabolic systems, and lists
some of the methods of data generation and the comprehensive resources of infor-
mation regarding metabolites that are presently available.
BIOCHEMICAL REACTIONS
8.1 Background
Metabolism is the sequential conversion of organic compounds into other com-
pounds. As chemical reactions, these conversions must of course obey the laws of
chemistry and physics. Intriguingly, thermodynamics tells us that for the unaided
conversion of one chemical compound into another to take place, the latter must be
in a lower-energy state than the former. Are biochemistry and metabolism at odds
with one of the fundamental principles of physics by converting food into metabolites
BIOCHEMICAL REACTIONS 233
of high energy content, such as amino acids or lipids? The situation is particularly puz-
zling in plants, where the “food” consists merely of carbon dioxide, CO2, which has
very low energy. How is the production of high-energy compounds possible?
The answer, which of course is fully compatible with the laws of thermodynam-
ics, consists of three components:
• Plants are able to utilize sunlight as energy for converting CO2 into metabolites
of higher energy content.
• In all organisms, the metabolic production of a desired high-energy molecule is
“paid for” by the conversion of another highly energetic molecule into a lower-
energy molecule. For instance, the conversion of glucose into glucose 6-phosphate
is coupled with the conversion of the high-energy molecule adenosine triphos-
phate (ATP) into the lower-energy molecule adenosine diphosphate (ADP).
• Highly energetic molecules in cells do not simply degrade into lower-energy
molecules, as thermodynamics might suggest. Instead, they can only move to
the lower-energy state if they transit an intermediate state of even higher energy.
This transition is therefore not automatic, but must be facilitated, which is
accomplished with enzymes (see Chapter 7 and Box 8.1). Indeed, the vast
majority of metabolic reactions require an enzyme, and they very often involve a
secondary substrate or co-factor that provides the energy for the creation of a
higher-energy molecule from a lower-energy molecule. The involvement of an
enzyme speeds up the reaction rate enormously and furthermore affords the
possibility of regulation, in which one or more metabolites influence the degree
of enzyme activity and thereby the rate of the metabolic reaction.
For the purposes of metabolic systems analysis, we often do not need to know
much more about the molecular mechanisms associated with an enzyme than that
the enzyme enables the transition from substrate to product and that its activity
can be regulated with various mechanisms that act on different timescales
BOX 8.1 THE FUNCTION OF AN ENZYME DEMONSTRATED WITH AN ANALOGY
The function of an enzyme is illustrated in Figure 1A, where the constrained: although the popcorn is in a pan high above the
original metabolite (typically called the substrate) with a high- floor, it cannot fall down to the lower-energy level of the floor.
energy state (red) must be briefly elevated to a transition state of Now we turn on the heat (activation energy), which releases the
even higher energy (blue), before it moves to the product state stored energy and allows each kernel to jump during its pop (the
with lower energy. The enzyme temporarily provides the necessary transition state), and it is indeed possible that some kernels will
activation energy EA. An intuitive analogy is popcorn (Figure 1B). jump over the rim of the pan onto the floor.
In its original state, consisting of cold corn kernels, its energy is
(A) (B)
EA
ES
energy content
EP
time
Figure 1 Processes overcoming thresholds of high energy. (A) The conversion of a metabolic substrate with energy ES into a product with lower
energy EP requires activation energy EA that exceeds a threshold (blue). (B) The situation in (A) is similar to popping corn, where individual kernels
may receive enough (heat) energy to jump out of the pan and onto the floor.
(see Chapter 7 and later in this chapter). What we do need to explore in detail is how
enzymatic processes are translated into mathematical models.
8.2 Mathematical Formulation of Elementary Reactions

To ease our way into this translation task, we begin with the simpler situation in
which a chemical compound degrades over time without the involvement of an
enzyme or co-factor. This situation is actually not very interesting biochemically,
but a good start for model development. It has two prominent applications. One is
radioactive decay, in which radionuclides spontaneously disintegrate and the col-
lective disintegration process is proportional to the presently existing amount. The
second is a diffusion process, in which the transported amount of material is pro-
portional to the current amount.
The proportionality in both cases gives us a direct hint for how to set up a describ-
ing equation: namely, the change at time t is proportional to the amount at time t.
Recalling Chapters 2 and 4, change is expressed mathematically as the derivative
with respect to time, the amount is a function of time, and proportional implies a
linear function. Putting the pieces together into an equation yields
dX
= X = −k ⋅ X . (8.1)
dt
The right-hand side of this ordinary differential equation (ODE) contains three
items that are reminiscent of the SIR model in Chapter 2. X is the amount (for exam-
ple, of the radionuclide), and one often refers to X generically as a pool (of mole-
cules). The second component is the rate constant k, which is always positive (or
zero) and constant over time. It quantifies how many units of X are changing per
time unit. Finally, the minus sign indicates that the change ( X ) is in the negative
direction. In other words, material disappears rather than accumulates. The formu-
lation in (8.1) as a description of a chemical process is actually based on consider-
ations of statistical mechanics and thermodynamics and was proposed more than
one hundred years ago by Svante Arrhenius (1859–1927).
As we discussed in Chapter 4, (8.1) is a linear differential equation that describes
exponential behavior. Suppose now that X is converted into Y and that the disap-
pearance of X is captured well by (8.1). Because all material leaving pool X moves
into pool Y, it is easy to see that the change in Y must be equal to the change in X,
except that the two have opposite signs. The dynamics of the two variables is there-
fore easily described as
X = −kX ,
(8.2)
Y = kX .
We are allowed to add the two differential equations, which yields
X + Y = 0 (8.3)
and has the following interpretation. The total change in the system, consisting of
the change in X plus the change in Y, equals zero. There is no overall change. This
makes sense, because material is just flowing from one pool to the other, while no
material is added or lost.
Many metabolic reactions involve two substrates and are therefore called
bimolecular. Their mathematical description is constructed in analogy to the one-
substrate case and leads to a differential equation where the right-hand side
involves the product of the two substrates. Specifically, suppose X1 and X2 are the
substrates of a bimolecular reaction that generates product X3. Then the increase
in the concentration of X3 is given as
X 3 = k3 X 1 X 2. (8.4)
Note that the substrates enter the equation as a product and not a sum, even though
one might speak of adding a second substrate or formulate the reaction as
X1 + X2 → X3. The reason for this formulation can be traced back to thermodynamics
and to the fact that the two molecules have to come into physical contact within the
physical space where the reaction happens, which is a matter of probability and
leads to the product.
Because X3 is the recipient of material and does not affect its own synthesis, the
right-hand side is positive and does not depend on X3, and its concentration contin-
ues to increase as long as X1 and X2 are available. For every molecule of X3 that is
produced, one molecule of X1 and one molecule of X2 are used up. Therefore, the
loss in either one substrate is
X 1 = X 2 = − X 3 = −k3 X 1 X 2. (8.5)
It is also possible that X3 is produced from two molecules of type X1 rather than from
X1 and X2. In this case, the describing equations are
X 1 = −2k3 X 12 ,
(8.6)
X 3 = k3 X 12 .
The product of X1 and X2 in (8.5) becomes X 12 in both equations of (8.6), and one says
that the process is of second order with respect to X1. The first equation in (8.6) fur-
thermore contains the stoichiometric factor 2, which does not enter the second
equation. The reason is that X1 is used up twice as fast as X3 is produced, because two
molecules of X1 are needed to generate one molecule of X3.
The mathematical formulations discussed here are the foundation of mass
action kinetics, which was introduced about 150 years ago [3–5]. According to this
widely used analytical framework, all substrates enter a reaction term as factors in a
product, where powers reflect the number of contributing molecules of each type.
The term also contains a rate constant, which is positive (or possibly zero) and does
not change over time.
Mass action formulations implicitly assume that many substrate molecules are
available and that they are freely moving about in a homogeneous medium. In many
cases, and especially in living cells, these assumptions are not really true, but they
do provide good approximations to realistic metabolic systems. They are frequently
used because more accurate formulations are incomparably more complicated. For
instance, one could more realistically consider the reaction between X1 and X2 as a
random encounter (stochastic) process. While intuitively plausible, a model of such
a process requires heavy-duty mathematics for further analysis [6, 7].
8.3 Rate Laws

Most biochemical reactions in a metabolic system are catalyzed by enzymes. About
a century ago, the biochemists Henri, Michaelis, and Menten proposed a mecha-
nism and a mathematical formula describing this process [8, 9]. They postulated
that the substrate S and the catalyzing enzyme E reversibly form an intermediary
complex (ES), which subsequently breaks apart and irreversibly yields the reaction
product P while releasing the enzyme unchanged and ready for recycling. The
diagram of the mechanism with typical rate constants for all steps is shown in
Figure 8.2. Modern methods of transient and single-molecule kinetics have shown
that this simple scheme is somewhat simplistic and that many reactions in fact
Figure 8.2 Conceptual model of an

enzyme-catalyzed reaction mechanism,
k1 k2
as proposed by Michaelis and Menten.
E + S E S E + P
A substrate S and an enzyme E reversibly
k–1
form a complex (ES), which may either
return S and E or lead to the product P,
thereby releasing the free enzyme, which
may be used again.
consist of entire networks of fast sub-reactions with different intermediary com-

plexes that form a multidimensional free-energy surface [10]. Nonetheless, the
Michaelis–Menten mechanism is a very useful conceptual framework, and an
enormous number of studies have measured its characteristic parameters.
A mathematical model for the mechanism is obtained straightforwardly by
setting up mass action functions (see Chapter 2). Specifically, each process term is
formulated by including all variables that are directly involved, as well as the appro-
priate rate constant. The result is
S = −k1 ⋅ S ⋅ E + k−1 ⋅ (ES ), (8.7)

i
(ES ) = k1 ⋅ S ⋅ E − (k−1 + k2 ) ⋅ (ES ), (8.8)
P = k2 ⋅ (ES ). (8.9)
These equations are easy to interpret. For instance, (8.7) says that the change in sub-
strate is governed by two processes. In the first, –k1SE, the substrate and enzyme
enter a bimolecular reaction with rate k1. The process carries a minus sign, because
substrate is lost. The second process, k−1(ES), describes that some of the complex
(ES) reverts to S and E, and this happens with rate k−1. The sign here is positive,
because the process augments the pool S.
An equation for E is not formulated, because the sum of free enzyme E and
enzyme bound in the complex (ES) is assumed to remain constant; this total enzyme
concentration is denoted Etot. Thus, once we know the dynamics of (ES), we can
infer the dynamics of E. It is usually assumed that the reversible formation of the
complex (ES) occurs much faster than the production of P, and that therefore k1 ≫ k2
and k−1 ≫ k2. The parameter k2 is sometimes also called the catalytic constant and
denoted by kcat. The production of P is assumed to be irreversible.
The differential equations in (8.7)–(8.9) do not have an explicit analytical solu-
tion but can of course be solved computationally with a numerical integrator, which
is available in many software packages. Since Henri, Michaelis, and Menten did not
have computers, they proposed reasonable assumptions that yielded a significant
simplification from the ODE system to an algebraic formulation. The assumptions
are that substrate is present in excess, relative to the enzyme (S ≫ E), and that the
enzyme–substrate complex (ES) is in equilibrium with the free enzyme and sub-
strate. As a more realistic variation of the latter, Briggs and Haldane [11] proposed
the quasi-steady-state assumption (QSSA), which entails that (ES) does not appre-
ciably change in concentration over time and that the total enzyme concentration
Etot is constant. These assumptions were made for in vitro experiments, where
indeed the solution can be stirred well and substrate easily exceeds the substrate
concentration many times.
With regard to metabolism in living cells, the QSSA is often wrong at the very
beginning of an experiment, when substrate and enzyme encounter each other for
the first time, and it becomes inaccurate at the end, when almost all substrate is
used up (see Figure 8.3; however, also see Exercise 8.7). In addition, the implicit
assumption of a well-stirred homogeneous reaction medium, which is a prerequi-
site for mass action processes, is of course not satisfied in living cells. In spite of
these issues, the QSSA is sufficiently close to being satisfied in many realistic situa-
tions, and a lot of experience has demonstrated that its benefits often outweigh its
problems (for further discussion, see [12, 13]).
If one accepts QSSA, life becomes much simpler. The left-hand side of (8.8) is
now equal to zero, because (ES) is assumed not to change. Furthermore, one defines
the rate of product formation as v P = P and introduces the maximum velocity of the
reaction as Vmax = k2 Etot and the Michaelis constant as KM = (k−1 + k2)/k1. Using
straightforward algebra, one can show that the rate of product formation can then
be expressed as
Vmax S
vP = (8.10)
KM + S
(See Exercise 8.5). The function in (8.10) is referred to as the Michaelis–Menten rate
law (MMRL) or the Henri–Michaelis–Menten rate law (HMMRL). A plot of the
MMRL shows a monotonically increasing function that approaches Vmax (Figure
8.3), which means that Vmax quantifies the fastest possible speed with which the
reaction can proceed. KM corresponds to the substrate concentration for which the
reaction speed is exactly half the maximum velocity. Its numerical value can vary
widely among different enzymes. In many cases, it is close to the natural substrate
concentration in vivo. Recall from Chapter 5 how the parameter values of the MMRL
can be obtained from in vitro measurements with simple linear regression.
Instead of using algebra, the MMRL can also be obtained from the following bio-
chemical arguments. The reversible reaction between substrate, enzyme, and inter-
mediate complex on the left-hand side of Figure 8.2 is governed by the mass action
formulation
k−1 E ⋅S
= kd = , (8.11)
k1 (ES )
where kd is called the equilibrium or dissociation constant. Because k2 is tradition-

ally assumed to be small in comparison to k1 and k−1, kd is approximately equal to
(k−1 + k2)/k1 and thus to KM.
The occupancy of an enzyme reflects how much of the enzyme is bound to the
substrate. In other words, it is the ratio (ES)/Etot. We can use (8.11) to rearrange this
ratio as
(ES ) (ES ) E ⋅ S / kd S / kd S S
= = = = ≈ . (8.12)
Etot E + (ES ) E + E ⋅ S / kd 1 + S / kd kd + S K M + S
Using this result, we immediately obtain the MMRL as
k ⋅ E ⋅S V ⋅S
v P = P = k2 ⋅ ( ES ) ≈ 2 tot = max . (8.13)
KM + S KM + S
Note that the solution of an ODE is typically presented as a plot of the dependent
variables as functions of time t. In (8.7)–(8.9), these are S, (ES), and P. By contrast,
(8.10), which is the form almost always found in biochemistry books and articles,
(A) (B)
10 2 Vmax
vP
P
S
reaction speed
concentration
5 1
(ES) (ES)
0 0
0 5 10 0 KM 5 10
time concentration of S
Figure 8.3 Plots of the Michaelis–Menten rate law. (A) Solution of the system of equations (8.7)–(8.9) in the form of concentrations versus time.
(B) The more familiar plot of reaction speed versus substrate concentration, (8.10). Parameters are k1 = 4, k–1 = 0.5, k2 = 2, KM = 0.625, Etot = 1, Vmax = 2,
S(0) = 9, and (ES)(0) = P(0) = 0.
formulates the speed of product formation vP as a function of the substrate concen-

tration S. The two plots are distinctly different (see Figure 8.3).
Thousands of investigations in biochemistry have used this formulation, and the
MMRL is even used as a “black-box” model in other contexts that have nothing to do
with biochemical kinetics, just because its shape fits many observed data. Of inter-
est is that KM is considered a property of an enzyme–substrate pair, so that KM values
from the literature or a database may be useful for an analysis even if they were
measured by different groups and under slightly different conditions. By contrast,
experimental measurements of Vmax can vary substantially, depending on the
experimental conditions and can seldom be assumed to have the same value under
different conditions.
A biochemical and mathematical extension of the MMRL is the Hill rate func-
tion [14], which describes biochemical processes involving enzymes with several
subunits. The mathematical consequence is that the substrate appears in the rate law
with a Hill coefficient, that is, with an integer power that reflects the number of sub-
units, which is typically between 2 and 4. With Hill coefficient n, the rate law reads
Vmax S n
v P ,Hill = . (8.14)
n
KM + Sn
In contrast to the MMRL, the Hill rate law with n > 1 is sigmoidal (Figure 8.4). For
n = 1, the Hill rate law is the same as the MMRL, and the S-shape becomes a simple
shoulder curve. The derivation of this rate law follows exactly the same steps as in
the case of the MMRL, except that one accounts for n substrate molecules and n
product molecules. For n = 2, (8.7) becomes
SH = −2k1SH2 E H + 2k−1 (ES )H , (8.15)
where the index H simply distinguishes the equation from (8.7) [15]. The algebraic
form (8.14) of the rate law can be derived again from the corresponding differential
equations, with arguments from either algebra or biochemistry, as we discussed for
the MMRL.
As described so far, the speed of an enzyme-catalyzed reaction depends exclu-
sively on the substrate concentration, the parameters KM and Vmax, and possibly n.
In reality, many reactions are controllable by inhibitors and activators that sense the
demand for a product and adjust the turnover rate of the reaction. Modulation of the
(A) (B)
10 Vmax 10 Vmax
vP1 vP4
vP3
reaction speed
reaction speed
5 vP2 5
0 0
0 10 20 0 10 20
concentration of S concentration of S
Figure 8.4 Plots of Michaelis-Menten (MMRL) and Hill rate laws (HRL). (A) Two MMRLs, vP1 and vP2, with the same Vmax = 10 and KM1 = 0.5 and KM2 = 2.5,
respectively. (B) Two HRLs, vP3 and vP4, with the same Vmax = 10 and KM = 2.5 and with Hill coefficients n = 2 and n = 4, respectively. Note that vP2 in (A)
corresponds to an HRL with Vmax = 10, KM = 2.5, and n = 1.
reaction speed by an inhibitor or activator may occur at several locations in the

Michaelis–Menten mechanism shown in Figure 8.2. For instance, an inhibitor may
compete with the substrate molecules for an active site on the enzyme. This may
happen if the inhibitor is structurally similar to the substrate and therefore fits into
the same binding site on the enzyme. It may also happen that the inhibitor forms a
rather stable complex with the enzyme–substrate complex (ES), which partially
prevents the reaction toward the product from occurring. If the inhibitor and the
substrate are structurally different, inhibition of the enzyme activity can be accom-
plished if the inhibitor binds to a different site than the active site on the enzyme. In
the process of binding, the catalytic activity of the enzyme at the active site is
reduced. This type of allosteric inhibition is quite frequent. It is also possible that
the allosteric effect is activating rather than inhibiting.
Mathematical descriptions differ for the various modes of inhibition and may in
effect increase the KM or decrease the Vmax value, or lead to a Hill-type function as in
(8.14). Thousands of articles and many books have discussed enzyme regulation
and described biochemical techniques for characterizing the regulatory mecha-
nisms. Since these investigations are firmly in the realm of biochemistry, we will not
discuss them much further. Good representations of mathematical models of
enzyme modulation can be found in [16–18].
Three types of inhibition are called competitive, uncompetitive, and noncom-
petitive. The terminology of the latter two is somewhat confusing, but has become
the standard. The three types are ultimately associated with the KM, the substrate
concentration in the denominator of the MMRL, and Vmax, respectively. Specifically,
in competitive inhibition, the inhibitor with concentration I competes with the sub-
strate for the binding site on the enzyme. The corresponding rate law is
Vmax S
v P , CI = .
 I  (8.16)
K M 1 + +S
 K I 
The new parameter KI is the dissociation constant of the inhibitor.

In uncompetitive inhibition, the inhibitor binds to the complex (ES). The corre-
sponding rate law is
Vmax S
v P , UI = .
 I  (8.17)
K M + S 1 +
 K I 
In noncompetitive inhibition, the inhibitor reduces the binding affinity between

enzyme and substrate without directly affecting the substrate-binding site. The cor-
responding rate law is therefore
−1
Vmax S  I 
v P , NI =  1+ . (8.18)
KM + S  K I 
A different type of inhibition is called allosteric regulation. In this case, the inhibi-
tor binds to the enzyme, but not at the substrate-binding site. Nonetheless, the dock-
ing of the inhibitor alters the substrate-binding site and thereby reduces the affinity
between enzyme and inhibitor. The general mathematical formulation is much more
complicated and is not based on the MMRL; a detailed derivation can be found in [18].
In reality, the inhibition may be a mixture of several mechanisms, which
increases the complexity of a mathematical representation. Circumventing this
issue, biochemical systems theory (BST) treats an inhibitor like any other variable
and represents it with a power-law approximation. Thus, if the degradation reaction
of X1 is inhibited by Xj, one sets up the model for X1 as
X 1 = production − β1 X 1h11 X j 1 j ,
h
(8.19)
and theory says that h1j is a negative quantity. If an inhibitor I is small in quantity,
it is sometimes replaced with (1 + I), and (8.19) then reads [19]
h1 I
 I 
X 1 = production − β1 X 1h11  1 + . (8.20)
 K I 
Thus, if the inhibitor concentration is reduced to 0, the MMRL is regained. Note that
this representation is similar to the noncompetitive rate law, although it is derived
purely through linearization in logarithmic space (Chapter 4).
Independent of the details of inhibition, data directly related to a traditional
enzyme kinetic investigation are typically presented as quantitative features of
enzymes that are characterized by parameters such as KM, inhibition and dissocia-
tion constants KI and KD, and sometimes Vmax values. Often, co-factors and modula-
tors of reactions are listed as well. Originally, these features were described in a
huge body of literature, but nowadays databases such as BRENDA [20] (see Box 8.2
below) and ENZYME [21] offer collections of many properties of enzymes and reac-
tions, along with pertinent literature references.
While few functions in biology have received as much attention as the MMRL
and Hill functions, one must caution that the two are approximations that are often
not really justified (see, for example, [12, 13, 16]). First and foremost, these func-
tions, just like mass action kinetics, assume reactions occurring in well-stirred,
homogeneous media, which obviously is not the case inside a living cell that is full
of materials and organelles and where reactions often happen at surfaces. Second,
the quasi-steady-state assumption is seldom truly valid, and neither is the amount
of enzyme always constant and much smaller than the substrate concentration, as
the MMRL assumes it to be. Finally, the kinetic functions fail for small numbers of
reactants, and ODEs or continuous functions have to be replaced with much more
complicated stochastic processes [6].
PATHWAYS AND PATHWAY SYSTEMS

8.4 Biochemistry and Metabolomics
If one had to describe the difference between traditional enzyme kinetics and
metabolomics, one might simplistically say that the former focuses primarily on
individual reactions and pathways, along with factors influencing them, while the
latter is primarily interested in (large) assemblies of reactions that collectively form
pathway systems. Expressed differently, the main focus of traditional biochemistry
and enzyme kinetics can be seen as targeted and local, while the focus of metabolo-
mics is comprehensive and global. This simplified distinction may suggest that
biochemistry is a subset of metabolomics, but this is not entirely so, because metab-
olomics looks at so many reactions simultaneously that details at the local level are
often only of secondary interest. As a case in point, an important component of
enzyme kinetic investigations is a detailed protocol for the purification of an enzyme
and the characterization of its affinity to different substrates. By contrast, typical
metabolomic studies attempt to determine two features. The first is the metabolic
profile of a biological sample, which consists of the relative (or absolute) amounts
or concentrations of all metabolites. The second feature is the characterization and
distribution of the relative (or absolute) magnitudes of all fluxes throughout the sys-
tem. Here, each flux is defined as the rate of turnover of molecules in a reaction step
or pathway or, expressed differently, as the amount of material flowing through a
metabolite pool or pathway at a given time. Some authors reserve the term flux for
pathways, while calling essentially the same quantity a rate when a single reaction is
being studied [22].
Studying hundreds of metabolites or fluxes at the same time usually mandates
that details of individual reaction steps or mechanisms of inhibition or activation be
ignored. This situation is comparable to the different foci in zoology and ecology,
where the former primarily targets the unique features of specific animals or animal
species, while the latter tries to understand the interactions among (many) species,
including animals, plants, and sometimes even bacteria and viruses.
PATHWAYS AND PATHWAY SYSTEMS 241
BOX 8.2 METABOLIC PATHWAY INFORMATION DISPLAYED IN BRENDA
The main focus of BRENDA [20] is quantitative information values for important quantities such as KM values, substrates,
on enzymes, which can be searched by name or Enzyme co-factors, inhibitors, and other features that can be very
Commission (EC) number, as well as by species. Figure 1 is a important for modeling purposes. All information is backed up
screenshot showing the results of a typical search. BRENDA offers by references.
Figure 1 Typical information provided

in BRENDA. Specifying the EC number
or name of trehalase offers a lot of
quantitative information that can be
very useful for pathway modeling
(see Chapter 11). The screenshot here
shows only a selection of results related
to KM values obtained under various
conditions. BRENDA also has direct
links to KEGG (see Box 8.3), BioCyc (see
Box 8.4), and other databases. (Courtesy
of Dietmar Schomburg, Technische
Universität Braunschweig, Germany.)
Of course, there is plenty of overlap between enzyme kinetics and metabolo-

mics. For instance, the flux through a pathway system is related to the enzyme activ-
ities and Vmax values of the individual reaction steps. One could surmise that the
analysis of the dynamics of a metabolic system might not require very detailed
information regarding each and every reaction step, as long as we know the turn-
over rate of the reaction and its modulators. However, once we start extrapolating
results from the analysis of one set of conditions to another, these details can become
crucially important. Thus, traditional biochemistry and metabolomics have their
own foci but also much common ground.
While a little naive, the above distinction between traditional enzyme kinetics
and metabolomics is useful for now, because it provides rationales and explana-
tions for the different classes of analytical methods in the two fields, and for the
types of data that are associated with them. Using painstaking analysis, traditional
biochemistry has yielded a solid understanding of most reactions in all major and
many minor metabolic pathways, and huge amounts of data related to enzyme
kinetics, such as KM and Vmax values, have been measured and are available under
their Enzyme Commission (EC) number in databases such as BRENDA [20]
(Box 8.2). In stark contrast, typical metabolomic data consist of comprehensive
metabolic concentration or flux profiles, either at a steady state or dynamically
changing in response to a stimulus.
8.5 Resources for Computational Pathway Analysis

The analysis of metabolic pathway systems typically follows one of two routes, which
are distinct and complement each other. The first is stoichiometric analysis.
It focuses on the connectivity within the pathway system and addresses which
metabolite can be converted directly into which other metabolite. The first step of
any stoichiometric analysis is therefore the collection of all known or alleged reac-
tions producing or degrading the metabolites in the pathway system of interest.
These reactions form a stoichiometric map consisting of nodes (metabolites) and
vertices (arrows: reactions). In mathematical terms, this map is a directed graph, for
which many methods of analysis are available. Chapter 3 discussed the most perti-
nent features of stoichiometric networks, along with analyses based on graphs and
linear algebra. That chapter also discussed possible transitions from these static
networks to dynamics networks, for instance, with methods of metabolic control
analysis (MCA) [22] or by taking into account that static networks can be consid-
ered steady states of fully dynamic systems, for example, within the framework of
biochemical systems theory (BST) (see, for example, [18, 23]).
As an example, consider glucose 6-phosphate, which is typically isomerized
into fructose 6-phosphate during glycolysis, but may also be converted into glucose
1-phosphate via the phosphoglucomutase reaction, reverted to glucose by glucose
6-phosphatase, or enter the pentose phosphate pathway through the glucose
6-phosphate dehydrogenase reaction. In a stoichiometric map, glucose 6-phos-
phate is represented as a node and the reactions are represented as arrows pointing
from glucose 6-phosphate to other nodes, representing the products of the
reactions.
Of enormous help for the construction of a stoichiometric map are freely acces-
sible databases such as KEGG (the Kyoto Encyclopedia of Genes and Genomes) [24]
and BioCyc [25] (Boxes 8.3 and 8.4), which contain collections of pathways, along
with information regarding their enzymes, genes, and plenty of useful comments,
references, and other pieces of information. The databases furthermore provide
numerous tools for exploring pathways in well-characterized organisms such as
Escherichia coli, yeast, and mouse, and also for inferring the composition of ill-char-
acterized pathways, for instance, in lesser-known microbial organisms.
In addition to KEGG and BioCyc, other tools are available for metabolic pathway
analysis. For instance, the commercial software package ERGO [26] offers a
bioinformatics suite with tools supporting comparative analyses of genomes and
metabolic pathways. Users of ERGO can combine various sources of information,
including sequence similarity, protein and gene context, and regulatory and expres-
sion data, for optimal functional predictions and for computational reconstructions
of large parts of metabolism. Pathway Studio [27] allows automated information
mining, the identification and interpretation of relationships among genes, pro-
teins, metabolites, and cellular processes, as well as the construction and analysis of
pathways. KInfer (Kinetics Inference) [28] is a tool for estimating rate constants of
systems of reactions from experimental time-series data on metabolite concentra-
tions. Another interesting database of biological pathways is Reactome [29, 30],
which contains not only metabolic pathways, but also information regarding DNA
replication, transcription, translation, the cell cycle, and signaling cascades.
The Lipid Maps project [31] provides information and databases specifically for
lipids and lipidomics, a term that was chosen to indicate that the totality of lipids
may be comparable to genomics and proteomics in scope and importance. Indeed,
the more we learn about lipids, the more we must marvel at the variety of functions
that they serve—from energy storage to membranes and to genuine mechanisms of
signal transduction. And not to forget: 60% of the human brain is fat.
Metabolic concentration and flux profiles are usually, but not always, measured
at a steady state. In this condition, material is flowing through the system, but all
metabolite pools and flux rates remain constant. At a steady state, all fluxes and
metabolite concentrations in the system are constant. Moreover, all fluxes entering
any given metabolite pool must in overall quantity be equal to all fluxes leaving this
pool. These are strong conditions that permit the inference of some fluxes from
knowledge of other fluxes. Indeed, if sufficiently many fluxes in the pathway system
could be measured, it would be possible to infer all other fluxes. In reality, the num-
ber of measured fluxes is almost always too small for a complete inference, but the
method of flux balance analysis (see [32] and Chapter 3) renders the inference pos-
sible, if certain assumptions are made.
The second type of representation of a pathway system also includes the stoi-
chiometric connectivity map, but in addition describes how the fluxes functionally
depend on metabolites, enzymes, and modulators (see Chapter 4). Mathematical
flux descriptions may consist of Michaelis–Menten or Hill rate laws, as discussed
earlier, but alternatives, such as power-law functions [23], linear–logarithmic func-
tions [33], and a variety of other formulations, are available and have their
BOX 8.3 METABOLIC PATHWAY REPRESENTATION IN KEGG
Chapter 11 will analyze a small metabolic pathway in yeast that Kyoto Encyclopedia of Genes and Genomes [24]. KEGG consists
is very important for the organism’s response to stresses such as of 16 main databases with information on genomes, pathways,
heat. The main compound of interest within this pathway is the functional aspects of biological systems, chemicals, drugs, and
sugar trehalose. Figures 1–3 are screenshots from KEGG—the enzymes.
Figure 1 Metabolic network associated

with trehalose dynamics. Trehalose is a
carbohydrate and can be found in KEGG’s
starch and sucrose pathway system. In
this representation, the metabolites are
the nodes. Associated with the reactions
(edges) are numbers in boxes, which are
the EC numbers that uniquely identify the
catalyzing enzymes. (Courtesy of Kanehisa
Laboratories.)
Figure 2 A more detailed view of the

synthetic and degradation pathways
of trehalose in KEGG. Clicking on
an item in a box leads to further
information (see Figure 3). (Courtesy of
Kanehisa Laboratories.)
Figure 3 Details provided by KEGG.

Clicking on an EC number in a pathway
representation as in Figure 2 leads
to biochemical and genome-related
features of the corresponding enzyme.
The information displayed here describes
the enzyme trehalase, which degrades
trehalose into glucose (see Chapter 11).
(Courtesy of Kanehisa Laboratories.)
BOX 8.4 METABOLIC PATHWAY REPRESENTATION IN BIOCYC
As mentioned in Box 8.3, Chapter 11 will require information on information. BioCyc also provides information on pertinent genes,
the sugar trehalose in yeast. Figures 1 and 2 are screenshots from a summary of pathway features, references, and comparisons with
BioCyc. Running the cursor over any of the entities shown on the other organisms. BioCyc furthermore offers sophisticated tools for
website generates names, synonyms, reactions, and other useful mining curated databases.
Figure 1 Screenshot from BioCyc.

Specifying the yeast Saccharomyces
cerevisiae as the organism of interest
and selecting trehalose biosynthesis as
the target pathway leads to a display
of the pathway. All pertinent items can
be clicked and guide the user to further
information. (Courtesy of Peter Karp, SRI
International.)
Figure 2 Details available in Biocyc.

The simplified pathway representation
in Figure 1 allows the user to
request more detail, which is shown
here. (Courtesy of Peter Karp, SRI
International.)
advantages and drawbacks. Because these dynamic–kinetic representations con-

tain much more information than simple stoichiometric representations, their con-
struction is more complicated and requires considerably more input data in the
form of kinetic information and/or metabolic time series.
8.6 Control of Pathway Systems

A key feature that is not addressed in stoichiometric models, but is present in
dynamic models, is the regulation and control of the flow of material throughout the
metabolic pathway system. Regulation allows the system to respond to changing

environments and changing demands where some metabolite is needed in greater
amounts than normal. For instance, by changing the activity of an enzyme at a
branch point where two pathways diverge, more material can be channeled into the
one or other pathway, depending on demand. We can easily see the importance of
controllability by studying what would happen without control. Imagine a water dis-
tribution system with pipes of fixed widths, where at each node the collective cross
section of all input pipes equals the collective cross section of all output pipes. If the
system is running at maximal capacity, the amount of water at each point is propor-
tional to the pipe size. If less water is entering the system, the amounts decrease
correspondingly, but without means of intervention it is impossible to steer the
water flow selectively into a direction where it is most needed for a particular
purpose.
The control of metabolic pathway systems occurs at several levels. The fastest is
a (temporary) change in enzyme activity, which is achieved at the metabolic level
itself. A premier example is feedback inhibition by the end product of a pathway: the
more product is generated the stronger becomes the inhibition of the enzyme at the
beginning of the pathway (Figure 8.5). Similarly, by inhibiting one pathway at a
branch point, the other branch receives more substrate. In many cases, the end
product not only inhibits its own branch, but also activates the competing branch
and/or inhibits a reaction step upstream of the branch. One possibility is a nested
feedback design where either end product inhibits its own branch and also the ini-
tial substrate production (Figure 8.6). These types of feedback patterns have been
analyzed in detail with regard to the efficiency of alternative designs (see, for exam-
ple, [18, 34]). The methods described in Chapters 2 and 4 allow us to set up models
for these types of pathways, and Chapter 5 provides recipes for finding appropriate
parameter values, if the right data are available.
The second manner of control happens at the proteomic level, where enzymes
can be activated or de-activated, just like other proteins—for instance, by covalent
modification in the form of phosphorylation, ubiquitination, or glycosylation, by
S–S bonding, or through protein–protein interactions. Of course, it is also possible to
degrade enzymes permanently by means of proteases.
The third option for controlling metabolism occurs at the level of gene expres-
sion, where transcription, translation, or both can be affected. At the level of tran-
scription, the gene coding for an enzyme may be up-regulated and lead to a larger
amount of enzyme and therefore to an increased turnover capacity. This mode of
control, which often relies on the mobilization or translocation of transcription fac-
tors, is very effective, but much slower than metabolic regulation such as end-
product inhibition, because it involves transcription and translation, which may
(A)
–
A B C D
(B)
1.0
D+
concentration
Figure 8.5 Linear pathway with feedback.

0.5 (A) Reaction scheme with feedback
inhibition of the initial step by the end
product. (B) Comparison of responses
to a sudden and persistent demand of
D–
metabolite D, starting at time t = 8. With
feedback (D+), the concentration of D
oscillates and converges to a level of about
0 two-thirds of the initial value. Without
0 25 50 feedback (D−), the concentration of D sinks to
time less than one-quarter.
C Figure 8.6 Branched pathway with several

feedback inhibition signals. Depending
on the strengths of the feedback signals,
material may be channeled preferentially
into either one of the two branches.
– –
(Adapted from Savageau MA. Biochemical
Systems Analysis: A Study of Function and
A B
Design in Molecular Biology. Addison-
– –
Wesley, 1976. With permission from John
Wiley & Sons.)
take on a timescale of the order of tens of minutes if not hours. It is also possible to
exert regulation at the level of translating RNA into proteins.
Finally, metabolic control can be implemented through the design of the path-
way and, in particular, with isozymes, that is, with distinct but similar enzymes cata-
lyzing the same reaction. For instance, in a situation as shown in Figure 8.6, it is not
uncommon to find two isozymes for the production of A, each of which is inhibited
by only one of the end products, C or D. In addition to feedback inhibition, these
isozymes may also be differently regulated at the genetic level. The two isozymes
may furthermore be physically separated by forming different supramolecular com-
plexes or through their location in different compartments.
In complicated situations such as stress responses, the regulation of metabolism
is the result of a combination of several modes of action, including direct metabolic
adjustments, the activation of transcriptional regulators, post-translational modifi-
cations, and various signaling events [35]. The immediate response may also occur
directly at the molecular level. For instance, it is known that temperature can change
the activity of enzymes that are involved in heat-stress response [36], and it has been
shown that this mechanism is sufficient for mounting an immediate cellular
response, whereas heat-induced changes in gene expression govern heat-stress
responses over longer time horizons (Chapter 11 and [37]).
METHODS OF METABOLOMIC DATA GENERATION

Throughout the history of biochemistry, measurement of concentrations and fluxes
has been as important as characterization of enzymes, modulators, and turnover
rates. The distinguishing feature of modern high-throughput metabolomics is the
simultaneous generation of large metabolite or flux profiles, which sometimes con-
sist of hundreds of data points obtained in a single experiment [38, 39]. Most of
these methods measure systems in a steady state, but some techniques also allow
characterization of changing profiles over a sequence of time points, following some
stimulus. It is worth noting in this context that there is a genuine difference in com-
plexity between metabolomics on one hand and genomics and proteomics on the
other. The genome consists of only four bases and the proteome of 20 amino acids,
and these are rather similar. By contrast, the metabolome consists of thousands of
metabolites, which are chemically very diverse and come in all sizes, chemical fami-
lies, and affinities to water or lipids.
Uncounted methods have been developed for measuring metabolites. Some of
these are quite generic, whereas others are specific for the metabolites in question.
Generically, metabolite analysis normally entails a sequence of processes [39, 40]:
1. Sample preparation and quenching of metabolic activity
2. An extraction procedure
3. Separation of metabolites
4. Quantification and/or profiling with methods such as nuclear magnetic reso-
nance (NMR) and mass spectrometry (MS)
5. Data analysis
METHODS OF METABOLOMIC DATA GENERATION 247
8.7 Sampling, Extraction, and Separation Methods

The key step in sample preparation is efficient rapid sampling [41]. This step is cru-
cial, because any changes induced by a stimulus or any intended or unintended
variation in the environment directly affect the metabolite concentrations in the
cell. For instance, intracellular metabolites in yeast typically have a turnover rate at
the order of one mole per second, and many metabolites respond to new conditions
within the millisecond range [39]. Thus, to yield valid snapshots of a metabolic sys-
tem, any rapid-sampling method must instantaneously quench all metabolic activ-
ity. This quenching is typically achieved through very fast changes in temperature,
for instance, by injecting the sample into liquid nitrogen, methanol, or a solution
containing ethanol and sodium chloride [42].
Once the sample has been quenched, intracellular metabolites must be sepa-
rated from metabolites of the external medium and extracted with a reagent such as
perchloric acid that disrupts the cell, releases the intracellular metabolites, and
denatures the enzymes, thereby preventing further degradation and biochemical
conversions. These initial steps of metabolite analysis are quite complex, and it is
virtually impossible to execute them without losing some of the metabolites [40]. As
a consequence, and because of the natural variability among cells, the reproduc-
ibility of metabolite extraction is often compromised. In some cases, it is possible to
obtain direct measurements in living cells, for instance, with fluorescence tech-
niques that allow visualization of specific metabolites such as glucose and NADH
in vivo, but it is not always clear how precise such direct measurements are. A prom-
ising alternative is sometimes in vivo NMR, which we will discuss in a moment.
Many methods have been developed for the quantitative measurement of intra-
cellular metabolites in extracts resulting from the sampling steps mentioned above.
The traditional option has been an enzymatic assay, which is specific and often very
precise. The disadvantages of this approach include relatively large sample volumes
and the need for separate assays for most metabolites [39]. A distinct alternative is
chromatography, which is one of the oldest and best-documented separation tech-
niques in biochemistry [40]. High-throughput metabolic studies often use gas
chromatography (GC), combined with mass spectrometry (GC-MS), or high-
performance liquid chromatography (HPLC). GC-MS has the advantage that it
offers high resolution—but only if the analyzed chemicals are volatile or can be
vaporized without decomposition, which is not always possible, in particular, if the
metabolites are large. In HPLC, a biological sample is pumped through a column,
separated, and measured with a detector. HPLC permits a wider range of com-
pounds to be measured than GC, but has a lower resolution.
Capillary electrophoresis (CE) encompasses a family of methods for separating
ionized molecules based on their charge and has become a valuable tool for metabolic
fingerprinting [43]. The separation takes place inside a small capillary connecting a
source vial with the biological sample in an aqueous buffer solution and a destination
vial. The sample enters the capillary owing to capillary action, and the analytes migrate
and separate when a strong electric field is applied. A detector close to the end of the
capillary measures the amount of each analyte and sends the data to a computer.
8.8 Detection Methods

Among the detection methods, mass spectrometry (MS) and nuclear magnetic
resonance (NMR) spectroscopy stand out. The main concepts of MS were dis-
cussed in Chapter 7. MS may be used as a stand-alone technique without prior sep-
aration, or in combination with a separation technique such as GC, HPLC, or CE
[44]. In the very powerful combination of CE and MS, the efflux from the capillary is
subjected to electrospray ionization, and the resulting ions are analyzed, for
instance, with a time-of-flight mass spectrometer.
The power of MS methods in metabolomics becomes especially evident when hun-
dreds of spectrograms are stacked up into metabolic landscapes. As a beautiful
example, Johnson and collaborators [45] combined liquid chromatography (LC) with a
specific high-resolution variant of MS, which was accurate and sensitive enough to
allow the discrimination of thousands of mass-to-charge (m/z) peaks (see Chapter 7).
Furthermore, because the method can be automated and is relatively cost-effective,
(A) 600 (B) 600 (C) 600

400 400 400
200 200 200
0 0 0
100 100 100
50 50 50
0 0 0
700 700 700
600 time 600 time 600 time
500 500 500
400 400 400
m/z 300 m/z 300 m/z 300
200 200 200
100 100 100
Figure 8.7 Metabolic landscapes of plasma samples from three human subjects (A–C). A typical MS spectrogram shows just one slice within these three-
dimensional plots, with mass over charge (m/z) as horizontal axis and ion intensity as vertical axis. The third axis (time) here indicates the retention time during
the sample separation with liquid chromatography. Analysis of the data in the figure shows that many features are common among the three individuals, but
also that subtle differences in metabolic states exist and can be detected. (Courtesy of Tianwei Yu, Kichun Lee, and Dean Jones, Emory University.)
it can be used to compare the metabolic profiles of individual subjects and thereby pro-
vides a powerful tool for personalized medicine (see Chapter 13). As an illustration of
the resolution power of the method, Figure 8.7 shows the metabolic profiles of plasma
samples from three individuals, as they change after stimulation with cystine. Many of
the peaks in such profiles are at first unknown and require further chemical analysis.
NMR spectroscopy exploits a physical feature of matter called nuclear spin. This
spin is a property of all atomic nuclei possessing an odd number of protons and neu-
trons, and can be regarded as a spinning motion of the nucleus on its own axis. Associ-
ated with this spinning motion is a magnetic moment, which makes the atoms behave
like small magnets. When a static magnetic field is applied, the atoms align parallel to
this field. Atoms possessing a spin quantum number of ½, which is the case for most
biologically relevant nuclei, separate into two populations that are characterized by
states with different energy levels and orientations. If an oscillating magnetic field is
applied in the plane perpendicular to the static magnetic field, transitions between the
two states can be induced. The resulting magnetization in the atoms can be recorded
and mathematically transformed into a frequency spectrum. The nucleus is to some
degree protected from the full force of the applied field by its surrounding cloud of elec-
trons, and this “shielding” slightly changes the transition frequency of the nucleus. This
effect on the frequency enables NMR to distinguish not only different molecules but
also the nuclei within each molecule. Because both excitation and detection are due to
magnetic fields, the sample is preserved and does not require prior treatment or sepa-
ration; thus, NMR is non-invasive and nondestructive (Figures 8.8 and 8.9).
circular pump controller
Figure 8.8 Basic experimental set-up

for an in vivo NMR experiment. A cell
solution is kept in a bioreactor, where pH,
pO2 pH temp.
temperature, and other physicochemical
parameters are controlled. A circulation
pump is used to move small quantities of
SOS
cell solution, via the blue tube, into the
detection zone, where a magnetic field is
applied and the spectrogram is measured.
The cell solution is returned to the bioreactor
through the black tube. Should the black
tube be clogged up with cells, the blue SOS
tube serves as a back-up to “save our system”
detection from spillage. The experiments can be done
zone cell suspension under aerobic or anaerobic conditions.
(Courtesy of Luis Fonseca and Ana Rute
bioreactor O2 Neves, ITQB, Portugal.)
METHODS OF METABOLOMIC DATA GENERATION 249
The most relevant nuclei detectable with NMR are 1H, 13C, 15N, 19F, and 31P. The β-glucose
lactate
nuclei 1H, 19F, and 31P are the predominant isotopes in nature, and NMR is useful
for analyses of mixtures, extracts, and biofluids. By contrast, 13C and 15N are rare.
13 α-glucose
C, for instance, has a natural abundance of only about 1%, which makes it par-
ticularly attractive for metabolic studies, where a 13C-enriched substrate is given to
cells and the fate of the labeled carbon atom is traced over time, thereby allowing
the identification of metabolic pathways and the measurement of carbon fluxes FBP
(see Case Study 3 in Section 8.12). In some cases, it is useful to couple 13C-NMR
with 31P-NMR, which permits additional measurements of time series of metabo- n)
mi
e(
lites such as ATP, ADP, phosphocreatine, and inorganic phosphate, and provides tim
100 80 60 40 20
an indication of the energy state and the intracellular pH of the cells. Like MS, NMR chemical shift (ppm)
can also be used to analyze metabolites in vitro. The advantages are less sample
preparation and better structural details, while the main disadvantage is lower Figure 8.9 Result of an in vivo NMR
sensitivity. experiment. The raw data consist of
frequency spectra, which are subsequently
converted into metabolite concentrations
8.9 Flux Analysis at different time points. Here, glucose in
the medium decreases over time, while
Stoichiometric and flux balance analysis (Chapter 3) require the experimental lactate accumulates. In addition to external
determination of metabolic fluxes. The fluxes fall in two categories. The first con- substrates and end products, internal
sists of fluxes entering or leaving the biological system. These are comparatively metabolites such as fructose bisphosphate
easy to measure with methods of analytical chemistry, for example, as uptake (dis- can be measured. (Courtesy of Ana Rute
Neves and Luis Fonseca, ITQB, Portugal.)
appearance) of substrate from the medium or the generation of waste materials,
such as lactate, acetate, or CO2. In contrast to these influxes and effluxes, the major-
ity of interesting processes happen inside cells, where they are difficult to measure.
It is not feasible to kill the cells for these measurements, because many internal
fluxes would presumably stop. In some cases, the analysis of external fluxes is
sufficient to infer the distribution of internal fluxes, but usually additional assump-
tions are required [46]. Furthermore, this inference fails if the system contains
parallel pathways, bidirectional reaction steps, or metabolic cycles. Finally, one has
to assume that energy-producing and energy-consuming reactions are known in
some detail.
One relatively recent method that addresses some of these issues is called meta-
bolic flux analysis and is based on isotopomers, which are isomers with isotopic
atoms in different positions [47]. For instance, in a typical isotopomer experiment
with 13C, a metabolite with three carbons can at most have 23 isotopomers, because
each carbon could be labeled or unlabeled. The various isotopomers are usually not
equally likely and instead are produced with a frequency distribution that contains
clues regarding the internal fluxes of a system. Specifically, one assumes that the
cellular system is in an open steady state, where material is fed to and metabolized
by the system, but where all metabolite concentrations and fluxes are constant.
Once the labeled material has been given to the cells, one waits until the label has
equilibrated throughout the metabolic system. Subsequently, the cells are har-
vested and metabolism is stopped, a cell extract is prepared, and the isotopomer
distribution for many or all metabolites is determined with NMR or MS techniques.
Evaluation of the results requires a mathematical model of the pathway that
describes how the labeled material is distributed throughout the system. The
parameters of this model are initially unknown and must be estimated from the iso-
topomer distributions. Once the model has been estimated, the internal fluxes can
be quantified. Isotopomer analysis can be very powerful, but the generation and
analysis of isotopomer data is not trivial. It requires solid skills in NMR spectroscopy
or MS, as well as in computational modeling and statistics. Also, there are many
challenges in the details. For instance, the time for the isotopomer equilibration to
stabilize is sometimes difficult to determine.
Proteome and genome data may be used as coarse substitutes for indirect
assessments of flux distributions, but caution is required. The abundance of enzymes
obtained with the methods of proteomics may be assumed to correlate (linearly)
with the overall turnover rates of the corresponding reaction steps, and changes in
gene expression may be assumed to correlate with amounts of proteins and enzyme
activities, but this is not always true. As is to be expected, these assumptions some-
times lead to good, but sometimes to rather unreliable, approximations of the actual
processes in living cells [48].
FROM DATA TO SYSTEMS MODELS

The variety and diversity of biochemical and metabolomic data offer distinct
approaches for the development of models. These have the potential for insights of
different types and are exemplified below with three case studies.
8.10 Case Study 1: Analyzing Metabolism in an Incompletely

Characterized Organism
It is obvious that some organisms are better understood than others. Much research
has focused on model organisms such as the bacterium Escherichia coli, the fruit fly
Drosophila melanogaster, and the mouse Mus musculus, with the expectation that
insights from these representatives are applicable to lesser-known organisms. In
many cases, such inferences are valid, but nature also has a way of varying details so
that, for instance, high percentages of genes still have unknown functions, even in
moderately well-characterized organisms.
A beautiful example of characterizing metabolism with a variety of modern
methods has been published in Science [35]. In this study, an international group of
researchers established a complete metabolic map of the small bacterium Myco-
plasma pneumoniae, which is the culprit in some types of human pneumonia. They
selected this particular organism mainly because of its comparatively small size and
simple organization: M. pneumoniae contains fewer than 700 protein-coding genes,
of which the authors, however, found only 231 to be annotated. Earlier studies had
characterized a number of metabolic aspects in this bacterium with biochemical
and computational methods [49, 50], and the results of these studies were integrated
here with new analyses.
The authors of [35] began by assembling a catalog of known reactions, using
KEGG as a starting point and assessing the activities of the various reactions with
information from the literature and some new annotations. Complementing this
information, they collected genome information, such as the co-occurrence of
genes in the same operon and the homology of sequences with known enzymes in
related organisms, and evaluated the likely functionality of inferred enzymes using
structural information regarding pertinent catalytic residues. This combined metab-
olomic and genomic data mining led to apparently disconnected pathways, but
also to suggestions for completing the metabolic map. As an example, the authors
used methods discussed in Chapter 3 to fill a critical gap in the seemingly discon-
nected ascorbate pathway. They succeeded by inferring from sequence homology,
predicted activity, and the position within the responsible operon that the organism
is likely to possess the critical enzyme i-ascorbate 6-phosphate lactonase, without
which the ascorbate pathway would not be functional. In other cases, the structural
analysis permitted the elimination of putative enzymes that lacked the catalytic resi-
dues necessary for substrate turnover. Finally, putative enzyme-catalyzed reactions
were eliminated from the metabolic map if their substrates were not produced by
any of the other pathways. The result of this information mining and manual cura-
tion was a metabolic map without gaps, isolated reactions, or open metabolic loops.
The map was then further refined with targeted laboratory experiments in order to
validate the presence of alternative pathways and to determine the primary direc-
tion in reversible pathways. These experiments included growth on different sub-
strates and 13C-labeled glucose.
Further analysis of the refined metabolic map, supported by additional wet
experiments, led to interesting insights. One finding confirmed the intuitive assump-
tion that genes are up- or down-regulated in functional, pathway-associated clus-
ters and in a dynamic fashion. Moreover, there appeared to be good concordance
between gene expression, protein abundance, and metabolic turnover. For instance,
the authors observed concomitant increases in mRNA, protein expression, and
turnover rate associated with glycolysis, following medium acidification. Closer
inspection of the results actually indicated, not surprisingly, a definite time delay
between gene regulation and changes in metabolite concentrations.
Experiments with 13C-labeled glucose showed that 95% of all glucose in
M. pneumoniae ultimately becomes lactate and acetate, which implies that the
FROM DATA TO SYSTEMS MODELS 251
lion’s share of glycolysis is used for energy generation. Another interesting finding
was that this simple organism contains many more linear pathways, and fewer
diverging and converging branch points, than more complicated organisms, and
is therefore less redundant. This decreased redundancy has direct implications for
the robustness of the organism, because some metabolites can only be produced
in a single manner. As a consequence, 60% of the metabolic enzymes in M. pneu-
moniae were found to be essential, which is about four times as many as in E. coli.
Finally, the authors concluded that M. pneumoniae is not optimized for fast
growth and biomass production, which is often assumed in flux balance analyses
as the main objective of a microbe, but possibly for survival strategies that sup-
press the growth of competitor populations and interactions with the host
organism.
8.11 Case Study 2: Metabolic Network Analysis

Microorganisms can produce valuable organic compounds in high quantities
and often exquisite purity. Examples include yeast cells converting sugars into
alcohol, fungi producing citric acid, and bacteria producing insulin. However,
many wild-type organisms produce these compounds only in small quantities,
and it is the task of metabolic engineers to improve yield. Traditionally, they
pursued this goal with long sequences of genetic manipulations, repeated fine-
tuning of the growth medium, and selection of the best strains [51]. As an alterna-
tive, modern methods of metabolic systems analysis may be pursued (see [12, 32,
52] and Chapter 14).
A representative example is the industrial production of the amino acid lysine,
using the microorganism Corynebacterium glutamicum. Lysine is a valuable addi-
tive for animal feeds, and many studies have therefore worked on improving its
yield. Most of these studies used stoichiometric and flux balance analysis (see
[53–55] and Chapter 3), but some groups also designed fully kinetic models of the
pathway [56, 57]. This option was feasible because the pathway structure of lysine
biosynthesis is well known, and databases such as KEGG, BioCyc, and BRENDA
contain information on all the enzymes involved.
Generally, the development of kinetic models is much more challenging than
a stoichiometric analysis, because the material balance at all metabolite nodes at
the steady state must be augmented with detailed rate equations. This process is
complicated for two reasons. First, the best-suited mathematical representations
of specific enzyme reactions are usually not known and any assumptions regard-
ing their appropriateness are difficult to validate. One may circumvent this chal-
lenge by selecting a canonical representation, but it is still not trivial to determine
to what degree such a representation is valid (see the discussions in Chapter 4
and [12, 58]). The second challenge is the determination of numerical values for
all parameters in the selected rate equations, such as the turnover rate of a pro-
cess or the Michaelis constant of an enzyme. Plenty of pertinent information
regarding numerous enzymes can be found in BRENDA and ExPASy, but this
information was often inferred from experiments in vitro, and it is unclear
whether the presented kinetic characteristics are numerically valid in vivo (see,
for example, [59]).
8.12 Case Study 3: Extraction of Dynamic Models from

Experimental Data
The vast majority of dynamic models of metabolic systems have been constructed
from the bottom up. In the first step, each reaction or process within the system is
assigned a mathematical representation, such as a mass action term, the MMRL, or
a power-law function. Second, data from the literature are used to estimate the
parameters of each rate function. Third, all representations are merged into a model
of the entire system. Fourth, with high hopes, one tests the integrated model against
validation data. Unfortunately, this test fails more often than not, and several rounds
of refinements and searches for better parameter values ensue. Even for moderately
sized systems, the process often takes months, if not years. However, the result is a
fully dynamic model that has a much higher potential than a static model. Examples
are manifold and include [60–62].
A distinct alternative is model development from the top down, using
time-series data. The group of Santos and Neves used in vivo 13C- and 31P-NMR
spectroscopy to measure metabolite concentrations in living populations of the
bacterium Lactococcus lactis [63, 64]. Because NMR is nondestructive, the group
was able to measure dense time series of the decreasing concentration of labeled
substrate, as well as the time-dependent concentrations of metabolic intermedi-
ates and end products, all within the same cell culture. By fine-tuning their meth-
ods, they were able to measure the key metabolites of glycolysis in intervals of 30
seconds or less, thereby creating hour-long time series of metabolite profiles. The
resulting data were successfully converted into computational models (see, for
example, [65, 66]).
Kinoshita and co-workers [67] measured time-course profiles of over 100
metabolites in human red blood cells, using capillary electrophoresis with subse-
quent mass spectrometry (CE-MS). Using these data, the authors developed a
mathematical pathway model, consisting of about 50 enzyme kinetic rate equa-
tions, whose formats were extracted from the literature. The model allowed
manipulations mimicking the condition of hypoxia, where the organism is
deprived of oxygen.
EXERCISES
8.1. Screen the literature or Internet for large reaction. If you have done Exercise (8.9), compare
metabolites that are not proteins, peptides, RNA, or the two types of inhibition.
DNA. Compare their sizes with those of proteins. 8.11. For a reaction with noncompetitive inhibition, plot
8.2. Describe (8.8) and (8.9) in words. vP,NI against S, and 1/vP,NI against 1/S, for several
8.3. Confirm by mathematical means that no material is inhibitor concentrations and compare the plots
lost or gained in the Michaelis–Menten process. with the corresponding plot from the uninhibited
reaction. If you have done Exercise (8.9) or (8.10),
8.4. Formulate a differential equation for E in the compare your results.
Michaelis–Menten process. Compare this equation
with (8.8) and interpret the result. 8.12. Extract from KEGG and BioCyc all pathways using
glucose 6-phosphate as a substrate. Are these
8.5. Derive (8.10) from the ODE system (8.7)–(8.9) or pathways the same in E. coli, baker’s yeast, and
find the derivation in the literature or Internet. For humans?
the derivation, make use of the quasi-steady-state
8.13. Mine databases and the literature to reconstruct
assumption and of the fact that the sum E + (ES) is
a metabolic pathway system describing the
constant.
production of lysine. Search for kinetic parameter
8.6. Simulate (8.7)–(8.10) and compare P with vp in values for as many steps as possible. Note: This is an
(8.10) for different initial values and parameter open-ended problem with many possible solutions.
values. Why are no initial values needed in (8.10)?
8.14. Given the known influxes (blue) and effluxes
8.7. Redo the analysis of Exercise 8.6, but assume that S (green) in Figure 8.10, compute the internal fluxes
is 1000 times larger than E, which might mimic (red) of the system. Is the solution unique? Discuss
in vitro conditions. Evaluate the validity of QSSA
in comparison with Figure 8.3.
8.8. Explain the details of the dynamics of the substrate 5
in the Hill model (8.15). Formulate equations for

8
(ES) and P. B C D
8.9. For a reaction with competitive inhibition, plot

20
vP,CI against S, and 1/vP,CI against 1/S, for several A G
inhibitor concentrations and compare the plots 7
with the corresponding plot from the uninhibited
E F
reaction.
8.10. For a reaction with uncompetitive inhibition, plot 10
H
vP,UI against S, and 1/vP,UI against 1/S, for several
inhibitor concentrations and compare the plots Figure 8.10 Generic stoichiometric map. The influxes (blue) and effluxes
with the corresponding plot from the uninhibited (green) are given in units of mmol L−1 min−1.
EXERCISES 253
issues of uniqueness generically and provide Xyl (X7)
examples and counterexamples. Is the system in CO2 V7,1

Figure 8.10 necessarily in a steady state? Is the
flux information sufficient to determine the
concentrations of the metabolites A, …, H? G6P (X6) V6,1 Ri5P; Xy5P; Ru5P (X1)
8.15. Design models for the pathway in Figure 8.5.

Represent the production of A as 1 · D−0.5. For all V1,2
simulations, start at the steady state (which you
have to compute first) and double the efflux from D
V5,6
at t = 10. In the first set of simulations, use the S7P (X2) GAP (X3)
MMRL for the reactions that use A, B, C, or D as
V3,0
substrate. Choose different combinations of Vmax
and KM values as you like. Repeat the simulations V2,5
with power-law rate laws. Summarize your findings F6P (X5)
in a report.
E4P (X4)
8.16. Design a power-law model for the pathway in V5,3
Figure 8.6. Ignore intermediates that are suggested V4,5
by dotted lines and set all rate constants equal to 1.
GAP (X3)
For all substrate degradation processes, use kinetic
orders of 0.5. For inhibition effects, start with kinetic V3,0
orders of −0.5. Subsequently, change the inhibition
effects one by one, or in combination, to other
negative values. Always start your simulations at the
Figure 8.11 Pentose phosphate pathway. Information on influxes and
steady state. Summarize your findings in a report.
effluxes permits the computation of internal fluxes. (From Wiechert W &
8.17. How could one model the effect of altered gene de Graaf AA. Biotechnol. Bioeng. 55 [1997] 101–117. With permission from
expression on a metabolic pathway? Discuss John Wiley & Sons.)
different options.
8.18. Compare the ascorbate pathway in M. pneumoniae
with that in E. coli. Report your findings. TABLE 8.1: NUMBER OF CARBON ATOMS
8.19. Find five pathways in E. coli that do not exist in IN EACH METABOLITE IN THE PATHWAY IN
M. pneumoniae. FIGURE 8.11
8.20. Explore how much citric acid is being produced Compound Pool Number of Carbon Atoms
annually and which microorganism is primarily X1 5
used for its large-scale industrial production.
X2 7
Search for models that have addressed citric acid
production and review their key features. X3 3
8.21. Compare kinetic parameters of the fermentation X4 4

pathway in yeast that were presented in the work of X5 6
Galazzo and Bailey [68] and Curto et al. [60] with X6 6
those that are listed in BRENDA and ExPASy. X7 5
Summarize your findings in a spreadsheet or a report.
8.22. Consider the model in Figure 8.11, which we have
already studied in Chapter 3 (Exercise 3.28). It
8.23. Using the model in Exercise 8.22 with the same flux
describes the pentose pathway, which exchanges
distribution, is it possible that the system can have
carbohydrates with different numbers of carbon
different steady-state concentrations? If not,
atoms, as shown in Table 8.1. For instance, two
explain. If the answer is yes, construct one or more
units of X1 are needed to produce one unit of X2 and
examples. Write a brief report on your findings.
X3, and each reaction V5,3 uses one unit of X5 to
generate two units of X3. Note that X3 appears in two 8.24. Using the results from Exercise 8.22, perform a
locations, but represents the same pool. Assume simulation that starts at the steady state and models
that all processes follow mass action kinetics and a bolus perturbation to X1. Make a prediction of
that the influx (V7,1) and the effluxes (V3,0 and CO2) the model responses before you simulate and
are 20 mmol L−1 min−1 and that all other fluxes have summarize your findings in a brief report.
a value of 10 mmol L−1 min−1. Confirm that these 8.25. Using the results from Exercises 8.22–8.24, perform
settings correspond to a consistent steady-state flux two simulations that start at the steady state and
distribution. Assume the following steady-state model either a persistent or a bolus perturbation to
concentrations: X1 = 10, X2 = 6, X3 = 12, X4 = 8, X7. Make a prediction of the model responses before
X5 = 8, X6 = 4, X7 = 20. Compute the rate constants you simulate and write a brief report on your
of all processes. findings.
REFERENCES
[1] Derechin M, Ramel AH, Lazarus NR & Barnard EA. Yeast [30] Joshi-Tope G, Gillespie M, Vastrik I, et al. Reactome: a
hexokinase. II. Molecular weight and dissociation behavior. knowledgebase of biological pathways. Nucleic Acids Res.
Biochemistry 5 (1966) 4017–4025. 33 (2005) D428–D432.
[2] Human Genome Project Information. http://www.ornl.gov/ [31] Lipid Maps: Lipidomics Gateway. http://www.lipidmaps.org/.
sci/techresources/Human_Genome/home.shtml. [32] Palsson BØ. Systems Biology: Properties of Reconstructed
[3] Guldberg CM & Waage P. Studier i Affiniteten. Forhandlinger i Networks. Cambridge University Press, 2006.
Videnskabs-Selskabet i Christiania (1864) 35. [33] Visser D & Heijnen JJ. Dynamic simulation and metabolic
[4] Guldberg CM & Waage P. Études sur les affinites chimiques. re-design of a branched pathway using linlog kinetics. Metab.
Brøgger & Christie, 1867. Eng. 5 (2003) 164–176.
[5] Voit EO, Martens HA & Omholt SW. 150 years of the mass [34] Alves R & Savageau MA. Effect of overall feedback inhibition
action law. PLoS Comput. Biol. 11 (2015) e1004012. in unbranched biosynthetic pathways. Biophys. J. 79 (2000)
[6] Gillespie DT. Stochastic simulation of chemical kinetics. Annu. 2290–2304.
Rev. Phys. Chem. 58 (2007) 35–55. [35] Yus E, Maier T, Michalodimitrakis K, et al. Impact of genome
[7] Wolkenhauer O, Ullah M, Kolch W & Cho K-H. Modelling and reduction on bacterial metabolism and its regulation. Science
simulation of intracellular dynamics: choosing an appropriate 326 (2009) 1263–1268.
framework. IEEE Trans. Nanobiosci. 3 (2004) 200–207. [36] Neves M-J & François J. On the mechanism by which a heat
[8] Henri MV. Lois générales de l’action des diastases. Hermann, shock induces trehalose accumulation in Saccharomyces
1903. cerevisiae. Biochem. J. 288 (1992) 859–864.
[9] Michaelis L & Menten ML. Die Kinetik der Invertinwirkung. [37] Fonseca LL, Sánchez C, Santos H & Voit EO. Complex
Biochem. Z. 49 (1913) 333–369. coordination of multi-scale cellular responses to
[10] Benkovic SJ, Hammes GG & Hammes-Schiffer S. Free-energy environmental stress. Mol. BioSyst. 7 (2011) 731–741.
landscape of enzyme catalysis. Biochemistry 47 (2008) [38] Harrigan GG & Goodacre R (eds). Metabolic Profiling. Kluwer,
3317–3321. 2003.
[11] Briggs GE & Haldane JBS. A note on the kinetics of enzyme [39] Oldiges M & Takors R. Applying metabolic profiling
action. Biochem. J. 19 (1925) 338–339. techniques for stimulus–response experiments: chances
[12] Torres NV & Voit EO. Pathway Analysis and Optimization in and pitfalls. Adv. Biochem. Eng. Biotechnol. 92 (2005)
Metabolic Engineering. Cambridge University Press, 2002. 173–196.
[13] Savageau MA. Enzyme kinetics in vitro and in vivo: Michaelis– [40] Villas-Bôas SG, Mas S, Åkesson M, et al. Mass spectrometry in
Menten revisited. In Principles of Medical Biology (EE Bittar, metabolome analysis. Mass Spectrom. Rev. 24 (2005) 613–646.
ed.), pp 93–146. JAI Press, 1995. [41] Theobald U, Mailinger W, Baltes M, et al. In vivo analysis
[14] Hill AV. The possible effects of the aggregation of the of metabolic dynamics in Saccharomyces cerevisiae: I.
molecules of hæmoglobin on its dissociation curves. J. Physiol. Experimental observations. Biotechnol. Bioeng. 55 (1997)
40 (1910) iv–vii. 305–316.
[15] Edelstein-Keshet L. Mathematical Models in Biology. McGraw- [42] Spura J, Reimer LC, Wieloch P, et al. A method for enzyme
Hill, 1988. quenching in microbial metabolome analysis successfully
[16] Schultz AR. Enzyme Kinetics: From Diastase to Multi-Enzyme applied to Gram-positive and Gram-negative bacteria and
Systems. Cambridge University Press, 1994. yeast. Anal. Biochem. 394 (2009) 192–201.
[17] Cornish-Bowden A. Fundamentals of Enzyme Kinetics, 4th ed. [43] Garcia-Perez I, Vallejo M, Garcia A, et al. Metabolic
Wiley, 2012. fingerprinting with capillary electrophoresis. J. Chromatogr.
[18] Savageau MA. Biochemical Systems Analysis: A Study of A 1204 (2008) 130–139.
Function and Design in Molecular Biology. Addison-Wesley, [44] Scriba GK. Nonaqueous capillary electrophoresis–mass
1976. spectrometry. J. Chromatogr. A 1159 (2007) 28–41.
[19] Berg PH, Voit EO & White R. A pharmacodynamic model for [45] Johnson JM, Yu T, Strobel FH & Jones DP. A practical approach
the action of the antibiotic imipenem on Pseudomonas in to detect unique metabolic patterns for personalized
vitro. Bull. Math. Biol. 58 (1966) 923–938. medicine. Analyst 135 (2010) 2864–2870.
[20] BRENDA: The Comprehensive Enzyme Information System. [46] Varma A & Palsson BØ. Metabolic flux balancing: basic
http://www.brenda-enzymes.org/. concepts, scientific and practical use. Biotechnology (NY)
[21] ENZYME: Enzyme Nomenclature Database. http://enzyme. 12 (1994) 994–998.
expasy.org/. [47] Wiechert W. 13C metabolic flux analysis. Metab. Eng. 3 (2001)
[22] Fell DA. Understanding the Control of Metabolism. Portland 195–206.
Press, 1997. [48] Griffin TJ, Gygi SP, Ideker T, et al. Complementary profiling of
[23] Voit EO. Computational Analysis of Biochemical Systems: gene expression at the transcriptome and proteome levels
A Practical Guide for Biochemists and Molecular Biologists. in Saccharomyces cerevisiae. Mol. Cell. Proteomics 1 (2002)
[24] KEGG (Kyoto Encyclopedia of Genes and Genomes) Pathway [49] Pachkov M, Dandekar T, Korbel J, et al. Use of pathway
Database. http://www.genome.jp/kegg/pathway.html. analysis and genome context methods for functional
[25] BioCyc Database Collection. http://biocyc.org/. genomics of Mycoplasma pneumoniae nucleotide
[26] ERGO: Genome Analysis and Discovery System. Integrated metabolism. Gene 396 (2007) 215–225.
Genomics. http://ergo.integratedgenomics.com/ERGO/. [50] Pollack JD. Mycoplasma genes: a case for reflective
[27] Pathway Studio. Ariadne Genomics. http://ariadnegenomics. annotation. Trends Microbiol. 5 (1997) 413–419.
com/products/pathway-studio. [51] Ratledge C & Kristiansen B (eds). Basic Biotechnology, 3rd ed.
[28] KInfer. COSBI, The Microsoft Research–University of Toronto Cambridge University Press, 2006.
Centre for Computational and Systems Biology. www.cosbi. [52] Stephanopoulos GN, Aristidou AA & Nielsen J. Metabolic
eu/index.php/research/prototypes/kinfer. Engineering: Principles and Methodologies. Academic Press,
[29] Reactome. http://www.reactome.org. 1998.
FuRTHER READING 255
[53] Koffas M & Stephanopoulos G. Strain improvement by [61] Hynne F, Danø S & Sørensen PG. Full-scale model of glycolysis
metabolic engineering: lysine production as a case study for in Saccharomyces cerevisiae. Biophys. Chem. 94 (2001) 121–163.
systems biology. Curr. Opin. Biotechnol. 16 (2005) 361–366. [62] Alvarez-Vasquez F, Sims KJ, Hannun YA & Voit EO. Integration
[54] Kjeldsen KR & Nielsen J. In silico genome-scale reconstruction of kinetic information on yeast sphingolipid metabolism
and validation of the Corynebacterium glutamicum metabolic in dynamical pathway models. J. Theor. Biol. 226 (2004)
network. Biotechnol. Bioeng. 102 (2009) 583–597. 265–291.
[55] Wittmann C, Kim HM & Heinzle E. Metabolic network analysis [63] Neves AR, Pool WA, Kok J, et al. Overview on sugar
of lysine producing Corynebacterium glutamicum at a metabolism and its control in Lactococcus lactis—the input
miniaturized scale. Biotechnol. Bioeng. 87 (2004) 1–6. from in vivo NMR. FEMS Microbiol. Rev. 29 (2005) 531–554.
[56] Dräger A, Kronfeld M, Ziller MJ, et al. Modeling metabolic [64] Neves AR, Ventura R, Mansour N, et al. Is the glycolytic flux in
networks in C. glutamicum: a comparison of rate laws in Lactococcus lactis primarily controlled by the redox charge?
combination with various parameter optimization strategies. Kinetics of NAD+ and NADH pools determined in vivo by 13C
BMC Syst. Biol. 3 (2009) 5. NMR. J. Biol. Chem. 277 (2002) 28088–28098.
[57] Yang C, Hua Q & Shimizu K. Development of a kinetic model [65] Voit EO, Almeida JO, Marino S, Lall R. et al. Regulation of
for l-lysine biosynthesis in Corynebacterium glutamicum and glycolysis in Lactococcus lactis: an unfinished systems
its application to metabolic control analysis. J. Biosci. Bioeng. biological case study. Syst. Biol. (Stevenage) 153 (2006)
88 (1999) 393–403. 286–298.
[58] Voit EO. Modelling metabolic networks using power-laws and [66] Chou IC & Voit EO. Recent developments in parameter
S-systems. Essays Biochem. 45 (2008) 29–40. estimation and structure identification of biochemical and
[59] Albe KR, Butler MH & Wright BE. Cellular concentrations genomic systems. Math. Biosci. 219 (2009) 57–83.
of enzymes and their substrates. J. Theor. Biol. 143 (1989) [67] Kinoshita A, Tsukada K, Soga T, et al. Roles of hemoglobin
163–195. allostery in hypoxia-induced metabolic alterations in
[60] Curto R, Sorribas A & Cascante M. Comparative erythrocytes: simulation and its verification by metabolome
characterization of the fermentation pathway of analysis. J. Biol. Chem. 282 (2007) 10731–10741.
Saccharomyces cerevisiae using biochemical systems theory [68] Galazzo JL & Bailey JE. Fermentation pathway kinetics and
and metabolic control analysis. Model definition and metabolic flux control in suspended and immobilized S.
nomenclature. Math. Biosci. 130 (1995) 25–50. cerevisiae. Enzyme Microbiol. Technol. 12 (1990) 162–172.
FuRTHER READING
Cornish-Bowden A. Fundamentals of Enzyme Kinetics, 4th ed. Savageau MA. Biochemical Systems Analysis: A Study of Function and
Wiley, 2012. Design in Molecular Biology. Addison-Wesley, 1976.
Edelstein-Keshet L. Mathematical Models in Biology. McGraw-Hill, Schultz AR. Enzyme Kinetics: From Diastase to Multi-Enzyme
1988. Systems. Cambridge University Press, 1994.
Fell DA. Understanding the Control of Metabolism. Portland Press, Torres NV & Voit EO. Pathway Analysis and Optimization in
1997. Metabolic Engineering. Cambridge University Press, 2002.
Harrigan GG & Goodacre R (eds). Metabolic Profiling. Voit EO. Computational Analysis of Biochemical Systems: A Practical
Kluwer, 2003. Guide for Biochemists and Molecular Biologists. Cambridge
Koffas M & Stephanopoulos G. Strain improvement by metabolic University Press, 2000.
engineering: lysine production as a case study for systems Voit EO. Biochemical systems theory: a review. ISRN Biomath.
biology. Curr. Opin. Biotechnol. 16 (2005) 361–366. 2013 (2013) Article ID 897658.
Signaling Systems
9
• Understand the fundamentals of signal transduction systems
• Discuss the advantages and disadvantages of discrete and continuous
models of signal transduction
• Describe the role of G-proteins
• Solve problems involving Boolean models
• Understand bistability and hysteresis
• Discuss quorum sensing
• Analyze differential equation models of two-component signal transduction
systems
• Analyze differential equation models of the MAPK signal transduction system
In cell biology, signal transduction describes a class of mechanisms through

which a cell receives internal or external signals, processes them, and triggers
appropriate responses. In many cases, the signals are chemical, but cells can also
respond to different kinds of physical signals, such as light, electrical stimuli, and
mechanical stresses. In fact, many different signals are often processed at the
same time [1]. If the cell receives an external chemical signal, the transduction
process usually begins with the binding of specific ligands to receptors on the
outer surface of the cell membrane. These receptors are transmembrane pro-
teins, and binding of a ligand to the extracellular domain of the receptor triggers
a change in the conformation or function of the receptor that propagates to the
intracellular domain of the receptor. This change mediates other changes within
the cell that ultimately result in an appropriate response to the original signal.
As a specific example, integrins are transmembrane proteins that transduce
information from the extracellular matrix or the surrounding tissue to the
inside of the cell and are involved in cell survival, apoptosis, proliferation, and
differentiation [2].
Good examples of signal transduction circuits can be found in an important
family of signaling molecules called G-proteins. These are located on the inside of
the cell membrane and are coupled to specific receptors that reach from the inside
to the outer surface of the cell and whose conformation changes when a ligand
binds to them. Specifically, upon binding to the extracellular domain of such a
G-protein-coupled receptor (GPCR) (Figure 9.1), the change in the intracellular
domain of the GPCR causes a subunit of the G-protein to be released and to bind to
a molecule of guanosine triphosphate (GTP), which leads to the dissociation of two
molecular subunits. The importance of this process led to the prefix “G,” which
258 Chapter 9: Signaling Systems
Figure 9.1 Example of a membrane-

spanning G-protein-coupled receptor
(GPCR_3ny8). The human β2-adrenergic
receptor (β2-AR) consists of a seven-helical-
bundle membrane protein. The two planes
represent the lipid bilayer. GPCRs represent
a large fraction of current pharmaceutical
targets, and β2-AR is one of the most
extensively studied. (Courtesy of Huiling
Chen and Ying Xu, University of Georgia.)
stands for “guanine-nucleotide-binding.” The activated G-protein subunits detach

from the receptor and initiate internal signaling processes that are specific to the
particular type of G-protein. For instance, the process may trigger the production of
cyclic adenosine monophosphate (cAMP), which can activate protein kinase A,
which in turn can phosphorylate and thereby activate many possible downstream
targets. Other G-proteins stimulate the generation of signaling compounds such as
diacylglycerol and inositol trisphosphate, which control the release of intracellular
calcium from storage compartments into the cytoplasm. After some recovery
period, hydrolysis removes the third phosphate group from GTP and readies the
G-protein for transduction of a new signal. The functionality of very many drugs is
associated with GPCRs. For their “discovery of G-proteins and the role of these
proteins in signal transduction in cells,” Alfred Gilman and Martin Rodbell received
the 1994 Nobel Prize in Physiology or Medicine.
The signal transduction processes inside the cell are mediated by chains of
biochemical reactions that are catalyzed by enzymes and often modulated by sec-
ond messengers, which are water-soluble or -insoluble molecules or gases such as
nitric oxide. Typical water-soluble, hydrophilic messengers are calcium and cAMP.
They are usually located in the cytosol and may be activated or deactivated by spe-
cific enzymes. Hydrophobic signaling molecules are often associated with the plasma
membrane. Examples include different phosphatidylinositols, inositol trisphosphate,
and the sphingolipid ceramide. The steps between receiving a signal at the mem-
brane and triggering a response in the cytosol or nucleus are often organized as a
cascade of events that amplify the incoming signal and filter out noise. The speed of
signal transduction depends strongly on the mechanism. Calcium-triggered electri-
cal signals occur at the order of milliseconds, while protein- and lipid-based cas-
cades respond at the order of minutes. Signals involving genomic responses happen
within tens of minutes, hours, or even days. Chapter 12 discusses the complex, cal-
cium-based events occurring during every beat of the heart.
Because signaling processes are ubiquitous in biology, an enormous literature
has been amassed over the past decades, and even a cursory review of all signal
transduction systems is far beyond the scope of this chapter. In contrast to the
wealth of experimental information, modeling studies are much fewer, although
their number has also been rising quickly in recent years. In this chapter, we will
discuss representative examples of models that illustrate certain aspects of signal
transduction systems. These models fall into two distinct classes with only scant
overlap. The first class considers signaling networks as graphs and tries (a) to
understand their functionality and (b) to deduce their causal connectivity from
experimental observations with methods of discrete mathematics and statistics.
The second class attempts to capture the dynamics of signaling networks with
ordinary or stochastic differential equation models.
Static ModelS of Signal tranSduction netWorkS 259
Static ModelS of Signal tranSduction

netWorkS
9.1 Boolean networks
An intuitive approach toward assessing signaling systems is a network with
hardwired connections and signaling events that happen on a discrete timescale.
Suppose the signaling system consists of 25 components. To be less vague, let us
imagine that these components are genes, and a that response is triggered if these
genes show a particular combination of expression, where certain genes are on
(meaning that they are expressed), whereas others are off (for instance, because
they are presently repressed). Furthermore, suppose that an expressed gene
ultimately leads to the production of a transcription factor or repressor that may
affect its own expression and/or that of other genes in the system.
Let us at once abstract and simplify the set-up by assembling the 25 genes in a
5 × 5 grid of boxes (Figure 9.2A) and let us name the genes according to their
position, such that G11 refers to the top left corner, G12 lies just to the right of G11, G21
is just below G11, and so on. An arbitrary decision says that a white box means the
gene is on and that black means off. Since we like numbers, we translate off into 0
and on into 1. Again, this setting is rather arbitrary. We can now formulate
statements such as the following:
for i, j = 1, …, 5,
1 if i = j,
Gij =  (9.1)
0 if i ≠ j,
which means that the genes on the northwest–southeast diagonal of the grid are on
(Figure 9.2B).
The beauty of these types of networks lies in the simplicity of defining their
dynamics. Time moves forward in a discrete fashion, which we easily formulate as
t = 0, 1, 2, … (see Chapter 4). Each Gij is now a recursive function of time, Gij = Gij(t),
and changes in the states of the network are determined by rules that state what
happens to each state during the transition from one time point to the next. As an
example of a very simple rule, consider the following situation:
for i, j = 1, …, 5,
1 if Gij (t ) = 0,
Gij (t + 1) =  (9.2)
0 if Gij (t ) = 1.
Put into words, this rule says that 0 switches to 1 and 1 switches to 0 at every time
transition (Figure 9.2C). In other words, the rule defines a blinking pattern, switching
back and forth between Figures 9.2B and C. This dynamics is independent of the
initial state of the network. In other words, we can define Gij(0) any way we want, as
long as all settings are either 0 or 1.
This type of network with on/off states and transition rules is often called a
Boolean network, honoring the nineteenth-century British mathematician and
philosopher George Boole, who is credited with inventing the logical foundation of
modern computer science. Much has been written about Boolean networks, and an
(A) (B) (C)
G11 G12 G13
Figure 9.2 An intuitive example of

G21 G22 G23
a Boolean network. The network is
G31 G32 represented by a grid with states that can
either be on or off. (A) Naming of grid boxes
in a systematic fashion. (B) On–off pattern
as defined by (9.1). (C) On–off pattern as
defined by (9.2), if the previous state was
that in (B).
excellent starting point for further exploration is Stuart Kauffman’s widely recog-
nized treatise The Origins of Order [3].
The rules in a Boolean network may be much more complicated than in the
illustration above. For instance, a rule could say that Gij switches if the majority of its
neighbors are 1. How could we formulate such a rule? Let us start with a node that is
not located at one of the sides of the grid. Thus, it has eight neighbors, which are
Gi −1, j −1 , Gi −1, j , Gi −1, j +1 , Gi , j −1 , Gi , j +1 , Gi +1, j −1 , Gi +1, j , Gi +1, j+1 .
Because each neighbor can only take values of 0 or 1, we may formulate the rule as
1 if Gij (t ) = 0 and sum of all neighbors > 4,

0 if Gij (t ) = 0 and sum of all neighbors ≤ 4,

Gij (t + 1) =  (9.3)
0 if Gij (t ) = 1 and sum of all neighbors > 4,
1 if Gij (t ) = 1 and sum of all neighbors ≤ 4.
The rule is applied iteratively at the transition from any time t to time t + 1, during
which the values of the neighbors of each node at time t determine the value of this
node at time t + 1. It should be noted that the rule in (9.3) may be written in differ-
ent, equivalent, ways, some of which will certainly be more compact and elegant.
For intriguing entertainment with a more complex network that has many
conceptual similarities with a Boolean network, explore the Game of Life, which was
created by John Conway in 1970 and attracted interest from an almost Zen-like
following of biologists, mathematicians, computer scientists, philosophers, and
many others. The game is still available on the Internet and even has its own
Wikipedia page.
Boolean networks can easily be interpreted as signaling systems, for instance, in
genomes. Instead of using a fixed grid, it is more common to locate the genes Gi as
nodes in a directed graph (see Chapter 3), where incoming arrows show the
influences of other genes (Gj, Gk, …), while outgoing arrows indicate which genes
are affected by the gene Gi. A typical illustrative example is given in Figure 9.3.
Because of evolutionary advantages with respect to robustness, gene networks
appear to be only sparsely connected, and most genes have only one or two upstream
regulators [4], which greatly facilitates the construction of gene regulatory network
models.
The signaling function is accomplished by defining a specific rule for each gene
Gi at the transition from time t to t + 1, which is determined by the expression state
of some or all other genes (or gene products) at time t. Under well-defined constel-
lations of the system, Gi will “fire” (become 1), while it is otherwise silent (Gi = 0).
As an example, suppose that at t = 0 only G1 in Figure 9.3 is on and all other genes are
off. The activation arrows and inhibition symbols suggest that, at t = 1, G2 will still be G1 G2
off, while G3 and G4 will turn on.
To capture the full dynamics of the network, it is necessary to define for each
gene whether it will be on or off at time t + 1, given an on–off pattern of all genes at
time t. These complete definitions are needed, because, for instance, the graphical
representation in Figure 9.3 does not prescribe what will happen to G3 if both G1 and G3
G2 are on. Will the activation trump the inhibition or will the opposite be true? The
state of G3 at t + 1 may depend on its own state at time t. In general, the rules may be
very complicated, which implies that rather complex signaling systems may be con-
structed and analyzed. As an example, Davidich and Bornholdt [5] modeled the cell
cycle regulatory network in fission yeast with a Boolean network. We will return to
G4 G5 G6
this example at the end of this chapter.
Two features of Boolean networks are of particular significance. First, there is no
real size limitation, and because the rules are usually defined locally, that is, for each
node, it is often relatively easy to construct very large dynamic processes with Figure 9.3 Representation of a generic
complicated on–off patterns. But even if one knows all the rules, it is sometimes gene interaction network as a graph.
difficult to predict with intuition alone how a network will evolve. Second, it is con- Arrows indicate activation while lines
ceptually easy to expand the number of possible responses per state. This expansion with terminal bars represent inhibition or
allows gray tones, which might correspond to several (discrete) expression levels of repression.
Signal tranSduction SySteMS Modeled With differential equationS 261
genes. The rules become somewhat more complicated, but the principles remain
the same. It is even possible to allow for a continuous range of each expression state.
With this expansion, the formal description becomes much more complicated and
resembles that of an artificial neural network except that an artificial neural network
allows modifications of the strengths of connections between nodes (see [6] and
Chapter 13). An expansion that is far from trivial is the permission to make the rules
themselves time-dependent or adaptive in a sense that they change in response to
the overall state of the network.
A limitation of traditional Boolean networks is indeed that typically they are
not adaptive and do not allow time-dependent rules, and also they have no
memory: as soon as a new state is reached, the previous state of the network is
forgotten. In many biological systems, such memory is important. Another limi-
tation is the strictly deterministic nature of these networks: given the initial state
and the transition rules, the future of the network is completely determined.
In order to overcome this particular limitation and to allow for biological
variability and stochasticity, Shmulevich and colleagues proposed probabilistic
Boolean networks, where the transitions between states occur with predefined
probabilities [7, 8]. This set-up may remind us of the Markov models that we
discussed in Chapter 4.
9.2 network inference

Natural signal transduction networks have the ability to function robustly in noto-
riously noisy environments. Whether it is gene expression, which can show
significant stochastic fluctuations, or a system of neurons, where action potentials
may fire spontaneously, a signal transduction system ultimately responds with a
high degree of reliability to the correct inputs while generally filtering out spurious
inputs.
The genuinely stochastic environments and internal features of signal transduc-
tion systems create many challenges for modeling and analysis. The biggest chal-
lenge is actually not the determination of whether a signal, corrupted by noise, will
trigger a response—usually, this merely requires a comparatively simple application
and evaluation of the rules governing conditional probabilities. The true challenge
consists of the opposite task, as we have already discussed in Chapter 3. Namely,
given only a set of input signals and the corresponding set of responses, is it possible
to infer or reconstruct the structure and numerical features of the inner workings of
the signal transduction network? Is it possible to conclude, with any degree of reli-
ability, that component X only fires if components Y and Z are both active? The
answers to such questions obviously depend on a number of factors, including the
amount of data and their degree of certainty, the expected noise, and, last but not
least, the size of the network: it does not take much imagination to expect that large
networks are much more difficult to infer than systems with just a handful of
components.
The topic of network inference has traditionally been addressed with two meth-
ods: either one designs a static graph model and uses methods of Bayesian network
inference, which we discussed in Chapter 3, or one constructs a system of differen-
tial equations and attempts to identify the connectivity and regulation of the
signaling system from sets of time-series data with inverse (parameter estimation)
methods (see [9, 10] and Chapter 5).
Signal tranSduction SySteMS Modeled With

differential equationS
9.3 Bistability and hysteresis
The construction of systems of differential equations that are geared toward captur-
ing the dynamics of signal transduction systems can take many forms. While the sky
is the limit in this approach, the key component is usually a module that permits
bistability. As the name suggests, such a module possesses two stable states, which
are separated by an unstable state. If a signal is low, the system assumes one stable H, PL
steady state, and if the signal exceeds some threshold, which is related to the 40 PL
unstable state, the system moves to the other stable steady state. The simplest
example of such a system is a switch that can be triggered by a signal. If the signal is H
present, the system is on, and when the signal disappears, the system turns off. 20
Bistable systems of different types are found in many areas of biology, examples
including classical genetic toggle switches [11, 12], cell cycle control, cellular signal
transduction pathways, switches to new programs during development, and the
0
switch between lysis and lysogeny in bacteriophages [13]. X
0 150 300
Bistable systems are easy to construct. For instance, in Chapter 4 we used a
polynomial of the form Figure 9.4 Bistability is often implemented
as a differential equation containing the
X = c (SS1 − X )(SS2 − X )(SS3 − X ). (9.4) difference between an S-shaped function
and a linear or power-law term. Shown
Here, the parameter c just determines the speed of a response, while SS1, SS2, and here are the graphs of a typical signal
response relationship in the form of a shifted
SS3 are three steady states. It is easy to confirm the latter, because setting X to any of
sigmoidal Hill function H (blue; the first two
the steady-state values makes X zero. For instance, for (SS1, SS2, SS3) = (10, 40, 100), terms on the right-hand side of (9.5)) and
and assuming that c is positive, SS1 = 10 and SS3 = 100 are stable steady states, while a power-law degradation function (green;
SS2 = 40 is unstable. the last term on the right-hand side of (9.5)).
A different approach to designing a system with a bistable switch is to combine a Their intersections indicate steady-state
sigmoidal trigger with a linear or power-law relaxation function that lets the response points. The most pertinent cases contain two
variable return to its off state. A minimalistic model consists of a single ODE, such as stable (yellow) and one unstable state (red).
20 X 4 TABLE 9.1: SCHEDULE OF

X = 10 + − 2 X 0.5 . (9.5)
100 4 + X 4 RESETTING X IN (9.5) AT SEVERAL
TIME POINTS t (SEE FIGURE 9.5)
This barely qualifies as a “system,” since it only has one variable but, notwithstand-
ing semantics, we may interpret the equation as the description of a quantity X that t X
is produced with a constant input of magnitude 10 and replicates or activates its 0 25
own production by means of a Hill term. Furthermore, X is consumed or degraded 20 10
according to a power-law function. 40 90
As always, the steady state(s) of the system can be assessed by setting (9.5) equal
100 60
to zero. We could try to solve the equation algebraically, but that is a bit messy. How-
ever, we can easily convince ourselves that a steady state is achieved if the two posi- 180 120
tive terms exactly balance the negative term. Thus, we can plot the two functions and 300 280
study where they intersect. The graph (Figure 9.4) indicates that there are actually 400 150
three intersection points, which correspond directly to the steady states. One occurs 480 50
for X ≈ 25, the second for X ≈ 100, and the third for X ≈ 210. Implementing the sys-
520 80
tem in PLAS and starting the solution with different initial values quickly shows that
the true steady states for X ≈ 25.41736 and X ≈ 210.7623 are stable. The third steady 550 10
state is actually at exactly X = 100, and it is unstable: starting slightly higher, even
for 100.1, the system diverges toward the higher stable steady state, whereas it X
300
approaches the lower state for smaller initial values, even if they are as close as 99.9.
An interpretation could be that the system is in either a healthy (low steady) state or
200
a diseased (high steady) state. Both are stable, and small perturbations are easily
tolerated in each state, whereas large perturbations may shift the system from health
to disease or vice versa. Biologically, the unstable state is never realized, at least not 100
for long, because it does not tolerate even the smallest disturbances.
As an example, suppose the system resides at the low steady state (X ≈ 25), but 0
0 300 600
time
external signals reset the value of X at a series of time points, as indicated in Table 9.1.
Figure 9.5 illustrates the responses of the system. Upon each resetting event, the Figure 9.5 Response of X in (9.5) to
system approaches the stable steady state in whose basin of attraction it lies. The external signals that reset the value of X
unstable state does not have such a basin, because it is repelling: solutions starting at the time points identified with dashed
even slightly higher than X = 100 wander toward the high steady state, and solutions red lines (see Table 9.1). The three dotted
starting lower than the unstable state approach the low steady state. lines indicate the three steady states. The
As an actual biological example of a bistable system, consider an auto-activation central steady state at X = 100 is unstable
and separates the two basins of attraction of
mechanism for the expression of a mutant T7 RNA polymerase (T7R), which was
the stable steady states at about 25 and 211.
investigated by Tan and collaborators [14]. A priori, the most straightforward repre- If reset above 100, the system approaches
sentation seemed to require merely a positive activation loop (Figure 9.6). However, the higher steady state, whereas resetting
closer observations indicated that the natural system exhibited bistability, which the below 100 causes the system to approach
simple activation motif is unable to do. The authors therefore combined two positive the low steady state.
activation cycles, one for the auto-activation and the second representing the meta-
bolic cost of the additional mechanism and subsequent growth retardation, as well
as dilution of T7R due to growth (Figure 9.7). Indeed, the merging of the two positive
+
X
loops generated bistability that was qualitatively similar to actual cellular responses.
With suitable scaling, Tan et al. [14] proposed the following equation to
represent the system:
Figure 9.6 A simple positive activation
loop. “Motifs” of this type are discussed in
i δ + α ⋅ (T 7 R) φ ⋅ (T 7 R)
(T 7 R) = − − (T 7 R). (9.6) Chapter 14.
1 + (T 7 R) 1 + γ ⋅ (T 7 R)
The first term on the right-hand side describes the production of T7R, including
dilution
self-activation, the second term represents utilization of T7R for growth, metabolic
costs, and dilution, while the last term accounts for intrinsic decay. With clever
numerical scaling, the authors were able to keep the number of parameters to a +
T7R growth
minimum, and the remaining parameter values were set as δ = 0.01, α = 10, φ = 20,
and γ = 10.
decay
Of course we can easily solve the differential equation (9.6), but of primary
metabolic burden
importance is whether the system has the potential for bistability. Thus, the aspect
of major interest is the characterization of fixed points, which requires setting the Figure 9.7 The combination of two auto-
differential equation equal to zero. The result is an algebraic equation that we can activation loops can lead to bistable
again evaluate by plotting the production term and the two degradation terms behavior. Here, a mutant T7 RNA polymerase
against T7R (Figure 9.8). This plot indicates that production and degradation inter- (T7R) is auto-active, but causes growth
sect three times, and a little bit more analysis shows that the system has two stable retardation. Growth, in turn, dilutes T7R.
steady states (for very high and very low values of T7R) and one unstable steady All effects together cause bitable behavior.
state for an intermediate value of T7R. See [14] for details.
With a little bit more effort, ODEs with sigmoidal components can be made more
interesting. For instance, consider a small two-variable system, which could be
interpreted as consisting of two genes that affect each other’s expression. A high 1000
signal S triggers the expression of Y, Y affects X, and X affects Y. The system is shown 100
in Figure 9.9.
10
As we have seen in Chapters 2 and 4, it is not too difficult to translate a diagram
process rate
like that in Figure 9.9 into differential equations. Of course there are many choices 1
for an implementation, and we select one that contains a sigmoidal function
0.1
representing the effect of Y on the expression of X. Specifically, let us consider the production
following system: 0.01 degradation
0
800Y 4
X = 200 + 4
0.0001 0.01 1 100
− 50 X 0.5 ,
16 + Y 4 (9.7)
T7R
Y = 50 X 0.5S − 100Y . Figure 9.8 Visualization of bistability.

Plotting the production and degradation
The signal is an independent variable that affects the system from the outside. rates of T7R in the system in Figure 9.7 and
Suppose the signal changes during our simulation experiment as shown in Table 9.2. (9.6) against T7R shows that the system has
Starting close to the steady state (X, Y) ≈ (20.114, 6.727), and following the signal three steady states. Two of these (green)
resetting schedule in the table, we obtain the response in Figure 9.10. are stable, whereas the center state (yellow)
Of greatest interest in comparison with the bistable system is that something is unstable, resulting in bistability of the
system.
truly new and intriguing is happening here. The system starts with S = 3, which cor-
responds to the low steady-state value X ≈ 20 and is to be interpreted as off. When
the signal is increased to 10, X shoots up to about 400 and is on. Fifteen time units
later, the signal goes back to S = 3, and we would expect the response to shut off
S
again. However, X does not return to its off state of about 20, but instead drops only
modestly to a value of about 337, which is still to be interpreted as on. At the end of
the experiment, the signal is back to S = 3 again, and X is off again, but this happens Y
only after the signal had been reduced temporarily to an even smaller value (S = 2.5). +
Thus, for a signal strength of S = 3, the system can be either on or off! Further analysis +
of this curious behavior implies is that the system has built-in memory: the state of X
the response variable depends not only on the current value of the external signal,
Figure 9.9 Artificial system with two
but also on the recent history of the signal and the system. Mathematically, this his-
components and an external signal. X and
tory is stored in the dependent variables. Indeed, if we look at signal S and response Y might be interpreted as the expression
X, it appears that there is memory, or hysteresis, as it is called in the jargon of the states of two genes. The system is affected
field. In conveniently structured models, conditions can be derived for the exis- by signal S, which could be interpreted as a
tence or absence of bistability and hysteretic effects (see, for example, [14–16]). An stressor or a transcription factor.
example of a relatively simple real-world system is a so-called two-component sig-

TABLE 9.2: SUSTAINED SIGNAL
naling system in bacteria, which we will discuss in the next section of this chapter.
STRENGTHS S BEGINNING AT
Hysteresis can be explored in greater detail with a two-phase simulation study TIME POINTS t
(Figure 9.11). In the first phase, we start with a steady state resulting from a low
value of S. We increase S slowly, in a stepwise fashion, let the system come sufficiently t S
close to the steady state, and then increase S again. For this illustration, let us 0 3
concentrate on the variable X. For values of S between 0 and 2.5, the steady-state
5 10
value of X, Xss, is close to 16 and changes very little, and on further raising S to about
3.3, Xss increases gradually to about 30. On raising S further, to about 3.4, Xss suddenly 20 3
jumps to about 365. Checking signal strength in the vicinity of 3.4 demonstrates that 40 2.5
this is a real jump and not a very fast, but gradual, increase in Xss. Further rises in S, 50 3
above 3.4, increase Xss a little more, but not much.
Now we enter the second phase, where we start with a high value of S and slowly
lower it in a stepwise manner, each time recording Xss. At first, the steady-state values
exactly trace those obtained from the first phase, and we would probably expect a
jump back for S ≈ 3.4. Surprisingly, this jump does not happen, and Xss decreases
only slowly. As a second surprise, Xss crashes from about 230 to about 17 once
S reaches a value close to 2.5. Again, this is not a fast, gradual decrease, but a real
jump. Lowering S further causes Xss to trace the early results of the first phase of the
simulation.
The important conclusion here is the following: for a signal with a strength S in
400
the range between about 2.6 and 3.3, the value of Xss depends on history of the X
system. The situation is depicted in Figure 9.11B, which shows the different steady
states of X for S in this range, which depend on whether S is increased from low
X, Y 200
values or decreased from high values. Note that there are no steady-state values on
the dashed lines. The response of Y is similar, although of lower magnitude, and we
Y
will return to it later in the chapter, when we discuss clocks.
0
A simple bistable system models a “hard” switch: a signal just a tiny bit above the 0 30 60
threshold will always trigger an on response, while a signal just below results in an
off response. These response patterns are entirely deterministic. Suppose now that time
the system operates close to the threshold value for Xss, that is, the unstable steady-
state point. This is a reasonable assumption, because otherwise the capability of Figure 9.10 The two-variable signaling
system in (9.7) shows hysteresis.
bistability of the system would be wasted. Let us suppose further that the environ-
The response variable assumes different
ment is mildly stochastic. If so, then every change in S across the threshold, up or values for the same input signal, depending
down, will make the system jump. Hysteresis softens this situation. In the range of S on the history of the system. See the text
between 2.6 and 3.3, the illustration system tends to remain on or off, depending for values of the signal, which changes at the
where it was just before, and noise in the signal is therefore easily tolerated. times indicated by the arrows.
(A) Xss 4 (B) Xss

450 450
400 400
350 350
3
300 300
250 250
S increasing
200 200 S decreasing
2
common
150 150
5
100 100
6
50 50
0 0
0 2 4 6 8 10 12 2.0 2.5 3.0 3.5 4.0
S S
1
Figure 9.11 Demonstration of hysteresis. (A) indicates a series of alterations in S and the corresponding responses in Xss. Increasing S from 0 to about 3
affects Xss only slightly. However, changing S from about 3.3 to 3.4 results in an enormous jump in Xss. Further increases in S do not change the new steady
state much. Decreasing S from 10 to 3.4 traces the same steady states. However, there is no jump at 3.4. Surprisingly, the jump happens between about
2.6 and 2.5. (B) shows a magnification of the system response plot for signal strengths around 3, where Xss depends on the recent history of the system.
The dashed arrows indicate jumps and do not reflect steady states.
Expressed differently, it takes a true, substantial change in signal to make the system
jump. Section 9.7 and Box 9.1 discuss further analyses and variants of this basic
hysteretic system.
In recent years, many authors have studied what happens if a signal and/or a
system are corrupted by noise. Such investigations require methods that allow
stochastic input, such as stochastic modeling, differential equations with stochastic
components, or stochastic Petri nets [13, 19, 20] (see Exercise 9.18).
BOX 9.1: ONE-WAY HYSTERESIS
An important step in the mammalian immune response is the notation (Figure 1C). Note that this diagram is actually quite
differentiation of B cells into antibody-secreting plasma cells. This similar to the small system in Figure 9.9 that we studied earlier.
step is triggered by an antigen that the host has not encountered We formulate the system with the following model, where the
before. The signaling pathway controlling this differentiation Hill function H is a function of X:
process is governed by a gene regulatory network, where several
genes and transcription factors play critical roles. Focusing on 4X 4
H = 1+ ,
the three key transcription factors Bcl-6, Blimp-1, and Pax5, 16 4 + X 4
Bhattacharya and collaborators [17] developed a mathematical X = 3.2H 2 + Y 0.5 S − 2 X , (1)
model capturing the dynamics of this differentiation system.
Y = 4 H − Y 0.5 .
They were specifically interested in understanding how dioxin
(TCDD; 2,3,7,8-tetrachlorodibenzo-p-dioxin) can interrupt the To examine the responses of the system to different magnitudes
healthy functioning of this signaling pathway. Dioxin is a toxic of the stimulus, we perform exactly the same analysis as for the
environmental contaminant that is produced in incineration system in (9.7). Namely, we increase S in repeated small steps
processes. It is also infamous as Agent Orange, which was used from a very low value to a high value, every time computing the
in the Vietnam War to defoliate crops and forests. A simplified stable steady state of the system and focusing in particular on the
diagram describing the functional interactions among the genes steady-state values Xss. The results are shown in Figure 2.
and transcription factors in the model is shown in Figure 1A. For small signal strengths, Xss is only mildly affected. For
The model that Bhattacharya et al. proposed is actually more S = 0, it has a value of about 1.6, which slowly increases to about
complicated, since they explicitly considered gene expression, 8 with higher values of S. Because we have some experience
transcription, and translation. Of interest is that the multiple with these types of systems, we are not too surprised that there
repression signals form a dynamic switch that, upon stimulation is a critical point, here for S ≈ 2.25, where the response jumps
with an antigen, ultimately directs a B cell to differentiate into a to a high level of about 62 (dashed arrow). For higher signal
plasma cell. The antigen considered here is a lipopolysaccharide strengths, the response increases further, but at a relatively slow
(LPS), a compound found on the outer membranes of many pace. Now we go backwards, decreasing S at every step a little
bacteria and a known trigger of strong immune responses in bit. Again, we are not surprised that the system does not jump
animals. Bhattacharya and collaborators analyzed in detail how back down to the lower branch at S ≈ 2.25, because we expect it
this pathway responds to deterministic and stochastic noise in the to make this jump at a smaller value. However, it never does. Even
LPS signal [17, 18]. reducing the signal to 0 does not move the system from the upper
To study the switching behavior of the deterministic system, branch to the lower branch. Mathematically, such a jump would
we make further drastic simplifications. In particular, we omit happen for about S ≈ −0.8, but of course that is biologically not
TCDD and its effect on the pathway. Furthermore, we leave out possible. In other words, the system has similar off steady states
Bcl-6 and replace the associated dual inhibition with an auto- for low signal strengths, jumps to on steady states at about 2.25,
activation process for Blimp-1, because two serial inhibitions but now is stuck with these on steady states, even if the signal
correspond to activation. Finally, we create a differentiation marker disappears entirely. Interpreting this observation for a B cell,
DM by combining Pax5 with IgM, with the understanding that IgM we are witnessing its irreversible differentiation into an antibody-
is high when Pax5 is low, and vice versa. The simplified signaling secreting plasma cell. The unstable steady states (red) cannot be
pathway is shown in Figure 1B. Defining X as Blimp-1, Y as the DM, simulated. However, they can be computed if we express S as a
and S as LPS, it is easy to transcribe the system into our typical function of Xss (see Exercise 9.34).
(A) LPS
TCDD
Figure 1 Simplified diagrams of the B-cell differentiation
pathway. (A) A slightly simplified version of the model proposed
Bcl-6 Blimp-1 Pax5 IgM by Bhattacharya et al. [17]. Lipopolysaccharide (LPS) stimulates this
system. Blimp-1 is a master regulator that represses the transcription
factors Bcl-6 and Pax5, which are typical for the B-cell phenotype. Pax5
inhibits the expression of immunoglobulin M (IgM), which mediates
(B) LPS (C) the immune response and can lead to the differentiation of the B cell
into an antigen-secreting plasma cell. The environmental pollutant
S Y dioxin (TCDD) interferes with the signaling pathway. (B) A drastic
reduction of the pathway, where TCDD is left out and Pax5 and IgM are
Blimp-1 DM X combined into a differentiation marker, DM. (C) The same diagram in
our typical notation, with X representing Blimp-1, Y the differentiation
marker, and S the lipopolysaccharide.
Xss Figure 2 The B-cell differentiation pathway exhibits “one-way

100 hysteresis.” For small signals, the system has a steady state on the
lower branch of the green hysteresis curve, which corresponds to the
phenotype of a B cell. When the signal exceeds 2.25 in magnitude,
the system jumps (dashed arrow) to a steady state on the upper
branch, which corresponds to the phenotype of an antibody-secreting
plasma cell. However, once the system has entered a steady state on
50 the upper branch, it can never return to a steady state on the lower
branch, even if the signal disappears entirely. Thus, the differentiation
step is irreversible. The red component depicts unstable steady states.
The dots indicate some simulation results.
S
0 1 2 3 4
9.4 two-component Signaling Systems

Two-component signaling (TCS) systems occur widely in archaea and bacteria, as
well as in some fungi and plants. They are absent in animals, which rather transduce
signals with three-component systems, protein kinase cascades, and/or lipid-based
systems. TCS systems allow organisms to sense changes in environmental factors
such as temperature and osmolarity, and to discern the direction of chemical
gradients. Once the signal has been received, it is processed and ultimately allows
the organism to respond in an appropriate fashion. In many cases, the TCS system
is integrated with other regulatory mechanisms, such as allosteric enzyme regula-
tion in the nitrogen assimilation of enteric bacteria [21]. For a recent collection of
articles on TCS systems, see [22].
A relatively well-understood variant of a bacterial TCS system is the signal trans-
duction protein CheY in Escherichia coli, which reacts to nutrients and other chemi-
cal attractants in the environment and affects the direction of the rotation of the
organism’s flagellar motor [23] (Figure 9.12). Binding of an attractant lowers the
activity of the receptor-coupled kinase CheA, which decreases the level of
phosphorylated CheY. This CheY-P in turn binds to a component of the motor
(see Chapter 7). The motor usually rotates counterclockwise, but changes direction
when CheY is phosphorylated. As a result, the bacterium tumbles and changes
direction. Thus, mediated by CheY, the cell controls its swimming behavior.
Figure 9.12 Structure of a CheZ dimer,

which is instrumental in chemotaxis in
E. coli. The phosphatase CheZ stimulates the
dephosphorylation of the response regulator
CheY by a so-far unknown mechanism. The
most prominent structural feature of the
CheZ dimer is a long four-helix bundle that
consists of two helices from each monomer.
(Protein Data Bank: PDB 1KMI [24].)
Figure 9.13 Signal transduction in a two-

signal
component system. The environmental
signal binds to the sensor S. A histidine
1 residue (His) on S receives a phosphate
group (Pi), which is quickly transferred to an
aspartic acid (Asp) residue on the response
S
regulator R, which binds to DNA and effects a
change in gene expression.
His
2 Pi ATP
ADP
Asp
R
Asp Pi
(A)
signal
The first component of a TCS system acts as the sensor or transmitter and S SP
consists of an unphosphorylated membrane-bound histidine kinase (Figure
9.13). If a suitable signal is present, a phosphate group is transferred from ATP to
R
a histidine residue on the sensor, resulting in ADP and a phosphohistidine resi-
due. Because all this happens on the same molecule, the kinase is sometimes Ph
called an autokinase and the process is called autophosphorylation. The bond
between the residue and the phosphate group is rather unstable, which allows RP
the histidine kinase to transfer the phosphate group to an aspartic acid residue
on the second component, the response regulator or receiver. The phosphate
response
transfer causes a conformational change in the receiver, which changes its affin-
ity to a target protein, or more typically to a specific DNA sequence, causing dif-
ferential expression of one or more target genes. The ultimate physiological (B)
signal
effects can be manifold.
Once the task of the TCS system has been accomplished, the response regulator
must be dephosphorylated in order to return it to its receptive state. The rate at S SP
which the aspartyl phosphate is released as inorganic phosphate is well controlled
and results in half-lives of the phosphorylated form of the regulator that may vary
R
from seconds to hours. Interestingly, two slightly different variants have been
observed in the dephosphorylation mechanism of the response regulator. In the
simpler case, the mechanism is independent of the state of the sensor, and the
sensor is called monofunctional. In the more complicated variant, the dephosphor- RP
ylation of the response regulator is enhanced by the unphosphorylated form of the
sensor protein, which implies that the sensor has two roles and is therefore response
bifunctional. The two variants are shown in Figure 9.14. In the following, we perform
simulations with both of them and refer the reader to the literature [15, 25] for a Figure 9.14 Two variants of a two-
further discussion of their comparative relevance. component signaling (TCS) system. The
It should be noted that the phosphorylation state of the regulator in both signal triggers the phosphorylation of a
variants of the design may change in the absence of a signal, which introduces a sensor S, which leads to phosphorylation of
certain level of noise into the system [15, 25]. Other variations of TCS are coupled a response regulator or receiver R, leading to
TCS systems and analogous systems containing more components. An example a response. (A) The monofunctional variant
includes a phosphatase (Ph), which removes
of the latter is the sporulation system Spo in organisms like Bacillus anthracis,
the phosphate from the phosphorylated
which consists of nine possible histidine sensors and two response regulators receiver RP. (B) In the bifunctional variant,
[26]. Other systems of a similar type are found in phosphorelays that regulate the sensor facilitates dephosphorylation of
virulence [22]. Signaling systems in plants also often contain more than two the phosphorylated receiver, a process that is
components. inhibited by the signal.
For a model analysis that is representative of these types of studies, we follow the signal
work of Igoshin et al. [15], which built upon earlier analyses by Alves and Savageau ATP ADP
[25] and Batchelor and Goulian [27]. Of the many questions that could be analyzed,
we focus on the shape of the signal–response curves and ask whether the TCS sys- kap
S SP
tem can be bistable and show hysteresis. Figure 9.15 shows a detailed diagram that
contains both the mono- and bifunctional variants (Figure 9.14A and B, respec- kad
tively) as special cases. It also contains intermediate complexes and shows the rate kd3 kd2 kd1
constants used in the model; their names are taken directly from Igoshin et al. [15]. R
By setting some parameters equal either to zero or to a different value, the mono- or kb3
kcat
kb1
bifunctional models are obtained. Ph~RP Ph
Following Igoshin and collaborators, we set the model up as a mass-action kb4 SP~R
model, so that the only parameters are rate constants, while all kinetic orders are 1 or S~R kd4
0. Translation of the diagram in Figure 9.15 leads directly to the following model: kb2 RP
kph kpt
⋅ S~RP
S = −kap (signal ) ⋅ S + kad (SP ) + kd 3 (S ~R) − kb 3S ⋅ R + kd 2 (S ~RP ) − kb 2S ⋅ (RP ),
⋅ Pi response
(SP ) = kapS − kad (SP ) − kb1 (SP ) ⋅ R + kd1 (SP~R),
⋅
(SP~R) = kb1 (SP ) ⋅ R − kd1 (SP~R) − kpt (SP~R), Figure 9.15 Detailed model diagram of the
⋅ TCS system. By setting kph = 0, one obtains
(S ~RP ) = kpt (SP~R) + kb 2S ⋅ (RP ) − kd 2 (S~RP ) − kph (S ~RP ),
the monofunctional variant.
⋅
(S ~R) = kb 3S ⋅ R − kd 3 (S ~R) + kph (S ~RP ),
R = −kb1 (SP ) ⋅ R + kd1 (SP~R) + kd 3 (S ~R) − kb 3S ⋅ R + kcat (Ph~RP ),
⋅
(RP ) = kd 2 (S ~RP ) − kb 2S ⋅ (RP ) + kd 4 (Ph~RP ) − kb 4 (Ph)⋅ (RP ),
⋅
(Ph~RP ) = kb 4 (Ph) ⋅ (RP ) − kd 4 (Ph~RP ) − kcat (Ph~RP ),
⋅
(Ph) = kd 4 (Ph~RP ) − kb 4 (Ph) ⋅ (RP ) + kcatt (Ph~RP ),
(9.8)
Igoshin and co-workers obtained most of the equations and parameter values from
Batchelor and Goulian [27], but added some reactions and the corresponding
parameters. A list of rate constants and initial values is given in Table 9.3; other ini-
tial values are zero. These parameter values were originally determined for the
EnvZ/OmpR TCS system. EnvZ is an inner membrane histidine kinase/phospha-
tase, for instance in E. coli, that senses changes in the osmolarity of the surrounding
medium. EnvZ thus corresponds to S in our model. It regulates the phosphorylation
state of the transcription factor OmpR, which corresponds to R in our notation.
OmpR controls the expression of genes that code for membrane channel proteins,
called porins, that are relatively abundant and permit the diffusion of molecules
across the membrane.
We begin our analysis with a typical simulation experiment, namely, we start the
system at its steady state and send a signal at time t = 100. Mechanistically, we just
change the independent variable signal in (9.8) from 1 to 3. Since we are not really
interested in intermediate complexes, we show only the responses in S, SP, R, and
RP, among which RP is the most important output to be recorded, because it is an
indicator of the response. The mono- and bifunctional systems show similar, yet
distinctly different, responses. In particular, the bifunctional system responds
noticeably faster (Figure 9.16). If the signal is stronger, (for example, signal = 20),
the two systems show very similar responses. Note that the two systems start at
different values, which correspond to their (different) steady states.
As a second test, we model a short signal. Specifically, we set signal = 20 at time
t = 100 and stop the signal at time t = 200 by resetting it to 1. Interestingly, the
two TCS systems now respond in a qualitatively different fashion (Figure 9.17).
TABLE 9.3: RATE CONSTANTS AND INITIAL VALUES FOR MONOFUNCTIONAL

(MF) AND BIFUNCTIONAL (BF) TCS SYSTEMS (DIFFERENCES ARE SHADED)
kap kad kb1 kd1 kb2 kd2 kb3 kd3 kb4 kd4 kpt kph kcat S(0) R(0) Ph(0)
MF 0.1 0.001 0.5 0.5 0.05 0.5 0.5 0.5 0.5 0.5 1.5 0 0.005 0.17 6.0 0.17
BF 0.1 0.001 0.5 0.5 0.05 0.5 0.5 0.5 0.5 0.5 1.5 0.05 0.025 0.17 6.0 0.17
(A) (B) Figure 9.16 Simulations with the two

6 0.2 6 0.2
RP RP variants of the TCS system. (A) Responses
R of the monofunctional variant to a signal
R
SP at time t = 100 (arrow). (B) Corresponding
R, RP
R, RP
S, SP
S, SP
3 SP 0.1 3 0.1
responses of the bifunctional design.
S S
0 0 0 0
0 1000 2000 0 1000 2000
time time
(A) (B)
6 0.10 6 0.10 Figure 9.17 Differences in responses of
R the two variants of the TCS system. The
RP
monofunctional (A) and bifunctional (B)
R, RP
R, RP
S
S, SP
S, SP
3 RP 0.05 3 0.05
S TCS systems show qualitatively different
R responses to a short, transient signal
SP SP
0 0 0 0 (arrows), with the former returning to its
0 1000 2000 0 1000 2000 steady state and the latter reaching a new
state that appears to be a steady state;
time time however, see Figure 9.18.
The monofunctional system responds to the signal and returns to its steady state as
soon as the signal stops. By contrast, the bifunctional system seems to assume a new
steady state. Will it eventually return to the initial state? The answer is difficult to
predict but easy to check with a simulation: we simply extend the simulation to
something like t = 12,000. Surprisingly, the system does not stay at this apparent
steady state (which really is none), nor does it return to the initial state. Instead, after 6 0.10
a while, it starts rising again (without any intervention from the outside) and RP S
assumes a new steady state (Figure 9.18). R, RP 3 0.05 S, SP
The short-signal experiment suggests at the very least that the bifunctional TCS R
SP
system permits bistability. But does it also permit hysteresis? Moreover, can we test 0 0
whether the monofunctional system exhibits similar responses? To answer these 0 6000 12,000
time
questions, we execute a series of experiments with varying signal strengths and study
the response in RP. We begin with the monofunctional variant, by running it from
Figure 9.18 Continuation of the simulation
the initial values in Table 9.3 (with all other initial values at zero) with a low signal of of the bifunctional system in Figure 9.17.
1 and computing its steady state. Now the real series of experiments begins. We initi- The presumed steady state is only a transient
ate the system at the steady state and reset the signal to a different value between 0 state, and the system shows a late response
and 4 at time t = 100. For signal strengths between 0 and about 1.795, the corre- to the signal by approaching the true steady
sponding values of the response variable RP increase slowly from 0 to about 2. state.
Intriguingly, a tiny further increase in the signal to 1.8 evokes a response of 5.57! Fur-
ther increases in signal do not change this response much. It is pretty clear that the
system is bistable.
Does the system also exhibit hysteresis? To find out, we perform a second set of
experiments, this time starting with the high steady state, which is characterized by
values that we obtain by running the original system with a high signal towards its 6
steady state. Beginning with a high signal of 4, we again see a response of about 5.6. 5
Now we slowly lower the signal from experiment to experiment, every time record- 4
ing the value of RP. Nothing much happens until 1.8, and the results are essentially RP 3
identical to those in the previous series of experiments. However, lowering the 2
1
signal to 1.795 does not cause a jump. In fact, the response is high until the signal
0
strength is reduced to about 1.23, where the response value suddenly drops to about 0 1 2 3 4
0.65. These results are a clear indication of hysteresis. The signal–response relation- signal
ship is shown in Figure 9.19.
Igoshin et al. [15] and Alves and Savageau [25] derived conditions for when the Figure 9.19 The signal–response
different TCS systems exhibit bistability and hysteresis and studied the effect of relationship of the monofunctional TCS
noise in the signal, which in the case of TCS systems appears to be of minor model shows clear hysteresis. For signal
significance. They also discussed the advantages and limitations of the mono- and strengths between about 1.23 and 1.8, the
response depends on whether the system
bifunctional variants in detail. As in many cases of this nature, one variant is more
was in a low or high state when the signal
efficient with respect to some physiologically relevant aspects of the response but hit. This type of phenomenon shows that the
less efficient with respect to other aspects, and the overall superiority of one over the monofunctional TCS system has memory.
other depends on the environment in which the organism lives. We will discuss The response is similar to that of system (9.5)
some methods for the analysis of such variants in Chapter 14. (see Figure 9.11).
signal Figure 9.20 Typical three-layer MAPK

cascade. At each level, the inactive form on
the left is phosphorylated (once or twice)
into the active form, which serves as enzyme
MAPKKK MAPKKK-P at the next layer. Phosphatases (PP1, PP2,
PP3 and PP3) return the active forms to their
inactive states by removing phosphate.
MAPKK MAPKK-P MAPKK-PP

PP2 PP2
MAPK MAPK-P MAPK-PP

PP1 PP1
response
9.5 Mitogen-activated Protein kinase cascades

In contrast to the TCS system in bacteria, most higher organisms use a more compli-
cated signal transduction system whose key component is the mitogen-activated
protein kinase (MAPK) signaling cascade. MAPK (usually pronounced “map-
kinase”) systems appear to be present in all eukaryotes and also in a few prokary-
otes, such as the biofilm-forming bacterium Myxococcus xanthus. Intriguingly,
MAPK cascades have a highly conserved architecture. The MAPK signaling cascade
receives signals from cell surface receptors or cytosolic events, processes them by
filtering out noise, usually amplifies them, and ultimately affects various downstream
targets, such as cytosolic proteins and nuclear transcription factors. The external
signal may consist of a mitogen or inflammatory cytokine, a growth factor, or some
physiological stress. This signal is transduced, for example, by a G-protein-coupled
receptor that activates the MAPK cascade. The ultimate result of the MAPK system
may be as different as inflammation, differentiation, apoptosis, or initiation of the
cell cycle.
The typical MAPK cascade is shown in Figure 9.20. It consists of three layers
where proteins are phosphorylated or dephosphorylated. The kinase nearest to the
signal source is generically called MAP kinase kinase kinase (MAPKKK). If it is
activated by a cytosolic or external signal, MAPKKK is phosphorylated to MAPKKK-P.
Being a kinase itself, MAPKKK-P activates phosphorylation of the MAP kinase
kinase (MAPKK) in the second layer. In fact, full activation of MAPKK requires
sequential phosphorylation at two sites: a tyrosine site and a threonine site.
The resulting MAPKK-PP in turn phosphorylates and thereby activates the kinase of
the third layer, the MAP kinase (MAPK). Again, full activation requires two
phosphorylation steps. The activated MAPK-PP can phosphorylate cytosolic targets
or can translocate to the nucleus, where it activates specific transcriptional pro-
grams. At each layer, a phosphatase can dephosphorylate the active forms into the
corresponding inactive forms. Two prominent examples of MAP kinases are the
extracellular-signal-regulated kinase (ERK) and the c-Jun N-terminal kinase (JNK),
where Jun refers to a family of transcription factors. The responses of the cascades
are quite fast: within the first 5 minutes of stimulation, ERK is activated up to 70%
[28], and within 10 minutes, significant amounts of activated ERK are translocated E
to the nucleus [29]. A mutated form of MAPKKK in the ERK pathway has often been
found in malignant melanomas and other cancers.
An obvious question is this: The three-layer cascade has been conserved through
evolution, so it is fair to assume that this design has advantages. But what are these, X XP XPP
for instance, in comparison with a simple or double phosphorylation at a single
layer? We use a model analysis to shed light on this intriguing question.
The key module of the cascade is the sequential phosphorylation at two sites,
Figure 9.21 Key module of the three-layer
which is catalyzed by the same enzyme from the next higher layer. Simplified and
MAPK cascade in Figure 9.20. It is tempting
detailed diagrams of this module are presented in Figures 9.21 and 9.22, where X to use Michaelis–Menten functions to model
stands for MAPK or MAPKK, E is the catalyzing enzyme, XP and XPP are singly or the two phosphorylation steps, but the two
doubly phosphorylated forms, and XE and XPE are complexes of X and XP with the steps compete for the same enzyme, whose
enzyme. concentration is not constant.
E Figure 9.22 Detailed representation of the

module in Figure 9.21. Accounting for the
two substrate–enzyme complexes XE and
k1 k4
XPE shows the dual role of enzyme E.
X XE XP XPE XPP
k2 k3 k5 k6
k7 k8
It is tempting to set up the two phosphorylation steps with Michaelis–Menten

rate functions, but such a strategy is not the best option, because (1) the enzyme
concentration is not constant, (2) the enzyme concentration is not necessarily
smaller than the substrate concentration, and (3) the two reaction steps are
competing for the same enzyme. Instead, it is useful to retain the mechanistic ideas
of the Michaelis–Menten mechanism, which postulates the formation of a
substrate–enzyme complex, and to formulate this mechanism in the basic format of 30 X
mass-action kinetics (see Chapters 2, 4, and 8). What we do not want to do is to rely
on the quasi-steady-state assumption that would simplify this system toward the
response
XPP
well-known Michaelis–Menten function, because in this format the enzyme 15 XE
concentration is no longer explicit, let alone dynamic. A direct translation of the XPE
diagram in Figure 9.22 into mass-action equations is straightforward: XP
0
⋅ 0 1 2 3 4
X = −k1 X ⋅ E + k2 ( XE ) + k7 ( XP ), time
⋅
( XE ) = k1 X ⋅ E − (k2 + k3 )( XE ),
⋅ Figure 9.23 Response of the MAPK module
( XP ) = k3 ( XE ) − k7 ( XP ) − k4 ( XP ) ⋅ E + k5 ( XPE ) + k8 ( XPP ), in Figure 9.22. At time t = 1, the signal is
⋅ increased. The result is single and double
( XPE ) = k4 ( XP ) ⋅ E − (k5 + k6 )( XPE ),
⋅ phosphorylation of X.
( XPP ) = k6 ( XPE ) − k8 ( XPP ).
(9.9)
30
To perform simulations, we specify more or less arbitrarily chosen values for the X
parameters and initial concentrations in Table 9.4. It is now easy to study the effects
response
of changes in the enzyme E on the variables of the module. At the beginning of the XE XPP
15
simulation, E = 0.01, and the initial conditions for the various forms of X are set such
that the system is more or less in a steady state. Suppose now that at time t = 1 the XPE
XP
signal E is increased to 10. This is a strong signal, and the system responds with phos-
phorylation on both sites. After a brief transition, the balance between X, XP, and 0
0 5 10
XPP is entirely switched, with very little unphosphorylated X left (Figure 9.23). Nota- time
bly, the amount of XPE is relatively large, because quite a lot of enzyme is available.
If the signal is reset to 0.01, the module returns to its initial state (Figure 9.24). Figure 9.24 Continuation of the response
It is also easy to study how strong the signal must be to trigger a response. If E is in Figure 9.23. At time t = 5, the signal is
set to 3, rather than 10, the response is rather similar to the previous scenario, decreased to its original low state, and X
although the response is not as pronounced (results not shown). If E is set to 1 returns to its unphosphorylated state.
instead, the response is much slower. X is reduced only to about 13.6, while XPP
assumes a value of about 3.2. In other words, no real switch is triggered. For even
weaker signals, the response is insignificant (Figure 9.25).
30
Now that we have a bit of a feel for the module, we can use it to implement a X
complete MAPK cascade (see Figure 9.20). Specifically, we set up the module in
triplicate, with additional variable names Y, YP, YPP, Z, ZP, and ZPP, which
response
15
represent the middle and bottom layers of MAPK and MAPKK phosphorylation,
respectively, and use the same parameter values and initial conditions as before. XP
XPP
XE
Regarding the first layer, which involves only one phosphorylation of MAPKKK, XPE
0
0 2 4
time
TABLE 9.4: KINETIC PARAMETER VALUES AND INITIAL CONDITIONS FOR THE
MAPK SYSTEM (9.9) Figure 9.25 Response of the MAPK
module in Figure 9.22 to a weak signal at
k1 = k4 k2 = k5 k3 = k6 k7 = k8 X(0) XE(0) XP(0) XPE(0) XPP(0) t = 1. While X responds to the signal with a
decrease, the doubly phosphorylated form
1 0.1 4 2 28 0.1 0.2 0.01 0.02 XPP remains low.
signal S Figure 9.26 Implementation of a model

of the MAPK cascade in Figure 9.20.
The MAPK and MAPKK layers are analogous,
MAPKKK layer
while the MAPKKK layer is simpler, since it

X = – k1 X•E•S + k2 (XE) + k7 (XP)
contains only one phosphorylation step.
(XE) = k1 X•E•S – (k2 + k3) (XE)
The kinetic rate constants may be different
(XP) = k3 (XE) – k7 (XP)
for each layer. See the text for a description
XP of simplifications in the equations for XP
and YPP.
Y = – k1 Y•(XP) + k2 (YE) + k7 (YP)
MAPKK layer
(YE) = k1 Y•(XP) – (k2 + k3) (YE)

(YP) = k3 (YE) – k7 (YP) - k4 (YP)•(XP) + k5 (YPE) + k8 (YPP)
(YPE) = k4(YP)•(XP) – (k5 + k6)(YPE)
(YPP) = k6 (YPE) – k8 (YPP)
YPP
Z = – k1 Z•(YPP) + k2 (ZE) + k7 (ZP)

MAPK layer
(ZE) = k1 Z•(YPP) – (k2 + k3) (ZE)

(ZP) = k3 (ZE) – k7 (ZP) – k4 (ZP)•(YPP) + k5 (ZPE) + k8 (ZPP) ZPP
(ZPE) = k4 (ZP)•(YPP) – (k5 + k6) (ZPE)
(ZPP) = k6 (ZPE) – k8 (ZPP)
response
we us a single Michaelis–Menten mechanism, again formulated in mass-action

representation, with X(0) = 28 and XE(0) = XP(0) = 0.1. The input to the first module
is signal S, and the output of this module is the singly phosphorylated form XP,
which catalyzes the two phosphorylation steps Y → YP and YP → YPP at the next
layer. Similarly, the output of the second layer, YPP, catalyzes the reactions Z → ZP
and ZP → ZPP at the bottom layer. ZPP triggers the actual physiological response,
but we simply use ZPP as output indicator of the cascaded system. The model imple- 30
mentation is shown in Figure 9.26, and we set the kinetic parameters to exactly the
X Y Z
same values as before (see Table 9.4).
response
Strictly speaking, this implementation is a simplification that assumes that the 15 XP

“enzymes” XP and YPP, as well as ZPP, remain constant, which is not really the case, YPP
since XP undergoes reactions with Y and YP, and YPP reacts with Z and ZP; the role ZPP
of ZPP is harder to quantify in general. Thus, for a more accurate representation, all
0
reactions utilizing or releasing XP, YPP, or ZPP should be included in the equations 0 8 16
governing these variables. For instance, the equation for XP should read time
⋅ Figure 9.27 Responses of the main

( XP ) = k3 ( XE ) − k7 ( XP ) − k1Y ⋅ ( XP ) + (k2 + k3 )(YE ) variables of the MAPK model to switches
−k4 (YP ) ⋅ ( XP ) + (k5 + k6 )(YPE ). (9.10) in signal. The X, Y, and Z variables respond
to an increase and decrease in signal in a
staggered manner (see the text for details).
If the signal is off, XP and YPP have rather small values, but in the case of greater
interest, namely if the signal is on and stays on, XP and YPP are constant, with values
of about 14.3 and 12.7 for the given parameters ki, and simulations show that
accounting explicitly for interactions with the lower levels slows down the responses 20 ZPP
of the cascade a little bit, but that ignoring them does not affect the results qualita-
tively. We therefore ignore these processes for simplicity. Exercise 9.30 assesses the
differences between models.
response
YPP
As a first simulation, suppose the external signal S starts at 0.01, switches to 50 at 10
t = 2, and switches back to 0.01 at t = 6. The responses of the most important XP

variables are shown in Figure 9.27. The strong signal causes rapid phosphorylation
at all layers. If the signal is not as strong (S = 5 at t = 2), X does not dip as close to zero 0
as before, but only to about 6. Otherwise the responses are similar (results are not 0 3 6
shown). For a relatively weak signal (S = 0.5 at t = 2), we see the benefit of the cas- time
caded system. While XP rises only to about 5, the ultimate response variable ZPP
Figure 9.28 Signal amplification by the
shows a robust response (Figure 9.28). In other words, we have solid signal amplifi-
MAPK cascade. Even for a relatively weak
cation in addition to a steeper response and clear signal transduction. If the signal signal, the responses at the three layers of
rises to only two or three times its initial value (S = 0.02 or S = 0.03 at t = 2), the the MAPK cascade are successively stronger,
response in ZPP is very weak. The results give us a first answer to our earlier question indicating amplification of the signal by the
of why there are three layers in the cascade: three layers permit improved signal cascaded structure.
amplification. At the same time, they effectively reduce noise [30]. It has also been 20
ZPP
shown that the three-layer design improves the robustness of the signal transduc-
tion system against variations in parameter values [31].
response
YPP
It is easy to investigate how the cascade responds to brief repeated signals. We 10
perform a simulation as before, but create an off–on sequence of signals. As an
example, if the signal switches between 0.01 and 0.5 every half time unit, beginning XP
with S = 0.5 at t = 2, we obtain the result shown in Figure 9.29. We can see that the
0
responses at the different layers are quite different. The first layer exhibits fast, but 0 8 16
very weak, responses in XP, whereas the ultimate, amplified response ZPP at the time
third layer in some sense smoothes over the individual short signals.
It should be clear that no attempt was made in the implementation of the cascade Figure 9.29 Responses of different layers
to estimate biologically relevant parameter values; our purpose was simply to of the MAPK cascade to brief, repeated
demonstrate the concepts of the dual-phosphorylation module and its role in the signals. The signal at the output layer
MAPK cascade. Huang and Ferrell [32] and Bhalla and Iyengar [33] discuss param- smoothes the responses at the MAPKK and
MAPKKK layers into a solid and sustained,
eter values and features of realistic MAPK cascades in some detail, and Schwacke
amplified signal.
and Voit [34] and Schwacke [35] provide a computational rationale for parameter
values of the cascade that are actually observed in nature and that lead to stronger
amplification than our ad hoc parameter set.
It should be noted that all signaling systems obviously have a spatial component,
which the typical ODE models ignore. In fact, it seems that some cascades are
organized along protein scaffolds, which can significantly affect their efficiency
[36–38]. Furthermore, MAPK cascades typically do not work in isolation. Instead,
the cell contains several parallel cascades, which communicate with each other via
crosstalk [39–41]. For instance, the output of the top layer in one cascade might
activate or inhibit the center layer in another. Crosstalk is thought to increase the
reliability and fidelity of signal transduction. It furthermore creates options for very
complex signaling tasks, such as an in-band detector, where the cascade only fires if
the signal is within a certain range, but neither weaker nor stronger. Crosstalk also
renders all logic functions (and, or, not) including the so-called exclusive or (xor)
mechanism possible. In the latter case, the cascade fires only if either one of two
signals is present, but it does not fire if both signals are on or if both signals are off
[33] (see also the discussion of bi-fans in Chapter 14). Detailed experimental and
theoretical work seems to indicate that normal and cancer cells may differ not so
much in their signaling components, but rather in the way in which these compo-
nents are causally connected within complex signaling networks [42].
It is interesting to note that cells often use the same or similar components of a
signaling cascade for different purposes. For instance, yeast cells employ Ste20p,
Ste11p, and Ste7p, for a number of different stress responses, including pheromone
exposure, osmotic stress, and starvation [38].
S2
9.6 adaptation
S1
The human body responds to many signals from the outside world on a daily basis. signal Rpeak
Our pupils constrict in bright sunlight and dilate when it is dark, we sweat to cool the
body, and we pull our hands back when we touch something hot. Many of these R2 Σ
responses are driven by the body itself, independent of our conscious control. R1 Δ
response
Intriguingly, we often get used to a signal, if it is repeated in short order or if it
persists for some time. We start breathing normally again a few days after moving to
time
high altitudes, and after a while we no longer really notice some noises, unless we
specifically focus on them. This phenomenon of a diminished biological response to Figure 9.30 Features of adaptation. For a
the same repeated signal is called adaptation. In the language of systems biology, low signal S1, the typical response is an off
the organism is initially in a homeostatic steady state and responds to occasional state, characterized by the response R1. For
signals by moving from an off state to an on state, but then returns to the off state short spikes in the signal, the response is
when the signal has vanished. However, if the same type of signal persists for some transient, and the system returns to R1 (not
while, the system returns from the on state to the off state anyway, even though the shown). In the case of adaptation, the system
off state normally corresponds to the absence of the signal. responds to the signal S2 by reaching Rpeak,
but then approaches a value R2 that is close
According to Ma and co-workers [43], adaptation can be characterized by two
to R1, even though the signal S2 persists.
features, which they call sensitivity and precision (Figure 9.30). The sensitivity Σ The response can be characterized by its
measures the normal strength of the response to a signal, that is, the relative differ- sensitivity Σ and precision Π; the latter is
ence between the output values of the on and off states in relation to the relative inversely related to the relative difference Δ
signal strength. The precision Π is defined as the inverse of the difference Δ between between R1 and R2.
the original homeostatic state and the state the system attains through adaptation. S
Specifically,
A
–
|R − R1 | / | R1 |
Σ = peak , (9.11)
| S2 − S1 | / | S1 |
B C
−1
 | R − R1 | / | R1 |  Figure 9.31 Three-node system for
Π = ∆ −1 =  2 . (9.12) studying adaptation. The system is adapted
 | S2 − S1 | / | S1 |  from Ma et al. [43], but represented in the
format we have been using throughout the
Ma’s group explored the question of what types of systems with three nodes are book.
capable of displaying adaptation. One such motif is shown in Figure 9.31 in our
notation. Whereas Ma used a Michaelis–Menten representation, we use an S-system
model with simplified notation, which here has the intriguing advantage that one
can actually compute conditions under which the system displays adaptation
(Exercise 9.31). A symbolic S-system model with positive variables A, B, and C can
be written down immediately:
A = α 1SB g1 − β1 Ah1 ,
B = α 2 A g 2 − β 2 B h2 , (9.13)
C = α 3 A − β 3C .
g3 h3
TABLE 9.5: PARAMETER VALUES
FOR THE ADAPTATION SYSTEM
Without loss of generality, we may suppose that the normal steady state is (1, 1, 1) (9.13)
for a signal S = 1, which immediately means that αi = βi for i = 1, 2, 3; convince your-
self that this is generally true. Table 9.5 provides typical parameter values. i αi = βi gi hi
To explore the responses of the system, we start at the steady state and send signals 1 2 −1 0.4
as brief “boluses” at times t = 5, 30 and 45. In PLAS, this is easily done with commands 2 0.1 1 0.05
like @ 5 S = 4, @ 5.1 S = 1. In all three cases, the output variable C shoots up, then
3 4 0.75 0.5
undershoots, before returning toward the homeostatic value of 1 (Figure 9.32).
The variables A and B are not of prime interest, but they return to 1 as well. These
responses are not surprising, because the steady state (1, 1, 1) is stable. Now the real
experiment begins. At t = 75, we reset the signal permanently to S = 1.2, which
corresponds to the relative change Ma and collaborators used in their paper. Although
the signal stays on at this value throughout the rest of the experiment, the system
adapts, and C approaches a value of 1.013, which corresponds to a deviation of 1.3%
from the original value. Ma and colleagues consider a difference below 2% as suffi-
ciently precise, so that our model is indeed both sensitive and precise. Note that the
internal variable B, which is of no real interest, shows a larger deviation.
A C
1.5
B
9.7 other Signaling Systems
The differential equation models discussed so far are paradigmatic signal transduction 1.0
systems in the narrowest sense. Many other model approaches have been proposed
under the same rubric of signaling systems. For instance, gene regulatory networks are
sometimes considered signaling systems, because the expression of one gene provides 0.5
0 50 100
the signal for other genes to be expressed or repressed, which is mediated through time
transcription factors, repressors, or small RNAs (see [10, 44] and Chapter 6).
A specific example is the genomic regulation of the cell cycle, which we have Figure 9.32 Responses of the three-node
already discussed in the context of Boolean network analysis [5]. For many years, system. Following transient signals at times
Tyson, Novak, and others have been using ordinary differential equations to study t = 5, 30, and 45, the variable of interest, C,
cell cycles and their control. In the case of yeast, these models have become so good shoots up and then returns to homeostasis.
that they capture essentially all alterations in cell cycle dynamics that stem from This response is to be expected, because the
mutants in any of the involved genes [45, 46]. Other approaches have been based on homeostatic steady state is stable. However,
in response to a persistent signal, here
stochastic differential equations [47], recursive models [48], hybrid Petri nets [49],
starting at t = 75, the system adapts, and C
or machine learning methods [50]. The juxtaposition of these diverse models of approaches a new state that is very close
gene regulatory networks demonstrates again that modeling can take many shapes to its homeostatic value. The time trends of
and forms, and that the choice of a particular model depends greatly on the purpose variables A and B are shown, but are not of
of the model and the questions that are to be answered. primary interest here.
Another fascinating class of signaling systems concerns circadian clocks. These

molecular clocks sense and interpret light or other environmental conditions and
allow organisms, ranging from prokaryotes to humans, to be active at appropriate
times during the day–night cycle. The oscillations are essentially self-sustained,
because they continue even if the organism is held in continuous light or continu-
ous darkness. However, they usually run out of phase after a while, indicating that
humans, for instance, have an intrinsic cycle duration that is usually slightly
longer than 24 hours. Because the clocks are based on biochemical events, one
might surmise that they would depend on the environmental temperature, but,
interestingly, this is not the case. An immense body of literature documents the
experimental and computational research that has been done on circadian
rhythms over the past 100 years, since J. S. Szymanski observed that animals
could maintain a 24-hour rhythm without external cues [51]. The relatively new
field of chronobiology focuses on issues of circadian clocks and, for instance,
tries to understand the effects on night shifts and jet lags on human health
[52, 53].
It is surprisingly simple to construct clocks from systems with hysteresis. As
a demonstration, let us return to the system in (9.7) and study the variable Y. Its
hysteresis plot is similar to that of X, although its numerical values are an order
of magnitude smaller. Importantly, we see the same type of jumping behavior
for the same values of S. This is no coincidence, because X and Y are tightly
interconnected.
In the earlier demonstration of hysteresis, we artificially raised S from low values
below 2 to high values above 4 and then lowered it back. This procedure suggests a
thought experiment where some force within or outside the system would do this
raising and lowering repeatedly. If so, we can quite easily convince ourselves that X
and Y should start to oscillate in response.
But we can do even better, namely, by formulating a model. So far, the steady
states of X and Y are functions of the independent variable S. Now, by making S a
dependent variable, we can have its dynamics driven by X or Y so that S becomes
high if X or Y is low, and low if X or Y is high.
To implement this plan, let us express S as a function of Yss, which here is feasible
with a little bit of algebra. For Y to be at a steady state, the second equation of (9.7)
must equal 0. Thus, we obtain
Yss
S= . (9.14)
0.5 X ss0.5
For X to be at a steady state, the first equation of (9.7) must equal 0, which yields
8Yss4
0.5 X ss0.5 = 2 + . (9.15)
16 4 + Yss4
Plugging (9.15) into (9.14), we obtain S as a function of Yss:
 8Y 4 
S = Yss /  2 + 4 ss 4  . (9.16)
 16 + Yss 
This function is shown in Figure 9.33A. Figure 9.33B flips this plot over, so that Yss
depends on S. Note, however, that Yss is not a function of S, because for values
around S = 3, Yss has three values. Nonetheless, this mapping is very useful, because
it shows the relationship between Yss and S. In fact, we can superimpose this plot on
Figure 9.34A and obtain Figure 9.34B. Two critical observations can be made. First,
the low and high branches of the plot in Figure 9.34A are exactly matched by the
mapping between Yss and S. Second, this mapping adds a piece to the curve, con-
necting the jump-off points of Yss as a function of S. This new piece, curving from
(S, Yss) ≈ (3.3, 9.1) upward to (S, Yss) ≈ (2.6, 19.7), exhibits unstable steady states
between the high and low stable steady states. All values on this piece are repelling,
and unless one starts exactly at such a point, the solution will go either to the upper
or to the lower steady state, both of which are stable.
(A) (B) Figure 9.33 Relationship between Yss

S Yss and the signal S. (A) A plot of (9.16), which
4.5 40
results from computing S from the steady-
4.0 35
state equations of (9.7). (B) The same plot
3.5 30 flipped over for S between 2 and 4. Although
3.0
25 S is a function of Yss, Yss is not a function of
2.5
20 S, because, for a value of S around 3, Yss can
2.0
15 have three different values.
1.5
1.0 10
0.5 5
0 0
0 10 20 30 40 2.0 2.5 3.0 3.5 4.0
Yss S
(A) (B) Figure 9.34 Hysteretic behavior of Yss in

Yss S increasing Yss S increasing response to the signal S. (A) The hysteresis
40 40 diagram of (9.7) with respect to Yss, which
S decreasing S decreasing
35 common 35 common is qualitatively similar to the diagram for Xss
Yss (Figure 9.11). (B) Superimposing Figure 9.33B
30 30
on this plot “fills in the blanks” by showing
25 25
the unstable steady states for signals
20 20 between about 2.6 and 3.3.
15 15
10 10
5 5
0 0
2.0 2.5 3.0 3.5 4.0 2.0 2.5 3.0 3.5 4.0
S S
Now let us return to the task of creating a clock. To minimize confusion, we

replace S with the dependent variable Z, which we define as
Z = 5Y −1 − 0.25 Z 0.5 . (9.17)

Other than exchanging Z for S, we retain the system in (9.7) without change.
The exact form of Z is actually not all that critical, but it does affect the shape of the
oscillations. The important feature on which we must insist is that large values of Y
strongly reduce the new signal and that low values strongly increase it. This require-
ment is met by Y’s exponent of −1. Also, Z should not wander off to infinity or zero,
and should instead have the potential for a steady state, which is accomplished
through the degradation term. With these settings, the signal becomes small as soon
as Y becomes big, and vice versa. A small signal causes Y to approach the low steady
state, but as soon as Y becomes small enough, the signal increases, causing Y to
approach the high steady state. As a consequence, X, Y, and Z never actually reach
their steady states and instead oscillate in the manner of a limit cycle (Figure 9.35),
which we discussed in Chapter 4, and which is a good model for a clock. It is easy to
check that the system is indeed a limit cycle by perturbing one of the variables.
For instance, even if X is externally reduced a lot (here, from about 270 to 150 at time
20, the system recovers very quickly (Figure 9.36).
This transition from bistability to limit cycles has been studied in a model of an
embryonic cell cycle oscillator in the frog Xenopus [54] and also with the so-called
repressilator system [55], which is discussed in Chapter 14.
(A) (B)
X, Y Z
320 4
160 2 Figure 9.35 Oscillations resulting from

making the signal dependent on Y. Owing
to the construction where high values of
Y Y lead to low values of Z, and low values of
Y lead to high values of Z, the system never
0 0
0 30 60 0 30 60 reaches a steady state and instead oscillates
time time in a sustained manner.
(A) (B) Figure 9.36 The oscillations resulting

X, Y Y from making the signal dependent on Y
320 30 1
form a limit cycle. If one of the variables
X is perturbed, the system quickly regains its
2
3 limit-cycle behavior. Here, X was reduced
at time t = 20 from about 270 to 150. The
160 15
perturbation leads to a slight phase shift, but
the original oscillation is regained. (A) shows
the oscillation in the time domain, while
Y
(B) shows the corresponding phase-plane
0 0 X
0 30 60 0 160 320 plot of Y versus X. Note that Z is too small to
time be recognizable.
While it is not surprising that clocks in mammals are quite complex, it is intrigu-
ing that it has been possible to create simple circadian rhythms even in vitro, without
transcription or translation [56]. All the self-sustained oscillator requires is a set of
three proteins, called KaiA, KaiB, and KaiC, from the cyanobacterium Synechococ-
cus elongatus, and an energy source in the form of ATP. The simple system is even
temperature-compensated, exhibiting the same oscillation period under tempera-
ture variations. In vivo, the Kai system is of course more intricate, and it has been
demonstrated how very slow, but ordered and fine-tuned, enzymatic phosphoryla-
tion of the Kai proteins allows the cyanobacterium not only to sustain oscillations, T ST
but also to respond appropriately to zeitgebers (timing signals) in the environment
[57]. Specifically, KaiC has two phosphorylation sites and can therefore be present KaiA
in four forms: unphosphorylated (U), phosphorylated only on a serine residue (S),
phosphorylated only on a threonine residue (T), or phosphorylated on both sites
(ST). These different forms cycle in different phases. KaiA enhances the autophos- KaiB
phorylation of KaiC, and KaiC dephosphorylates in the absence of KaiA. KaiB in
turn impedes the activity of KaiA (Figure 9.37).
A conversion of the diagram into a kinetic model allows further analyses of
the oscillator and of various perturbations and mutations [57]. Under baseline U S
conditions, the model oscillates as shown in Figure 9.38. Using first-order kinet-
ics, but with a nonlinear influence of the KaiA concentration, the model of Rust Figure 9.37 Schematic of the Kai
and collaborators [57] reads oscillation system associated with
circadian rhythms. Green arrows indicate
T = kUT (S )U + kDT (S )D − kTU (S )T − kTD (S )T , T (0) = 0.68, activation and red lines with bar show
inhibition. See the text for details. (Data from
D = kTD (S )T + kSD (S )S − kDT (S )D − kDS (S )D, D(0) = 1.36, Rust MJ, Markson JS, Lane WS, et al. Science
318 [2007] 809–812.)
S = kUS (S )U + kDS (S )D − kSU (S )S − kSD (S )S, S (0) = 0.34,
A = max{0,[KaiA ] − 2S },
A
k XY A(S )
k XY (S ) = k XY
0
+ .
K + A(S ) (9.18)
80
U-KaiC
% KaiC
40 total KaiC-P
T-KaiC
S-KaiC
Figure 9.38 Oscillations of the Kai system
D-KaiC
in Figure 9.37. (From Rust MJ, Markson
0 JS, Lane WS, et al. Science 318 [2007]
0 50 100 809–812. With permission from the American
time (h) Association for the Advancement of Science.)
TABLE 9.6: RATE PARAMETER VALUES FOR THE CLOCK SYSTEM (9.18)
UT TD SD US TU DT DS SU
k 0
0 0 0 0 0.21 0 0.31 0.11
kA 0.479077 0.212923 0.505692 0.0532308 0.0798462 0.173 −0.319385 −0.133077
Here, D refers to the double-phosphorylation state ST. Suitable kinetic parameter

values are given in Table 9.6. The total concentration was taken as 3.4, and the
amount of unphosphorylated KaiC was thus U = 3.4 − T − D − S. Furthermore,
[KaiA] = 1.3 and K = 0.43 [57].
Maybe most intriguing is the precision of this relatively simple oscillator in the
actual organism. In spite of asynchronous cell division, the clocks of a cell and its
offspring run with very little deviation over several weeks, even in the absence of
external cues. Many other processes in the organism are driven by the Kai oscilla-
tions, by virtue of rhythmic changes in the supercoiling state of the bacterial DNA
and subsequent global changes in gene expression [58].
As a final example of a widespread signal transduction system, consider cell-to-
cell signaling among microbes, which leads to quorum sensing [59]. This phenom-
enon allows groups of cells or organisms to coordinate their responses, such as
swarming or aggregation, based on the current population density and the emission
of specific signaling molecules. Once emitted, these molecules diffuse and bind to
dedicated receptors on the surfaces of other microbes, which leads to changes in the
expression of target genes, including those responsible for the production of the sig-
naling molecule. Thus, triggering the signal transduction event requires a certain
concentration of the molecule, which is only achieved if a quorum of bacteria is
sending out the signal. Quorum sensing has been observed in bioluminescence and
biofilms and is of great importance and concern for the virulence of pathogens,
such as Pseudomonas aeruginosa and Staphylococcus aureus, some of which have
become multidrug-resistant. Understanding the mechanistic, molecular details of
quorum sensing may point to means of controlling these pathogens and offer new
avenues of synthetic biology, with the goal of engineering cells with rewired
metabolic pathways or other desirable properties [60] (see also Chapter 14).
eXerciSeS
9.1. Formulate rules, corresponding to (9.3), for nodes 9.6. For a 5 × 5 grid with 25 nodes and initially arbitrary
with fewer than eight neighbors. on–off states, is it possible to create rules such that no
9.2. Create a 5 × 5 grid with 25 nodes. Starting with on–off pattern is ever repeated? Either create such
an arbitrary combination of on–off states, we rules or prove that this is not possible.
established a simple rule producing a blinking 9.7. Suppose each node may have one of three states,
dynamics in which every state switched from on to such as red, blue, or green. Construct a blinking
off or from off to on at every transition. Is this rule pattern, in which each state cycles through the three
unique, or can other rule sets produce blinking? Is it colors.
possible for each state to stay on (off) for two time 9.8. Search the literature or the Internet for at least three
units, then switch to off (on) and stay off for two time biological examples that were modeled with
units, and then to switch back and forth every two Boolean methods. Write a brief summary report.
time units? Explain your answer.
9.9. How could one set up a Boolean network model that
9.3. For a 5 × 5 grid with 25 nodes and initially arbitrary remembers the two previous states? Sketch out a
on–off states, create rules such that eventually all plan and implement a small network.
states are zero except for the center state, which
9.10. Review the principles and challenges of Bayesian
should be constantly on. Are the rules unique?
inference. Write a one-page summary report.
9.4. For a 5 × 5 grid with 25 nodes, define rules and
9.11. Review what is known about switches between lysis
initial conditions such that an on column moves
and lysogeny in bacteriophages. Write a one-page
from left to right, then wraps around by starting on
summary report.
the left again.
9.12. Discuss the stability of the steady states of the
9.5. Establish rules that make an on diagonal move in
differential equation
the southwest direction and then start again in the
northeast corner for the next diagonal wave. X = 2 X (4 − X 2 )
referenceS 279
from a mathematical and a biological point of view. 9.25. Study bistability and hysteresis in the MAPK
Discuss what happens if the multiplier 2 is replaced cascade. Can a single module, as shown in
with −2. Figure 9.22, be bistable and/or hysteretic? Is the
9.13. Perform simulations with the bistable system in complete MAPK cascade bistable? Is it hysteretic?
(9.5) to find out in which range(s) the system is most Investigate these questions with simulations.
sensitive to noise. Interpret your findings. Summarize your answers in a report.
9.14. Construct a new “bi-unstable” system with two unstable 9.26. Create a small pathway model with two sequential
states and a stable state in between. Demonstrate the reactions that are each modeled as regular
responses of the system to perturbations similar to Michaelis–Menten rate laws. Compare its
those shown in Table 9.1 and Figure 9.5. responses with those of the double-phosphorylation
system at the center and bottom layers of the
9.15. Demonstrate with simulations that the model in
MAPK cascade, where the two phosphorylation
(9.6) is bistable.
steps compete for the same, limited amount of
9.16. Determine whether the model in (9.4) can exhibit enzyme.
hysteresis.
9.27. Investigate what happens if the MAPK cascade
9.17. Implement the two-variable hysteretic system in receives repeated brief signals. Discuss your
(9.7) and explore how strong a perturbation must be findings.
to move the response incorrectly from off to on.
9.28. Investigate what happens if the phosphatases in the
Study persistent and temporary perturbations.
MAPK cascade have very low activity.
Explain your findings.
9.29. Kholodenko [61] discussed the consequences of
9.18. Create a sequence of signals that are corrupted by
potential negative feedback, in which MAPK
noise. Study its effects on the bistable system in (9.7).
inhibits the activation of MAPKKK. Implement
9.19. Test whether the bifunctional variant of (9.8) also such a feedback mechanism in the model of the
exhibits bistability and hysteresis. MAPK cascade discussed in the text and study
9.20. Igoshin and collaborators [15] emphasized the the implications of different strengths of this
importance of the phosphatase Ph in the bifunctional feedback.
model. Explore bistability and hysteresis if there is 9.30. The MAPK model, as used in the text, assumes that
no phosphatase Ph. For this purpose, use the same XP and YPP are constant. Study the consequences of
model as in the text, but set kb2 = 0.5 and set the making them dependent on their next lower level in
initial value of Ph equal to zero. Repeat the experi- the cascade.
ment with values of Ph between 0 and 0.17. Summa-
9.31. Derive conditions on the parameter values in (9.13)
rize your findings in a report.
under which the system exhibits adaptation.
9.21. Study the responses of the two TCS systems (mono- Assume that the initial steady state is (1, 1, 1), that
and bifunctional) to repeated brief signals. the signal is permanently increased from 1 to 1.2,
9.22. Implement the MAPK cascade shown in Figure 9.26 and that the output variable C approaches the
and study the effects of noise on the signal–response steady-state value of 1.02. Solve the system for the
relationship. steady state and derive the conditions.
9.23. Explore the responses of a MAPK cascade with 9.32. Explore the numerical features and functions in
parameter values that are closer to reality (cf. [32, 34, (9.17) and their effect on the limit-cycle clock.
35]), namely, k1 = k4 = 0.034, k2 = k5 = 7.75, 9.33. Use (9.18) to analyze biologically relevant scenarios.
k3 = k6 = 2.5, k7 = k8 = 1; X(0) = 800, Y(0) = 1000, To get started, consult the original literature.
Z(0) = 10,000, and E = 1; all other initial values are 0. Document your findings in a report.
9.24. Explore the effects of increasing the rate of phos- 9.34. Compute the red portion of the hysteresis curve in
phorylation at the three layers. Use either the model Figure 2 of Box 9.1 by formulating S as a function
in the text or the model in Exercise 9.23. of Xss.
referenceS
[1] Ma’ayan A, Blitzer RD & Iyengar R. Toward predictive models [6] Russell S & Norvig P. Artificial Intelligence: A Modern
of mammalian cells. Annu. Rev. Biophys. Biomol. Struct. Approach, 3rd ed. Prentice Hall, 2010.
34 (2005) 319–349. [7] Shmulevich I, Dougherty ER, Kim S & Zhang W. Probabilistic
[2] Cheresh DE (ed.). Integrins (Methods in Enzymology, Volume Boolean networks: a rule-based uncertainty model for gene
426). Academic Press, 2007. regulatory networks. Bioinformatics 18 (2002) 261–274.
[3] Kauffman SA. Origins of Order: Self-Organization and [8] Shmulevich I, Dougherty ER & Zhang W. From Boolean
Selection in Evolution. Oxford University Press, 1993. to probabilistic Boolean networks as models of genetic
[4] Leclerc R. Survival of the sparsest: robust gene networks are regulatory networks. Proc. IEEE 90 (2002) 1778–1792.
parsimonious. Mol. Syst. Biol. 4 (2008) 213. [9] Chou IC & Voit EO. Recent developments in parameter
[5] Davidich MI & Bornholdt S. Boolean network model predicts estimation and structure identification of biochemical and
cell cycle sequence of fission yeast. PLoS One 3 (2008) e1672. genomic systems. Math. Biosci. 219 (2009) 57–83.
[10] Nakatsui M, Ueda T, Maki Y, et al. Method for inferring and [32] Huang C-YF & Ferrell JEJ. Ultrasensitivity in the mitogen-
extracting reliable genetic interactions from time-series activated protein kinase cascade. Proc. Natl Acad. Sci. USA
profile of gene expression. Math. Biosci. 215 (2008) 105–114. 93 (1996) 10078–10083.
[11] Monod J & Jacob F. General conclusions—teleonomic [33] Bhalla US & Iyengar R. Emergent properties of networks of
mechanisms in cellular metabolism, growth, and biological signaling pathways. Science 283 (1999) 381–387.
differentiation. Cold Spring Harbor Symp. Quant. Biol. [34] Schwacke JH & Voit EO. Concentration-dependent effects on
28 (1961) 389–401. the rapid and efficient activation of MAPK. Proteomics 7 (2007)
[12] Novick A & Weiner M. Enzyme induction as an all-or-none 890–899.
phenomenon. Proc. Natl Acad. Sci. USA 43 (1957) 553–566. [35] Schwacke JH. The Potential for Signal Processing in
[13] Tian T & Burrage K. Stochastic models for regulatory networks Interacting Mitogen Activated Protein Kinase Cascades.
of the genetic toggle switch. Proc. Natl Acad. Sci. USA Doctoral Dissertation, Medical University of South Carolina,
103 (2006) 8372–8377. Charleston, 2004.
[14] Tan C, Marguet P, &You L, Emergent bistability by a growth- [36] Kholodenko BN & Birtwistle MR. Four-dimensional dynamics
modulating positive feedback circuit. Nat. Chem. Biol. of MAPK information processing systems. Wiley Interdiscip.
5 (2009) 942–948. Rev. Syst. Biol. Med. 1 (2009) 28–44.
[15] Igoshin OA, Alves R & Savageau MA. Hysteretic and graded [37] Muñoz-García J & Kholodenko BN. Signalling over a distance:
responses in bacterial two-component signal transduction. gradient patterns and phosphorylation waves within single
Mol. Microbiol. 68 (2008) 1196–1215. cells. Biochem. Soc. Trans. 38 (2010) 1235–1241.
[16] Savageau MA. Design of gene circuitry by natural selection: [38] Whitmarsh AJ & Davis RJ. Structural organization of MAP-
analysis of the lactose catabolic system in Escherichia coli. kinase signaling modules by scaffold proteins in yeast and
Biochem. Soc. Trans. 27 (1999) 264–270. mammals. Trends Biochem. Sci. 23 (1998) 481–485.
[17] Bhattacharya S, Conolly RB, Kaminski NE, et al. A bistable [39] Jordan JD, Landau EM & Iyengar R. Signaling networks: the
switch underlying B cell differentiation and its disruption by origins of cellular multitasking. Cell 103 (2000) 193–200.
the environmental contaminant 2,3,7,8-tetrachlorodibenzo-p- [40] Hanahan D & Weinberg RA. The hallmarks of cancer.
dioxin. Toxicol. Sci. 115 (2010) 51–65. Cell 100 (2000) 57–70.
[18] Zhang Q, Bhattacharya S, Kline DE, et al. Stochastic modeling [41] Kumar N, Afeyan R, Kim HD & Lauffenburger DA.
of B lymphocyte terminal differentiation and its suppression Multipathway model enables prediction of kinase inhibitor
by dioxin. BMC Syst. Biol. 4 (2010) 40. cross-talk effects on migration of Her2-overexpressing
[19] Hasty J, Pradines J, Dolnik M & Collins JJ. Noise-based mammary epithelial cells. Mol. Pharmacol. 73 (2008)
switches and amplifiers for gene expression. Proc. Natl Acad. 1668–1678.
Sci. USA 97 (2000) 2075–2080. [42] Pritchard JR, Cosgrove BD, Hemann MT, et al. Three-kinase
[20] Wu J & Voit EO. Hybrid modeling in biochemical systems inhibitor combination recreates multipathway effects of a
theory by means of functional Petri nets. J. Bioinf. Comp. Biol. geldanamycin analogue on hepatocellular carcinoma cell
7 (2009) 107–134. death. Mol. Cancer Ther. 8 (2009) 2183–2192.
[21] Jiang P & Ninfa AJ. Regulation of autophosphorylation [43] Ma W, Trusina A, El-Samad H, et al. Defining network
of Escherichia coli nitrogen regulator II by the PII signal topologies that can achieve biochemical adaptation.
transduction protein. J. Bacteriol. 181 (1999) 1906–1911. Cell 138 (2009) 760–773.
[22] Utsumi RE. Bacterial Signal Transduction: Networks and Drug [44] Beisel CL & Smolke CD. Design principles for riboswitch
Targets. Landes Bioscience, 2008. function. PLoS Comput. Biol. 5 (2009) e1000363.
[23] Bourret RB & Stock AM. Molecular Information processing: [45] Tyson JJ, Csikasz-Nagy A & Novak B. The dynamics of cell cycle
lessons from bacterial chemotaxis. J. Biol. Chem. 277 (2002) regulation. Bioessays 24 (2002) 1095–1109.
9625–9628. [46] Tyson JJ & Novak B. Temporal organization of the cell cycle.
[24] Zhao R, Collins EJ, Bourret RB & Silversmith RE. Structure and Curr. Biol. 18 (2008) R759–R768.
catalytic mechanism of the E. coli chemotaxis phosphatase [47] Sveiczer A, Tyson JJ & Novak B. A stochastic, molecular
CheZ. Nat. Struct. Biol. 9 (2002) 570–575. model of the fission yeast cell cycle: role of the
[25] Alves R & Savageau MA. Comparative analysis of prototype nucleocytoplasmic ratio in cycle time regulation. Biophys.
two-component systems with either bifunctional or Chem. 92 (2001) 1–15.
monofunctional sensors: differences in molecular structure [48] Vu TT & Vohradsky J. Inference of active transcriptional
and physiological function. Mol. Microbiol. 48 (2003) 25–51. networks by integration of gene expression kinetics modeling
[26] Bongiorni C, Stoessel R & Perego M. Negative regulation and multisource data. Genomics 93 (2009) 426–433.
of Bacillus anthracis sporulation by the Spo0E family of [49] Matsuno H, Doi A, Nagasaki M & Miyano S. Hybrid Petri
phosphatases. J. Bacteriol. 189 (2007) 2637–2645. net representation of gene regulatory network. Pac. Symp.
[27] Batchelor E & Goulian M. Robustness and the cycle of Biocomput. 5 (2000) 341–352.
phosphorylation and dephosphorylation in a two-component [50] To CC & Vohradsky J. Supervised inference of gene-regulatory
system. Proc. Natl Acad. Sci. USA 100 (2003) 691–696. networks. BMC Bioinformatics 9 (2008).
[28] Schoeberl B, Eichler-Jonsson C, Gilles ED & Müller G. [51] Szymanski JS, Aktivität und Ruhe bei Tieren und Menschen.
Computational modeling of the dynamics of the MAP kinase Z. Allg. Physiol. 18 (1919) 509–517.
cascade activated by surface and internalized EGF receptors. [52] Cermakian N & Bolvin DB. The regulation of central and
Nat. Biotechnol. 20 (2002) 370–375. peripheral circadian clocks in humans. Obes. Rev. 10(Suppl. 2)
[29] Pouyssegur J, Volmat V & Lenormand P. Fidelity and spatio- (2009) 25–36.
temporal control in MAP kinase (ERKs) signalling. Biochem. [53] Fuhr L, Abreu M, Pett P & Relógio A. Circadian systems
Pharmacol. 64 (2002) 755–763. biology: when time matters. Comput. Struct. Biotechnol.
[30] Thattai M & van Oudenaarden A. Attenuation of noise J. 13 (2015) 417–426.
in ultrasensitive signaling cascades. Biophys. J. 82 (2002) [54] Ferrell JE Jr, Tsai TYC & Yang Q. Modeling the cell cycle: Why do
2943–2950. certain circuits oscillate? Cell 144 (2011) 874 – 885.
[31] Bluthgen N & Herzel H. How robust are switches in intracellular [55] Elowitz MB & Leibler S. A synthetic oscillatory network of
signaling cascades? J. Theor. Biol. 225 (2003) 293–300. transcriptional regulators. Nature 403 (2000) 335–338.
further reading 281
[56] Nakajima M, Imai K, Ito H, et al. Reconstitution of circadian [59] Fuqua WC, Winans SC & Greenberg EP. Quorum sensing in
oscillation of cyanobacterial KaiC phosphorylation in vitro. bacteria: the LuxR–LuxI family of cell density-responsive
Science 308 (2005) 414–415. transcriptional regulators. J. Bacteriol. 176 (1994) 269–275.
[57] Rust MJ, Markson JS, Lane WS, et al. Ordered phosphorylation [60] Hooshangi S & Bentley WE. From unicellular properties to
governs oscillation of a three-protein circadian clock. Science multicellular behavior: bacteria quorum sensing circuitry and
318 (2007) 809–812. applications. Curr. Opin. Biotechnol. 19 (2008) 550–555.
[58] Woelfle MA, Xu Y, Qin X & Johnson CH. Circadian rhythms of [61] Kholodenko BN. Negative feedback and ultrasensitivity can
superhelical status of DNA in cyanobacteria. Proc. Natl. Acad. bring about oscillations in the mitogen-activated protein
Sci. USA 104 (2007) 18819–18824. kinase cascades. Eur. J. Biochem. 267 (2000) 1583–1588.
further reading
Bhalla US & Iyengar R. Emergent properties of networks of Kholodenko BN, Hancock JF & Kolch W. Signalling ballet in space
biological signaling pathways. Science 283 (1999) 381–387. and time. Nat. Rev. Mol. Cell Biol. 11 (2010) 414–426.
Hasty J, Pradines J, Dolnik M & Collins JJ. Noise-based switches Ma’ayan A, Blitzer RD & Iyengar R. Toward predictive models of
and amplifiers for gene expression. Proc. Natl Acad. mammalian cells. Annu. Rev. Biophys. Biomol. Struct. 34 (2005)
Sci. USA 97 (2000) 2075–2080. 319–349.
Huang C-YF & Ferrell JEJ. Ultrasensitivity in the mitogen-activated Sachs K, Perez O, Pe’er S, et al. Causal protein-signaling networks
protein kinase cascade. Proc. Natl Acad. Sci. USA 93 (1996) derived from multiparameter single-cell data. Science
10078–10083. 308 (2005) 523–529.
Igoshin OA, Alves R & Savageau MA. Hysteretic and graded Schwacke JH & Voit EO. Concentration-dependent effects on the
responses in bacterial two-component signal transduction. rapid and efficient activation of MAPK. Proteomics 7 (2007)
Mol. Microbiol. 68 (2008) 1196–1215. 890–899.
Jordan JD, Landau EM & Iyengar R. Signaling networks: the origins Tyson JJ, Csikasz-Nagy A & Novak B. The dynamics of cell cycle
of cellular multitasking. Cell 103 (2000) 193–200. regulation. Bioessays 24 (2002) 1095–1109.
Kauffman SA. Origins of Order: Self-Organization and Selection in Utsumi RE. Bacterial Signal Transduction: Networks and Drug
Evolution. Oxford University Press, 1993. Targets. Landes Bioscience, 2008.
Population Systems
10
• Identify and discuss standard approaches to studying populations
• Distinguish models for homogenous and stratified populations
• Design and analyze models for age-structured populations
• Develop and analyze models for competing populations
• Understand how to incorporate population growth into larger systems models
Trends in the sizes of populations have fascinated humans for a long time. Hunters
and gatherers were not only interested in the size of their own population, but had a
strongly vested interest in the populations of animals and plants in their surround-
ings, upon which their existence depended [1]. Kings, emperors, and other politi-
cians throughout the ages found population sizes important for forecasting food
needs and for levying taxes. Apparently, the art of quantifying population sizes can
be traced back at least to the Paleolithic period of 30,000 years ago, when early
humans in Central Europe used tally sticks to keep track of changes in populations;
exponential growth was recorded as early as 4000 years ago by the Babylonians
[1–3]. Thus, throughout recorded history, population growth has been measured,
documented, predicted, and analyzed.
This chapter takes a very broad view of populations. While we might immedi-
ately think of the human world population or the population growth in our home
town, populations may also consist of free-living bacteria, viruses, and healthy or
tumor cells. One might even include unconventional populations, such as the
alleles of a gene within a human or animal population [4] or a population of mole-
cules that is converted into a population of different molecules by the action of an
enzyme.
POPULATION GROWTH
The earliest rigorous mathematical descriptions of population sizes and their trends
are often attributed to the British clergyman and economist Thomas Robert Malthus
[5], who formulated the law of exponential growth, and to the Belgian mathematician
Pierre-François Verhulst [6], who proposed the sigmoidal logistic growth function
that we often encounter in microbial populations (see Chapter 4). An enormous
number of other growth functions were subsequently proposed for diverse popula-
tions, ranging from humans, animals, and plants to bacteria, cells, and viruses, and
the literature on growth functions is huge. Most of these are relatively simple nonlin-
ear functions, some discrete, some continuous, while others can be represented as
284 Chapter 10: Population Systems
the solutions of differential equations [1]. One might find it odd that population sizes,
which are clearly discrete integers, are often modeled with differential equations,
which could predict something like 320.2 individuals. The simple rationale for this
strategy is that differential equations are often easier to set up and analyze than the
corresponding discrete models. The situation is in reality even more complicated,
because processes like growth that evolve over a long time period are almost always
affected by stochastic events. Yet differential equations that capture the average
growth behavior are often more useful than fully stochastic models, because they are
so much easier to analyze and understand.
Previous chapters have made it clear that biological systems, even at the subcel-
lular level, are exceedingly complex. Since higher organisms consist of very many
cells, one should therefore expect that the growth of organisms, and of populations
of organisms, is several orders of magnitude more complex. In some sense, this
would be true if we wanted to keep track of every cell or even every biological mol-
ecule in a population of organisms. However, while molecular and intracellular
details are of course crucial, the study of populations focuses on a different level,
where cells, individuals, or organisms are the units of choice, and processes occur-
ring within these units are not considered.
One rationale supporting this simplification is the fact that intracellular pro-
cesses are much, much faster than the growth dynamics of individuals and organ-
isms. In fact, they are so fast that they are always essentially in a steady state, so that
their fast dynamics is not of importance for the much slower dynamics of the growth
and aging of individuals and organisms. As a result, there are often only a few pro-
cesses that are occurring on just the right timescale to affect growth [1]. Further-
more, population studies by their nature consider large numbers of individuals,
which leads to an averaging of individual variability. Thus, in simplified growth
models, all individuals or members of the population are essentially the same, and
one studies their collective behavior.
The logistic growth function, which we discussed in some detail in Chapter 4, is
a great example of these approximation and averaging effects. By describing mil-
lions of cells, aspects of individuality and diversity, which can actually be quite sig-
nificant even in bacteria, are averaged away, and the populations follow simple,
smooth trends (as, for example, in Figure 10.1) that are characterized by just a very
few population parameters such as the growth rate and the carrying capacity (see
below). This approach might seem overly crude, but it is very powerful, as the
analogy with an ideal gas shows: while it is impossible to keep track of every gas
molecule and its interactions with other molecules or the wall of the container,
physicists have developed surprisingly simple gas laws that relate volume, pressure,
and temperature to each other in a very accurate manner.
10.1 Traditional Models of Population Growth

The exponential and logistic growth functions are not the only growth functions. In 50
fact, very many growth laws have been proposed over the past 200 years, either as
size of colony (cm2)
40
explicit functions or as ordinary differential equations (ODEs). One example is
30
the Richards function [7]
20
−1
N R (t ) = K {1 + Q exp[−α ν(t − t 0 )]} , ν
(10.1) 10
0
0 1 2 3 4 5
which is very flexible in shape and has been used widely in plant science, agricul- age (days)
ture, and forestry. Clearly, NR is a function of time, t. N0 = NR(t0) = NR(0) is the popu-
lation size at some time point declared as 0, K is the carrying capacity, α and ν are Figure 10.1 Growth of a bacterial
positive parameters and Q = (K/N0)ν − 1. The various parameters permit shifting colony. The colony, measured by its
and stretching the sigmoid curve in different directions. area on a Petri dish, was assessed over
Another growth function, which is often used in actuarial sciences, is the time (red dots). The logistic function
Gompertz growth law [8] N(t) = 0.2524/(e−2.128t + 0.005125) (blue line;
see Chapter 4) models the data very well.
(Data from Lotka A. Elements of Physical
N G (t ) = K exp[−b exp(−rt )], (10.2) Biology. Williams & Wilkins, 1924.)
POPULATION GROWTH 285
where K and r are again the carrying capacity and the growth rate, respectively, and
b is an additional positive parameter.
It has turned out that it is beneficial for many purposes to represent growth func-
tions as ODEs. For instance, the logistic growth function N(t) = 0.2524/
(e−2.128t + 0.005125) (see Figure 10.1 and Chapter 4) may be written as
r
N = rN − N 2 . (10.3)
K
For the specific example in Figure 10.1, the corresponding parameter values are
r = 2.128, K = 49.25, and N0 = 0.2511.
Mathematically, this conversion is often not too difficult (Exercise 10.2), and
from a modeling point of view, the ODE format facilitates the incorporation of the
growth process into more complicated dynamic systems [9]. Specifically, the ODE
formulation treats the growth process like any other process in a system, and (10.3)
could be interpreted as a small system that describes the balance between one aug-
menting term, rN, and one diminishing term, (r/K)N 2. As a consequence, the growth
process may be modulated by other components of the system. For instance, sup-
pose that bacteria are growing in a human lung during pneumonia. Drug treatment
with penicillin does not kill bacteria, but prevents them from proliferating. Thus,
this effect would be incorporated into the positive growth term of the population
model (10.3). By contrast, the body’s immune system kills bacteria, and a pneumo-
nia model would include this killing in the negative degradation term of (10.3).
Another example with wide applicability involves ODEs that describe in a rather
intuitive fashion how different populations living in the same environment interact
and affect their growth characteristics. The typical approach for assessing this
dynamics is the formulation of logistic functions with added interaction terms. We
will discuss this example later in this chapter.
In many cases, ODE representations of growth processes are also more intuitive
than an explicit growth function, because they can be interpreted in comparison
with exponential growth, for which we have a good intuitive feel. In (10.3), the first
term on the right-hand side represents unabated exponential growth, whereas the
second term is sometimes interpreted as a diminishing crowding effect, which is
proportional to the square of the number of individuals in a population. As an alter-
native, one may formulate (10.3) as
 K −N 
N = r  N. (10.4)
 K 
Mathematically, (10.3) and (10.4) are exactly equivalent, but (10.4) suggests the
interpretation of a growth rate that depends on the population density. This growth
rate is almost equal to r for very small populations, where N is much smaller than K,
and decreases toward 0 if the population reaches the carrying capacity K.
Similarly, the Richards function (10.1) may be written as
 N  ν

N R = α 1 −  R   NR , (10.5)
  K  
which again can be interpreted as a variant of an exponential process with a

population-density-dependent growth rate, which consists of α and the term in
square brackets.
The Gompertz function (10.2) may also be written in ODE form with a density-
dependent rate:
 K 
N G = r ln  NG . (10.6)
 N G 
Alternately, the same Gompertz function may be formulated as a set of two ODEs:
N G = RN G ,
(10.7)
R = −rR.
While this representation may seem unnecessarily complicated, it is actually easier

to interpret than (10.2) and (10.6). In the first equation, the new variable R is easily
identified as the rate of an exponential growth process. The second equation shows
that this growth rate is time-dependent; namely, it represents an exponential decay
of the type R0e−rt (see Chapter 4). Thus, different representations may be mathemati-
cally equivalent but can have different biological interpretations. Furthermore, the
embedding of functions into differential equations can be a tool for comparing and
classifying growth laws [1, 9–11].
10.2 More Complex Growth Phenomena

Although many growth laws have been collected over time, there is still no guaran-
tee that an observed dataset can be accurately modeled by any of them. As a case in
point, it is still being debated what the maximal sustainable size of the human popu-
lation is (Box 10.1). If we knew the growth function of the world population, we
could easily determine the maximal size.
In some cases, the parameters of growth functions are relatively easy to assess.
For instance, exponential growth is characterized by a growth rate parameter, which
often is directly related to features such as the division time of cells (see [12, 13] and
Chapters 2 and 4). However, other population parameters are not directly related to
parameters that can be measured in individuals or even in small, growing popula-
tions. For instance, the bacterial colony in Figure 10.1 approaches a final size of
about 50 cm2, but this value cannot be inferred from studying individual bacteria or
subpopulations, because it combines genetic, physiological, and environmental
factors in some unknown combination. Trees in a planted plot of land grow accord-
ing to a widely acknowledged “3/2 rule” that relates their size to the planting den-
sity. This relationship of a power-law function with negative exponent 3/2 has been
observed uncounted times, but there is no true rationale for the relationship or the
specific number [14, 15].
A different unsolved problem regarding population dynamics is the formula-
tion and analysis of models for so-called metapopulations. These metapopula-
tions sometimes consist of hundreds or thousands of different species of
microbes that at once compete for space and nutrients and rely on each other for
survival. Indeed, metapopulations are the rule rather than the exception,
because we can find them in “environments” from our mouth to soil and sewage
pipes. We will discuss some emerging approaches to studying such metapopula-
tions in Chapter 15.
Typical population studies ignore spatial considerations, even though they may
be very important. Intriguing examples are the growth of tumors, the incremental
spread of introduced pests, such as the red fire ant Solenopsis invicta (see Exercise
10.32), and the spread of a disease throughout a country. A complicating aspect of
human populations is that humans behave in complex and often unpredictable
ways. For instance, imagine the task of predicting the trends in a population of peo-
ple with diabetes under different proposed health-care schemes. A relatively recent
approach to studying such cases, at least via computer simulation, is agent-based
modeling, where individuals are modeled with heuristic rules that guide their deci-
sions and interactions with others [16, 17]. Typically, all individuals have the same
rule set, but it is also possible to account for stratified populations that consist of
different subpopulations. Agent-based models are well suited for spatial phenom-
ena, and the actions of individuals can easily be made to depend on their present
states, their neighbors, and their local environment. While simulations with these
types of models can be very powerful, mathematical analyses are often difficult. We
will discuss this approach in Chapter 15.
POPULATION GROWTH 287
BOX 10.1: TRENDS IN THE GROWTH OF THE HUMAN POPULATION
The human world population has been steadily growing throughout using estimates of current population sizes (Figure 1) and birth rates
recorded history. The big question on many people’s minds is whether (Figure 2). Even with the best data available, the resulting predictions
or when this growth will slow down or even stop. Census bureaus vary tremendously (Figure 3).
around the world try to answer this enormously important question,
Figure 1 Estimated human world

population in 2009. Current population
distributions, combined with country-
specific birth rates, are used to predict
future population trends. (Courtesy
of Roke under the Creative Commons
Attribution–Share Alike 3.0 Unported
license.)
1336 M
400 M
50 M
Figure 2 Estimated world net birth

rates in 2007. Birth rate maps, combined
with actual population distributions, are
used to predict future population trends.
(Courtesy of Homunculus_2 under the
Creative Commons Attribution–Share
no
12.0
14.0
16.0
20.0
22.0
24.0
26.0
28.0
30.0
32.0
>34.0
<–6.0
–4.0
–2.0
0.0
2.0
4.0
6.0
8.0
10.0
18.0
data
per 1000 population per year
15,000 estimated
14,000 UN high
13,000 UN medium
UN low
12,000
actual
11,000
10,000
millions of people
9,000
8,000
7,000
6,000
5,000
4,000
3,000
Figure 3 Recent history and differing forecasts for the growth of
2,000
the human world population. The red, orange, and green curves
1,000 reflect the United Nations’ high, medium, and low estimates. (Courtesy
0 of Loren Cobb Creative Commons Attribution–Share Alike 3.0
1800 1840 1880 1920 1960 2000 2040 2080 Unported license.)
POPULATION DYNAMICS UNDER EXTERNAL 50
PERTURBATIONS
population size
25
The dynamics of a single unperturbed population is usually not very complicated, as
we have seen in the previous section. For instance, it does not really matter all that
much how small the initial population size is: in a realistic model, the population 0
increases toward some final, stable size, which is easily determined through simula- 0 4 8
tion (Figure 10.2). We can also often compute this final size with elementary math- time
ematical methods. If the growth process is formulated as an explicit function of time,

Figure 10.2 Shifted logistic functions.
we consider the growth function for time going to infinity. For example, in the case of
Different initial values shift the logistic
the Gompertz function (10.2), the exponential term exp(−rt) approaches zero, which function (10.3) in the horizontal direction,
causes the term exp[−b exp(−rt)] to approach 1, so that NG(t) approaches K. but eventually all curves reach the same
If the growth process is represented by a differential equation, we know from final value, K. The red curve N(0) = 0.2511
earlier chapters how to proceed: the time derivative is set equal to zero and the corresponds to the process in Figure 10.1.
resulting algebraic equation is solved. For instance, the steady states Nss of the logis- Initial values for the curves, from left to right,
tic function (10.3) are determined by solving the equation are 20 × 0.2511; 5 × 0.2511; 0.2511; 0.2511/5;
0.2511/20.
r 2
0 = rN ss − N ss . (10.8)
K
The problem has two solutions. First, for Nss = 0, the population is “empty” and
remains that way. Second, for the case Nss ≠ 0, we are allowed to divide by r and by
250
Nss, which leads to 1 – Nss/K = 0, which gives Nss = K as the carrying capacity.
Is it possible that the population ever exceeds this value? Sure. It could be that
population size
individuals immigrate from other areas, thus creating a higher value. It might also 125
happen that the conditions in the previous years were very favorable, thereby per-
mitting the population to grow beyond its typical carrying capacity. In contrast to
growth from a small size, which is S-shaped, the decrease in population size looks 0
like an exponential function, which approaches the same steady state, K, as before 0 1 2
(Figure 10.3). time
In nature, many organisms are subjected to predation, and bacteria are no

Figure 10.3 Logistic functions above their
exception. For instance, some ameba and other protozoans eat bacteria. As Richter
carrying capacity. If the logistic function is
[18] pointed out in the context of whale hunting, the dynamics of a species depends initialized above K, it decreases toward K in
critically on the type of predation. In many models, predation is assumed to be pro- a monotone fashion that does not resemble
portional to the current number of individuals, while it occurs with a constant rate the S-shaped growth from low initial values.
in another typical case. It is easy to explore the consequences. If predation is pro- Initial values, from bottom to top, are 75;
portional to population size, we formulate the dynamics as 1.2 × 75; 1.5 × 75; 2 × 75; 3 × 75.
r
N 1 = rN 1 − N 12 − pN 1 , (10.9)
K
where we have added the index 1 merely for easy discussion and comparisons. In
the case of constant predation, the formulation is
r
N 2 = rN 2 − N 22 − P. (10.10)
K
Let us suppose that the common parameters in the two models are the same as
before and that both populations start with an initial size of 50. When we more or less
arbitrarily set the first predation parameter as p = 0.75 and select P = 25, the two pop-
ulations have very similar dynamics, including roughly the same steady state (Figure
10.4A), and one might think that the type of predation does not really matter. How-
ever, the two populations exhibit distinctly different responses to perturbations. For
example, if the two populations are at some point reduced to 15, maybe owing to
other predators, disease, or some environmental stress, the first population recovers
quite quickly, while the second population never recovers, and eventually dies out
(Figure 10.4B). Stability analysis allows the computation of the perturbation thresh-
old above which the second population recovers and below which it goes extinct.
In typical models of population growth, all parameters are constant, and the
dynamics is therefore fully determined. In reality, the growth rate may depend on
ANALYSIS Of SUBPOPULATIONS 289
(A) (B)
50
N2
N1
N1
population size
25
N2
0
0 2.5 5.0 0 2.5 5.0
time time
Figure 10.4 Comparison of two logistic growth processes exposed to different types of predation. Without perturbations (A), the two functions
(10.9) and (10.10) are quite similar. However, if some external factor reduces both populations to a sufficiently small value, the population with proportional
predation (N1) recovers, while the population exposed to constant predation (N2) becomes extinct. In (B), both populations are reduced to 15 at time 2.5.
environmental factors, such as temperature, which of course is a function of time, or

it may depend on the current population size [18]. Furthermore, the parameters
usually fluctuate stochastically within some range. For very large populations, such
variations are not problematic, but they can become detrimental if a population is
very small.
ANALYSIS Of SUBPOPULATIONS
While the growth of a homogeneous population is interesting in many respects,
the range of likely dynamics is somewhat limited. More interesting are the dynamic
behaviors of stratified populations containing different categories of individuals
and the interactions between populations. An example of the dynamics of subpop-
ulations was rather extensively discussed in Chapter 2, where models described
how susceptible individuals become infected, recover, and possibly become sus-
ceptible again. Uncounted variations of these SIR models have appeared in the
mathematical modeling literature, and we will not discuss them here again. A vari-
ation of such subpopulation models accounts for the different ages of individuals
that are exposed to a disease vector. Specific questions in such a study ask whether
a disease outbreak can be controlled and at what age it is most efficacious to vac-
cinate individuals. If many age classes, along with different health and treatment
states, are considered, these models can consist of hundreds of differential equa-
tions [19]. Interestingly, the dynamics of the actual disease-causing agents, such as
viruses, is often not explicitly accounted for in these models.
An interesting example of a simple, yet fairly realistic, growth model with age
classes was proposed by Leslie [20] many decades ago. It is framed as a linear recur-
sive model of the type we discussed in Chapter 4. The population is categorized
into subpopulations of individuals with different ages and different proliferation
rates. All individuals have the same potential lifespan, but they may of course die
earlier. For humans with a maximum age of 120, one could select m = 12 age classes
of τ = 10 years each and set a corresponding recording interval of τ years. In other
words, population sizes are exclusively documented only every τ years, but not in
between.
The Leslie model is most easily formulated as a matrix equation, which for four
age classes has the form
 P1   α1 α2 α3 α 4   P1 
 P2   σ1 0 0 0   P2 
  =   . (10.11)
 P3  0 σ2 0 0   P3 
P   σ3 0   P4  t
 4  t +τ  0 0
Let us dissect this equation. The vector on the left contains P1, . . ., P4 at time t + τ.
These quantities are the sizes of the subpopulations in age classes 1, 2, 3, 4, respec-
tively. Almost the same vector appears on the far right, except that it captures the
population sizes at time t. Thus, the recursive character of the equation is the same
as in the simple case of exponential growth (Chapter 4), and writing (10.11) in
vector and matrix notation yields
Pt +τ = LPt . (10.12)
Connecting the two vectors is the Leslie matrix L. It consists mainly of zeros, except
for quantities α in the first row and quantities σ in the subdiagonal. The α ’s repre-
sent the reproduction rate for each age class, given as the number of offspring pro-
duced per individual in this age class within the time period τ. The parameters σ
represent the fractions of individuals surviving from one age class into the next and
have values between 0 and 1. Writing out the matrix equation shows these features
more intuitively. For instance, the equation for P2 is obtained by multiplying the
elements in the second row of the matrix with the vector on the right-hand side,
which yields
P2 ,t +τ = σ 1P1,t + 0 ⋅ P2 ,t + 0 ⋅ P3 ,t + 0 ⋅ P4 ,t = σ 1P 1,t . (10.13)
Expressed in words, age class 2 at time t + τ contains exactly those individuals that
were in age class 1 at time t and survived time period τ. Age classes 3 and 4 have a
similar structure. (Please write down the equations to convince yourself.) Now let us
look at age class 1. There is no class 0, so that there cannot be survival σ0 into class 1.
Instead, all individuals in class 1 come exclusively from reproduction, and the rate
of this process differs for the various age classes. For instance, individuals in class 3
produce offspring at a rate of α3 per time period τ. These birth processes are reflected
in the first equation, describing P1, t + τ , which is
P1,t +τ = α 1P1,t + α 2 P2 ,t + α 3 P3 ,t + α 4 P4 ,t . (10.14)
How does this equation account for the fact that only a certain percentage (quanti-
fied by σ1) of individuals of class 1 survive? Only indirectly. First, according to the
set-up of the model, no individual can stay in class 1 (although this class can have
offspring that again would be in the same class). Furthermore, those surviving (σ1P1,t
individuals) show up in class 2, and those dying (namely, (1 − σ1)P1,t individuals)
are no longer explicitly accounted for; they disappear from the population and from
the model.
Because the process is so rigidly structured, we can write the Leslie equation for
any numbers of age classes:
 α1 α2 … α m−1 αm 
 P1  σ1  P1 
 P2  0 … 0 0  P2 
 
  = 0 σ2 … 0 0   . (10.15)
    
P   P 
 m  t +τ   mt
0 0 … σ m−1 0 
This recursion again moves forward by τ time units at a time. It is also possible to
jump nτ steps at once. Namely, we simply use the nth power of the matrix:
Pt +nτ = Ln Pt . (10.16)
Leslie models of this type are interesting tools for exploring trends in subpopula-
tion, even though they make many explicit and implicit assumptions (for examples,
ANALYSIS Of SUBPOPULATIONS 291
see Box 10.2). For instance, there is no mention of population sizes at time points
inside the intervals, that is, between t + kτ and t + (k + 1)τ for k = 0, 1, 2, . . .. For
simplicity, males are usually not included, because they do not bear children.
Importantly, the growth process is independent of the current population size. In
other words, the parameters α and σ do not depend on time or on the population
size—which is at odds with many real populations. As a consequence, the popula-
tion either always keeps growing or always keeps decreasing, but it almost never
reaches a stable size that is not zero (for an exception, see Exercise 10.11). Of course,
it is mathematically possible to make α and σ dependent on time or population size,
but then the model is no longer linear, and certain computations become mathe-
matically much more difficult. For pure simulation purposes, these extensions are
easy to implement and explore.
The Leslie model uses average survival and reproduction rates and therefore
is deterministic. As a consequence, every simulation with the same parameters
and initial sizes of subpopulations yields exactly the same results. In a variation
that is conceptually simple but not trivial in its implementation and analysis, one
could consider randomness in survival and reproduction [12, 13]. For instance,
BOX 10.2: TYPES OF AGE STRUCTURES WITHIN POPULATIONS
In the Leslie model, the birth rates within a population typically Geographic (January 2011) contrasted typical population pyramids
differ by the age of the mother but they are constant over time resulting from trends in birth and death rates. Representative
within each age class. Similarly, death rates are usually higher for pyramids are shown in Figure 1. China’s actual population
older individuals, but remain constant over time. In reality, neither structure is shown in Figure 2. It is a mixture of the pyramids in
birth nor death rates are constant. As an illustration, National Figure 1.
Figure 1 Types of population

80 years
pyramids. Depending on the balance
60 years between birth and death rates, which
40 years changes over time in a population, the
age distribution within a population
20 years
assumes a different shape.
Wide base, narrow top: Tapered pyramid: In many Western Aging population
representative of high representative of high societies the birth where death rate
birth and death rates. birth rate and rate falls and life exceeds birth rate.
lowered death rate. expectancy rises.
(years)
males females
age
91–95
86–90
81–85
76–80
71–75
66–70
61–65
56–60
51–55
46–50
41–45
36–40
31–35
26–30
21–25
16–20 Figure 2 China’s 2009 population,
11–15 stratified by 5-year age groups and
6–10
gender. The effects of China’s one-
0–5
child policy, introduced in 1978, are
70 60 50 40 30 20 10 0 10 20 30 40 50 60 70 visible in this population pyramid of
population (millions) 2009.
one might imagine cells that are moving through the cell cycle and enter its dif-
ferent phases in a probabilistic fashion. Typical quantities to be assessed would
include the numbers of cells in each phase and the distribution of transition
times throughout the cell cycle. This seemingly harmless change from determin-
istic to random events moves the model into the realm of stochastic processes,
which are much more complicated than the deterministic Leslie model. Good
texts include [21–23].
INTERACTING POPULATIONS
10.3 General Modeling Strategy
In the simplest case of an interaction system, the individuals of two populations do
not influence each other much, except that they might compete for the same
resources. The simplest example may be a mixture of two exponentially growing
bacterial populations with different growth rates. A more interesting example is a
tumor growing within the tissue where it originated. Often the growth rate of tumor
cells is not necessarily faster than that of healthy cells, but tumor cells undergo cell
death at a much lower rate. We can easily model this situation with a system of
healthy (H) and tumor (T) cells of the form
H = rH H − mH H 2 ,
(10.17)
T = rT T − mT T 2 .
The parameters rH, rT, mH, and mT represent growth and mortality rates of healthy
and tumor cells, respectively. Suppose both growth rates are the same (rH = rT = 1),
the mortality rates are mH = 0.01 and mT = 0.0075, and initially only 0.1% of the cells
are tumor cells (H(0) = 1, T(0) = 0.001). With these settings, it is easy to simulate the
system. Even though the growth rates are the same and the number of tumor cells is
at first miniscule, the lowered mortality allows the tumor eventually to take over
(Figure 10.5).
More realistic than the coexistence in this example is the situation where two
or more populations truly interact with each other. The most straightforward case
is competition, which reduces the carrying capacities of the populations. The
usual approach is a model of logistic growth processes, which are expanded with
interaction terms. If the interactions are very strong, one can imagine that not
every population can survive. The typical questions asked in these situations are
therefore: Who will survive? Is coexistence possible? How long does it take to
reach equilibrium? What do the transients from initiation to termination look
like? Interestingly, some of these questions can be answered with a high degree
of generality if only two populations are involved. The methods for this task are
discussed next.
200
10.4 Phase-Plane Analysis
T
population size
Every time we run a simulation study, we have to ask ourselves how much the
results depend on the particular parameter values we had chosen. If we changed 100
H
the values in some way, would the results be fundamentally different? For instance,
if population A survives in a simulated scenario, is that a general outcome or is it
merely the result of our current choice of initial numbers, growth rates, and other 0
characteristics? For complicated models, these types of questions can be difficult 0 10 20
to answer. However, if we are dealing with just two populations, qualitative phase- months
plane analysis is a very effective tool. In principle, qualitative analysis could also
be done for systems with more species, but a key contributor to our intuition would Figure 10.5 Dynamics of two coexisting
populations. In the long run, a higher
no longer be of help, namely a two-dimensional visualization. So, let us suppose we
growth rate or lower mortality rate is more
have two interacting populations whose dynamics is described with two ODEs. influential than a larger initial population
These equations consist of logistic growth functions of the type in (10.3) and (10.4), size, as demonstrated here with a simplified
which furthermore contain one interaction term each. These interaction terms are model of healthy (H) and tumor (T) cells
typically of the same form but with different parameter values, because the two growing in the same tissue (see (10.17)).
INTERACTING POPULATIONS 293
species respond differently to the effect of the interaction. Specifically, we use the 50 N1
following pair of equations as an example:
population size
N2
25
 K − N 1 − aN 2 
N 1 = r1N 1  1  ,
 K1
(10.18) 0
 K − N 2 − bN 1  0 50 100
N 2 = r2 N 2  2  .
 K2 time (days)
Figure 10.6 Two populations, N1 and N2,

In order to facilitate both mathematical and computational analyses, we select the
are competing for the same resources.
following, more or less arbitrary, parameter values: r1 = 0.15, K1 = 50, a = 0.2, N2 grows faster, but competition from N1
r2 = 0.3, K2 = 60, b = 0.6, N1(0) = 1, N2(0) = 1. eventually leads to a decrease in population
It is easy to solve this system numerically for N1 and N2 as functions of time; a size. Neither population reaches its carrying
plot of the solution is shown in Figure 10.6. Some details are probably expected, capacity.
others may be surprising. First, the dynamics of N1 is similar to that of a logistic func-
tion that is not affected by outside forces. N2 grows faster, owing to its larger growth
rate, but begins to fade as population N1 grows. Owing to the competition, neither
population reaches its carrying capacity. In particular, N2 only reaches a final size of
about 34, while its carrying capacity is 60.
The key concept of the following type of qualitative analysis is that we do not take
explicit account of time and instead plot the behavior of one population against the
behavior of the other. Thus, we look at a coordinate system where the axes represent 50 t = 30
N1 and N2 (Figure 10.7). If we kept a record of N1 and N2 over time, the dataset would t = 40
t = 20
have the form shown in Table 10.1. We can easily see from Figure 10.7 that time is
t = 50
represented in a drastically different way than before, namely merely as labels along
a curve that describes how N2 depends on N1. In fact, we could double the speed of t = 60
population size N2
the system, by multiplying all rate constants in (10.18) by 2, and the plot would look
25
exactly the same, except that the time labels would be different. Expressed differ-
ently, time has moved to the background, and the focus is on the relationship
between N1 and N2. Even though time no longer appears explicitly, this phase plot t = 10
indicates the speed of the processes indirectly through the distance between con-
secutive time points. In Figure 10.7, the speed is greatest between t = 10 and t = 20,
and the process almost comes to a halt beyond t = 60. 0
t=0
0 25 50
Intriguingly, we can analyze this relationship generically with minimal informa-
population size N1
tion about the parameter values of the system. We begin with the steady states of the
system. These are determined by setting both time derivatives in (10.18) equal to
Figure 10.7 Phase-plane plot of two
zero and solving for N2 as a function of N1 (or the other way around). The specific interacting populations. N2 is plotted
form of the equations makes this very easy. First, we can convince ourselves rather against N1 for the 11 time points in
easily that (N1, N2) = (0, 0) is a steady state. This makes sense, because if no indi- Table 10.1. The population sizes between
viduals exist, there is no dynamics. This case is important to remember, but let us times t = 70 and t = 100, computed from
assume for the following that N1 and N2 are strictly positive. If we divide the two (10.18), cluster close to those at t = 60.
TABLE 10.1: SIZES OF TWO COMPETING POPULATIONS, N1 AND N2, OVER TIME
Time N1 N2
0 1 1
10 4.06 14.40
20 11.96 46.38
30 23.58 47.96
40 33.76 42.27
50 39.34 38.04
60 41.70 35.81
70 42.61 34.80
80 42.96 34.37
90 43.10 34.20
100 43.15 34.13
equations by N1 and N2, respectively, which we are allowed to do because we assume

that these variables are not zero, we obtain two simpler equations in two unknowns,
which we can solve directly:
 K − N 1 − aN 2  K1 − N1
0 = r1N 1  1  ⇒ N 2 = ,
 K1 a
(10.19)
 K − N 2 − bN 1 
0 = r2 N 2  2  ⇒ N 2 = K 2 − bN 1 .
 K2
These equations are valid for any parameter values. Using the numerical values
from before, the nonzero steady-state solution is N1 = 43.18, N2 = 34.09. The system
contains two further steady states, where either N1 or N2 is zero and the other vari-
able is not.
In addition to the steady states, we can explore situations where only one of the
time derivatives is zero. In fact, these situations are modeled by the two equations in
(10.19) when they are considered separately. In the first case, the equation given by
N 1 = 0 is called the N1-nullcline, where “null” refers to zero and “cline” comes from
the Greek root for slope. In general, this function is nonlinear, but the N1-nullcline
in our example is represented by a linear function, namely N2 = (K1 − N1)/a. It has
a negative slope, and intersects the vertical axis at K1/a, and the horizontal axis at
K1. Similarly, the N2-nullcline is determined from setting the time derivative of N2
equal to zero, and the result is N2 = K2 − bN1, which is again a straight line with
negative slope. It intersects the vertical axis at K2, and the horizontal axis at K2/b.
The two nullclines intersect at the nonzero steady state, because at this point both
time derivatives are zero. Figure 10.8, which is intentionally not drawn to scale to
indicate the generic nature of the plot, shows the two nullclines and the two steady-
state points.
Even without specifying parameter values, we can find out more about the system
by simply looking at the phase plot with two nullclines. For instance, for every point
exactly on the N1-nullcline, the derivative of N1 is by definition zero. Therefore, any
trajectory of the ODE system must cross this nullcline vertically, because the value of
N1 does not change. Furthermore, N 1 is less than zero on the right side of the nullcline
and greater than zero on the left side. We can confirm this by moving any point on the
nullcline a tiny bit to the right: the value of N1 increases very slightly, which makes the
right-hand side of the first differential equation slightly negative. Thus, all derivatives
N 1 on the right of the N1-nullcline are negative, and all those on the left are positive.
Analogous arguments hold for the N2-nullcline, where all trajectories must cross hori-
zontally. Taking these pieces of information together, the plot thus consists of four
areas that are characterized by the signs of the two derivatives (see Figure 10.8). Every
K1/a •
N1 > 0
•
N2 < 0
•
N1 < 0
•
N2 < 0
population size N2
•
N1 = 0
K2
•
•
N2 = 0 N1 < 0
•
•
N2 > 0 Figure 10.8 Phase-plane plot for two
N1 > 0
• competing populations N1 and N2. The
N2 > 0
nullclines (red and blue lines) intersect at the
0 nonzero steady state (yellow circle). At other
0 K1 K2/b
steady states (green circles), N1, N2, or both
population size N1
populations are zero.
Figure 10.9 Coarse visualization of the

K1/a
dynamics of two competing populations.
Sketching “generic” trajectories in the phase-
plane plot shows that the two competing
populations N1 and N2 approach the nonzero
steady state (yellow circle), where both
survive and coexist. Trivial mathematical
population size N2
exceptions are cases where N1, N2, or both

are zero at the beginning.
K2
0
0 K1 K2/b
population size N1
trajectory that satisfies the pair of differential equations (10.18) must move according
to these derivatives. This allows us to sketch all possible trajectories on the phase
plane. We do not know exactly what these trajectories look like, but we do know their
most significant features, namely, the direction in which they move and whether they
cross a nullcline in a vertical or horizontal direction (Figure 10.9). Two special cases
arise if one of the two variables is zero (see Exercises 10.19 and 10.20). Finally, the
phase plot indicates that all trajectories move away from (0, 0): this trivial steady state
is unstable.
In the example we have just discussed, the two nullclines intersect at a single
point, the nontrivial steady state. It may also happen that they do not intersect in the
positive quadrant, which is the only area of interest. For instance, suppose K1 > K2/b
and K1/a > K2. The situation is shown in Figure 10.10. Now there are only three
areas, with different combinations of signs of the two derivatives. Exactly the same
arguments hold as before, but now all trajectories converge to N1 = K1, N2 = 0, with
the exception of trajectories that start exactly on the vertical axis (that is, N1(0) = 0).
The interpretation is that, unless there are no individuals in population N1 at all, N1
will eventually survive and grow to its carrying capacity, while N2 will become
extinct. Examples are invasive species, which start with a few organisms but, if
unchecked, lead to the extinction of the corresponding native species. The slopes
and intersections of the two nullclines permit two additional situations, which are
left for Exercises 10.21 and 10.22.
K1/a
population size N2
K2
Figure 10.10 Depending on the

parameter values, the two populations
cannot coexist and one will eventually go
extinct. If the parameter values are such that
the nullclines do not intersect in the positive
0 quadrant, one of the two populations (here
0 K2/b K1
N1) survives at its carrying capacity, whereas
population size N1
the other (here N2) disappears.
Arguably the most famous class of examples where this type of analysis has been
used concerns predator–prey systems. These systems were first discussed with
mathematical rigor almost one hundred years ago by Alfred Lotka, a physical chem-
ist and biologist, and Vito Volterra, a mathematician. Lotka used this approach
originally for the analysis of chemical reactions, but it is better remembered for
population studies in ecology (see also Chapter 4). In its simplest form, the Lotka–
Volterra (LV) model describes the interactions between one prey species N and a
predatory species P. The equations are nonlinear and quite similar to those we
discussed above:
N = αN − βNP ,
(10.20)
P = γ NP − δP.
All parameters are positive. In the absence of predators, N grows exponentially with
growth rate α, whereas the presence of predators limits this growth through the
term βNP. The form of a product of the two variables with a rate constant is based on
arguments similar to those discussed in the context of chemical mass action kinet-
ics (Chapter 8), where the product is used to describe the probability of two (chemi-
cal) species encountering each other in a three-dimensional space. The predator
population P is assumed to rely exclusively on N for food; without N, the predator
population decreases exponentially toward extinction. Again, the growth term is
formulated as a product, but the rate γ is usually different from the rate β in the prey
equation, because the magnitude of the effect of the same process (predation) is
significantly different for prey and predators.
With typical parameter values (α = 1.2, β = 0.2, γ = 0.05, δ = 0.3, N0 = 10, P0 = 1), 50
N and P exhibit ongoing oscillations, in which P follows N with some delay and typi- N
population size
cally with a much smaller magnitude (Figure 10.11).
Linearization and stability analysis demonstrates that (0, 0) is a saddle point and 25
that the interior, nontrivial steady state is a center (see Chapter 4). Moreover, one
P
can show that the general solution outside (0, 0) consists of closed orbits around the
interior steady-state point, which correspond to sustained oscillations with the 0
0 30 60
same amplitude [24, 25]. However, these oscillations are not limit cycles, and every time
perturbation, even if it is very small, moves the system to a different orbit or causes
a species to die. Figure 10.11 Idealized dynamics of
Many important features of a phase-plane analysis remain the same if the predator (P) and prey (N) populations.
nullclines are nonlinear. In particular, the slope of a variable right on its nullcline is The dynamics of the Lotka–Volterra system
zero, so that trajectories again transect nullclines either horizontally or vertically. (10.20) consists of ongoing oscillations.
In reality, such oscillations are only seen
Also, the intersections of nullclines are steady-state points. The main difference is
when the two species coexist in a complex
that the nullclines may intersect more than once. Thus, the system may have two environment.
interior intersection points, where neither of the variables is zero, one of which is
typically stable whereas the other is unstable. Of course, there could be more than
two intersections, depending on the model structure.
As an example, let us consider the system
N 1 = N 1 (10 + 1.5 N 1 − 0.1 N 12 − N 2 )/150,

(10.21)
N 2 = N 2 (50 + 0.3 N 12 − 7 N 1 − N 2 ) /150,
where each population is augmented and reduced through two processes each. The
nullclines are easily computed by setting the first or the second equation equal to
zero and expressing N2 as a function of N1. They are shown in Figure 10.12A. As
before, the axes are additional nullclines, and the system allows for the extinction of
either N1 or N2. Also, (0, 0) is a steady-state point, and we will see that it is unstable.
To explore the system behavior within the different areas created by the nullclines,
it is useful to compute the vector field of the system. This representation consists of
solving the equations many times for a very short time period, each time starting
with initial values that are taken from a grid of relevant (N1, N2) pairs. Such a repre-
sentation is shown in Figure 10.12B. Relatively long lines correspond to fast changes,
whereas lines that almost look like dots represent very slow changes; rationalize
why this is so. True dots are steady-state points. A sample of trajectories is shown in
Figure 10.12C; the arrows have been added to indicate the direction of each
(A) N2 (B) N2 Figure 10.12 Phase-plane analysis of

50 50 the system in (10.21). In contrast to the
earlier systems, the nullclines in this case
are nonlinear, thereby allowing two interior
steady-state points. (A) shows the nullclines
(including the axes) and the steady-state
points (yellow). (B) depicts the vector field
of the system, which summarizes the flow
25 25 of all trajectories. A sample of trajectories
is shown in (C), together with a separatrix
(dashed), which separates trajectories
ending either at (0, 50) or at (14.215, 11.115).
(D) superimposes (A), (B), and (C) for a
complete qualitative picture. Note that all
trajectories transect the nullclines either in
0 N1 0 N1
0 5 10 15 20 0 10 20 horizontal or vertical direction. The steady
state at (7.035, 15.60) is a saddle point.
Moving exactly on the separatrix toward
(C) N2 (D) N2 this point would end in this point. However,
50 50 deviating even slightly to the left or right of
the separatrix pushes any trajectory further
away.
25 25
0 N1 0 N1
0 5 10 15 20 0 10 20
trajectory. The dashed line is called the separatrix, which is Latin for “she who sepa-
rates.” Why this line is allegedly female may be pondered, but is not important here.
What is important is that this line divides the phase plane into sections with differ-
ent behaviors: Starting to the left of the separatrix, the system will eventually
approach the steady state where N1 becomes extinct and N2 is 50. Starting to the
right, the system will spiral into the coexistence state (N1, N2) ≈ (14.215, 11.115). Not
surprisingly, the separatrix runs through the saddle-point steady state at (7.035,
15.60) and into the unstable point (0, 0). Of note is that every perturbation away
from (0, 0), no matter how slight, can lead to the extinction of either N1 or of N2 or
to coexistence. This observation highlights the existential risks faced by small
populations.
10.5 More Complex Models of Population Dynamics

Obviously, all these models are very much simplified and sometimes distant from
reality. For instance, predators very rarely rely on a single prey species and are
instead participants in complex food webs where many predator species feed on
many prey species. Other complications in population modeling result from com-
plex dependences, such as the pollination of plants by specific insect species.
Another issue of great import is the dynamics for small values. In the differential
equation system, even 0.1 individuals of prey may recover and later provide food for
the predators. In reality, the species would die out. Furthermore, all systems in
nature are subject to numerous environmental influences, some of which can
greatly stabilize or destabilize the system. For instance, refuges for prey can drasti-
cally change the simple dynamics exhibited in Figure 10.11. Finally, as soon as the
spatial expansion of a population becomes important, as is prominent in the spread
of epidemics, one has to resort to entirely different models, such as partial differen-
tial systems or agent-based models (Chapter 15).
TABLE 10.2: PARAMETER VALUES FOR THE LOTKA–VOLTERRA SYSTEM

(10.22) PLOTTED IN FIGURE 10.12
a1 1 b11 –1 b21 0 b31 –2.33 b41 –1.21
a2 0.72 b12 –1.09 b22 –1 b32 0 b42 –0.51
a3 1.53 b13 –1.52 b23 –0.44 b33 –1 b43 –0.35
a4 1.27 b14 0 b24 –1.36 b34 –0.47 b44 –1
In spite of the numerous immediate and more indirect caveats, LV models have
found much application in ecology and many other areas. In higher dimensions, a
typical formulation for n variables is
 n

X i = ai X i ×  1 +
 ∑ b X  .
j =1
ij j (10.22)
As in the lower-dimensional case of a predator–prey system, each right-hand side of

this system of ODEs consists of a linear term and a sum of binary (second-degree)
terms, which describe interactions between two species (i ≠ j) or the crowding term
within a species (i = j) that we encountered several times in logistic growth functions
and competition models. Because of the linear structure inside the parentheses, this
model allows steady-state analyses and parameter estimation for arbitrarily large
systems (Chapters 4 and 5). It is even possible to include environmental factors in
these types of models [26].
Higher-dimensional LV systems may have very interesting mathematical fea-
tures. As a case in point, they can exhibit deterministic chaos. (see Chapter 4). As a
reminder, this term indicates that the system is entirely deterministic and no ran-
domness is included in the model structure. At the same time, the system oscillates
in complicated patterns that are not predictable without the model and never repeat
themselves. As a numerical example, consider a four-variable LV system of the
type (10.22) with the parameters in Table 10.2, which was proposed by Sprott
and colleagues [27, 28] and exhibits extremely complicated, chaotic dynamics
(Figure 10.13); all four initial values were chosen as 1.
1.0
X1 X2
X3 X4
X1, X2, X3, X4
0.5
Figure 10.13 Even small dynamic systems

may exhibit deterministic chaos. The
form of a Lotka–Volterra model (10.21) is
deceptively simple. In reality, such a system
may exhibit very complicated dynamics,
such as the chaotic behavior of the four-
variable system shown here. (Adapted from
Sprott JC, Vano JA, Wildenberg JC, et al. Phys.
Lett. A 335 [2005] 207–212, with permission
from Elsevier, and from Vano JA, Wildenberg
0 JC, Anderson MB, et al. Nonlinearity 19
0 150 300 [2006] 2391–2404, with permission from IOP
time Publishing.)
EXERCISES 299
While higher numbers of variables often make systems more complicated, other
factors can contribute to the complexity of a model. For instance, some or all param-
eters could be made time-dependent. As so often, there is hardly any limit to the
range of possibilities. A rather recent topic of strong interest is the characterization
and analysis of metapopulations, consisting of hundreds or thousands of microbial
species that live in the same space and simultaneously compete for resources and
support each other through an intricate division of labor. We will discuss these com-
plicated population systems in Chapter 15.
EXERCISES
10.1. Fit the data from Table 10.3 (taken from Figure 10.1)
TABLE 10.4: GROWTH OF THE HUMAN WORLD
with Richards and Gompertz functions, as well as with POPULATION
the generalized growth equation proposed by
Savageau [17], which has the form Year World population P log P
1000 25,400,000 7.404833717
N S = αN Sg − βN Sh ,
1100 30,100,000 7.478566496
with positive parameter values. Use fitting methods 1200 36,000,000 7.556302501
discussed in Chapter 5. 1300 36,000,000 7.556302501
1400 35,000,000 7.544068044
TABLE 10.3: GROWTH OF A BACTERIAL POPULATION* 1500 42,500,000 7.62838893

1600 54,500,000 7.736396502
Age of colony (days) Colony size (area in cm2)
1700 60,000,000 7.77815125
0 0.24
1800 81,300,000 7.910090546
1 2.78
1900 155,000,000 8.190331698
2 13.53
1910 175,000,000 8.243038049
3 36.3
1920 186,000,000 8.269512944
4 47.5
1930 207,000,000 8.315970345
5 49.4
1940 230,000,000 8.361727836
*Data from Lotka A. Elements of Physical Biology. Williams & Wilkins, 1950 240,000,000 8.380211242
1924.
1960 3,042,389,609 9.483214829
1970 3,712,813,618 9.569703148
10.2. Derive the ODE format of the logistic growth law 1980 4,452,686,744 9.648622143
(10.3) from the explicit form in Figure 10.1, 1990 5,288,828,246 9.723359464
including the appropriate initial value. Compute
2000 6,088,683,554 9.784523403
the derivative of the explicit function and express
it in terms of N and suitable parameters. 2010 6,853,019,414 9.835881962
10.3. For small ν, the Richards growth function can be 2020 7,597,238,738 9.880655774
represented approximately as the ODE 2030 8,259,167,105 9.916936253
 K 
N R = αNR ln  ν.
 NR  10.5. Use simulations or stability analysis to find the
perturbation threshold above which the population
Assess the quality of this approximation for different with constant predation, (10.10), recovers and
parameter values. below which it goes extinct. Can you find a similar
threshold for the population with proportional
10.4. It is sometimes said that the world population is
predation, (10.9)? Argue whether this is possible or
growing “super-exponentially.” Use Table 10.4,
not and/or execute simulations.
which was assembled from US Census Bureau
datasets, to analyze this claim. Is it possible to 10.6. Provide arguments and/or simulation results for
predict from these data what the carrying capacity why small populations are more strongly affected
of the human population on Earth might be? Try to by (random) perturbations than large populations.
use methods from Chapter 5 to represent the data 10.7. Discuss why it is important to use age classes of the
with one of the growth models described in this same length and to match this length with the
chapter. reporting time period τ in a Leslie model.
10.8. For n = 2, n = 3, and n = 5, show that the matrix enter the second year to produce offspring.
equation (10.16) generates the same numbers as This number continues to decrease, as
using (10.15) twice, three times, or five times, indicated in Table 10.5, which was compiled
respectively. Think before you execute the matrix from data in [29]. Fortunately for the species,
equation five times. the breeding females lay between 25 and 150 eggs
10.9. Construct a Leslie model where the population size per year.
is constant but not zero. Execute simulations to Use the information in Table 10.5 to construct a
show how robust this model is to changes in one of Leslie model. Study whether the population will
the age classes or in one of the parameters. increase, decrease, or remain constant with these
10.10. Select values for α1, . . ., α4 and σ1, . . . , σ3 and settings. Explore what happens if the survival rate of
compute the total size of the population as a eggs is 8% or 15%. Execute simulations to determine
function of time. whether the order matters mathematically and/or
biologically in which 10 years of 8% and 10 years of
10.11. Construct cases where the Leslie model shows
20% survival occur? Discuss and interpret your
either growth or decline. Construct a case where the
simulation results.
population stays exactly the same over time. Is there
only one solution to this problem?
10.12. In the Leslie model, make α and/or σ dependent on TABLE 10.5: LIFE STATISTICS OF THE ROLY POLY
time. Perform simulations and study the
Age class Number of eggs Number of eggs
consequences. (years) or breeding females* produced
10.13. In the Leslie model, make α and/or σ dependent on
0 10,256 0
population size. Perform simulations and study the
consequences. 1 126 3,470
10.14. Population experts claim that delaying the first 2 98 4,933

birth by a few years would have a significant impact 3 22 1,615
on the human population growth explosion in 4 2 238
developing countries. Set up a Leslie model to test
this claim. State which assumptions (realistic or *The entry for age class 0 (<1 year old) represents the total number
of eggs. Entries for later age classes are breeding females.
unrealistic) you made.
10.15. The roly poly or pillbug, Armadillidium vulgare
(Figure 10.14), is a terrestrial crustacean in the 10.16. Make predictions regarding the dynamics of
grasslands of California. It can live up to 4 or two coexisting bacterial populations that are
sometimes even 5 years. The world is a dangerous growing exponentially with different growth rates.
place for the roly poly: of roughly 10,000 eggs To what degree do the initial population sizes
hatched in a year, only about 11% survive to enjoy matter?
their first birthday. Of these, about 48.5% are
10.17. What happens if the mortality rate of tumor cells in
female, of which only about 23% breed. As a
(10.17) is 0? Predict the outcome and confirm or
consequence, only about 126 breeding females
refute the predictions with simulations. Adjust other
model parameters if the results appear to be
unrealistic. Summarize your findings.
10.18. Study what happens to the phase-plane plot if
one multiplies both equations in the population
system (10.18) by the same positive factor.
Explore the same with a negative factor. In all
cases, make predictions, study the phase-plane
plot, run simulations, and discuss biological
relevance.
10.19. Compute steady-state points in the population
system (10.18) when either N1 or N2 is zero and the
other variable is not. Interpret these points in the
context of logistic growth functions.
10.20. Use conceptual arguments, rather than a formal
mathematical derivation, to discuss the stability of
the two steady states in (10.19) when one variable is
zero and the other is not.
Figure 10.14 Roly poly. A typical population of the Californian roly
10.21. Discuss the two nullcline situations not analyzed in
poly or pillbug, Armadillidium vulgare, exhibits complex birth and the text. Sketch the nullclines and trajectories in
death patterns. (Courtesy of Franco Folini under the Creative Commons each case and make predictions about survival and
Attribution–Share Alike 3.0 Unported license.) extinction.
REfERENCES 301
10.22. Select parameter values for the four combinations of N 1 = 3N 1 (10 − N 2 + 1.5N 1 − 0.1N 12 ),
nullclines and explore different trajectories with
simulations. N 2 = 0.7 N 2 (20 − N 1 − N 2 ).
10.23. Study what happens in the predator–prey model 10.30. Create a figure like Figure 10.12 for the following
(10.20) if an external factor alters the value of N or P system:
at time t = 20. Discuss your findings.
10.24. Perform a qualitative analysis of the predator–prey N 1 = 0.3N 1 (500 − 10 N 2 + 70 N 1 − 3N 12 ),
model (10.20). Discuss coexistence and extinction.
N 2 = 0.7 N 2 (80 − 2 N 1 − N 2 ).
10.25. Linearize the predator–prey model (10.20) at the
trivial and nontrivial steady states and locate their 10.31. Create a figure like Figure 10.12 for the following
stability patterns on the diagram in Figure 4.17. system:
10.26. Implement the chaotic Lotka–Volterra system and
study the effects of changing the initial value of X1. N 1 = 2 N 1 (10 − N 2 + 1.5N 1 − 0.1N 12 ),
Summarize your findings in a report.
N 2 = 2 N 2 (50 − 7 N 1 − N 2 + 0.3N 12 ).
10.27. Perform a phase-plane analysis for the following
system: 10.32. Around the year 1918, the first black fire ant
Solenopsis richteri reached the shores of the United
 100 − 2 N 2 − N 1  States through the port of Mobile in Alabama. In the
N 1 = 0.1N 1   ,
 15 late 1930s, it was joined by the more aggressive red
fire ant, Solenopsis invicta. Without significant
 60 − N 1 − N 2 
N 2 = 0.7 N 2   . predators, the two ant species spread quickly,
 45 honoring the red ant’s Latin species name meaning
“undefeated.” In 1953, the United States Department
10.28. Perform a phase-plane analysis for the following
of Agriculture found that 102 counties in 10 states
system:
had been invaded. By 1996, about 300,000 acres
throughout the South were infected [30]. How
N 1 = 0.6 N 1 (50 − N 1 − 0.833N 2 ), would you design one or more models describing
N 2 = 0.2 N 2 (30 − 0.75N 1 − N 2 ). the past spread and predicting future invasion into
new territories? What kinds of data would be
10.29. Perform a phase-plane analysis for the following needed for what kind of model? Search the literature
system: for model studies of the phenomenon.
REfERENCES
[1] Savageau MA. Growth of complex systems can be related to [11] Voit EO. Recasting nonlinear models as S-systems. Math.
the properties of their underlying determinants. Proc. Natl Comput. Model. 11 (1988) 140–145.
Acad. Sci. USA 76 (1979) 5413–5417. [12] Voit EO & Dick G. Growth of cell populations with arbitrarily
[2] Neugebauer O & Sachs A (eds). Mathematical Cuneiform distributed cycle durations. I. Basic model. Math. Biosci.
Texts, p 35. American Oriental Society/American Schools of 66 (1983) 229–246.
Oriental Research, 1945. [13] Voit EO & Dick G. Growth of cell populations with arbitrarily
[3] Bunt LNH, Jones PS & Bedient JD. The Historical Roots of distributed cycle durations. II. Extended model for correlated
Elementary Mathematics, p 2. Prentice-Hall, 1976. cycle durations of mother and daughter cells. Math. Biosci. 66
[4] Segel LA. Modeling Dynamic Phenomena in Molecular and (1983) 247–262.
Cellular Biology. Cambridge University Press, 1984. [14] White J. The allometric interpretation of the self-thinning rule.
[5] Malthus TR. An Essay on the Principle of Population. J. Theor. Biol. 89 (1981) 475–500.
Anonymously published, 1798 (reprinted by Cosimo Classics, [15] Voit EO. Dynamics of self-thinning plant stands. Ann. Bot.
2007). 62 (1988) 67–78.
[6] Verhulst PF. Notice sur la loi que la population poursuit [16] Bonabeau E. Agent-based modeling: methods and
dans son accroissement. Corr. Math. Phys. 10 (1838) techniques for simulating human systems. Proc. Natl Acad. Sci.
113–121. USA 14 (2002) 7280–7287.
[7] Richards FJ. A flexible growth function for empirical use. J. Exp. [17] Macal CM & North M. Tutorial on agent-based modeling and
Bot. 10 (1959) 290–130. simulation. Part 2: How to model with agents. In Proceedings
[8] Gompertz B. On the nature of the function expressive of the of the Winter Simulation Conference, December 2006,
law of human mortality, and on a new mode of determining Monterey, CA (LF Perrone, FP Wieland, J Liu, et al., eds),
the value of life contingencies. Philos. Trans. R. Soc. Lond. 115 pp 73–83. IEEE, 2006.
(1825) 513–585. [18] Richter O. Simulation des Verhaltens ökologischer
[9] Savageau MA. Growth equation: a general equation and a Systems. Mathematische Methoden und Modelle. VCH, 1985.
survey of special cases. Math. Biosci. 48 (1980) 267–278. [19] Hethcote HW, Horby P & McIntyre P. Using computer
[10] Voit EO. Cell cycles and growth laws: the CCC model. simulations to compare pertussis vaccination strategies in
J. Theor. Biol. 114 (1985) 589–599. Australia. Vaccine 22 (2004) 2181–2191.
[20] Leslie PH. On the use of matrices in certain population [26] Dam P, Fonseca LL, Konstantinidis KT & Voit EO. Dynamic
mathematics. Biometrika 33 (1945) 183–212. models of the complex microbial metapopulation of Lake
[21] Pinsky MA & Karlin S. An Introduction to Stochastic Modeling, Mendota. NPJ Syst. Biol. Appl. 2 (2016) 16007.
4th ed. Academic Press, 2011. [27] Sprott JC, Vano JA, Wildenberg JC, et al. Coexistence and chaos
[22] Ross SM. Introduction to Probability Models, 10th ed. in complex ecologies. Phys. Lett. A 335 (2005) 207–212.
Academic Press, 2010. [28] Vano JA, Wildenberg JC, Anderson MB, et al. Chaos in
[23] Wilkinson DJ. Stochastic Modelling for Systems Biology, 2nd low-dimensional Lotka–Volterra models of competition.
ed. Chapman & Hall/CRC Press, 2012. Nonlinearity 19 (2006) 2391–2404.
[24] Hirsch MW, Smale S & Devaney RL. Differential Equations, [29] Paris OH & Pitelka FA. Population characteristics of the
Dynamical Systems, and an Introduction to Chaos, 3rd ed. terrestrial isopod Armadillidium vulgare in California
Academic Press, 2013. grassland. Ecology 43 (1962) 229–248.
[25] Kaplan D & Glass L. Understanding Nonlinear Dynamics. [30] Lockley TC. Imported fire ants. In Radcliffe’s IPM World Textbook.
Springer-Verlag, 1995. University of Minnesota. http://ipmworld.umn.edu/lockley.
fURTHER READING
Edelstein-Keshet L. Mathematical Models in Biology. McGraw-Hill, Murray JD. Mathematical Biology II: Spatial Models and Biomedical
1988. Applications, Springer, 2003.
Hassell MP. The Dynamics of Competition and Predation. Edward Ptashne M. A Genetic Switch: Phage Lambda Revisited, 3rd ed.
Arnold, 1976. Cold Spring Harbor Laboratory Press, 2004.
Hoppensteadt FC. Mathematical Methods of Population Biology. Renshaw E. Modelling Biological Populations in Space and Time.
Cambridge University Press, 1982. Cambridge University Press, 1991.
May RM. Stability and Complexity in Model Ecosystems. Princeton Savageau MA. Growth of complex systems can be related to the
University Press, 1973 (reprinted, with a new introduction properties of their underlying determinants. Proc. Natl Acad.
by the author, 2001). Sci. USA 76 (1979) 5413–5417.
Murray JD. Mathematical Biology I: Introduction, 3rd ed. Springer, 2002.
Integrative Analysis of
Genome, Protein, and
Metabolite Data: A Case
Study in Yeast
11
• Discuss the steps for converting diverse experimental data into a
computational model
• Describe how available data and research goals define the focus and scope of
a model
• Match different data types with suitable modeling methods
• Identify the main components of the heat stress response in yeast
• Describe responses on different timescales and their impact on modeling
analyses
• Explain the role of trehalose in heat stress responses
• Set up a metabolic model of the heat stress response from metabolic data
• Set up a metabolic model of the heat stress response based on gene
expression data
• Discuss the advantages and limitations of time-series data for metabolic
modeling
In the first part of the book, we learned the basics of the wondrous world of model-
ing in biology and gained a glimpse into the structure of networks and systems.
Then, in the second part, we discussed various types of data at different levels of the
hierarchy of biological organization, their information content, limitations, and
peculiarities, and, in particular, their value for a deeper understanding of how sys-
tems are designed and how they operate. We saw that biological data can take many
forms and that some are clearly more useful for systems analyses than others. It is
now time to study some concrete applications of models, because these, more than
any general discussions, reveal where the real challenges are and how they might be
overcome. Also, as any modeler learns over time, actual biological systems have a
304 Chapter 11: Integrative Analysis of Genome, Protein, and Metabolite Data: A Case Study in Yeast
way of surprising us with aspects that would have never occurred to us before we
embarked on their analysis. Many hypotheses and draft theories have come crash-
ing down when they were applied to real (or, as some theoreticians like to think,
“dirty”) data.
This chapter addresses some basic issues of multiscale analysis. The system we
discuss is comparatively simple, and the multiple scales do not span the entire spec-
trum from molecules to the world oceans (see Chapter 15), but merely include
genes, proteins, and metabolites. Nevertheless, the small case study is sufficient for
reaching the accurate conclusion that it is often problematic to slice biological sys-
tems for analysis into isolated levels of organization. We will see that focusing solely
on the genomic or the metabolic level masks important clues as to how the system
as a whole is organized. The case study also shows that even apparently simple sys-
tems require a lot of information regarding their components, processes, and other
aspects, and that it is difficult to discern a priori whether particular details can be
omitted without much loss or whether they provide clues for important features of
the function and regulation of the system. Putting the simplicity of the case study in
this chapter into perspective, one will easily imagine how any multilevel system
becomes much more complicated when the range of scales is widened, and we will
discuss one such situation, the modeling of the heart, as a case study in Chapter 12.
ON THE ORIGIN OF MODELS

It is usually difficult to pinpoint where the ideas of a new research study originated.
There are intriguing cases such as the dream of a snake catching its own tail, of mon-
keys chasing each other in a circle, or of dancing ballerinas that allegedly inspired
the German chemist Friedrich August Kekulé von Stradonitz to propose a ring
structure for the carbon atoms in a benzene molecule. Alas, in many other cases,
new studies seem to have come from a spark of unknown origin or from ongoing
discussions augmented by a loose collection of numerous tidbits of intellectual con-
tributions. The situation is without doubt even more complicated in interdisciplin-
ary studies, where one side usually does not fully understand the inner workings of
the other. So, for the sake of argument, let us just imagine a fictitious water cooler
encounter between an experimentalist E and a modeler M:
E: Have I ever told you about my heat stress stuff? I think it’s pretty cool. Got
some great data. I’m sure you could model them!
M: Hmm. I’m really sort of busy. What’s so special about heat stress?
E: It’s interesting.
M: I guess it would be nice to figure out how to respond to these recent heat
waves.
E: No, no, we do yeast.
M: Yeast? Why is that interesting?
E: Well, because yeast is a good model.
M: Model for what?
E: Responses to stresses in all kinds of cells and organisms, including humans.
M: Yeah, it would be nice to know how to deal with stress! Let’s have coffee, and
you tell me more.
E: Well, let me start with an intriguing molecule: trehalose …
As in a biochemical reaction that requires the collision of molecules, two scien-

tists bump into each other, creating a spark that might quickly dissipate or lead to a
long-term collaboration. Intriguingly, it is almost impossible to define conditions or
predictors for either outcome, but it appears that every new collaborative project
demands a certain leap of faith. And if the project is interdisciplinary, there must be
two leaps and they must be rather large.
In our encounter, the modeler eventually becomes tempted to pursue the proj-
ect, and the first order of business is now to gauge whether there is even enough
common ground and what the prospects of success might be. The experimenter’s
ON THE ORIGIN OF MODELS 305
statement “I’m sure you could model them!” is certainly not enough, because exper-
imentalists often do not have a sufficient feel for modeling to judge the suitability of
their data for a computational analysis. Furthermore, fathoming the potential and
probability of success is not easy for anybody, and while experience certainly helps,
it is a good idea early on to define what exactly the purpose of the model analysis
might be. A crucial component of this assessment is limiting the scope, which is
easier said than done, because the creation of a focus often requires extended dis-
cussions and hours of literature browsing. In our case, the modeler quickly learns
from the experimentalist that any cellular stress response involves very many steps
at various levels of biological organization: the response is clearly physiological, and
dissecting it reveals that genes are up-regulated; the activities of existing proteins
might be changed, new proteins are synthesized, and some proteins are deacti-
vated; there are most certainly shifts in the cell’s metabolic profile; and it is to be
expected that the overall response is controlled by signal transduction.
It is impossible to capture and integrate all these components simultaneously
and in detail with today’s modeling methods. Even though teragrid and cloud com-
puting have become a reality and computer scientists have produced very efficient
algorithms for executing mega-size discrete-event simulations, biological systems
are typically so ill-characterized in detail that the bottleneck for modeling analyses
is often not computational power but the lack of enough quantitative biological
information. It is therefore necessary for the experimentalist and the modeler to
spend sufficient time to understand each other’s intellectual context, milieu, con-
cepts, and ways of thinking; the suitability, quality, and quantity of data; the poten-
tially applicable modeling techniques; a realistic scope; and some targets that one
might attain with reasonable likelihood of success. The following is a possible sam-
ple of lead questions:
• What is heat stress? What is heat shock? Does the study have broader
significance?
• What is trehalose?
• Is the response natural? Does it occur under conditions that are typical for the
organism?
• Which biological level is mainly responsible? Genomic, proteomic, metabolic,
physiological?
• Is the heat response found in many/all creatures?
• What information is available?
• What types of data are available? How reliable are they?
• What is already known? What is not known, but significant?
• Has anyone already constructed pertinent mathematical models? Are they
good?
• What aspects have been modeled?
• Which questions can be answered with the models?
• What is the main purpose of a new model analysis?
• Is it feasible to limit the scope without losing the big picture?
• Can the system be modularized without compromising its natural function?
• Is it possible to foresee whether a model would have a good chance of success?
An experienced experimentalist will be able to give at least partial answers to the

biological questions, but the situation is complicated because there is so much
known and it is quite evident that not all information is equally important for the
modeling effort. What should the modeler be told, without being totally over-
whelmed? There are also aspects that the experimentalist might not know in sufficient
detail and that the modeler will have to find out independently. Let us fast-forward
and suppose that the modeler is indeed interested in working with the experimental-
ist on one or more mathematical models that explain certain aspects of the heat
stress response. The basis from which to start the new project is common knowledge
of the heat response in yeast, and in particular the role of trehalose, which the
experimentalist had identified as interesting for the ongoing investigation.
A BRIEF REVIEW OF THE HEAT STRESS RESPONSE

IN YEAST
If the temperature of the medium is changed from 30 °C to 39 °C, several events are
triggered within the yeast cells (for reviews, see [1–4]). These events are most cer-
tainly interconnected, but it is so far not clear what controls their coordination. A
quick foray into the literature shows that there is a fine, although not always
observed, distinction between heat stress and heat shock, with the first usually
referring to increased heat conditions that can still be considered physiological
and the latter referring to even higher temperatures, at which the cells barely sur-
vive. The experimentalist is most interested in heat stress, which thereby refines the
scope of the study.
Maybe the most astounding response happens at the metabolic level: this is the
production of the disaccharide trehalose, which essentially consists of two glucose
molecules bound together (see later). Under normal physiological growth condi-
tions, yeast cells contain merely trace concentrations of trehalose and its immediate
precursor trehalose 6-phosphate (T6P). However, once the temperature rises above
37°C, the amount of trehalose may rise to an amazing concentration of up to 1 g of
trehalose per 1 g of protein, or 40% of the entire cell biomass [5]. Mutants that lack
the ability to form trehalose or that overproduce the trehalose-degrading enzyme
trehalase are overly sensitive to environmental stresses, including heat, which indi-
cates that the dramatic trehalose production is apparently vital for an effective
response. Maybe of greatest importance, trehalose protects proteins against dena-
turation and reduces the formation of unnatural protein aggregates. In a similar
role, it preserves the integrity of DNA, lipids, and membranes. Beyond protecting
macromolecules and membranes, the trehalose pathway is involved in the control
of glucose utilization by counteracting the accumulation of free cytoplasmic glu-
cose and sugar phosphates, which would otherwise result in the depletion of cyto-
plasmic phosphate and adenosine triphosphate (ATP).
A different metabolic response is the dramatic activation of specific lipids, called
sphingolipids, which have been implicated in a variety of fundamental decisions at
the cellular level, including differentiation, apoptosis, and the response to various
stresses, including heat [6, 7]. Within a few minutes of heat stress, key enzymes of the
pathway of sphingolipid biosynthesis exhibit increased activity. Several intermedi-
ates of the pathway increase several-fold in quantity, and some remain elevated for
two hours or more. The role of sphingolipids in the response to heat stress appears to
be twofold. First, they are important components of plasma membranes, and
because the stress response requires the transport and mobilization of proteins
and metabolites, the additional quantities of sphingolipids might be required to sus-
tain membrane integrity. Second, sphingolipids serve as second messengers in
signal transduction processes. As one example, sphingolipid metabolites activate
the transcription of a number of genes, including some of those associated with
trehalose synthesis [8].
At the genome level, the response to heat is widespread, and many genes are up-
regulated within a few minutes [9]. Not surprisingly, almost all genes coding for
enzymes associated with the trehalose cycle are up-regulated, but two details are
intriguing. First, the degree of up-regulation differs dramatically, spanning a range
from essentially unchanged to over 100-fold. Second, the genes coding for treha-
lose-degrading enzymes, trehalases, are robustly up-regulated as well. Why would
that happen, if the cell needs large amounts of trehalose? We will return to this ques-
tion later. It should also be noted that some genes are strongly down-regulated.
Examples are ribosomal protein transcripts. In fact, there is a transient decrease in
overall translation initiation, coupled with transient cell cycle and growth arrest. It
is possible that the combined effects of the decrease in transcript and protein syn-
thesis reduce biosynthetic costs while the cell adapts to stress [1, 10].
Proteins are involved in stress responses as well [11]. Several dozen yeast pro-
teins are transiently induced by heat stress. Roughly a third of them are heat shock
proteins, which mediate protein–protein interactions and the transport of proteins
through membranes. They are apparently involved in the recovery from stress and,
in particular in the correct refolding of proteins that had unfolded during heat.
A BRIEF REVIEW OF THE HEAT STRESS RESPONSE IN YEAST 307
Some of these proteins respond very strongly. For instance, the heat shock protein
Hsp90 is induced ten- to fifteenfold. Of particular importance is also Hsp104, which
seems to be associated with accumulating trehalose and with the removal of dena-
tured and unsalvageable protein fractions, for instance through the proteasome
(see [12] and Chapter 7). In addition to heat shock proteins, several enzymes are
activated. We will study later the manner in which this activation appears to be cru-
cial to the metabolic heat stress response, and specifically to the huge increase in
trehalose in response to strong rises in temperature.
Further aspects of the heat response might be called physiological. These include
changes in potassium channels and the mobilization of transcription factors, such
as the zinc finger proteins Msn2p and Msn4p, which control the expression of most
of the enzymes of carbon metabolism, as well as antioxidant defense proteins of the
heat stress response. These responses seem to be obligatory, because MSN2/MSN4
double mutants are unable to accumulate trehalose under heat conditions. The
importance of this mechanism is underscored by inbuilt redundancy: failure in one
of the transcription factors is compensated by the other. Translocation of Msn2p
and Msn4p from the cytosol to the nucleus occurs within minutes of heat stress and
has been considered to be the rate-determining step in the activation of the heat
stress response. In the nucleus, Msn2p and Msn4p interact with stress response
elements (STREs), which are found in some of the genes associated with the treha-
lose cycle, among others. It appears that this interaction between Msn2/4p and the
STREs is modulated by sphingolipids [8, 13].
This brief summary indicates that the heat stress response is very important to the
cell, because its control spans all hierarchical levels of biological organization. These
control mechanisms do not act in isolation, and there are checks and balances, as
one would expect to find in a robust regulatory design (Figure 11.1). For instance,
most of the mechanisms outlined above lead to an increase in trehalose. However,
the cell also possesses mechanisms to slow down or block trehalose biosynthesis.
Particularly important among them is the Ras–cAMP–PKA pathway, which in
unstressed cells keeps the transcription of trehalose-associated enzymes and the
concentration of trehalose at low levels. The pathway can also activate the trehalose-
degrading enzyme trehalase. In direct opposition to the effects of heat, cyclic adenos-
ine monophosphate/protein kinase A (cAMP/PKA) blocks STREs and causes
movement of Msn2/4p from the nucleus back to the cytosol. Furthermore, it seems
that PKA prevents the accumulation of Msn2/4 in the nucleus. These processes
involve the enzyme Yak1 kinase. In addition, most genes that are induced by Msn2/4p
after heat stress are repressed by an excess of cAMP. Thus, any heat stress response
Yak1 heat
kinase
Msn2/4
sphingo-
lipids
Figure 11.1 Tentative structure of the
cAMP/ dynamic system of checks and balances
STRE
PKA regulating the heat stress response. The
TPS1/2
nucleus shaded yellow square represents the cell
cytosol nucleus. The blue arrows indicate gene
expression or metabolic reactions, the
green arrows indicate activating effects,
T6P trehalose the blunted red lines indicate inhibitory
effects, the orange arrow indicates genetic
trehalase activation and enzyme deactivation, and
the purple arrow indicates the interaction
between transcription factors (Msn2/4) and
stress response elements (STRE). See the text
for further information.
must overcome or control intracellular cAMP levels. Interestingly, T6P inhibits one of
the early glycolytic enzymes and indirectly affects the activation of the cAMP-
dependent signaling pathway. Finally, trehalase is a substrate of PKA and therefore
not only degrades trehalose but also induces the PKA pathway. While some of these
details may not be crucial for the following analyses in this chapter, they serve to
demonstrate that there is a fine-tuned balance between antagonistic pressures of
the direct and indirect effects of heat on one hand and the activity of the Ras–cAMP–
PKA signaling pathway on the other (for further details, see [2, 14, 15]).
11.1 The Trehalose Cycle

The experimentalist and modeler agree that the modeling effort should initially CH2OH OH
focus on the role of trehalose in the heat stress response. It is therefore necessary to O
OH
explore the basic features of this metabolite and, in particular, to study its biosyn- OH HOH2C
O
thesis and degradation. HO O OH
Trehalose (α,α′-trehalose, α-d-glucopyranosyl α-d-glucopyranoside, or α,α- OH
1,1-diglucose) (Figure 11.2) is a disaccharide (two sugars) that is produced from
glucose in several enzymatic steps. Under normal conditions, most glucose enters Figure 11.2 α,α′-trehalose. The molecule is
glycolysis and the pentose pathway. Excess glucose is typically stored in the form of a disaccharide; that is, it consists of two sugar
the highly branched polysaccharide glycogen. However, under heat conditions, a (glucose) molecules.
sufficient amount of glucose must be devoted to the synthesis of trehalose. These
changing demands already imply that the cell has to solve a complex control task.
As in many cases of metabolic networks, information regarding pathway con-
nectivity can be found in databases such as KEGG [16] and BioCyc [17], as well as in
the original literature. A simplified diagram is presented in Figure 11.3. More diffi-
cult is the determination of the regulatory structure, which requires sophisticated
text mining and a thorough study of the primary literature. A summary of the path-
way with regulation is given in Figure 11.4. Furthermore, information on gene
expression in response to heat stress can be visualized as in Figure 11.5.
Yeast cells can take up glucose from the medium with one or more of about 20
hexose transporters (HXTs), which are specialized for different conditions. The glu-
cose is instantaneously converted into glucose 6-phosphate (G6P), and there is
speculation that G6P might inhibit glucose transport into the cell (see, for example,
[18]). The phosphorylation of glucose to G6P is catalyzed by three isozymes: hexo-
kinase I, hexokinase II, and glucokinase; their corresponding genes are HXK1,
HXK2, and GLK1. It is commonly assumed that this phosphorylation step is essen-
tially irreversible. The three isozymes control each other through complex
interactions that determine which kinase is most active under specific environmen-
tal conditions. For instance, during exponential growth on glucose, hexokinase
medium
cell
glucose
trehalose
(internal)
trehalose
6-phosphate
pentose
phosphate glucose glucose glycogen
UDPG
pathway 6-phosphate 1-phosphate
Figure 11.3 Connectivity of the trehalose

cycle. Trehalose is the product of a pathway
fructose
1,6-bisphosphate branching off from glycolysis. Degradation
of trehalose returns glucose, which can enter
glycolysis or other pathways. (Adapted from
Voit EO. J. Theor. Biol. 223 [2003] 55–78. With
A BRIEF REVIEW OF THE HEAT STRESS RESPONSE IN YEAST 309
medium Figure 11.4 Regulation of the trehalose

cell
cycle. Green arrows show flow of material;
thicker arrows indicate the main flux. Red
– and blue arrows indicate inhibition or
activation, respectively. (Adapted from Voit
glucose EO. J. Theor. Biol. 223 [2003] 55–78. With
(internal) trehalose permission from Elsevier.)
– –
trehalose
+ + 6-phosphate
pentose
glucose glucose
phosphate UDPG glycogen
6-phosphate 1-phosphate
pathway +
–
fructose
1,6-bisphosphate
I and glucokinase are down-regulated, whereas hexokinase II is induced and

becomes the dominant catalyst. The enzymes are rather stable and can remain
active for several hours in vivo. Because our main focus here is on the trehalose
cycle, it should be noted that T6P is a strong inhibitor of hexokinase II at physiologi-
cal concentrations, but that it inhibits hexokinase I only weakly and glucokinase not
at all. Again, things are much more complicated than they might appear at first.
G6P is arguably the most important branch point of carbohydrate metabolism,
because it is here where material is channeled into glycolysis by isomerization to
fructose 6-phosphate, into the pentose phosphate pathway by means of the G6P
dehydrogenase reaction, or into other pathways, such as the trehalose cycle (see
Figures 11.3–11.5). Under normal conditions, roughly two-thirds to three-quarters
of G6P is used for glycolysis, 5–10% is channeled toward the pentose phosphate
pathway (PPP), about 20% is reversibly converted into glucose 1-phosphate (G1P),
and the remaining smaller quantities move into other pathways, for instance,
toward glycogen and trehalose. Intriguingly, this distribution changes drastically
under heat stress conditions. The gene for G6P dehydrogenase is up-regulated
about four- to sixfold and the genes for the pathways toward trehalose and glyco-
gen are up-regulated very strongly (ten- to twentyfold). By contrast, the genes
HXT1-17 medium
cell
glucose NTH1/2, ATH1

(internal) trehalose
HXK1, TPS2
HXK2
GLK1 trehalose
6-phosphate
TPS1
pentose ZWF1 glucose glucose
phosphate
6-phosphate 1-phosphate
UDPG glycogen Figure 11.5 Changes in the expression
pathway of genes associated with the trehalose
PGM1/2 UGP1 GSY1/2
cycle in response to heat stress. Different
PFK1/2 thicknesses of arrows indicate degrees
of gene expression, while dashed arrows
GPH1
represent no change. (Adapted from Voit
fructose EO. J. Theor. Biol. 223 [2003] 55–78. With
1,6-bisphosphate permission from Elsevier.)
coding for phosphofructokinase (PFK1/2), the first committed step toward glycol-
ysis, are apparently not affected at all.
The formation of trehalose occurs in several steps (see Figures 11.3–11.5).
Phosphoglucose mutase reversibly converts G6P into G1P, which secondarily
undergoes a reaction with uridine triphosphate to form uridine diphosphate glu-
cose (UDPG). A subsequent reaction, which is catalyzed by T6P synthase (TPS1),
synthesizes trehalose 6-phosphate (T6P) from G6P and UDPG. T6P phosphatase
(TPS2) quickly converts T6P into trehalose through phosphoric ester hydrolysis.
Indeed, TPS1 and TPS2 form a complex, together with two regulatory subunits,
TPS3 and TSL1 (trehalose phosphate synthase and trehalose synthase long chain)
that are co-regulated and induced by heat. At least in bacteria, and presumably in
yeast, the T6P synthase reaction is irreversible. Its gene, TPS1, is repressed by
glucose, and the degree of repression determines the concentration and activity
of the trehalose production complex. In contrast, G6P and UDPG induce treha-
lose production. TPS1 induction after heat stress is reflected in a corresponding
production of T6P.
Through O-glycosyl bond hydrolysis, trehalose can be split into two molecules of
glucose. This reaction is catalyzed by one or more of three trehalases, NTH1, NTH2,
or ATH1, which are located in different cellular compartments and exhibit different
activity profiles during growth. As discussed earlier, the trehalase step is important
for trehalose degradation, and trehalase deficiency leads to trehalose accumulation.
Furthermore, successful recovery from heat stress appears to require rather rapid
removal of trehalose, because high concentrations of trehalose impede the refold-
ing of proteins that had partially denatured during heat stress [19, 20]. Conversely,
any overproduction of trehalase results in insufficient trehalose concentrations
under stress and may increase mortality.
Taken together, the biosynthesis and degradation of trehalose begin and end
with glucose, and thereby form the trehalose cycle:
2 glucose → 2 G6P; 1 G6P → 1 UDPG;

1 G6P + 1 UDPG → 1 T6P; 1 trehalose → 2 glucose
The trehalose cycle provides microorganisms with an amazing tolerance to heat,

as well as to cold, osmotic pressure, and water shortage, and this tolerance is
achieved effectively and cheaply. This combination of fortuitous features has
piqued the interest of biotechnologists, who are beginning to replace some of the
established preservation techniques for valuable organic materials, such as icing,
liquid nitrogen, and vacuum drying, with much more subtle and less damaging
metabolic methods involving trehalose [21, 22]. The most obvious and direct
application is the optimization of freezing and freeze-drying of yeast cells, for
instance, in the food industry. More generally, because of its unique ability to
stabilize molecules, its mild sweetness, high solubility, and low hygroscopicity,
and, last but not least, a price that has become affordable through genetic modi-
fications of microorganisms, trehalose has become an important biotechnologi-
cal target for manufacturing dried and processed food and making dry powder
from foods and liquids. Currently, over 20,000 food products, made by 7000 com-
panies around the world, contain trehalose [23]. For the same reasons, trehalose
is of interest for the production of cosmetics, such as lipsticks. Furthermore, tre-
halose treatment permits efficient freeze-drying of platelets and other blood
cells, which is of great importance because it could partially alleviate storage
problems that are the cause of a chronic worldwide shortage of human blood
platelets. Trehalose may also be used for the storage of vaccines, active recombi-
nant retroviruses, and mammalian and invertebrate cells at low temperature,
which is, for instance, needed for experiments in space laboratories.
MODELING ANALYSIS OF THE TREHALOSE CYCLE

11.2 Design and Diagnosis of a Metabolic Pathway Model
Given the connectivity and regulatory structure of the pathway, it is straightforward
to formulate symbolic equations, especially if one uses the default of a canonical
MODELING ANALYSIS OF THE TREHALOSE CYCLE 311
modeling approach, such as biochemical systems theory (BST; see Chapter 4 and
[24–26]). One begins by defining dependent variables for all metabolite pools of
interest and constructs their differential equations such that all fluxes entering or
exiting these pools are formulated as power-law functions. Each of these functions
contains all variables that directly affect the corresponding flux, raised to an appro-
priate power, called the kinetic order, and in addition each flux term has a rate
constant that is positive. If other types of functions are known (or can be assumed)
to be appropriate representations, then they may also be used in the model con-
struction. Functional forms such as Michaelis–Menten and Hill rate laws can often
be found in the literature (see Chapter 8).
For the design of a model for the trehalose cycle, several starting points are pos-
sible. One could start from scratch, consulting pathway databases such as KEGG
[16] and BioCyc [17] and extracting kinetic data from a database such as BRENDA
[27]. However, it is of course easier to construct a model by adapting an already
existing precursor. In this case, Galazzo and Bailey [18] performed targeted experi-
ments on glycolysis in yeast and set up and parameterized a mathematical pathway
model, using Michaelis–Menten rate laws and their generalizations. While they did
not specifically consider the production of trehalose, owing to the low flux through
this pathway under standard conditions, their model can directly serve as a baseline
that is to be expanded toward the trehalose cycle. Further simplifying our task, Curto
and co-workers [28] converted Galazzo and Bailey’s model into BST models, which
others later used for various purposes, such as illustrating model design and optimi-
zation methods [25, 29, 30]. Some of these models were even expanded to include
the trehalose cycle [2, 31, 32], thereby making our life unusually easy. Furthermore,
several datasets are available that characterize gene expression associated with heat
stress in yeast [2, 9, 31–33]. As an additional data source, which we will discuss
later in more detail, Fonseca et al. [34] measured time series of relevant metabolite
concentrations.
A model of the trehalose cycle (see Figures 11.3–11.5) in generalized mass-
action (GMA) form, as adapted from earlier studies, is as follows:
Glucose : X 1 = 30 X 2−0.2 − 90 X 10.75 X 6−0.4 + 2.5 X 70.3 , X 1 = 0.03;

G6P : X 2 = 90 X 10.75 X 6−0.4 + 54 X 2−0.6 X 30.6 − 23 X 20.75
− 3 X 20.2 − 7.2 X 20.4 X 3−0.4 − 0.2 X 1−0.3 X 20.3 X 40.3 , X 2 = 1;
G1P : X 3 = 7.2 X 20.4 X 3−0.4 + 0.9 X 2−0.2 X 4−0.2 X 50.25 − 54 X 2−0.6 X 30.6
− 11X 30.5 − 12 X 2−0.2 X 30.8 X 4−0.2 , X 3 = 0.1;
UDPG : X 4 = 11X 30.5 − 0.2 X 1−0.3 X 20.3 X 40.3 − 3.5 X 20.2 X 40.4 , X 4 = 0.66;
Glycogen : X 5 = 3.5 X 20.2 X 40.4 + 12 X 2−0.2 X 30.8 X 4−0.2 − 0.9 X 2−0.2 X 4−0.2 X 50.25
− 4 X 50.25 , X 5 = 1.04;
T6P : X 6 = 0.2 X 1−0.3 X 20.3 X 40.3 − 1.1X 60.2 , X 6 = 0.02;
Trehalose : X 7 = 1.1X 60.2 − 1.25 X 70.3 , X 7 = 0.05. (11.1)
Here, the units are micromoles for concentrations and micromoles per minute for
fluxes. Note that the model contains not only trehalose and T6P, in addition to the
glycolytic metabolites, but also the pathway toward glycogen, which competes with
the production of trehalose for the shared substrate UDPG. We will see later that the
dynamics of glycogen is strongly affected by heat stress. It is at this point unclear to
what degree the reaction between G1P and glycogen is reversible, and the model
above includes both directions.
Before the model is used for further analyses, it should be subjected to the stan-
dard battery of diagnostic tests that, if successful, will increase our confidence in its
reliability. Two typical tests explore the steady state, along with the eigenvalues
and sensitivity values of the model (see Chapter 4). Indeed, with the given initial
values, the system is very close to a steady state and all eigenvalues have negative
real and zero imaginary parts, confirming local stability. The sensitivities and
gains of metabolites and fluxes with respect to rate constants and kinetic orders are
unremarkable, with the vast majority of them being smaller than 1 in magnitude, 2 ~
X5
indicating that most perturbations in parameters or independent variables are ~
~ ~
readily attenuated. Only a few sensitivities are larger than 1 in magnitude, but none X2 X3 X1
have values high enough to be worrisome. Because sensitivities and gains are char-
normalized concentration
acteristics associated with very small changes, it is also useful to spot-check the
local and structural stability of the model with simulations of temporary or persis- ~
X7
1
tent increases in some of the variables or parameters. For instance, one might initi-
ate the system at the steady state but with one of the initial values doubled or ~
X4
halved. Again, the results are unremarkable for our model, and the system very
~
quickly recovers from such perturbations. Interestingly, the recovery from a bolus X6
in UDPG takes the longest, but even here the system is essentially back to its steady
state in less than 5 minutes. 0
As a specific example, suppose the influx of glucose into the system is augmented 0 5 10
time
by a constant bolus of 10 units between times 1 and 5, with the system starting at
time 0 in its steady state. As a comparison, the natural influx has about 30 units. As
Figure 11.6 Responses of the trehalose
Figure 11.6 indicates, all variables react to the bolus with increasing concentra-
pathway model to a glucose bolus
tions, which return to normal once the bolus is removed. Note that the variables are between times 1 and 5. All variables are
normalized with respect to their steady-state values, which moves the steady state to normalized with respect to their steady-
(1, 1, . . ., 1). state values. Variable names are the same
There is no limit to the kinds of simulations that one can perform with models of as in (11.1), with the tilde (~) denoting
this type. One interesting case is a persistent excess or shortage of glucose. It has normalization.
been speculated that trehalose is a carbohydrate storage molecule, but that glyco-
gen is preferred because it is energetically superior. The simulation is simple: per-
manently increase or decrease the glucose influx to the system. As an example, if the
influx is doubled, the glycolytic flux increases by 80% and the trehalose concentra-
tion by about 40%. In contrast, glycogen increases fivefold. Under glucose shortage,
the glycogen storage pool becomes quickly depleted.
Another set of simulation experiments helps us explore the response of the sys-
tem to enhanced or decreased enzyme activities, which could be due to natural or
artificial genetic alterations. For instance, we can easily test and confirm that muta-
tions in TPS1 or TPS2 lead to reduced trehalose production.
In addition to exploring the effects of temporary or persistent perturbations,
one can use the model to study the specific roles of the various regulatory signals
that are present in the trehalose cycle. The tool for these types of explorations of
so-called design principles is the method of mathematically controlled compari-
son, which we will discuss in detail in Chapter 14. In a nutshell, one performs two
analyses in parallel: one with the observed system and one with a system that is
identical except for one feature of interest. Differences in responses between the
two systems can then be attributed to the differing feature. An analysis of this type
shows that each regulatory feature in the trehalose system offers the pathway a
slight advantage, and that all signals together comprise a much more effective con-
trol system [2].
11.3 Analysis of Heat Stress

The model as formulated so far does not contain variables or parameters that rep-
resent heat. So, how can we study the effects of heat stress? Several strategies may
be adopted, and we begin with the simplest, which focuses on metabolites and
enzymes. The metabolites of interest are already in the model as dependent vari-
ables, but the enzymes of the pathway that catalyze the different reactions convert-
ing glucose eventually into trehalose and other products are not obvious. However,
the model implicitly accounts for them, because they directly affect the rate con-
stants. For instance, the term 1.1X 60.2 in the equations for T6P and trehalose
describes the conversion of T6P into trehalose, which is catalyzed by the enzyme
T6P phosphatase. The rate of this reaction has a numerical value of 1.1, and this
quantity is in truth the product of a rate constant and the enzyme activity. If the
enzyme activity is doubled, the rate constant is doubled. This linear relationship is
all we need in the given situation, and we do not really need to know how exactly
the value 1.1 is split between the rate constant and the enzyme activity.
TABLE 11.1: INDEPENDENT VARIABLES IN MODEL (11.2)

Catalytic or transport step Enzyme or transporter Variable
Glucose uptake HXT X8
Hexokinase/glucokinase HXK1/2, GLK X9
Phosphofructokinase PFK1/2 X10
G6P dehydrogenase ZWF1 X11
Phosphoglucomutase PGM1/2 X12
UDPG pyrophosphorylase UPG1 X13
Glycogen synthase GSY1/2 X14
Glycogen phosphorylase GPH X15
Glycogen use GLC3 X16
α,α-T6P synthase TPS1 X17
α,α-T6P phosphatase TPS2 X18
Trehalase NTH X19
Nevertheless, when many enzyme activities are to be changed, it is convenient

to display them explicitly in the equations. This is easily accomplished by defining
an independent variable for each enzyme, include it in the appropriate reaction
steps, which sometimes appear in two or more equations, and set its value equal to
1 for normal conditions. Table 11.1 and a slightly expanded set of equations, (11.2),
illustrate these changes. Note that these equations are exactly like (11.1), except
that the independent variables for all pertinent enzymes are now explicitly shown.
As long as these have their normal value of 1, the two sets of equations are numeri-
cally the same.
Glucose : X 1 = 30 X 2−0.2 X 8 − 90 X 10.75 X 6−0.4 X 9 + 2.5 X 70.3 X 19 , X 1 = 0.03;

G6P : X 2 = 90 X 10.75 X 6−0.4 X 9 + 54 X 2−0.6 X 30.6 X 12 − 23 X 20.75 X 10
− 3 X 20.2 X 11 − 7.2 X 20.4 X 3−0.4 X 12 − 0.2 X 1−0.3 X 20.3 X 40.3 X 17 , X 2 = 1;
P : X 3 = 7.2 X 20.4 X 3−0.4 X 12 + 0.9 X 2−0.2 X 4−0.2 X 50.25 X 15 − 54 X 2−0.6 X 30.6 X 12
G1P
− 11X 30.5 X 13 − 12 X 2−0.2 X 30.8 X 4−0.2 X 15 , X 3 = 0.1;
UDPG : X 4 = 11X 30.5 X 13 − 0.2 X 1−0.3 X 20.3 X 40.3 X 17 − 3.5 X 20.2 X 40.4 X 14 , X 4 = 0.66; 3
Glycogen : X 5 = 3.5 X 20.2 X 40.4 X 14 + 12 X 2−0.2 X 30.8 X 4−0.2 X 15 − 0.9 X 2−0.2 X 4−0.2 X 50.25 X 15
− 4 X 50.25 X 16 , X 5 = 1.04;
2
T6P : X 6 = 0.2 X 1−0.3 X 20.3 X 40.3 X 17 − 1.1X 60.2 X 18 , X 6 = 0.02;
relative activity
Trehalose : X 7 = 1.1X 60.2 X 18 − 1.25 X 70.3 X 19 , X 7 = 0.05.

(11.2) 1
Interestingly, Neves and François [35] reported that the three enzymes directly
associated with the trehalose dynamics exhibit activities that depend on the ambi-
ent temperature. In fact, they presented a temperature–activity plot for each enzyme
0
(Figure 11.7). According to these plots, the activities of T6P synthase and T6P phos- 20 30 40 50 60
phatase increase roughly 2.5-fold, if the temperature in the medium is raised from a temperature (°C)
normal temperature (between 25°C and 30°C) to heat stress conditions of about
39°C. By contrast, the activity of trehalase falls to about half its original value. This Figure 11.7 Effect of temperature on
information can be entered directly into our model. Namely, the corresponding activities of enzymes producing (green)
and degrading (red) trehalose. Light green,
enzyme terms (X17, X18, and X19) are multiplied by 2.5, 2.5, and 0.5, respectively. As
trehalose 6-phosphate synthase; dark green,
an illustration, the term 0.2 X 1−0.3 X 20.3 X 40.3 X 17 represents the T6P synthase reaction trehalose 6-phosphate phosphatase; red,
and appears in the equations for G6P, UDPG, and T6P; for the heat stress scenario, trehalase; lines are smoothed, interpolating
the enzyme activity X17 is changed from 1 to 2.5. The term 1.1X 60.2 X 18 appears in the trends. (Data from Neves MJ and François J.
equations for T6P and trehalose, and again the enzyme activity X18 is set to 2.5. Biochem. J. 288 [1992] 859–864.)
2 200 Figure 11.8 Differential response to

changes in reported heat-stress-sensitive
~ enzymes. When the activities of enzymes
X7
reported in [35] are adjusted to reflect heat
~ ~ ~ stress conditions, trehalose (X7) increases
X3 X1 X2
about 150-fold, while the other metabolites
remain more or less unchanged or decrease.
1 100
All quantities are presented as relative to the
original steady-state values, as indicated by
the tildes. Note the different scales along
~ ~ ~ the axes.
X6 X5 X4
0 0
0 5 10 0 50 100
time time
Finally, the term 1.25 X 70.3 X 19 is directly part of the equation for trehalose, and
appears as 2.5 X 70.3 X 19 in the equation for glucose; to represent heat stress condi-
tions, the enzyme activity X19 is set to 0.5. Everything else in the model remains
the same.
We can easily solve the model with the new parameter values, starting with the
original steady state, and study the responses in metabolites. The simulation shows
that trehalose increases tremendously to about 7.5 mM, which corresponds to
roughly 150 times its level under normal temperature conditions. At the same time,
G6P and G1P remain essentially unchanged, while the remaining metabolites
decrease to between 50% and 75% of their original values (Figure 11.8). As before,
the sky is the limit for variations on these types of simulations. As a final note on this
set of studies, one could ask how quickly the enzyme activities change in response
to heat stress. This question is not all that pertinent here, as long as the enzyme
reacts with a similar speed, because a common time delay of δ minutes would sim-
ply move all transients by that much to the right.
11.4 Accounting for Glucose Dynamics

All simulations so far have assumed that the glucose concentration in the medium
is so high that the cells can take up glucose at a constant rate, which in the glu-
cose equation has the value 30. In a laboratory experiment, this situation may be
feasible, but it might not be realistic in nature. We can adapt the model with rea-
sonable effort to account for depletion of substrate in the medium. It is quite
obvious that the constant rate of 30 must be replaced with a function that accounts
for glucose utilization. As before, it is useful to replace the constant rate of 30 with
the product of the concentration of the external glucose (GLU) and a rate constant
α. For the sake of argument, let us assume the external concentration is GLU = 100,
so that the new rate constant is α = 30/(GLU) = 0.3, which numerically leads to
the same model we had before, because α ⋅ (GLU) = 30. Thus, the glucose uptake
term in the equation of X1 becomes 0.3 ⋅ (GLU ) ⋅ X 2−0.2. The advantage that we have
gained is that we can now represent GLU as an independent or dependent vari-
able, which may be constant or not. In particular, if we define GLU through a dif-
ferential equation, we can model its change over time, while we still have the
option of setting (GLU ) = 0 , which corresponds to a constant glucose concentra-
tion in the medium.
The new option of interest is now the depletion of glucose by the cells. For
instance, we could set
) = − p ⋅ 0.3 ⋅ (GLU ) ⋅ X 2−0.2 .

(GLU (11.3)
In this formulation, glucose disappears from the medium, which is seen in the
minus sign. Furthermore, the depletion of glucose is proportional to the uptake by
the cells, with a proportionality constant p. This parameter p reflects the ratio
between the volume of the medium and the volume of the cytosol, and is usually not 2
equal to 1, because the external concentration is based on the volume of the

medium, whereas the concentration inside the cells needs to reflect the intracellular
volume. Even seemingly trivial aspects like concentrations can become compli-
cated! To permit us to do some simulations, suppose p = 0.025, and we begin with
~
normal temperature conditions and an initial concentration of external glucose of 1 X3 ~ ~
X4, X6
100. The result is shown in Figure 11.9. All metabolites decrease in concentration, ~
X7
with glycogen (X5) being affected most. This detail makes sense, because glycogen is ~
X5
a storage compartment for rainy days. ~
X1 ~
It is now easy to study what happens under heat stress and simultaneous glu- X2
cose depletion. Namely, we use the most recent variant of the model, which 0
0 50 100
accounts for changes in GLU, and multiply the trehalose enzymes by 2.5 and 0.5, time
respectively, as before. Some results are shown in Figure 11.10. The shape of the
transients is somewhat different and, most prominently, trehalose reaches a maxi- Figure 11.9 Under normal temperature
mum, before decreasing. If we were to extend the simulation to longer time win- conditions, a slow depletion of
dows, we would find that trehalose approaches 0 at about t = 450. However, we external glucose leads to decreases in
must be careful with such extensions, because cells under stress usually turn off all metabolites. Note that the storage
all processes that are not absolutely essential and enter something like a state of compartment of glycogen (X5), suffers the
suspended animation. greatest loss. All quantities are normalized
with respect to their steady-state values.
11.5 Gene Expression

Uncounted papers have been written about changes in gene expression under stress
conditions, including heat stress and heat shock in yeast. Many of these papers
address an assortment of heat shock proteins, whose genes are up-regulated under
adverse conditions. However, changes in expression are also observed in a large
number of genes coding for enzymes, and, indeed, several of the enzymes of the
trehalose cycle are up-regulated in response to heat treatment. These changes in
expression, which are usually measured with microarray and DNA chip experi-
ments, are very interesting, because they can be obtained for many genes in single
experiments, whereas metabolic data are often derived from different experiments
and sometimes even from different types of experiments, so that one has to ask how
compatible these diverse data really are.
Gene expression data are not without problems either, for two main reasons.
First, technical challenges and experimental noise lead to uncertainties regarding
the accuracy of the results, as we discussed in Chapter 6. More importantly, it is not
clear whether gene expression is truly correlated with the observed numbers of
mRNA molecules and the corresponding amounts of active proteins in a more or
less linear fashion [36]. In spite of these uncertainties, gene expression data are
very valuable owing to their large information content from single experiments,
and they can be employed, at the very least, as coarse tools for providing clues
regarding the coordinated responses that cells mount when under stress. A large
collection of gene expression data from yeast under various stresses is available in
2 200
~
X7
1 100
~ ~ ~
X1 X3 X2
~
X4
~ ~ Figure 11.10 Simulation of a scenario that
X6 X5
0 0 combines glucose depletion and heat
0 50 100 0 50 100
stress. Most notably, trehalose (X7) reaches a
time time maximum before decreasing.
the Stanford Yeast Database [9, 33, 37]. Similar data, specifically for heat stress, 1.2
~
X1
were also presented in [31]. ~
X4
If one assumes, in first approximation, a direct proportionality between gene ~
X7
expression and protein prevalence, then it is possible to associate enzyme activ-
ities with gene expression and to make inferences from the genome level to the
~
metabolic level. For instance, the gene ZWF1 codes for the enzyme G6P dehy- 0.6 X2 ~
X3
drogenase, which catalyzes the reaction that diverts carbon from glycolysis ~
X6
~
X5
toward the pentose phosphate pathway. Suppose we wanted to investigate the
effects of fourfold up-regulation of ZWF1 on the trehalose cycle, under the
assumption that none of the other reactions are affected. Presuming direct pro-
portionality between gene expression and enzyme activity, we start with the 0
0 1 2
original model in the form of (11.2), multiply the corresponding enzyme term time
X11 in the equation for G6P by 4, and solve the system. As one might expect, all
concentrations rapidly fall, because material is irreversibly channeled out of the Figure 11.11 Effect of the pentose
model system. Interestingly, the transients are all slightly different, and X1 even phosphate pathway (PPP). An increase
shows a small, initial overshoot that is at first glance unexpected (Figure 11.11); in the activity of G6P dehydrogenase (X11)
see Exercise 11.10. channels material into the PPP and thereby
Using this association strategy, one can easily change enzyme activities accord- leads to decreases in all system variables.
ing to observed changes in gene expression. We begin by extracting pertinent infor-
mation from the Stanford database [37]. It is presented in Table 11.2 in the form of
the relative expression change in each gene of the trehalose cycle as a power of 2.
Thus, a value of 3 indicates a 23 = eight-fold change, while −1 corresponds to half
the expression level as under normal conditions. Blank entries indicate missing
measurements.
The expression data clearly demonstrate that the yeast cells initiate a coordi-
nated, large-scale response within minutes of the beginning of heat stress. In fact,
other studies have demonstrated changes within 2 minutes. Most of the involved
genes rise in expression and eventually return to the normal baseline. Furthermore,
all genes seem to reach their maximal change in expression roughly between 15 and
20 minutes after heat stress.
TABLE 11.2: CHANGES IN SELECTED YEAST GENES DURING HEAT STRESS*
ORF Gene Enzyme or step 5 10 15 20 30 40 60 80
YPR026W ATH1 Vacuolar trehalase 3.03 2.94 1.39

YCL040W GLK1 Glucokinase 3.21 4.38 5.32 5.12 4.59 2.19 1.99
YPR160W GPH1 Glycogen phosphorylase 3.76 5.07 6.78 5.99 2.83
YFR015C GSY1 Glycogen synthase 2.97 2.96 4.02 3.29 1.76 1.05 0.12 0.4
YLR258W GSY2 Glycogen synthase 2.96 4.04 4.42 4.24 3.45 2.67 1.59 1.61
YFR053C HXK1 Hexokinase I 3.57 4.7 5.72 4.84 4.53 1.96 0.75 1.65
YGL253W HXK2 Hexokinase II 0.5 0.4 0.64 0.15 −0.76 −1.09 −0.32 −0.09
YDR001C NTH1 Neutral α, α-trehalase 2.26 2.79 3.52 3.24 3.33 2.22 1.56 1.43
YBR001C NTH2 Neutral α, α-trehalase 1.45 1.75 2.17 2.5 2.12 1.66 1.45 1.19
YKL127W PGM1 Phosphoglucomutase I 0.07 −0.12 0.06 0.08 −0.71 −0.32 0.12 −0.36
YMR105C PGM2 Phosphoglucomutase II 4.6 6.05 7.22 6.83 6.59 4.8 2.45 2.28
YGR240C PFK1 Phosphofructokinase I −0.19 −0.9 −0.41 −0.68 −0.64 −0.15 −0.55 −0.45
YMR205C PFK2 Phosphofructokinase II −0.06 −1 0.29 −0.22 0.07 0.03 0.19
YBR126C TPS1 T6P synthase 2.06 3.46 3.69 3.44 3.1 1.81 1.06 0.91
YDR074W TPS2 T6P phosphatase 3.57 3.4 4.16 4.24 3.31 1.78 1.89 1.76
YMR261C TPS3 T6P synthase regulator −0.09 0.15 1.73 −0.09
YML100W TSL1 T6P synthase regulator 4.88 4.86 6.28 6.38 5.7 4.75 3.66 3.31
YKL035W UGP1 UDPG pyrophosphorylase 2.23 3.05 4.25 3.36 1.92 0.99 0.99
YNL241C ZWF1 G6P dehydrogenase 0.68 1.27 2.34 1.95 1.71 1.33 0.53 0.53
*Data extracted from [37].
Limiting our analysis to the trehalose cycle, we find that most genes are up-
change in gene expression (log2)

5 NTH1
regulated within a few minutes and remain up-regulated for some while. In particu- NTH2
lar, the trehalose-producing and -degrading genes follow similar trends (Figure 4 TPS1
11.12). Interestingly, the degree of up-regulation varies very widely (see Table 11.2), 3
TPS2
and some genes, such as the two coding for phosphofructokinase (PFK1 and PFK2),
2
are not up-regulated much in response to heat stress, while others are expressed
very strongly. For instance, glycogen phosphorylase (GPH1), which uses glycogen 1
to produce G1P, reaches an up-regulation level of almost 7, which really means
0
27 = 128-fold! Also intriguing is that the gene coding for hexokinase 2 (HXK2) is not 0 20 40 60 80
changed much at all, while its isozymes hexokinase 1 and glucokinase, which cata- time of heat stress
lyze the same phosphorylation reaction of glucose to G6P, are strongly up-regulated.
Another interesting detail is the fact that the expression patterns of TPS1 and TPS2 Figure 11.12 Heat-stress-induced
are very similar. As a consequence, the pool of T6P remains more or less constant changes in the expression of genes
throughout the heat stress response. This constancy is important, because T6P is directly associated with trehalose.
NTH1/2, trehalase; TPS1, T6P synthase; TPS2,
toxic at high concentrations [38, 39]. T6P phosphatase. TPS2 rises to a level of
At first glance, many details of these patterns in expression may not make much 24.24 = 18.9, which means an almost twenty-
sense. However, they can be explained to some degree by means of the metabolic fold increase in gene expression over normal
model and the assumption that enzyme activities and gene expression levels are temperature conditions about 20 minutes
directly correlated. For simplicity of analysis, let us focus on the maximal deviation after heat stress begins.
from the normal state, which occurs roughly 15–20 minutes into the heat stress and
convert this maximum level of gene expression directly into changes in enzyme
activity. Thus, if a gene is up-regulated eightfold, we multiply the activity of the cor-
responding enzyme by 8. Because most processes contribute to the degradation of a
substrate as well as the synthesis of a product, two terms are often affected. For
instance, the hexokinase reaction from glucose to G6P appears in the first two equa-
tions as 90 X 10.75 X 6−0.4 X 9, once with a minus sign and once with a plus sign. One fur-
ther complication comes from the fact that several reactions, such as the conversion
of glucose into G6P, can be catalyzed by several isozymes. A particularly dramatic
example is the uptake of external glucose by the cells, for which yeast uses roughly
20 different transporters. If several enzymes catalyze the same reaction, their
involvement may be apportioned, for instance, by looking at their affinity under the
most pertinent conditions or by weighing their importance by the number of cor-
responding mRNA molecules per cell [32].
Clearly, the conversion of gene expression into enzyme activity and the weight-
ing of isozymes are drastic simplifications, and results from analyses based on such
assumptions must be considered with caution. Nevertheless, when a comprehen-
sive analysis of this type was executed [32], three insights were gained:
1. If one were to cluster all genes by their degree of up-regulation, as is sometimes

done in bioinformatics for identifying shared control of ill-characterized pathway
systems, it would appear that the glycolytic genes have nothing to do with each
other, because they are expressed at such different degrees and would therefore
be categorized in distinct classes. Therefore, clustering of genes by expression,
while often beneficial, must be done with caution.
2. The up-regulation profile as observed after about 15–20 minutes of heat stress is
not very intuitive. However, if this pattern is translated into enzyme activities, it
appears that the response is well coordinated. The system produces plenty of
trehalose, ATP (via glycolysis) for energy, and NADPH (within the pentose path-
way) for an adequate redox balance. At the same time, the pathway system does
not accumulate unnecessary or toxic intermediates, such as T6P.
3. It is possible to use the model for testing alternative hypothesized up-regulation
patterns and assess their performance with respect to trehalose, ATP, and
NADPH production, as well as the generation of intermediates. For instance, the
cell would have to expend less effort by just up-regulating one or two key
enzymes. Analysis of such hypothetical, alternative up-regulation profiles dem-
onstrates that they do not seem to perform better than the observed profile and
are in many respects inferior.
Many variations on this theme can be implemented. Foremost, the analysis

described in the preceding paragraphs considered just one gene expression level
(at 15–20 minutes) and only assessed the resulting steady state of the system under
conditions of unlimited glucose in the medium. In expanded investigations, one

could study transients and delays, change gene expression dynamically according
to the published microarray data, compare constant glucose availability and glu-
cose depletion, or combine some or all of these aspects. In addition, one could
account for metabolic changes brought forth by temperature-dependent changes
in enzyme activities, as we discussed before. The next section describes one option
for such a combined analysis. The interested reader should also study [10] and [31]
for different analyses of similar data on heat stress in yeast.
MULTISCALE ANALYSIS
In the first part of the analysis, we used the temperature-dependent changes in key
enzymes to model the heat stress response. While the results looked reasonable,
very few concrete data were available for comparison. Furthermore, this strategy
entirely ignored the fact that a significant component of the heat stress response is
controlled at the genomic level. In the second part of the analysis, we concentrated
on gene expression and ignored direct temperature effects on enzymes. This strat-
egy also led to some insights, but suffered from the uncertainties surrounding the
assumption that gene expression can be converted directly and quantitatively into
enzyme activities. This assumption was necessary because flux measurements over
time were not available.
11.6 In Vivo NMR Profiles

New methodologies are capable of obtaining additional types of information.
Specifically, it is possible to measure entire time series of concentrations in vivo,
using NMR spectroscopy (see Chapter 8). Fonseca and collaborators [34] used this
methodology to study heat stress in yeast. They grew the cells to a stable population
size and then provided them with a bolus of 13C-labeled glucose at a favorable tem-
perature of 30°C. While the bolus was taken up by the cells, the investigators
measured selected metabolites that were produced from the glucose substrate,
including trehalose, G6P, fructose 1,6-bisphosphate (FBP) as an indicator of
glycolysis, and final products such as ethanol, acetate, and glycerol (Figure 11.13).
A second bolus of labeled glucose was given some while after the first bolus had
been taken up, but at this time the yeast cells were under heat stress. Finally, the
temperature was returned to its normal value of 30°C and another set of measure-
ments was taken. Typical results are shown in Figure 11.14. Significant features of
this experimental set-up are that glucose input is not constant but the external glu-
cose is quickly depleted, and the cells are not growing under the given conditions.
Inspection of the results from this experiment leads to several insights. First, the
cells quickly take up the glucose and essentially immediately convert it into G6P,
FBP, trehalose, as well as end products like ethanol, acetate, and glycerol. G6P is not
seen in Figure 11.14, because it requires labeling with glucose labeled on carbon 6,
but it shows a time profile similar to that of FBP (Figure 11.15). Second, production
of trehalose is much greater during heat stress. Instead of a peak of 4 mM under
normal conditions, the cells generate new trehalose up to about 8 mM. While this
doubling is significant, the final amount seems low in comparison with ample data
from the literature that suggest much higher amounts. Under recovery conditions,
trehalose production is similar to that under normal conditions during the first
phase. Closer analysis also shows that the uptake of glucose is slightly faster under
heat stress conditions.
Some of these and earlier results make intuitive sense, but some are rather sur-
prising. We know that much of the heat stress response is controlled at the genome
level. However, if genetic control were the driver, it would be impossible to see
increased trehalose production within 2 minutes of glucose availability, because
transcription and translation together take about 20 minutes in yeast. But if treha-
lose can be produced in large amounts without any specific action or intervention of
genes, why do the cells expend the physiological costs of up-regulating the genes of
the trehalose cycle at all? Furthermore, we have already noted that three of the tre-
halose enzymes are heat-dependent, with T6P synthase and T6P phosphatase about
MULTISCALE ANALYSIS 319
Figure 11.13 Model structure used to

Glc X1
analyze time-series data from in vivo
outside NMR experiments. Measured quantities are
indicated by ellipses. (Adapted from Fonseca
– F1 inside L, Sánchez C, Santos H & Voit EO. Mol. BioSyst.
F11
7 [2011] 731–741. With permission from The
Glc X2 Tre X8 F10 Royal Society of Chemistry.)
–
F2 T6P X7
F9
–
F15 F3 F5
+ glycogen
PPP X12 G6P X3 G1P X4 UDPG X5
F7 X6
F4 F6
–
F12
– F8
leakage F14
FBP X9
X11
inside
F13 outside
EtOH + Ac + Gly
X10
2.5 times as active and trehalase half as active as under optimal temperature condi-
tions (see Figure 11.7 and [35]). These effects of temperature make intuitive sense,
because all three lead to increased trehalose production, which is a desirable goal.
But are they sufficient to mount an effective response as we observe it in Figure
11.14? A question of this type is difficult to answer with hard thinking or even with
wet experiments alone, which suggests the use of a modeling analysis.
In principle, such an analysis is straightforward. Once we have a model that fits
the data under normal conditions, we can simply change the three enzyme activi-
ties in the model (by adjusting the corresponding rate constants) and solve the dif-
ferential equations to see whether the heat response is modeled with sufficient
accuracy. Furthermore, we can explore whether other processes might be directly
affected by heat stress. For instance, closer inspection of Figure 11.14 seems to
indicate that the uptake of glucose from the medium occurs slightly faster under
30°C 39°C 30°C

300
60 glucose
Glc, Ac, Gly (mM)
200
ethanol (mM)
40 acetate
ethanol
100
20
glycerol Figure 11.14 In vivo NMR profiles of
metabolites during the consumption
0 0 of three pulses of glucose (65 mM). This
experiment measured glucose, ethanol,
15
acetate, and glycerol in the medium, and
FBP and trehalose inside the cells. Each
FBP, trehalose (mM)
trehalose pulse of glucose was supplied under normal

10
(30ºC), heat stress (39ºC), and recovery
(normal temperature) conditions. The pale
5 FBP orange band indicates the heat stress period.
(Adapted from Fonseca L, Sánchez C, Santos
0 H & Voit EO. Mol. BioSyst. 7 [2011] 731–741.
0 10 20 30 40 50 60 70 80 90 100 110 With permission from The Royal Society of
time (min) Chemistry.)
300 Figure 11.15 Metabolic profile of G6P,

glucose FBP, and end products obtained from an
60
ethanol experiment with [6-13C]glucose. The blue
Glc, Ac, Gly (mM)
200 and pale orange blocks indicate optimal
ethanol (mM)
40 and heat stress conditions. (Adapted
acetate from Fonseca L, Sánchez C, Santos H &
100 Voit EO. Mol. BioSyst. 7 [2011] 731–741.
20 glycerol
With permission from The Royal Society of
Chemistry.)
0 0
glucose 6P
15
FBP, G6P (mM)
10
5 FBP
0
0 10 20 30 40 50 60 70
time (min)
heat stress conditions. In fact, it is not difficult to perform a comprehensive simula-

tion study in which one, two, several, or all enzyme activities are made inducible by
heat [34].
Before we embark on such modeling of the heat stress response, let us return to
an earlier question, namely, why the cell might up-regulate any of the pertinent
genes if a metabolic response alone is sufficient. An additional set of experiments,
performed specifically to shed light on this question [34], may provide some clues.
Because metabolic effects of changes in gene expression only begin to materialize
about 20–30 minutes after the initiation of heat stress at the earliest, Fonseca grew
cells and exposed them to 39°C heat stress during the last 40 minutes of growth,
thereby allowing them time to adjust their gene expression profile. He then repeated
the earlier experiment of glucose boluses under normal and then under heat
conditions.
The results for these preconditioned cells are shown in Figure 11.16 (right col-
umn) and should be compared with those from normally grown control cells (Fig-
ure 11.16, left column). The differences are striking. First, glucose uptake is much
slower (note the different timescales). In the normally grown population, the glu-
cose bolus at 30°C is depleted within about 8 minutes, while the preconditioned
population needs almost twice as long. At 39°C, the uptake by the control cells is
essentially the same as at 30°C, but in the preconditioned cells it has sped up from
about 13 minutes to about 8 minutes. Second, the FBP profiles are distinctly differ-
ent. For the control cells, similar peaks of about 18 mM are reached within 2 min-
utes at both temperatures. By contrast, the peaks in the preconditioned cells are
barely recognizable. Third, and maybe most importantly for this case study, the tre-
halose profiles are dramatically different. In the control cells, trehalose reaches
peaks of about 4 mM at 30°C and 10 mM at 39°C, whereas in the preconditioned
cells, the peaks reach about 20 mM (30°C) and over 90 mM (39°C). These compara-
tive results suggest that preconditioning by heat has significant effects later in life,
when the cells are again exposed to heat stress. The preconditioning seems to pre-
pare them for a redistribution of fluxes and, especially, a much more forceful
response by the trehalose cycle.
11.7 Multiscale Model Design

Now let us design a model that integrates these findings and possibly explains
them. The earlier models in (11.1) and (11.2) had the correct basic structure, but it
is advantageous for our purposes here to make the separate roles of genes and
enzymes explicit. Moreover, the NMR experiments do not include all components
of the earlier model, such as UDPG and T6P, but do include others, such as etha-
nol and acetate (see Figure 11.13), so that it is best to set up a correspondingly
control cells preconditioned cells Figure 11.16 Experimental time courses

250
30°C 39°C 30°C 39°C (symbols) and model fits for key
60 representatives of the trehalose cycle.
200
end products (mM)

The dark blue symbols and curves in the top
150 row show glucose in the medium, while the
glucose (mM)
40
dark red symbols and curves show all end
100 products combined. Their concentration is
20 much higher than that of glucose, because
50
they are trioses (three-carbon sugars), while
0
glucose is a hexose (six-carbon sugar).
0
The profiles of FBP (lighter red curves and
20 symbols in the center row) are dramatically
different for control cells and preconditioned
FBP, G6P (mM)
15 cells. G6P (green curves and symbols in

the center row) was not measured for the
10 preconditioned cells, but the profile is
expected to mirror that of FBP. Note the
5
striking differences in trehalose production
0 between control and preconditioned
cells (bottom row). Also note the different
timescales for control and preconditioned
75
cells, which clearly indicate that the control
trehalose (mM)
cells take up glucose more efficiently.

50 (Adapted from Fonseca L, Sánchez C, Santos
H & Voit EO. Mol. BioSyst. 7 [2011] 731–741.
25 With permission from The Royal Society of
Chemistry.)
0
0 5 10 15 20 25 30 10 20 30 40 50
time (min) time (min)
adjusted model. This situation is quite common: as soon as the focus of an analy-
sis shifts, it might be advisable to set up an amended model that contains exactly
those components of the system for which we have data. This switch does not
mean that one model is better than the other; it just means that it is often easier to
develop models that more closely match the new focus. The new model design
does not have to start from scratch, though, because information used for one
model can often be used in the alternative model as well. Furthermore, compari-
sons of model results in overlapping aspects can serve as some sort of quality con-
trol. Ultimately, it would of course be desirable to have one model covering all
different conditions, but such a high standard is not often achievable, at least not
in the beginning.
The new model has to address four scenarios, namely control and precondition-
ing both at 30°C and at 39°C. Our hypothesis is that the switch from 30°C to 39°C for
either control or preconditioned cells is due to heat-induced changes in the activities
of some of the enzymes. A sub-hypothesis is that changes in the three enzymes iden-
tified by Neves and François [35] are sufficient, while the alternative sub-hypothesis
postulates that more than these three enzymes must change in activity to launch the
observed response.
The change from control to preconditioning is complicated, because the govern-
ing mechanisms have not been identified experimentally. Nonetheless, we can
hypothesize that heat preconditioning affects the changes in the regulation of perti-
nent genes, which in turn have a longer-term effect on the amounts of correspond-
ing enzymes. To test this hypothesis with the model, we can introduce into each
power-law term of the model a factor representing the amount of enzyme. Under
normal conditions, this factor is set to 1, while it takes a so-far unknown value for
heat-preconditioned cells. Thus, each term in the model equations contains, as
always, a rate constant and the contributing variables, raised to appropriate kinetic
orders, and, in addition, a factor for heat-induced changes in enzyme activity and a
factor for preconditioning.
The second issue with the original model is the fact that the parameter values
had been collected from different sources and experiments executed under
different conditions. In particular, the literature data corresponded to
constant-glucose conditions, whereas the new time-series data are obtained dur-
ing glucose utilization. Furthermore, several parameter values of the original
model had to be assumed, based on collective experience with these types of path-
ways, but could not be tested further, since no suitable data were available. Because
of these uncertainties, it is doubtful that the original parameter values are optimal
for the new, rather different sets of experiments. Instead, we can use the new time-
series data and deduce parameter values directly with the inverse methods dis-
cussed in Chapter 5. This estimation procedure is not at all trivial, but because we
are at this point not interested in algorithmic aspects, we just skip these issues and
focus on the results (see [34] for details).
The new model has a GMA structure that is similar to that in (11.2) but contains
a slightly different set of variables, as well as new parameter values deduced from
the new time-series data (Table 11.3). The symbolic form of the model is as
follows:
B γ 1 τ P1 X 1h1 X 3hr1 Q 1(T −30)/10 ,


B γ 2 τ 2 X 2 2 X 7 2 ,
P h hr
− F1 Vext ,  (T −30 )/10

B γ 3 τ X 3 Q 3
P3 h3
 ,
( F1 + 2 F11 − F2 ) Vint , 
B γ 4 τ X 4 ,
P4 h4

( F2 + F4 − F3 − F9 − F12 − F15 ) Vint , B γ 5 τ P5 X 4h5 ,
( F − F + F − F + F ) V , 
 3 4 6 5 8 int B γ 6 τ P6 X 5h6 ,
( F5 − F6 − F7 − F9 ) Vint , 
 B γ 7 τ P7 X 5h7 X 3hr3 ,
( F7 − F8 ) Vint , 
X =  F = B γ 8 τ P8 X 6h8 X 3hr4 X 5hr5 ,
( F9 − F10 ) Vint ,  (T −30 )/10
B γ 9 τ X 3 X 5 X 2 Q 9
P9 h9 h10 hr6
 ,
( F10 − F11 ) Vint , B γ τ P10 X h11 Q (T −30)/10 ,
  10 7 10
( F12 − F13 − F14 ) Vint , B γ 11 τ P11 X 8h12 Q11(T −30 )/10

,
2 F V , 
 13 ext B γ 12 τ X 3 ,
P12 h13
 F14 Vint ,  (T −30 )/10

B γ 13 τ 13 X 9 14 Q11
P h
 ,
 F15 Vint, 
B γ 14 τ X 9 ,
P14 h15
0.05 F .
 12
(11.4)
The left set of equations in (11.4) are the differential equations, where, for instance,
the first row is to be read as X 1 = − F1 /Vext . The right set of equations are all the fluxes,
sorted according to Figure 11.13. The fluxes contain the usual rate constants γ and
the dependent variables Xi with their kinetic orders, h and hr, which characterize
substrate dependence and regulatory influences, respectively. The equations con-
tain several other quantities: Vext is the volume of the cell suspension (50 mL), Vint is
the total intracellular volume (7.17 mL), and B is the biomass in the reactor (3013 and
2410 mg of dry weight under normal temperature and heat, respectively). The fac-
tors τ Pi represent changes in protein amounts due to altered gene expression. We set
τ = 1 for cells grown under control conditions and τ = 2 for heat-preconditioned
cells, so that all τ Pi correspond to powers of 2, as is typical in gene expression stud-
ies. The values of the different Pi were estimated from the measured time series
together with the other parameters. The temperature dependence of each affected
enzyme is explicitly modeled as Qi(T −30) 10. Each Qi is a typical temperature coefficient
(Q10) for enzymatic reaction i, which depends on the actual ambient temperature T
(either 30 or 39°C) and represents the change in enzymatic activity brought about by
a 10°C increase in temperature.
The model fits the data under normal temperature and heat preconditioning
quite well (see Figure 11.16). In addition to confirming this fit, we can use the model
TABLE 11.3: MODEL PARAMETERS (RATE CONSTANTS AND KINETIC ORDERS) AND PROTEIN CHANGES OBTAINED FROM
TIME-SERIES DATA BY MEANS OF OPTIMIZATION*
Flux Model step Rate constant γ Kinetic orders for substrates† Kinetic orders for regulators‡ Fold change (2Pi) Qi
1 HXT 2.87 × 10−5 0.526 (30°C) 0.472 (39°C) −0.002 (G6P) 0.7 1.57
2 HXK 1.90 × 10 −4
0.510 −0.209 (Tre6P) 9.2
3 PGMF 5.66 × 10−6 0.400 20.7 1.48
4 PGM R
3.13 × 10 −5
0.471 17.3
5 UGPF 3.58 × 10−5 0.767 16.2
6 UGP R
1.31 × 10 −5
0.159 26.0
7 GSY 9.43 × 10−7 0.459 0.000 (G6P) 0.9
8 GPH 6.94 × 10 −8
0.311 −0.002 (G6P) −0.001 (UDPG) 61.8
9 TPS1 1.19 × 10−6 0.659 (G1P) 0.625 (UDPG) 0.000 (Glc) 21.5 2.48
10 TPS2 3.24 × 10 −6
0.361 14.2 2.35
11 NTH 1.99 × 10−7 0.082 4.9 0.42
12 PFK 2.89 × 10 −5
0.693 1.0
13 “FBA”§ 6.13 × 10−5 0.369 1.2 1.26
14 Leakage 5.54 × 10 −6
0.672 4.1
15 ZWF 1.45 × 10−7 0.693 1.0
*From Fonseca L, Sánchez C, Santos H & Voit EO. Complex coordination of multi-scale cellular responses to environmental stress. Mol. BioSyst. 7 (2011)
731–741. With permission from The Royal Society of Chemistry.
†
The parentheses show the temperature associated with the kinetic order for glucose transport or the substrate associated with each kinetic order.
‡
The parentheses show the regulating metabolite.
§
”FBA” designates the collection of enzymatic steps between fructose 1,6-bisphosphate aldolase and the release of end products.
to explore the questions we posed earlier. First, are the heat-induced changes in
enzyme activities, which we discussed earlier, sufficient for an adequate heat stress
response? The answer is “Yes” and “No.” It is “No” in a sense that the observed
changes in the trehalose-associated enzymes (TPS1, TPS2, and NTH1/2) alone are
not sufficient to explain the observed trehalose response, because the model indi-
cates a much weaker response than that observed (Figure 11.17). However, the
answer is a conditional “Yes” if the measured changes in these enzymes are accom-
panied by a slight reduction in glucose uptake (to about 60% hexose transporter
activity; see also Figure 11.16), a 50% increase in phosphoglucomutase and a slighter
(about 25%) collective increase in glycolytic activity (see Figure 11.17). Interestingly,
in vitro activity assays of glycolytic enzymes show similar trends [34].
The data analysis with the model also makes predictions regarding alterations in
protein amounts due to preconditioning. No direct data are available for compari-
sons. However, as we discussed earlier in the chapter, one may translate alterations
in gene expression into amounts of protein, at least as a crude substitute. The com-
parison of the quantitative changes in protein levels with literature data is quite
favorable (Table 11.4), even though they were obtained with entirely different
methods.
Table 11.4 exhibits one striking difference, namely in glycogen synthase. In the
time-series (bolus) experiments, the activity of this enzyme is slightly decreased,
while it is strongly increased under constant-glucose conditions. The reasons for
this difference are unclear, but it could be that the constant glucose supply was more
than the cells needed at the time, and therefore allowed them to store some of the
material as glycogen. Also, one must keep in mind that the constant-glucose experi-
ments were performed with growing cells under steady-state conditions, while the
dynamic experiments used resting cells. Finally, the inferred change in FBP used for
non-glycolytic pathways, which in the time-series experiments increased about
fourfold, has no analog in the mRNA study.
Summarizing the results from the time-series analysis, the model seems to cap-
ture the responses of the yeast cells under different conditions quite well. In particu-
lar, it shows that heat-induced changes in enzyme activities are sufficient to mount a
control cells preconditioned cells Figure 11.17 Testing the assumption that
250
30°C 39°C 30°C 39°C exclusively TPS1, TPS2, and NTH1/2 are
60 affected by heat [35]. Experimental time
200
end products (mM)

courses (symbols) and model fits for key
150 representatives of the trehalose cycle, under
glucose (mM)
40
the assumption that only the three enzymes
100 TPS1, TPS2, and NTH1/2 are affected by heat.
20 Symbols and colors are the same as in Figure
50
11.16. (Adapted from Fonseca L, Sánchez
0
C, Santos H & Voit EO. Mol. BioSyst. 7 [2011]
0
731–741. With permission from The Royal
20 Society of Chemistry.)
FBP, G6P (mM)
15
10
75
trehalose (mM)
50
25
0
0 5 10 15 20 25 30 10 20 30 40 50
time (min) time (min)
short-term response, but preconditioning makes such a response much stronger and
effective. Furthermore, the cellular adjustments between heat conditions and
between normally grown and preconditioned cells seem to be reasonable.
11.8 The Trehalase Puzzle

Have all questions been addressed here? In biology, the answer to such a ques-
tion is of course seldom “Yes.” Here, we still have an unresolved puzzle surround-
ing the role of trehalase in the heat stress response. There is clear evidence that
TABLE 11.4: CHANGES IN ENZYME AMOUNTS BETWEEN NORMALLY GROWN

AND PRECONDITIONED CELLS, COMPARED WITH HEAT-INDUCED CHANGES IN
mRNA LEVELS REPORTED IN THE LITERATURE
Enzyme Change in amount inferred from time data* Change in mRNA†
Forward reaction Reverse direction
Hexokinase 9 — 8
Phosphoglucosemutase 21 17 16
UDPG pyrophosphorylase 16 26 16
Glycogen synthase 0.9 — 16
Glycogen phosphorylase — 62 50
T6P synthase 21 — 12
T6P phosphatase 14 — 18
Neutral trehalase 5 — 6
Phosphofructokinase 1 — 1
*Expressed as ratio between protein amount after preconditioning and protein amount in normally
grown cells.
†
Data from [2].
the genes coding for trehalase (NTH1/2 and ATH1) are immediately and strongly
up-regulated under heat stress. This up-regulation seems at first counterintuitive,
because it suggests an increase in the degradation of a metabolite that is
clearly needed. We have also seen that the activity of trehalase is directly reduced
by heat. Thus, there are concurrent but apparently counteracting changes in
response to heat stress. One way to reconcile the opposing changes is to have a
closer look at the timing of these processes. The direct reducing effect of heat on
the activity of the enzyme very quickly allows the necessary accumulation of
trehalose. After about 20 minutes for transcription and translation, the increased
gene expression results in a higher amount of enzyme, which apparently com-
pensates for the heat-induced reduction in enzyme activity and keeps trehalose
production and degradation in balance. More importantly, as soon as the heat
stress ceases, the heat-induced reduction in activity stops, while the amount of
enzyme is still increased. This combination of a large amount of enzyme with
uninhibited activity allows the cell to remove trehalose very quickly. This removal
of unneeded trehalose is apparently crucial for a healthy reentry into normal
physiological operation [19, 20].
While this seems reasonable, nature is once again not that simple. In both nor-
mally grown and preconditioned cells, the trehalose degradation profile does not
seem to differ much under normal temperature, heat, and recovery conditions,
although it might be slightly slower during heat stress. Intrigued by these observa-
tions, Fonseca and co-workers [34] performed an experiment in which first a bolus
of [2-13C]glucose (labeled on carbon position 2) and then a bolus of [1-13C]glucose
were given during heat stress. NMR spectroscopy can distinguish these two forms,
and analysis of the data revealed the puzzling observation that the “new” treha-
lose appears to be degraded independently of the previously generated trehalose
(Figure 11.18). This distinction also seems to occur during recovery, where it
appears that the new trehalose is degraded quickly, while the old trehalose is
degraded very slowly and in an essentially linear pattern that continues the
dynamics that began during heat stress (see Figure 11.14). Our mathematical
model assumes that trehalase does not differentiate between the two “types” of
trehalose, and therefore it cannot offer an explanation. A possible reason for the
differential degradation could be that, under heat conditions, trehalose binds to
membranes [40] or associates with hydration cages that form around proteins
during heat stress [41]. It could be that it is more difficult for trehalase to access the
“old” trehalose, which was formed under heat stress. Further dedicated experi-
ments will be necessary to provide definite answers to these questions. But would
the effort of designing new experiments be worthwhile, or is the logic of the argu-
ment flawed from the onset?
Instead of waiting for further experimental evidence, let us formulate a specific
hypothesis regarding the intriguing degradation patterns of different pulses of tre-
halose and test with a model analysis whether this hypothesis could be true, at least
40 80
glucose
30 60
Glc, Ac, Gly (mM)
ethanol (mM)
20 ethanol 40
10 acetate 20
glycerol
0 0
Figure 11.18 Metabolic profile obtained
10 by supplying yeast cells with differently
labeled glucose isotopomers. In the first
trehalose (mM)
6
pulse, unlabeled glucose was used under
normal conditions (data not shown). Then
4 trehalose cells were subjected to heat stress and two
2 pulses of glucose were supplied, using [2-13C]
0 and [1-13C]glucose. (Adapted from Fonseca
0 10 20 30 40 50 60 70
L, Sánchez C, Santos H & Voit EO. Mol. BioSyst.
7 [2011] 731–741. With permission from The
time (min) Royal Society of Chemistry.)
in principle. Thus, suppose the generated trehalose can be free (TF) or bound (TB) heat
to intracellular structures, such as membranes, and that the enzyme trehalase (E)
primarily (or exclusively) degrades TF (Figure 11.19). Suppose further that the
association (binding) constant ka between enzyme and substrate is normally ka
much smaller than the dissociation constant kd, which itself has a low value. If so, trehalose
production
then most trehalose at optimal temperature is in the TF state, and its degradation TF TB
follows something like a power-law or Michaelis–Menten process. Now let us make
the assumption that heat stress increases ka, so that it greatly exceeds kd. This rever-
sal of association and disassociation causes most trehalose to assume the TB state, kd
which, similar to Styrofoam packaging, protects membranes and other structures
from disintegration. The conversion of TB into TF is slow owing to the value of kd, trehalose
degradation
which is now relatively low in comparison with ka. Because the concentration of TF
is low, the degradation process is substrate-limited and essentially independent of Figure 11.19 Simplified model diagram for
E. Indeed, the concentration of TF is governed by the small kd and therefore more exploring the trehalase puzzle. The specific
or less constant, because there is a lot of TB. A second trehalose pulse under heat hypothesis tested here is that trehalose may
stress leads to twice as much TB and therefore to roughly a doubled amount of exist in two states, namely, as free (TF) or as
kd × TB that becomes TF and can be degraded. Overall, this mechanism would lead bound (TB) trehalose.
to a process that appears to degrade the two trehalose pulses independently. By
contrast, if the second trehalose pulse is given under cold conditions, it is immedi-
ately degraded according to some biochemical process, while the “old” trehalose is
still mostly bound.
An intuitive speculation like that in the previous paragraph might be attractive,
but it is prone to logical or numerical inconsistencies. However, given the specificity
of the hypothesis, it is not difficult to formulate it as a small mathematical model of
the association/disassociation system. Reality is most likely much more complex,
but simplicity in such a model analysis is key. After all, the analysis is only supposed
to provide proof of concept that the distinction of TF and TB could possibly explain
the intriguing trehalose degradation pattern in Figure 11.18. Thus, we translate the
diagram of Figure 11.19 directly into a model, which may have the following simple
form:
i
(TF ) = kd (TB) − ka (TF ) + Bolus − βE ⋅ (TF )h,
i
(11.5)
(TB) = ka (TF ) − kd (TB).
Here, the conversions between TF and TB are simple linear processes with rates ka
and kd, the new production of trehalose is represented by the independent variable
Bolus, and the degradation of TF is modeled as a power-law function with rate con-
stant β, trehalase as enzyme E, and TF as substrate with a typical kinetic order h. Let
us suppose the initial values are TF(0) = TB(0) = 0.1, that kd = 0.01, and that
ka = 0.002 or ka = 5 for optimal and heat stress conditions, respectively. For normal
conditions, we set E = 1, but we are not even planning to change the enzyme amount
or activity, so that we could really eliminate E from the system altogether. Further-
more, we define free trehalose degradation with β = 0.15 and h = 0.1 or h = 0.8 for
optimal and heat stress conditions, respectively, and define for convenience the
total trehalose concentration TT = TF + TB. This variable does not require a differ-
ential equation. Notice that all these settings are arbitrary, but in line with typical
parameters in power-law modeling (see Chapter 4).
The model can be used to analyze the responses to two boluses. For instance,
one might start the system under heat stress with a bolus of freshly synthesized tre-
halose. In the first simulation, a second bolus is given after 60 minutes, while heat
continues. The second simulation is similar, but, with the second bolus, the tem-
perature is set to optimal temperature conditions (Figure 11.20). Indeed, the simple
model shows trehalose degradation patterns very similar to those observed (see
Figures 11.14 and 11.18). Of course, this model output is no proof that two trehalose
states exist, but it provides and affirms a possibly testable hypothesis that could lead
to an explanation of the intriguing pattern.
CONCLUDING COMMENTS 327
(A) Figure 11.20 Results of model simulations

50 of trehalose degradation under the
hypothesis of free (TF) and bound (TB)
trehalose (see Figure 11.19, (11.5), and the
trehalose concentration
TT discussion in the text). Two boluses were
administered (indicated by suffixes1 and
2), and TF, TB, and total trehalose (TT) were
25 TB2 plotted. (A) corresponds to Figure 11.18
and represents the situation where both
boluses are given under heat stress, while
TB1 (B) corresponds to Figure 11.14, where the
second bolus is given under cold conditions.
TF1 TF2
(Adapted from Fonseca L, Sánchez C, Santos
0 H & Voit EO. Mol. BioSyst. 7 [2011] 731–741.
(B) With permission from The Royal Society of
50 Chemistry.)
trehalose concentration
TT
25
TB1
TF2
TF1 TB2
0
0 60 120
time (min)
CONCLUDING COMMENTS
Although heat stress has been studied extensively over the past few decades, many
questions regarding the overall coordination of stress responses continue to puzzle
the scientific community. One reason for the ongoing challenge is that typical stress
responses are systemic. That is, they involve a wide repertoire of mechanisms at dif-
ferent levels of biological organization, and these mechanisms do not operate in
isolation but are interdependent, functionally interwoven, and synergistic. Further-
more, the responses occur on different, yet overlapping, timescales.
We have discussed specific aspects of the heat stress response in yeast, along
with small, simplified models that can aid our intuition and help us explain certain
aspects, including sometimes confusing observations. Although this system is com-
paratively small and well understood, we have seen how challenging a detailed
understanding of a coordinated response is. As a case in point, we have seen that
even a narrow focus on the metabolic aspect of the response cannot be validly
addressed without looking at processes in the realms of genes and proteins. Fur-
thermore, the responses in the present case depend critically on the history of the
cells. If these have been preconditioned by heat stress earlier in life, their metabolic
responses turn out to be much stronger than if the cells were naive. This observation
and its broader implications should give us reason to be very cautious with sweep-
ing statements regarding cellular responses or diseases, because we often have no
good information regarding the history of a cell or organism, although this history
may have a critical role in its future fate.
As we continue to generate increasingly more comprehensive and sophisti-
cated data, it is becoming ever clearer that the unaided human mind will soon
reach its limits in integrating the numerous and diverse pieces of quantitative
information into a functioning response and that only computational models will
be able to organize this information to provide a fuller understanding of complex
response systems.
EXERCISES
11.1. Implement the model (11.1) in a software 11.13 Predict what the influence of the size of the external
program, such as MATLAB•, Mathematica•, glucose pool is on the dynamics of the model
or PLAS. Check whether the system has one or (11.2). Check your predictions with simulations.
more steady states and whether (any of ) these 11.14. Discuss which assumptions must be made,
steady states are stable. Revisit the meaning of explicitly and implicitly, when one substitutes
eigenvalues and their real and imaginary parts in heat-induced changes in enzyme activities with
this context. Investigate the sensitivities of the corresponding changes in gene expression.
system and recall what exactly the magnitude and
11.15. Study the article “Complementary profiling of
sign of a sensitivity or gain means.
gene expression at the transcriptome and
11.2. Perform perturbation experiments with the model proteome levels in Saccharomyces cerevisiae” by
(11.1) in order to develop a feel for how it Griffin and collaborators [36] and discuss implica-
responds to slight alterations in numerical tions for the modeling efforts in this chapter. Look
features. Investigate transient features of the in particular for genes associated with glycolysis.
model with simulations of temporary or persistent
11.16. Speculate why glycogen phosphorylase might be
increases or decreases in some of the variables or
so strongly up-regulated under heat stress
parameters.
conditions.
11.3. Expand your program to normalize all variables
with respect to their steady-state values. What is 11.17. Discuss the impact of the half-lives of mRNAs and
the main advantage of this normalization? enzymes on the dynamics of the heat stress
response.
11.4. Simulate excess and shortage of glucose with the
model (11.1) and report your findings. 11.18. Carry out simulations that combine alterations in
gene expression. Assume that changes in gene
11.5. Simulate mutations in TPS1 or TPS2 with the expression as shown in Table 11.5 can be directly
model (11.1) and report your findings. Search the translated into corresponding changes in enzyme
literature for data and compare them with your activities. Explore what happens if only one or two
modeling results. enzyme activities are changed at a time.
11.6. Simulate mutations leading to decreased activities Summarize your findings in a report.
(80%, 20%) of all enzymes, one at a time and two 11.19. Combine changes in gene expression with
at a time. Discuss to what degree single mutations heat-induced changes in enzyme activities, as
are predictive of double mutations. observed by Neves and François [35]. Assume at
11.7. How would you simulate alterations in enzymatic first a constant supply of external glucose. In a
steps that are catalyzed by two or more enzymes, second set of simulations, consider glucose
such as hexokinase I, hexokinase II, and depletion.
glucokinase? Search the literature for responses to
11.20. Several studies in the literature [2, 31, 32] have
deletions of one of several enzymes catalyzing the
analyzed the expression of various genes coding
same reaction. Write a brief summary report.
for enzymes of the glycolytic pathway and the
11.8. It has been observed that TPS1 deletion mutants trehalose cycle using methods described in the
are unable to grow on glucose substrate. Search text, as well as other techniques. Summarize the
the literature for details. Discuss how this highlights of these investigations in a report.
phenomenon can be demonstrated with the
11.21. Temperature-induced changes in enzymes occur
model or why such an analysis is not possible.
very rapidly, while the expression of genes and the
11.9. Why is the trehalase reaction in the equations of synthesis of new proteins take some while.
glucose and trehalose represented with different Discuss the implications of these timescale issues
terms? for models of the heat stress response.
11.10. Explain with logic and with simulations why X1 in 11.22. Set up a model to test the hypothesis that
(11.2) shows a small, initial overshoot when the trehalose may be present in free or bound form.
amount of enzyme ZWF1 in the equation for G6P Use the model (11.5). First set β and Bolus = 0 to
is multiplied by 4 (see Figure 11.11). model the absence of both input and degradation.
11.11. Using the model (11.2), test whether changes in Study what happens if ka = 0.01 or ka = 5. Set
only one or two enzyme activities would result in Bolus = 5 for 3 time units and study the effect. Set
appreciable increases in trehalose. Report how the Bolus = 5 for 5 time units and activate the treha-
metabolic profiles are affected by each change. lase reaction by setting β = 0.15 and h = 0.1.
11.12. Using the model (11.2), explore different 11.23. Use the model (11.5) under heat stress conditions
alterations in enzyme activities. Assemble a table and compare the results with those under optimal
showing which alterations lead to significant temperature. In a similar study, start with heat
increases in trehalose. Add one sentence per conditions and then give a second glucose bolus
simulation describing other effects on the system. under either heat or cold conditions.
REFERENCES 329
TABLE 11.5: FOLD INCREASES IN ENZYME ACTIVITIES USED FOR MODELING THE
HEAT STRESS RESPONSE
Gene of enzyme Heat-induced fold increase
Variable Catalytic or transport step
or transporter in enzyme activity
HXT X8 Glucose uptake 8
HXK1/2, GLK X9 Hexokinase/glucokinase 8
PFK1/2 X10 Phosphofructokinase 1
ZWF1 X11 G6P dehydrogenase 6
PGM1/2 X12 Phosphoglucomutase 16
UPG1 X13 UDPG pyrophosphorylase 16
GSY1/2 X14 Glycogen synthase 16
GPH X15 Glycogen phosphorylase 50
GLC3 X16 Glycogen use 16
TPS1 X17 α, α-T6P synthase 12
TPS2 X18 α, α-T6P phosphatase 18
NTH X19 Trehalase 6
11.24. Read some of the literature on heat shock them into one of the models discussed here
proteins and design a strategy for incorporating or into a different type of heat stress response
them into one of the models discussed here or model.
into a different type of heat stress response 11.26. Draft a model for the network in Figure 11.1 and
model. develop a strategy for implementing it. Without
11.25. Explore the role of sphingolipids in heat actually designing this model, make a to-do list of
stress and design a strategy for incorporating steps and information needed.
REFERENCES
[1] Hohmann S & Mager WH (eds). Yeast Stress Responses. yeast to environmental changes. PLoS Comput. Biol. 6 (2010)
Springer, 2003. e1000674.
[2] Voit EO. Biochemical and genomic regulation of the [11] Palotai R, Szalay MS & Csermely P. Chaperones as integrators
trehalose cycle in yeast: review of observations and of cellular networks: changes of cellular integrity in stress
canonical model analysis. J. Theor. Biol. 223 (2003) and diseases. IUBMB Life 60 (2008) 10–18.
55–78. [12] Hahn JS, Neef DW & Thiele DJ. A stress regulatory network
[3] Paul MJ, Primavesi LF, Jhurreea D & Zhang Y. Trehalose for co-ordinated activation of proteasome expression
metabolism and signaling. Annu. Rev. Plant Biol. 59 (2008) mediated by yeast heat shock transcription factor. Mol.
417–441. Microbiol. 60 (2006) 240–251.
[4] Shima J & Takagi H. Stress-tolerance of baker’s-yeast [13] Dickson RC, Nagiec EE, Skrzypek M, et al. Sphingolipids are
(Saccharomyces cerevisiae) cells: stress-protective molecules potential heat stress signals in Saccharomyces. J. Biol. Chem.
and genes involved in stress tolerance. Biotechnol. Appl. 272 (1997) 30196–30200.
Biochem. 53 (2009) 155–164. [14] Görner W, Schüller C & Ruis H. Being at the right place at
[5] Hottiger T, Schmutz P & Wiemken A. Heat-induced the right time: the role of nuclear transport in dynamic
accumulation and futile cycling of trehalose in transcriptional regulation in yeast. Biol. Chem. 380 (1999)
Saccharomyces cerevisiae. J. Bacteriol. 169 (1987) 5518–5522. 147–150.
[6] Hannun YA & Obeid LM. Principles of bioactive lipid [15] Lee P, Cho BR, Joo HS & Hahn JS. Yeast Yak1 kinase, a
signaling: lessons from sphingolipids. Nat. Rev. Mol. Cell Biol. bridge between PKA and stress-responsive transcription
9 (2008) 139–150. factors, Hsf1 and Msn2/Msn4. Mol. Microbiol. 70 (2008)
[7] Jenkins GM, Richards A, Wahl T, et al. Involvement of yeast 882–895.
sphingolipids in the heat stress response of Saccharomyces [16] KEGG (Kyoto Encyclopedia of Genes and Genomes) Pathway
cerevisiae. J. Biol. Chem. 272 (1997) 32566–32572. Database. http://www.genome.jp/kegg/pathway.html.
[8] Cowart LA, Shotwell M, Worley ML, et al. Revealing a [17] BioCyc Database Collection. http://biocyc.org/.
signaling role of phytosphingosine-1-phosphate in yeast. [18] Galazzo JL & Bailey JE. Fermentation pathway kinetics and
Mol. Syst. Biol. 6 (2010) 349. metabolic flux control in suspended and immobilized
[9] Gasch AP, Spellman PT, Kao CM, et al. Genomic expression Saccharomyces cerevisiae. Enzyme Microb. Technol.
programs in the response of yeast cells to environmental 12 (1990) 162–172.
changes. Mol. Biol. Cell 11 (2000) 4241–4257. [19] François J & Parrou JL. Reserve carbohydrates metabolism
[10] Vilaprinyo E, Alves R & Sorribas A. Minimization of in the yeast Saccharomyces cerevisiae. FEMS Microbiol.
biosynthetic costs in adaptive gene expression responses of Rev. 25 (2001) 125–145.
[20] Nwaka S & Holzer H. Molecular biology of trehalose and [31] Vilaprinyo E, Alves R & Sorribas A. Use of physiological
the trehalases in the yeast Saccharomyces cerevisiae. constraints to identify quantitative design principles
Prog. Nucleic Acid Res. Mol. Biol. 58 (1998) 197–237. for gene expression in yeast adaptation to heat shock.
[21] Paiva CL & Panek AD. Biotechnological applications of BMC Bioinformatics 7 (2006) 184.
the disaccharide trehalose. Biotechnol. Annu. Rev. 2 (1996) [32] Voit EO & Radivoyevitch T. Biochemical systems analysis
293–314. of genome-wide expression data. Bioinformatics
[22] Lillford PJ & Holt CB. In vitro uses of biological 16 (2000) 1023–1037.
cryoprotectants. Philos. Trans. R. Soc. Lond. B Biol. [33] Eisen MB, Spellman PT, Brown PO & Botstein D. Cluster
Sci. 357 (2002) 945–951. analysis and display of genome-wide expression patterns.
[23] Hayashibara. The sugar of life. Time, November 22 (2010) Proc. Natl Acad. Sci. USA 95 (1998) 14863–14868.
Global 5; see also http://www.hayashibara.co.jp/. [34] Fonseca L, Sánchez C, Santos H & Voit EO. Complex
[24] Savageau MA. Biochemical Systems Analysis: A Study of coordination of multi-scale cellular responses to
Function and Design in Molecular Biology. Addison-Wesley, environmental stress. Mol. BioSyst. 7 (2011) 731–741.
1976. [35] Neves MJ & François J. On the mechanism by which a heat
[25] Torres NV & Voit EO. Pathway Analysis and Optimization in shock induces trehalose accumulation in Saccharomyces
Metabolic Engineering. Cambridge University Press, 2002. cerevisiae. Biochem. J. 288 (1992) 859–864.
[26] Voit EO. Computational Analysis of Biochemical Systems: [36] Griffin TJ, Gygi SP, Ideker T, et al. Complementary profiling
A Practical Guide for Biochemists and Molecular Biologists. of gene expression at the transcriptome and proteome
Cambridge University Press, 2000. levels in Saccharomyces cerevisiae. Mol. Cell. Proteomics 1
[27] BRENDA: The Comprehensive Enzyme Information System. (2002) 323–333.
http://www.brenda-enzymes.org/. [37] Stanford_Gene_Database. http://genome-www.stanford.
[28] Curto R, Sorribas A & Cascante M. Comparative edu/yeast_stress/data/rawdata/complete_dataset.txt.
characterization of the fermentation pathway of [38] Sur IP, Lobo Z & Maitra PK. Analysis of PFK3—a gene involved
Saccharomyces cerevisiae using biochemical systems theory in particulate phosphofructokinase synthesis reveals
and metabolic control analysis. Model definition and additional functions of TPS2 in Saccharomyces cerevisiae.
nomenclature. Math. Biosci. 130 (1995) 25–50. Yeast 10 (1994) 199–209.
[29] Schlosser PM, Riedy G & Bailey JE. Ethanol production in [39] Thevelein JM & Hohmann S. Trehalose synthase: guard
baker’s yeast: application of experimental perturbation to the gate of glycolysis in yeast? Trends Biochem.
techniques for model development and resultant changes Sci. 20 (1995) 3–10.
in flux control analysis. Biotechnol. Prog. 10 (1994) [40] Lee CWB, Waugh JS & Griffin RG. Solid-state NMR study
141–154. of trehalose/1,2-dipalmitoyl-sn-phosphatidylcholine
[30] Polisetty PK, Gatzke EP & Voit EO. Yield optimization of interactions. Biochemistry 25 (1986) 3739–3742.
regulated metabolic systems using deterministic branch- [41] Xie G & Timasheff SN. The thermodynamic mechanism
and-reduce methods. Biotechnol. Bioeng. 99 (2008) of protein stabilization by trehalose. Biophys. Chem.
1154–1169. 64 (1997) 25–34.
FURTHER READING
Eisen MB, Spellman PT, Brown PO & Botstein D. Cluster analysis and Hohmann S & Mager WH (eds). Yeast Stress Responses. Springer,
display of genome-wide expression patterns. Proc. Natl Acad. 2003.
Sci. USA 95 (1998) 14863–14868. Vilaprinyo E, Alves R & Sorribas A. Use of physiological constraints
Fonseca L, Sánchez C, Santos H & Voit EO. Complex coordination to identify quantitative design principles for gene expression
of multi-scale cellular responses to environmental stress. in yeast adaptation to heat shock. BMC Bioinformatics
Mol. BioSyst. 7 (2011) 731–741. 7 (2006) 184.
Gasch AP, Spellman PT, Kao CM, et al. Genomic expression programs Vilaprinyo E, Alves R & Sorribas A. Minimization of biosynthetic costs in
in the response of yeast cells to environmental changes. Mol. adaptive gene expression responses of yeast to environmental
Biol. Cell 11 (2000) 4241–4257. changes. PLoS Comput. Biol. 6 (2010) e1000674.
Görner W, Schüller C & Ruis H. Being at the right place at the right Voit EO & Radivoyevitch T. Biochemical systems analysis of
time: the role of nuclear transport in dynamic transcriptional genome-wide expression data. Bioinformatics 16 (2000)
regulation in yeast. Biol. Chem. 380 (1999) 147–150. 1023–1037.
Hannun YA & Obeid LM. Principles of bioactive lipid signaling: Voit EO. Biochemical and genomic regulation of the trehalose cycle
lessons from sphingolipids. Nat. Rev. Mol. Cell Biol. 9 (2008) in yeast: review of observations and canonical model analysis.
139–150. J. Theor. Biol. 223 (2003) 55–78.
Physiological Modeling:
The Heart as an Example 12
• Understand the basics of the anatomy and physiology of the heart
• Discuss the electrochemical processes during contraction and relaxation
• Formulate black-box models of heartbeats
• Analyze the role of perturbations and pacemakers in oscillators
• Explain the biology and mathematical formulation of action potentials
• Understand the principles and challenges of multiscale modeling
The heart is an incredible machine. A little bigger than a fist and weighing just about
half a pound, it pumps blood without ceasing, beating roughly 100,000 times every
day, between two and three billion times in a normal lifetime. It keeps on working
without us thinking about it, but beware if it skips even a few beats in a row! Many
societies have given the heart exalted roles, associating it with life, love, and stress.
Just think: what would Valentine’s Day be without that heart shape? The fact that it
is really quite different from the heart’s true anatomy (Figure 12.1) is easily forgotten
on February 14th.
Virtually unlimited sources of general and specific information are readily avail-
able about all aspects of the heart, from its cells and tissues to its anatomy, from its
normal electrical and mechanical functioning to the plethora of possible problems
and pathologies. The easiest access to matters of the heart is probably the large body
of textbooks, as well as an enormous variety of websites (for example, [1–5]), which
we will not specifically cite in the following.
The role of the heart is to move blood through the body and to supply the cells
with most of what they need. The blood carries oxygen and nutrients, and transports
unwanted chemicals away. The heart accomplishes this feat by a regular pumping
action that moves blood through a network of about 60,000 miles of arteries, veins,
and capillaries. Every minute, it pumps about 5 liters of blood, for a total of over 7000
liters per day. If the heart stops, we are in trouble. Coronary heart disease is respon-
sible for about half of all deaths in Western countries, and about 700,000 people die
from heart attacks every year in the United States alone.
This chapter discusses several aspects of the function of the heart with mathe-
matical models in the form of vignettes. Some of the models are very simple, others
are more complicated, and some of the simulation models that have been developed
for academic and commercial purposes are extremely complex, and we will only
mention them. To some degree, the usefulness of these models correlates with their
complexity, and the most sophisticated simulation models have become accurate
enough to allow realistic studies of the normal heart and some of its diseases.
332 Chapter 12: Physiological Modeling: The Heart as an Example
Nonetheless, the simpler models are of interest as well, because they give us insights
into the heart’s macroscopic oscillatory patterns and its basic electrical and chemical
properties, without overwhelming us with too much detail.
The strategy of the chapter is as follows. It begins with a general background
section that is rather superficial and by no means does justice to the amazing
features of the heart; a more detailed, very accessible introduction can be found in
[1]. Next, we introduce some simple black-box models that describe and analyze
heartbeat-like oscillations. Looking a little deeper into the physiology of the heart,
we find that its function is directly associated with oscillations in the dynamics of
calcium. Thus, the next step in our modeling efforts is to describe calcium oscilla-
tions with a minimalistic ordinary differential equation model and to explore how
far such a model can lead us in understanding heart function. Indeed, we will see
that quite simplistic models are sufficient to describe stable calcium oscillations.
However, these models are of course very limited and, for instance, do not account
for the electrochemical processes that we know to exist in the heart. We therefore
dig deeper, trying to understand how chemistry and the laws of electrophysiology
govern the contractions in individual heart cells, and how the activities of all heart
cells are coordinated in a fashion that results in proper heart pumping. We stop
short of the most sophisticated simulation models of the heart, because they are
naturally very complicated. Nonetheless, we do briefly discuss their underlying
principles. Thus, we will start at the macroscopic level, dive into metabolism,
electrophysiology, chemistry, and genetics, and ultimately return to the effects of
molecular changes on the functioning of the whole heart.
This gradual procedure of starting at a high level with a rather coarse model and
then studying selected aspects with a finer resolution and with an increasing level of
Figure 12.1 Model of the human heart. The
sophistication mimics the way we learn many things when growing up. Early on, we heart is a complicated muscle that encloses
distinguish static items from things that move, we learn to differentiate between four chambers. With each beat, the heart
living and engineered machines that move, we become able to distinguish cars from dynamically changes its shape and thereby
trucks, and cheap cars from expensive cars, and in some cases we learn to discern pumps blood to the lungs and throughout
the year a car was made even if the differences from one year to the next are very sub- the body. (Courtesy of Patrick J. Lynch under
tle. In modeling, a similar manner of learning has been proposed in biomechanics, the Creative Commons Attribution 2.5
where one might begin with a macroscopic, high-level assessment of locomotion, Generic license.)
and subsequently embed increasingly finer-grained models into these first
coarse models [6]. In this biomechanical context, the high-level models are called
templates. They help the modeler organize anchors, which are more complex
models representing anatomical and physiological features in a detailed fashion
and can be incorporated into the template models in a natural manner.
Numerous models of the heart have been developed over the past century that
could serve as templates or anchors. In fact, many investigators have devoted their
careers to heart modeling. Some giants in the field include James Bassingthwaighte,
Peter Hunter, Denis Noble, Charles Peskin, and Raimond Winslow. Moreover, the
electrochemical processes in heart muscle cells show a number of resemblances
with nerve cells, and descriptions of the dynamics of beating heart cells have been
inspired directly by experimental and computational research on nerve cells, which
began more than half a century ago with the pioneering work of Alan Lloyd Hodgkin
and Andrew Huxley and was extended and further analyzed by leaders in the field of
neuroscience, including Richard FitzHugh, Jin-Ichi Nagumo, and John Rinzel (see,
for example, [7]). Indeed, 50 years ago, Noble [8, 9] showed with the first computer
model of its kind that action potentials in the heart can be explained with equations
similar to those proposed by Hodgkin and Huxley.
HIERARCHY OF SCALES AND MODELING APPROACHES

It is quite obvious that the function of the heart spans many scales in size, organiza-
tion, and process speed. Beginning with the most evident scale of a whole organ, the
heart is a peristaltic pump that moves blood. This pump is composed of tissues,
which need to work in a tightly coordinated fashion. The tissues consist of several
types of cells, which have their specific functions within the concerted effort of
making the heart beat. Within and between the cells, there is a lot of controlled
movement of molecules and especially ions, which are directly associated with the
HIERARCHY OF SCALES AND MODELING APPROACHES 333
electrical activity of the heart. Inside the cells, we find metabolites and proteins,
some of which are found in many cells, while others are specific. Finally, there are
genes, some of which are only expressed in heart cells and can cause problems if
they are mutated.
Ideally, complementary models of all these aspects would seamlessly span the
entire spectrum from electrical and chemical features to intracellular mechanical
and metabolic events, from cells to cell aggregates and tissues, and from the differ-
ent tissues to a complete picture of the normal physiology and pathology of the
heart. However, such a complete multiscale story has not been written, and for the
demonstration in this chapter, we will select some aspects, while unfortunately
having to minimize—or even omit—others. For instance, we will not discuss in any
length very important issues of tissue mechanics and blood flow, three-dimensional
models of the anatomy and physiology of the heart, four-dimensional changes in
the heart over short and long time horizons, or mechanisms of adaptation and
remodeling after infarction. Good reviews of pertinent research questions, model-
ing needs, and accomplishments at different hierarchical levels of organization
have been presented within the framework of the Physiome, an international effort
to understand normal and abnormal functioning of the heart through computa-
tional modeling and to translate the insights gained from modeling into advances in
clinical applications (see, for example, [10–15]). Like other large-scale efforts, such
as the Virtual Liver Network [16], the Lung Physiome [17], and the Blue Brain Project
[18], the target of model development for these complex, multiscale systems is not
necessarily a single all-encompassing supermodel that would account for every
detail, but rather a set of complementary models of different dimensionality and
sophistication that explain certain aspects. The reason for not targeting a single
model is that the relevant time and size scales in the heart and other organs are so
vastly different that it seems impossible, at least with current methods, to merge
them in a meaningful fashion [10]. Nonetheless, the various models should be able
to talk to each other, which the Physiome Project facilitates by using cell- and
tissue-specific mark-up languages (XMLs), such as CellML, SBML, and FieldML
[13, 19, 20].
12.1 Basics of Heart Anatomy

Before we discuss different modeling approaches, we need at least a coarse descrip-
tion of our subject. The human heart is a muscle that contains four chambers,
namely, the left and right atria and the left and right ventricles (Figure 12.2).
esophagus
trachea
left common
brachiocephalic artery carotid artery
left subclavian
artery
superior vena cava aorta
left pulmonary
right pulmonary arteries arteries
right pulmonary veins left pulmonary

veins
left atrium
bronchi
right atrium
semilunar valves
bronchi atrioventricular
atrioventricular (mitral) valve
(tricuspid) valve Figure 12.2 Diagram of the parts of
left ventricle the human heart. In healthy humans,
septum the left chambers (atrium and ventricle)
are completely separated from the right
right ventricle
chambers. During the cardiac cycle, blood
inferior vena cava
runs through the four chambers in a well-
coordinated fashion. (Courtesy of ZooFari
descending
esophagus under Creative Commons Attribution-Share
aorta
The left and right sides of a human heart become strictly separated during embry-
onic development, so there is normally no direct blood flow between them. With
each heartbeat, blood flows through the chambers in a well-controlled, cyclic man-
ner that consists of two phases called diastole and systole, which are Greek terms
for dilation and contraction, respectively. Because the process is cyclic, it does not
really matter where we start. So, let us begin with the phase of early diastole. At this
point, the heart is relaxed and begins to receive blood in two states. The first is
deoxygenated, that is, relatively low in oxygen content, because it returns from all
parts throughout the body, where oxygen had diffused from the bloodstream into
the various tissues. The second state of blood is oxygenated. It comes from the lungs,
where hemoglobin molecules in the red blood cells have been loaded with oxygen.
The right atrium exclusively receives the deoxygenated blood through two
branches of the vena cava, which collect blood from the various tissues of the body.
Meanwhile, the left atrium exclusively receives oxygenated blood through the
pulmonary vein that comes from the lungs. The atria fill quickly, and because the
atrioventricular (AV) valves are open (see Figure 12.2), so do the ventricles. These
are in a relaxed state where their wall tissue is thin and the plane of the AV valves
shifts back up toward the base of the heart. A following contraction by the atria
pushes additional blood into the ventricles. During the second phase, the ventricles
begin to contract and their walls thicken. The AV valves close and the semilunar
outflow valves open. Deoxygenated blood is sent through the pulmonary artery on
its way to the lungs, while oxygenated blood enters the aorta, the largest artery in the
body, from where it is distributed throughout the body, first through increasingly
smaller arteries, and finally through tiny capillaries that are found in all parts of the
body and release oxygen to adjacent cells. The capillaries turn into small and then
larger veins, which ultimately return the deoxygenated blood to the right atrium.
The total volume of the heart does not change significantly during the heartbeat.
12.2 Modeling Targets at the Organ Level

At the macroscopic level of an organ, the heart acts like a peristaltic pump—or
perhaps it is better to say that the heart is a combination of two pumps that drive
different portions of the circulation of blood. Unlike most engineered pumps, the
heart is highly sensitive to its mechanical surroundings and constantly adjusts influx
and efflux in response to the current requirements of the body and various stresses.
These adjustments occur through autonomic controls that do not require conscious
input. In fact, very few people are capable of conscientiously affecting their
heartbeat, although mental stresses such as worry or fear can easily alter it.
The adjustment mechanism of the heart to different pumping demands is char-
acterized by the Frank–Starling law [21]. This law is derived from a nonlinear par-
tial differential equation (PDE), developed in the nineteenth century and called
the Young–Laplace equation, which generically relates pressure differences in fluids
to the shape of the wall retaining the fluid. Broadly, it states that increased pressure
requires increased wall thickness to maintain a stable wall tension. Applying these
relationships to the heart, the Frank–Starling law formalizes how increasing the
venous blood input to the ventricle stretches the ventricular wall, enhances contrac-
tility, and elevates the diastolic pressure and volume of the ventricle, which in turn
leads to an increased stroke volume [22]. Thus, the heart uses an adaptive response
mechanism that adjusts each ventricular output to its inflow and ensures the
balance between the outputs of the two ventricles over time.
The balancing of venous input and cardiac output occurs without external regu-
lation. The heart accomplishes this control task by adjusting the force of contraction
proportionally to changes in the lengths of its muscle fibers. This adjustment is
related to a decreased spacing of filaments within the myocytes, the formation of
cross-bridges between filaments during each contraction of the heart muscle, and
an increased sensitivity of the filaments to calcium (see later in this chapter
and [23]).
The physical relationship between pressure and wall thickness expressed in
the Frank–Starling law has been used to explain the gradual thickening of arteries
and of the left-ventricular wall in response to persistently high blood pressure.
The unfortunate consequence of this trend is that a thicker left ventricle is stiffer
(A) (B) (C) Figure 12.3 Images of ventricular

tachycardia and fibrillation, which are
potentially life-threatening arrhythmias
that originate in one of the ventricles
of the heart. (A) Transverse sections from
a sophisticated MRI scan of a rabbit heart,
proceeding from apex to base. (B) High-
resolution reconstruction of the rabbit
ventricular structure. (C) Orientation of
(D) (E) (F) (G) fibers on the rabbit ventricular structure.
(D) Simulation of ventricular tachycardia,
15 in the form of a single spiral wave, in the
5 rabbit ventricular structure. (E) Simulation of
–5 ventricular fibrillation, in the form of multiple
25 spiral waves, in the rabbit ventricular
structure. (F) Simulation of ventricular
45
fibrillation in a reconstructed canine heart.
65
1 cm (G) Experimental ventricular fibrillation in a
–85 canine wedge preparation recorded through
mV
optical mapping. Images of arrhythmias
in motion can be found at the interactive
website http://thevirtualheart.org. (Courtesy
than normal, so that filling it with blood requires elevated pressure. Over time, this of Elizabeth M. Cherry, Alfonso Bueno-Orovio
altered pressure can lead to a decline in performance during diastole and eventually and Flavio H. Fenton, Cornell University. With
to diastolic heart failure. permission from Elsevier.)
Beyond the important, yet somewhat generic, insights gained from the Frank–
Starling law, modern methods of fluid mechanics, engineering, computing, and
modeling permit detailed and very sophisticated analyses of the spatial organiza-
tion and three-dimensional architecture of the heart and their impact on proper
functioning. These analyses are crucial for an understanding of healthy and
perturbed blood flow in the atria and ventricles, as well as in the coronaries—the
arteries and veins that provide the heart muscle with oxygen and nutrients and
remove breakdown products. These vessels are rather narrow and prone to blockage,
which can lead to atherosclerosis and heart attacks. Modeling blood flow as well as
the mechanical deformation of the heart during the cardiac cycle requires systems
of PDEs and finite element approaches, which subdivide a three-dimensional
body into very small cubes and quantify changes within and between these cubes
over time. Finite element models faithfully reflect the anatomy of the heart, including
the orientation of fibers within the layers of the heart wall. They are fundamental to
predicting whole-organ function based on features of the heart tissue that vary from
person to person, for instance when the wall thickens owing to disease [15, 24].
Researchers at the École Polytechnique Fédérale de Lausanne (EPFL) have devel-
oped a very sophisticated three-dimensional supercomputer model of coronary
blood flow that is able to predict heart attacks [25].
In addition to mechanical features and issues of blood flow, comprehensive
organ-level models need to address the electrical activation of heart contractions,
which initiate at the sino-atrial node and spread to the Purkinje fibers and the entire
myocardium (heart muscle), as discussed in detail later. Fascinating pictures and
movies of electrical activity waves can be found in [13, 26] and on many websites; an
example is shown in Figure 12.3. Finally, all cells in the heart need energy, which
requires efficient transport of metabolites and oxygen between the coronaries and
the myocardium, and this needs to be taken into account in realistic organ models.
12.3 Modeling Targets at the Tissue Level

The pumping action of the heart is accomplished with contractile tissues that con-
tain specialized cells, called cardiac cells, myocardial cells, or cardiomyocytes.
These cells are in many ways similar to skeletal muscle cells, but, in contrast to the
latter, the coordinated pumping action of the heart does not require our conscious
input. Also, the heart muscle is very resistant to fatigue, because the cardiomyocytes
contain many mitochondria, which are responsible for respiration and energy
supply, and a large number of myoglobins, which are proteins that are able to store
oxygen. Furthermore, the heart muscle is richly equipped with coronary blood
vessels, which provide oxygen and nutrients to its cells. If the heart does not receive
sufficient amounts of oxygen, an infarct results, leading to decreased cardiac

function, tissue damage, and possibly death.
Models of the heart at the tissue level account for structural details and for
properties of electrical conductance. Prominent structural aspects include the
mechanics of the deformation of the heart tissue during contractions [11]. Because
the heart contains roughly 1010 myocytes, a key concept of tissue-level modeling is
the approximation of the discrete aggregate of these large numbers of cells and their
interactions with a collective representation. This so-called continuum approxi-
mation permits the application of approaches gleaned from the physics of elastic
materials, which deals with forces, stresses, strains, deformations, and distortions,
and helps explain shape changes during the cardiac cycle as well as the spread of
electrical signals.
The electrical signals affecting the heart tissue and its billions of cardiomyocytes
are coordinated by a primary pacemaker region in the wall of the right atrium, called
the sinoatrial (SA) node. We will discuss this node several times in the following, so
it suffices to state here that the SA node cells beat on their own, and that their
oscillation in electric membrane charge is transduced to the excitable myocytes of
the atria and ventricles. This traveling signal, originating at the SA node, ultimately
makes the heart contract in a well-controlled manner. The SA node consists of a
group of modified cardiac myocytes that spontaneously contract at the order of 100
times per minute.
The electrical signals generated in the SA node cause the atria to contract. They
also spread along dedicated conduction channels in the walls of the atria and
quickly reach the atrioventricular (AV) node, which constitutes the electrical con-
nection between the atria and ventricles. Signals from the SA node reach the AV
node with a delay of about 0.12 s. This delay may sound trivial, but is very important
because it ensures the correct timing between the ejection of blood from the atria
and the contraction of the ventricles. The AV node contains a control system that
slows down conduction if the node is stimulated too frequently. This mechanism,
together with a refractory period where further triggering is not possible (see later),
prevents contractions of the ventricles that are too rapid and could be hazardous.
Signals from SA and AV nodes subsequently run through the conductive bundle of
His, named after the Swiss cardiologist Wilhelm His, and to the Purkinje fibers,
which branch out throughout the walls of the ventricles. The bundle and Purkinje
fibers are specialized to conduct electrical signals rapidly to the myocardium. AV
node and Purkinje cells can fire autonomously, but because their rate is slow, the
electrical signals are normally dominated by the SA node.
The beating rate of the SA node is constantly modified by impulses from the
autonomic nervous system (ANS), which generally acts as a control system that
maintains homeostasis in the body, so that the ultimate heart rate under resting
conditions is only about 70 beats per minute, but may go up to 200 beats during
strenuous exercise. ANS activities occur without conscious control, sensation or
thought. The heart rate can also be affected by noradrenaline (also known as norepi-
nephrine), the most common neurotransmitter of this system, and the hormone
adrenaline (also known as epinephrine). We all know that sudden excitement or
stress can lead to a burst in adrenaline (and noradrenaline) and sudden increases in
heartbeat that get us ready for “fight or flight.” Alas, phenomena like the ANS-
controlled pacemaker activity in the SA node are usually more complex than they
seem at first. As an analogy, consider the act of focusing on near or distant objects,
which requires the combined voluntary actions of the extra-ocular muscles to move
the eyes and of smooth muscles, controlled by the ANS, to change the shape of the
lens. Clearly, the focusing task combines voluntary and involuntary actions.
Similarly, the heartbeat is not entirely controlled by the ANS, and, for instance,
mobilization of the cardiovascular system is driven in advance of exercise to ready
the musculature for motor activity.
As in the case of whole-organ studies, the spatial and mechanical properties of
the heart tissue are of great importance at the tissue level, and computational
(finite element) analyses are crucial for understanding the design principles
underlying a successful electrical propagation from the pacemaker cells to the sur-
rounding, excitable heart muscle [27]. For instance, it is important that the electri-
cal coupling between pacemaker cells and excitable cells in the ventricles
suppresses electrical interactions at inappropriate times. This suppression depends

on a variety of factors, such as the intercellular coupling among the heart’s pace-
maker cells and differences in their membrane properties, as well as the commu-
nication among the surrounding cells, which depends on the number and spatial
orientation of gap junctions connecting them. Thus, analyzing the spatial organi-
zation of the heart at the tissue level is important for a deep understanding of the
correct propagation of electrical signals throughout the heart, the fine-tuned tim-
ing of contractions of the chambers, the pathological responses of the heart close
to areas of infarction, and even for correct interpretations of ECG readings
(see later).
12.4 Modeling Targets at the Cell Level

The cells of the heart have intriguing specific features, but of course they also con-
tain generic components that would have to be included in a comprehensive
model of the heart. Not surprisingly, heart cells express different genes than other
cells, their metabolism requires more energy, they contain proteins some of which
are uncommon in other cells, and some of their signaling systems are germane to
the function of the heart. Like skeletal muscle cells, most cells of the heart muscle
are contractile, and the mechanics of the responsible filaments constitutes an
interesting target for modeling. The electrical activity of these cells is tightly asso-
ciated with the flux of ions, so that ion transport and electrochemical models
demand special attention (see later). Finally, specifically adapted genomic, pro-
teomic, and metabolic models are required for understanding the particular
features of the heart. Some interesting modeling work in this context is presented
in [28, 29].
Of crucial importance is that the heart contains a number of different cell types,
two of which are of particular relevance here, namely, cells of the SA node and excit-
able cardiomyocytes. The cells in the SA node are modified cardiomyocytes and
constitute only about 1% of all heart cells, but they are of immense importance. SA
node cells exhibit a number of physiological differences in comparison with
ventricular cells. For instance, while they do contain contractile filaments, they do
not contract. The most important distinguishing feature is the ability of SA node
cells to beat (depolarize) spontaneously (discussed in detail later), which has led to
SA node cells being termed pacemaker cells. This electrical depolarization, which
is caused by ion fluxes across the cell membrane, makes the membrane more posi-
tive and results in an action potential, which ultimately spreads throughout the
heart muscle, as we will discuss in more detail later. The spontaneous depolariza-
tion in SA node cells occurs at a rate of about 100 times per minute, which, however,
is constantly modified by the ANS. Thus, SA and AV node cells are clearly different
from most other cardiomyocytes, but there are also significant differences within
the latter group, for instance, with respect to the characteristics of their resting
potentials and the shapes of their action potentials [30].
Interestingly, the transition between cells in the SA node and the peripheral
atrial cells is smooth, with one cell type gradually changing into the other within a
region surrounding the SA node. While this spatial transition might seem to be a
subtle detail, it has turned out to be very important. Three-dimensional tissue
reconstructions have shown how cell types and their communication channels are
distributed in these transition regions and how this spatial heterogeneity is crucial
for proper signal propagation. In fact, even within the SA node, differences can be
found between central and peripheral regions in terms of both morphology and
action potentials [27].
The spreading of the electrical impulse from cell to cell occurs through gap junc-
tions, which are specialized sets of half-channels connecting the cytoplasm of two
neighboring cells and permitting the free flow of ions and some other molecules
(see Chapter 7). Gap junctions are of utmost importance to the heart, because they
allow the efficient passing of ion-based electrical signals from one cell to another
and thereby enable the coordinated depolarization and contraction of the heart
muscle. In addition to the spreading of electrical charge through gap junctions, the
electrical signal is transduced through the bundle of His and Purkinje fibers, which
consist of specialized cell types that conduct five to ten times as fast as signal trans-
duction through gap junctions.
The receivers of the electrical impulses from the SA and AV nodes and the Pur-
kinje fibers are the regular, excitable cardiomyocytes. These elongated cells
slightly differ in structure and electrical properties between the atria and the ven-
tricles, but their important common property is that they are contractile. This
ability to contract is due to large numbers of protein filaments, or myofibrils,
which are a few micrometers long and similar to those found in skeletal muscle
cells. The contraction itself is a complex process, which is ultimately accom-
plished by the sliding motion of two myofibrils, actin and myosin. In simplified
terms, the myosin filaments contain numerous hinged heads, which form cross-
bridges with the thinner actin filaments (Figure 12.4). By changing the angle of
the hinges, the actin and myosin filaments slide against each other in a ratcheting
manner, for which the energy is provided by hydrolysis of adenosine triphosphate
(ATP). As a result of the sliding action, the entire cell contracts and becomes
shorter and thicker. A review of some of the pertinent modeling activities is given
in [31].
In reality, the myofibrils are components of a complicated multiprotein complex
that includes the important proteins titin and troponin. The details of the sliding
motion are quite involved, and it suffices here to state that they are coupled to
oscillations in the cytosolic calcium concentration (see later and, for example, [1,
32]). When the calcium concentration increases, the cell contracts. When the con-
centration decreases, the filaments slide in opposite directions, and the cell relaxes.
This tight correspondence is called excitation–contraction (EC) coupling [14].
Because the filament movement requires a lot of energy, cardiomyocytes contain
many mitochondria.
Nobel Laureate Andrew Huxley proposed the first sliding filament model. While
this model is not as prominent as his model of nerve action potentials, it has served
as a basis for an entirely different class of models than the electrical models of nerve
and heart cells.
Importantly, the contraction process in cardiomyocytes is initiated by electrical
impulses and the flow of ions, as we will discuss later in greater detail. This electrical
excitation in a cardiomyocyte is followed by a refractory period during which the
cell normally does not respond to further electrical excitations. After this period, the
cells are ready to fire again.
The membranes of cardiomyocytes change in their electrical charge once
during every heartbeat. As a consequence, an electrical charge spreads through-
out the heart and to the surrounding tissues, and ultimately causes very subtle
electrical changes on the skin. These changes can be measured over a period of
time with electrodes attached to the skin and are interpreted with methods of
electrocardiography. Because the electrical impulses move as waves throughout
the heart and are modified by the tissues of the body, the shape of the resulting
actin myosin actin
contraction relaxation
Figure 12.4 Simplified representation of

the sliding action of myofilaments, which
causes a cardiomyocyte to contract. The
myosin filaments (brown) contain hinged
heads (tan) that form cross-bridges with the
thinner actin filaments (green). A change in
the angle of the hinges leads to sliding and
contraction. In reality, this process is much
more complicated.
SIMPLE MODELS OF OSCILLATIONS 339
Figure 12.5 A 12-lead ECG of a healthy

young man. Electrical currents generated
during the cardiac cycle spread throughout
the body and can be measured by an array
of electrodes placed on the skin. See the
text for further explanation. (Courtesy of
Wikimedia Commons, File:12leadECG.jpg.)
electrocardiogram (ECG; or EKG from the German analog) depends on the

three-dimensional architecture of the heart and the exact locations of the elec-
trodes on the body. In order to assess heart function as comprehensively as fea-
sible, it is therefore common practice to place several sets of electrodes at defined
locations on the chest and limbs. The ECG displays the voltages between pairs of
electrodes, and their spatial as well as temporal interpretation gives clues regard-
ing the overall health status of the heart, as well as abnormal rhythms that may be
due to imbalances in electrolytes or to damage to the heart tissue in specific loca-
tions; some details will be discussed later. The ECG can also identify, and to some
degree locate, tissue damage in the case of a myocardial infarction (MI—a heart
attack). Since the ECG only measures electrical activity, it does not truly measure
the heart’s ability to pump sufficient quantities of blood. The amounts of blood
flowing through the atria and ventricles can be measured with an ultrasound-
based echocardiogram or with methods of nuclear medicine. Figure 12.5 shows
a 12-lead ECG of a healthy young man.
Thus, we started with organ-level models, pointed to modeling needs at the
tissue level, discussed features of heart cells, and have now returned to their effects
as they are manifest at the whole-body level.
SIMPLE MODELS OF OSCILLATIONS

The information we have discussed so far is already sufficient for some very simple
mathematical heart models. Let us start by focusing on the most obvious—the oscil-
latory pattern. This oscillation can be observed in many forms, for instance in the
ECG, but also in the increasing and decreasing volume of blood in a ventricle, cyclic
changes in electrical voltage, or rhythmic changes in the deformation of the
ventricular wall. All these derivative oscillations ultimately originate from the
autonomous depolarization and repolarization of the cells in the SA node, which
result in a relatively regular oscillation in voltage (Figure 12.6). Here the term
“autonomous” means that the cells exhibit sustained oscillations without needing
repeated triggering events.
An intriguing aspect of these oscillations is that small perturbations are easily
tolerated. In other words, if something temporarily disturbs or modulates the
rhythm or the amplitude of the oscillation, the healthy heartbeat soon returns to the 0
original rhythm. Of course, this tolerance is absolutely necessary for the heart to
operate in a healthy way. Just imagine going to a horror movie, where a gruesome –50
mV
zombie attack makes your heart race. For a moment, you may enjoy the excitement,
but you would certainly like your heart to go back to normal after a while and with-
0 1 2 3
out conscious prodding. s
Figure 12.6 Oscillatory change in the

12.5 Black-Box Models of Oscillations voltage of an SA node cell. Different cell
types and cells in different locations of the
As we discussed in Chapter 4, tolerance to perturbations is a hallmark feature of heart exhibit distinctly different shapes
stable limit-cycle oscillations. We mentioned different formats, including the van (see Figure 12.12), and these are again very
der Pol oscillation [33, 34], which was proposed many decades ago as a model for different from the various patterns in an ECG
heartbeats, and return here to the very flexible limit cycles based on S-systems [35]. (see Figure 12.5) [1, 29].
(A) Figure 12.7 Stable S-system oscillations.

1.2 (A) Variable X1 of System A in (12.1), which
resembles the oscillations in the SA node
(see Figure 12.6). (B) Variable X1 of System B
X1 in (12.1), which is similar to a simplified ECG
trace (see Figure 12.5).
1.0
0.8
0 1.5 3.0
time
(B)
3.0
X1
1.5
0
0 25 50
time
Specifically, we consider the two examples in Figure 12.7, which correspond to the
model equations
X 1 = 2.5( X 2−6 − X 13 X 2−3.2 ), X 1 = 0.9,

System A:
X 2 = 1.001⋅ 2.5(1 − X 1−5 X 2−3 ), X 2 = 1.2 ;
(12.1)
X 1 = 1.005( X 2−8 − X 17 X 25 ), X 1 = 0.25,
System B:
X 2 = X 16 X 25 − X 1−1 X 2−4 , X 2 = 2.5.
Systems A and B lead to oscillations that are vaguely reminiscent of the dynamics in
the SA node (see Figure 12.6) and of ECG traces (see Figure 12.5), and we can easily
confirm that both oscillations are indeed stable, namely, by briefly perturbing them
and watching them return to their original oscillations.
If stable oscillations are repeatedly “poked” with the “wrong” frequency, bad
things can happen. Mathematically speaking, a stable limit cycle that is exposed to an
oscillating stimulus pattern may become chaotic: although everything in the system is
entirely deterministic, the oscillatory behavior is no longer predictable, unless one
actually solves the equations. The analogy with heart physiology is evident: repeated
spurious electrical signals can affect the system, and the result may be an unpredict-
able beating pattern. Specific examples are blockages in the AV node or the fast con-
ducting bundles, which may lead to repeated additional electrical signals that can
interfere with the usual electrical patterns in the heart and cause arrhythmia and even
death. At the same time, one must be careful with the interpretation of irregularities:
while chaotic oscillations may be a sign of disease, a fetal heartbeat that is too regular
is often a sign of trouble [36–38]. Indeed, loss of variability in oscillations has been
associated with different pathologies, frailty, and aging (see Chapter 15 and [39, 40]).
As a numerical example, suppose the stable oscillatory System B in (12.1) is
subjected to a small regular oscillation. We model this idea by adding 0.01(s + 1) to
both equations, where s is the value of a sine function, which we discussed in the
form of ordinary differential equations (ODEs) (see (4.25) and (4.72) in Chapter 4),
so that s + 1 oscillates between 0 and 2. In addition, we introduce an extra parameter
a, which permits us to change the frequency of the sine oscillation. Thus, our
heartbeat together with spurious electrical signals is given by the following four-
variable model:
X 1 = 1.005( X 2−8 − X 17 X 25 ) + 0.01(s + 1), X 1 = 0.25,

X 2 = X 16 X 25 − X 1−1 X 2−4 + 0.01(s + 1), X 2 = 2.5,
(12.2)
s = ac , s(0) = 0,
c = −as , c (0) = 1.
If a = 1, the frequency of the sine system is the same as before. Larger values mean
faster oscillations, smaller values slower oscillations. If a = 0, the sine system does
not oscillate at all (the reader should confirm this with mathematical arguments
and with simulations!), which permits a good comparison between systems with
and without spurious stimuli. The system is in this case almost the same as in (12.1),
except that both equations have an extra term +0.01, which is apparently negligible.
The oscillations of this system (without spurious oscillations) are shown in
Figure 12.8A for a long time horizon. In Figure 12.8B–F, the frequency parameter is
varied; the sine oscillation itself is visualized for clarity with a ten-fold magnification
of its amplitude. If a = 10, the sine oscillation is so fast that it does not change the
appearance of the heart oscillation (this is not shown, but is very similar to
Figure 12.8A). In Figure 12.8B, with a = 1, the heartbeat is not quite even.
Figure 12.8C, with a = 0.5, spells trouble, with quite arrhythmic oscillations, which
seem to recover after about 700 time units. However, the heart rate is much increased.
Further lowering the parameter a seems to remedy the situation, but if we look
(A) (B)
3.0 X1 0.1(s+1) 3.0 X1 0.1(s+1)
1.5 1.5
0 0
0 500 1000 0 500 1000
t t
(C) (D)
3.0 X1 0.1(s+1) 3.0 X1 0.1(s+1)
1.5 1.5
0 0
0 500 1000 0 500 1000
t t
(E) (F)
3.0 X1 0.1(s+1) 3.0 X1 0.1(s+1)
1.5 1.5
0 0
0 500 1000 0 500 1000
t t
Figure 12.8 Simulation of a perturbed heartbeat, ultimately leading to fibrillation and death. If a stable oscillator is regularly “poked,” for instance
with a sine function, the results may be negligible or detrimental, depending on how the frequencies of the two oscillators relate to each other. The
frequency of the sine oscillator is determined by a. See the text for details.
1 Figure 12.9 Effectiveness of a simple

pacemaker. An irregular heartbeat, here
modeled as a chaotic Lotka–Volterra (LV)
X1 oscillator, becomes regular through the
effect of an appropriate sine function.
Only the first variable of the LV system is
shown, along with a scaled sine function,
0.5 which oscillates between −0.05 and 0.05,
starting at t = 8. For better visibility, the sine
function is not shifted in this graph, although
it is shifted by 1 in the model.
p sin t
0
0
10 20
time
closely, the heart goes into double beats. For a = 0.40, the heart starts skipping beats,
and even very slight further decreases in a cause the heartbeat to stop. The smaller
the value of a, the faster this happens (see Figure 12.8D–F).
We can see that even simple models are able to capture the essence of the heartbeat
and point to some very serious problems. Because these models are so simple, they are
well suited for rigorous mathematical analysis. Indeed, many researchers have studied
limit cycles in this and other contexts and explored their vulnerability with respect to
input oscillations of an unfavorable frequency (see, for example, [41, 42]).
One may also use simple models to explore whether a highly irregular heartbeat
can be made regular by a pacemaker, and how robust such a correction might be.
Pacemakers are implanted electronic devices that send out electrical signals with
the goal of correcting troublesome oscillations. If adjusted correctly, a modern
pacemaker can ensure that the heart beats with the right speed at the right time.
To assess the function of a very simplistic pacemaker, we model the irregular
heartbeat with the chaotic Lotka–Volterra (LV) oscillator from Chapter 10. To
resemble the timing of a heartbeat, we multiply all equations by 50, which results
in a rather erratic heartbeat with about 60–70 beats per minute. We model the
overly simplistic pacemaker with a sine oscillator, represented as a pair of ODEs
(s = q c = −qs; see above and Chapter 4) and add to each equation in the LV-
system a term of the form p(s + 1). An example result, with q = 6, is shown in
Figure 12.9. Initially we set p = 0, which corresponds to the unaffected heartbeat.
At t = 8, we reset p = 0.05. We can see that it takes several beats, but the pace-
maker indeed tames the chaos. In reality, pacemakers are much more sophisti-
cated and respond to different demands of the body that require the heart beat
faster or more slowly. Exercise 12.5 explores this topic in more detail.
12.6 Summary of Black-Box Oscillation Models

Rather simple oscillation models can represent some of the phenomena that have
been observed in well-functioning and diseased hearts, such as arrhythmias, beat
skipping, the function of pacemakers, and the potentially detrimental effects of
spurious electrical signals. So, we have to ask ourselves whether these simple mod-
els are sufficient. As is so often the case, the answer depends on the purpose of the
modeling effort. For instance, if we want to demonstrate arrhythmias, the stable
black-box oscillation models might be most instructive, because nothing distracts
from the simplicity of their mathematical structure. Or, if we plan to use the heartbeat
as an input module for a larger model, and we are not specifically interested in
mechanistic details, a small ordinary differential equation model might be a good
choice. As an example, suppose we want to study the metabolic responses of cells to
oxygen supply, which changes during the cardiac cycle. In this case, the mechanisms
that fundamentally lead to oscillations in blood oxygen may not be particularly
relevant. In fact, since larger models always require more effort in terms of parameter
estimation and diagnostics, investing more time on the oscillator may not be worth
the effort. Also, as mentioned earlier, a lot can be learned mathematically about
small limit cycle oscillators, and many books have used simple models like those
above for analyzing different types of oscillations, questions of structural stability,
and bifurcation points in nerves, hearts, and other oscillating systems, which would
not be possible in larger models [7, 41, 42].
While small models have strong advantages, they obviously have limitations as
well. For instance, if we want to dissect a disease pattern into possible molecular,
physiological, or electrochemical causes, the small models simply cannot provide
adequate answers. It is not even possible to ask the right questions, because these
models do not contain the relevant components, such as a membrane, concentra-
tions of ions, or gated channels, which are very important for proper functioning, as
we will see later in this chapter. In order to explain the processes that are ultimately
responsible for a regular heartbeat, and for the purposes of studying and treating
heart disease, we need to dig deeper, down to the real molecular and physical
drivers, formulate them quantitatively, and then use mathematical models and
computation to integrate results from molecular and electrochemical levels to gain
insight into macroscopic function or failure. We begin with an exploration of calcium
dynamics, which is crucial for proper function, develop a model for this particular
aspect, and then work toward more comprehensive models that include chemical
and electrical processes. To venture toward such more sophisticated models, we
need to understand more about the biology of the heart: what makes the heart tick?
12.7 From a Black Box to Meaningful Models

We have seen that it is surprisingly easy to construct oscillations and even limit-
cycle oscillations as black-box models. The question is now whether we can
construct a simple yet biologically meaningful model that exhibits sustained and
robust oscillations. One approach, which we will pursue here, is to study the
movement of ions during the cardiac cycle.
The chain of contraction and relaxation events in heart cells, including SA node
cells, is closely associated with the periodic movement of calcium ions between
three locations, namely the extracellular fluid, the cytosol, and an intracellular
compartment, called the sarcoplasmic reticulum (SR). The cycle begins when the
membrane of a cell depolarizes, which is associated with an influx of sodium and a
relatively small amount of calcium from the extracellular space into the cytosol. The
calcium binds to proteins in the membrane of the SR, where it triggers a mechanism
called calcium-induced calcium release (CICR), which results in the flow of large
amounts of calcium from the SR into the cytosol [43, 44]. In regular cardiomyocytes,
this calcium binds to troponin and leads to the sliding action of the myofibrils actin
and myosin, as discussed before, and this sliding action causes contraction of the
myocyte. Calcium now disengages from the myofibrils. Most of it is pumped back
into the SR, and some leaves the cell through specialized calcium exit carriers that
are located on the cell’s membrane and simultaneously export calcium and import
sodium. The cell relaxes and is ready for a new contraction.
In regular cardiomyocytes, the initial depolarization is triggered by electrical sig-
nals coming from the SA node or from neighboring cells, as discussed earlier. In
contrast, the membrane in an SA node cell depolarizes spontaneously, starting from
a minimum charge of about −60 mV (see Figure 12.6). While the initial trigger is dif-
ferent, the release of calcium from the SR also happens in the SA node [44]. Thus,
both SA node cells and ventricular cardiomyocytes depend on the movement of
calcium among three compartments.
The question then becomes whether a mathematical model of the three-
compartment calcium dynamics can in principle lead to sustained, stable oscilla-
tions, as observed in cardiomyocytes. Because contractions in heart and skeletal
muscle, as well as calcium bursting patterns in organs such as the pancreas, are of
great clinical importance, it is not surprising that there is plenty of literature on
models of calcium oscillations. As a sort of proof of principle, we only look at one
minimalistic model that captures the calcium dynamics nicely and quite easily, and
serves our purposes of providing a biologically motivated oscillator. This model was
proposed by Somogyi and Stucki [45, 46], and many more complex models followed
later; a review is given in [47].
[Ca2+]ext = constant rEC Figure 12.10 Simplified diagram of

calcium flow associated with a myocyte.
The illustration shows the processes and
symbols used in the Somogyi–Stucki model,
y = [Ca2+]cyt ctSCf(y) which leads to stable calcium oscillations.
The red area represents the cytosol of the
cell, while the pale area represents the
sarcoplasmic reticulum. The notation is
explained in the text.
pCS rSC
pCE
x = [Ca2+]SR
The Somogyi–Stucki model describes a dynamical system that consists of three

physical spaces and calcium movements among them (Figure 12.10). The spaces are
the cytosol, the SR inside the cell, and the extracellular space. The model assumes
that the extracellular space contains such a high concentration of calcium, [Ca2+]ext, in
comparison with the cytosol that it is considered an independent, constant variable.
The calcium concentrations in the SR and the cytosol are denoted by x = [Ca2+]SR and
y = [Ca2+]cyt, respectively, pump rates by p, simple transport rates by r, and channel
transport by ct, with subscripts indicating source and target. According to Somogyi
and Stucki, the following assumptions are made in the construction of the model:
1. The extracellular space serves as an inexhaustible source for an influx of
Ca2+ = [Ca2+]ext at a constant rate rEC.
2. Ca2+ is pumped from the cytosol into the extracellular space at a constant rate
pCE.
3. Ca2+ is pumped from the cytosol into the SR at a constant rate pCS.
4. Ca2+ leaks from the SR into the cytosol proportionally to the concentration differ-
ence between the SR and the cytosol and at a rate rSC.
5. Ca2+ is actively pumped from the SR into the cytosol. The process depends on the
concentration difference between the SR and the cytosol, and its rate is a sigmoi-
dal function of the concentration in the SR.
The equations for this small dynamic system are easily set up according to the
guidelines discussed in Chapters 2 and 4. Namely, because we are interested in
time-dependent processes, we use differential equations, and if we ignore spatial
details, for reasons of simplicity, these equations are ODEs. Thus, we formulate an
ODE for the two dependent variables, x and y, and we include on the right-hand
sides all relevant variables, functions, and rates. To implement this strategy, we
make the simplifying assumption that all processes can be adequately represented
as linear functions. This assumption is typical for transport processes and for
processes whose turnover rates are proportional to their substrate concentrations.
The only exception is f(y), which describes the channel transport. For this function,
Somogyi and Stucki use a Hill function with Hill parameter h, which saturates at the
maximal rate (ct)SC. Thus, the nonlinear channel transport process is modeled as
yh
f ( y ) = (ct )SC . (12.3)
K h + yh
Taken together, x is affected by transport into and out of the SR, and y is affected by
transport into and out of the cytosol, and the mathematical formulation is
yh
x = pCS y − rSC ( x − y ) − (ct )SC ( x − y ),
K h + yh
(12.4)
yh
y = rEC + rSC ( x − y ) + (ct )SC ( x − y ) − pCE y − pCS y .
K + yh
h
ELECTROCHEMISTRY IN CARDIOMYOCYTES 345
As in so many cases, it is not easy to pull suitable parameter values out of a hat (see 2
Chapter 5). Since we are not really interested in parameter estimation here, we sim-
ply take values similar to those used by Somogyi and Stucki [45, 46], namely, x
pCS = 4.5, rSC = 0.05, (ct )SC = 3, 1
x, y
h = 4, K = 0.1,
rEC = 0.075, pCE = 1.5.
y
With these values, the system becomes 0
0 50 100
time
y4
x = 4.5 y − 0.05( x − y ) − 3 ( x − y ), Figure 12.11 Oscillations generated by
0.14 + y 4 the Somogyi–Stucki system (12.5). Even
(12.5) without external cues, the system oscillates
y4
y = 0.075 + 0.05( x − y ) + 3 (x − y ) − 6 y. in a stable manner. The variables x and y
0.11 + y 4
4
represent calcium concentrations in the SR
and the cytosol, respectively.
The initial values turn out to be rather unimportant, and we set x0 = 2, y0 = 0.1.
The result of a simulation with these conditions is shown in Figure 12.11. We see
that this rather simple model of the calcium dynamics in a cell is indeed able to
oscillate. In fact, perturbations demonstrate that it oscillates all by itself in a stable
manner. Our proof of principle is complete.
ELECTROCHEMISTRY IN CARDIOMYOCYTES
The oscillations in the heart are ultimately due to ion movements and electrical
charges, but we have not really discussed how these two are related. The key
to understanding these relationships lies in transport processes at the cell mem-
brane and the membrane of the SR. A superb description of these cross-membrane
processes can be found in [48]; much of the following is excerpted from this source.
The transport processes across the membranes are tied to two distinctly differ-
ent, yet intimately related, gradients, namely, a concentration gradient and an
electrical (charge) gradient. While the cell contains many types of molecules, the
concentrations of interest here are those of ions. To refresh our memory, ions are
atoms or molecules carrying electrical charges that result from a difference between
their numbers of electrons and protons. In the case of an atom, the dense nucleus
contains protons, which are positively charged, and neutrons, which are electrically
neutral. The nucleus is surrounded by a cloud of negatively charged electrons. In a
negatively charged ion, also called an anion, the cloud contains more electrons than
there are protons in the nucleus. The most important examples here include ionized
atoms such as chloride ion and some larger molecules such as negatively charged
proteins. Positively charged ions, called cations, have one or several more protons in
the nucleus than electrons in the shells of their electron clouds. The most important
examples for membrane activities are sodium, potassium, and calcium ions.
In the case of a cell that is surrounded by a watery medium and separated from
the medium by a porous membrane, ions can be transported in and out through the
membrane either by means of specific transport mechanisms or, to some degree,
through leakage. As we know from many real-life situations, differences within
gradients tend to equalize if left alone. Expressed in the terminology of physics, the
total energy of the system approaches the minimum that is achievable in the given
situation. Hot coffee in a cool room gets cold, and eventually reaches the same
temperature as its surroundings. Sugar poured into a glass of water dissolves through
random molecular movements, and its concentration eventually becomes homoge-
neous. This natural tendency to equalize concentration differences also applies to
concentrations inside and outside a cell. However, in the process of equalizing the
concentrations of ions, it is possible that the charges become unbalanced. Thus, the
total energy of the system is composed of two components: the chemical energy
associated with molecular movement within the concentration gradient and the
electrical energy associated with charges. Nature tends to minimize the sum of
these two components, and, given enough time, the sum becomes zero. We will
discuss this aspect later more rigorously.
The gatekeeper of ion movements between the cytosol and the extracellular
space is the cell membrane. It consists of a lipid bilayer that creates a substantial
barrier for polar molecules, and ions in particular, because of strong attractive elec-
trostatic forces between ions and water molecules. This impediment to free ion flow
is crucially important, because it allows an excitable cell, such as a cardiomyocyte,
to create and retain concentrations and electric charges in its cytosol that differ
drastically from the conditions in the surrounding extracellular fluid. The cell man-
ages the creation of concentration gradients across membranes, and thus the
storage of potential energy and the possibility of electrical signal transduction,
through transmembrane channel proteins and carrier proteins that are often
specialized to transport particular ions or water-soluble organic molecules into or
out of the cell.
Channel proteins form hydrophilic, water-filled pores that cross the lipid
bilayer and allow select ions to travel through the membrane. This travel is very
fast, with up to 100 million ions being able to pass through a channel within 1 s
[48]. With such speed, channel transport can be 100,000 times faster than trans-
port facilitated by carrier proteins, a property that is of great import in electric
signaling. The most characteristic feature of channels is that transport must always
occur down the concentration gradient. In the case of ions, this passive transport
results in a change in membrane potential, which in turn may influence further
ion transport. Because most cells exhibit a negative charge toward the inside of
the membrane, the entry of positively charged ions is favored. A fascinating fea-
ture of channel proteins is their fine-tuned specificity, which is a consequence of
their molecular structure. A pertinent example is the potassium leak channel,
which conducts K+ ions 10,000 times better than it does Na+. This difference is
intriguing, because a sodium atom is slightly smaller than a potassium atom.
Nature solved this problem with a potassium leak channel protein that creates a
specially charged ion selectivity filter, which almost exclusively prevents sodium
from traversing the membrane [48].
In contrast to simple aqueous pores, carrier proteins and ion channels are
controllable by the cell. Specifically, they may be opened and closed through
electrical or mechanical means, as well as through the binding of ligands. Most
important among the channels are voltage-gated and receptor-gated channels that
are respectively controlled by voltage and by binding of molecules such as the
neurotransmitter acetylcholine.
Carrier proteins, also called pumps or permeases, are transporters that bind a
specific ion and, while undergoing a transformational change, actively facilitate its
transport through the membrane. Carrier proteins are specific with respect to the
molecules they transport and require energy, which may be supplied, for instance,
in the form of ATP. Their dynamics resembles that of enzyme-catalyzed reactions
and shows a concentration dependence resembling that of a Michaelis–Menten
process (see Chapters 4 and 8), with a characteristic binding constant KM and satu-
ration at a maximal rate Vmax for high ion concentrations. Some carrier proteins also
permit passive transport in either direction, as long as it occurs down the concentra-
tion gradient, but their main role is active transport against the electrochemical
gradient.
Carrier proteins come in several variations. In the first case, uphill transport
(against the concentration gradient) is coupled to downhill transport of another
molecule. If both the ion and the other molecule are transported in the same
direction, the transport protein is called a symporter. If they are transported in
opposite directions, we talk about an antiporter. For example, moving an ion down
the gradient releases free energy that can be used to move a larger molecule, such as
a sugar, up the gradient. The second type of a carrier protein is a uniporter. In this
case, only one ion is transported and the process is coupled to the hydrolysis of ATP.
It is also possible that an antiporter is combined with ATP hydrolysis. As a further
possibility, which, however, is not of relevance here, bacteria may use light as an
energy source for ion transport against the gradient.
The most important transport process in our context is the Na+–K+ pump,
which hydrolyzes ATP. Because Na+ is pumped out of the cell, against a steep
electrochemical gradient, while K+ is pumped in, the transporter is an antiporter,

and, since it uses ATP, it is sometimes called a Na+–K+ ATPase. The pump can also
run in reverse, where it generates ATP from ADP. The Na+–K+ pump functions in
such a manner that during each cycle three Na+ ions are pumped out, while only two
K+ ions are pumped in. As a consequence, the pump changes the charge balance
and is therefore called electrogenic, that is, it alters the membrane potential. How-
ever, the electrogenic change is small in comparison with other processes and con-
tributes little to the electrical state of the cell’s membrane.
A second crucial antiporter protein for the transport of ions during the cardiac
cycle is the sodium–calcium exchanger (Na+–Ca2+ exchanger), which removes cal-
cium from the cell, in particular after an action potential. It allows sodium to flow
across the plasma membrane, down its electrochemical gradient, and thereby
transports calcium ions in the opposite direction. The removal of calcium is very
fast, with several thousand calcium ions crossing the membrane per second. This
speed is necessary for effective signal transduction, but it is also expensive: the
removal of each calcium ion in this process requires the import of three sodium
ions. The transport through the Na+–Ca2+ exchanger is electrogenic, and strong
depolarization of the membrane can actually reverse the direction of ion exchange,
which is possible under the influence of some toxins and drugs. Alterations in the
activity of the Na+–Ca2+ exchanger can have serious consequences and lead to ven-
tricular arrhythmias [49].
An additional pump that removes calcium is the plasma membrane Ca2+ ATPase
(also known as the PMCA), which has a higher affinity for calcium than the Na+–Ca2+
exchanger, but a much lower transport capacity. Because of its higher affinity, the
Ca2+ ATPase is effective when the calcium concentration is low, and thus it comple-
ments the Na+–Ca2+ exchanger. The Ca2+ ATPase is well understood in terms of its
structure and function [48].
Normally, the Na+–K+ exchanger sets up a sodium gradient that drives the Na+–
2+
Ca exchanger to extrude calcium. The foxglove plant (Digitalis purpurea) produces
the toxic substance digitalis, which is used as a medication to control arrhythmias of
the heart. Digitalis functions by poisoning the Na+–K+ exchanger and thereby
decreasing the Na+ gradient, which in turns reduces Ca2+ extrusion. The accumu-
lated intracellular Ca2+ promotes contraction [50].
The channels are the mediators for transfer of ions into and out of cardiomyo-
cytes, and because ions are electrically charged, their transport can lead to changes
in the electrical state of the cell and, specifically, its membrane. How is it possible
that these transport processes lead to autonomous oscillations?
12.8 Biophysical Description of Electrochemical Processes at the

Membrane of Cardiomyocytes
Ion transport and changes in the cell’s electrical state constitute one of the rather
rare cases in biological systems analysis where processes are so close to processes in
physics that the corresponding laws of physics can be used directly for explanations,
or at least as an inspiration. Here we borrow heavily from electrical theory. As dis-
cussed before, nature tends to minimize energy, which in the case of excitable cells
consists of two components, namely, the chemical energy associated with molecu-
lar movement within the concentration gradient and the electrical energy associ-
ated with charges. Given enough time without outside influences, the sum of the
two sources of energy approaches zero.
Formulating this fundamental phenomenon in the terminology of physics leads
to the Nernst equation, which quantifies electrochemical gradients. One begins
with the free-energy change per mole ion, ΔGconc, which is associated with the con-
centration gradient across the membrane. Nernst found that this quantity can be
computed as −RT ln(Co/Ci), where R is the gas constant, T is the absolute tempera-
ture in Kelvin, and Co and Ci are the concentrations of ions outside and inside the
cell, respectively. Moving an ion across the membrane into the cell causes a charge-
related free-energy change of size ΔGcharge. This change depends on the charge z of
the ion and the voltage V between the inside and outside of the cell. Specifically, the
change in energy ΔGcharge is given as zFV per mole ion, where F is Faraday’s constant.
A resting cell approaches a state where the total energy is minimal. This state is
therefore characterized by
C 
− RT ln  o  + zFV = 0. (12.6)
 Ci 
It is typically more useful to turn this equation around in order to compute the
equilibrium potential or voltage at rest as
RT  Co 
V= ln . (12.7)
zF  Ci 
Furthermore, the logarithm with base 10 (log10) is often preferred over the natural
logarithm (ln), and if we are considering a human cell, the temperature is more or
less constant at 37°C, which corresponds to 310.15 K. Thus, with R = 1.987 cal K−1
mol−1, F = 2.3 × 104 cal V−1 mol−1 and ln 10 = 2.3026, the equilibrium potential for a
singly charged ion is
C 
V ≈ 61.7 log10  o 
 Ci 
[mV ]. (12.8)
For the example of a typical cardiomyocyte, the intra- and extracellular potassium
concentrations, [K]i and [K]o, are about 140–150 and 4–5 mM, respectively, so that
the equilibrium potential for potassium, VK, is roughly between −90 and −100 mV.
The same type of computation holds for other ions, and the laws of physics allow
us to compute the total potential by an appropriate summation of concentrations, if
we account for the fact that different ions have different permeabilities through the
membrane. Specifically, the membrane potential VM is the equilibrium potential at
which the voltage gradients and the concentration gradients of all ions are in
balance, so that there is no flux of ions in either direction. This membrane potential
is formulated as the Goldman–Hodgkin–Katz equation, which for a combined
potential including K+, Na+, and Cl− reads
RT  PK [K + ]o + PNa [Na + ]o + PCl[Cl − ]i 

VM = ln . (12.9)
zF  PK [K + ]i + PNa [Na + ]i + PCl[Cl − ]o 
Note that the subscripts i and o for chloride are exchanged, since chloride is an
anion, and that the different quantities P quantify the permeability for each ion and
depend on the properties of the membrane [7].
Although they are very important, ions like hydrogen and magnesium are often
ignored in this computation, because their concentrations are orders of magnitude
smaller than those of K+, Na+, and Cl− [48]. The membranes of most excitable cells
are much more permeable to K+ than to Na+ and Cl−, which results in a resting mem-
brane potential that is closer to the equilibrium potential for K+ than to that for Na+
and Cl− [51].
12.9 Resting Potentials and Action Potentials

When an excitable cardiomyocyte is relaxed, its membrane is negatively charged at
about −90 mV. This state is called the resting potential and describes the electrical
potential across the cell membrane, when the outside is considered 0 mV. Electrical
stimulation of sufficient strength, or a change in ion balance, changes this state by
causing ion channels to open. When positively charged Na+ and Ca2+ ions enter the
cell, the membrane begins to depolarize, that is, its charge becomes more positive.
Intriguingly, only relatively few Na+ ions are needed to trigger depolarization. After
a brief period of time, K+ ions begin to leave the cell, the cell repolarizes (its
membrane potential becomes more negative), and the cell returns toward its
negative charge. Na+/Ca2+ and Na+/K+ exchangers reestablish the resting potential.
One cycle of up-and-down changes in voltage constitutes an action potential. While (A)
some aspects of action potentials are the same for excitable cardiomyocytes and for
action potential (mV)

cells of the SA node, there are also significant differences. We discuss the two cell 0
types separately. 0 3
An intriguing feature of SA node cells is that they do not have a true resting
potential (Figure 12.12A). Instead, they spontaneously depolarize and therefore
can generate repeated action potentials, a phenomenon called automaticity. In –50 4 4
comparison with other cardiomyocytes, which cycle through five phases during
each action potential (see below), SA node cells have only three phases (4, 0, and 3). 0 1
Phase 4 is characterized by the most negative membrane potential, about −60 mV. s
This potential is less negative than in ventricular cardiomyocytes and sufficiently (B)
1
high for the activation of slow inward-pointing, mixed sodium–potassium currents,

2
called funny currents. These funny currents initiate the depolarization of the mem- 0
brane and thereby control the spontaneous activity of the SA node [52]. Once the
0
potential reaches about −50 mV, transient T-type calcium channels open and Ca2+
rushes into the cell, thereby depolarizing the cell even further. At −40 mV, additional –50 3
L-type calcium channels open and increase the Ca2+ influx further. At the same time,
potassium ceases to move out of the cell. During the following phase 0, more Ca2+ 4 4
enters the cell through L-channels, while the T-channels begin to close. Overall, the –100
0 0.2
depolarization takes longer than in other cardiac cells. In phase 3, the SA node cells s
repolarize. Potassium channels open, leading to an increased outward flow of K+. (C)
40
L-channels close, and the depolarizing inward Ca2+ currents subside. The action

potentials in atrioventricular (AV) node cells are similar to those in the SA node.
Of note is that the controlling funny currents can also be activated by cyclic 0
nucleotides such as cyclic adenosine monophosphate (cAMP) and by G-protein-

coupled receptor signaling (see Chapter 9). This metabolic signaling option –40
provides a mechanism for the autonomic nervous system to stimulate the beating of
the heart [53]. –80
In contrast to SA and AV node cells, ventricular and atrial cardiomyocytes, as
0 0.2 0.4 0.6
well as Purkinje cells, have a true resting potential, which is associated with diastole s
and characterized by a selective permeability to different ions. Specifically, the
membrane is permeable to K+ ions but essentially impermeable to other ions. As a Figure 12.12 Action potentials in heart
consequence, the resting potential of about −90 mV is close to the equilibrium cells [1, 29, 54]. (A) Action potential of a
potential with respect to K+. Cells in this resting state generate an action potential spontaneously depolarizing SA node cell.
only when externally stimulated. Under normal conditions, this stimulus ultimately (B) Action potential of a typical ventricular
originates at the SA node and spreads throughout the heart muscle, as we discussed cardiomyocyte. Phases 1 and 2 of the latter
before. Once triggered by adjacent cells, the action potential in these cells develops are missing in the action potential in the SA
node. (C) Canine cardiac action potential
very quickly, and its overall appearance is distinctly different from that of SA node
(green) in a healthy animal and in an animal
cells (Figure 12.12B). Starting again with phase 4, the cells have a resting membrane with congestive heart failure (red).
potential of about −90 mV. When the cell receives an action potential from a neigh-
boring cell, it begins to depolarize (phase 0). As soon as the charge exceeds a thresh-
old of about −70 mV, the process is followed by further, rapid depolarization that is
mediated by an increase in fast inward Na+ currents. During this phase, the cell
begins to contract. At the same time, the K+ channels close. Subsequently, the fast
Na+ channels are inactivated, and outward K+ channels begin to open, leading to a
brief hyperpolarization, before the cell begins to repolarize (phase 1). At this time,
contraction is still in progress. As we discussed before, the sodium influx changes
the membrane such that calcium begins to flow in as well, thereby triggering an
amplification through the calcium-induced calcium release mechanism and result-
ing in the flow of large amounts of calcium from the sarcoplasmic reticulum into the
cytosol [43], which we also mentioned before and which we will discuss in detail
later. Owing to the influx of calcium, the initial repolarization is delayed, which can
be seen in a plateau of the action potential (phase 2). The contraction ends, and the
cell begins to relax. The full repolarization occurs during phase 3, where the potas-
sium efflux is high, the Ca2+ channels are inactivated, and the membrane potential
becomes more negative. To restore the resting potential, Na+–K+ and Na+–Ca2+
exchangers, as well as the plasma membrane Ca2+ ATPase, become active. To retain
a long-term balance, the amount of Ca2+ leaving the cell equals the amount of trig-
ger Ca2+ that entered at the start of the cycle. Between phase 0 and most of phase 3,
the so-called effective refractory period, the cell cannot be excited again. This mech-
anism normally prevents spurious signals from spreading throughout the heart
QRS Figure 12.13 Diagram of an ECG

complex trace, showing depolarization and
PR ST
repolarization of the atria and ventricles
segment segment
in the heart. Standardization of the
recording speed allows measurements of
PR the heart rate from the intervals between
R
interval different waves. See the text for further
explanation.
QT
interval
T
1 mV
Q
S
0 0.2 0.4 0.6 0.8

time (s)
muscle and ensures that there is sufficient time to fill the heart with blood and eject
it in a controlled manner.
It should be noted that the actual shapes of action potentials vary among differ-
ent regions of the heart, which is important for the correct propagation of electric
activity and normal cardiac function throughout the heart muscle [29]. Some heart
diseases are associated with a change in the pattern of a normal action potential
(Figure 12.12C).
The different phases of the action potential can be directly related to features of
an ECG (Figure 12.13). The PR interval between the onset of the P wave and Q has a
length of 120–200 ms and measures the time between the onset of atrial and ven-
tricular depolarization. Within this interval, the P wave reflects the depolarization
spreading from the SA node throughout the atria, which lasts for 80–100 ms. During
the brief isoelectric period afterwards, the impulse travels to the AV node. The QRS
complex is the phase of rapid ventricular depolarization, which lasts for 60–100 ms.
The ST segment represents the isoelectric period during which the entire ventricle is
depolarized. It corresponds more or less to the plateau phase in the ventricular
action potential. The T wave represents ventricular repolarization. This process
takes longer than the depolarization during the P wave. The QT interval corresponds
roughly to the average ventricular action potential and ranges between 200 and
400 ms, but is shorter at high heart rates. Even seemingly minute abnormalities in
any of these constituents of the ECG trace are often signs of various pathologies.
12.10 Models of Action Potentials

The first scientists to study action potentials in excitable cells quantitatively were
Alan Hodgkin and Andrew Huxley. They did their legendary work on axons of squid
neurons in the early 1950s [55] and were awarded the Nobel Prize for their efforts in
1963. A key to their success was the recognition of a direct analogy between the
functioning of a nerve cell and a simple electrical circuit. A few years later, Noble [8,
9, 56] and others [57, 58] demonstrated that the Hodgkin–Huxley model for nerve
cells could be adapted to represent cardiac action potentials, which have a much
longer repolarization phase than neurons.
Hodgkin and Huxley exploited the analogy with electrical circuits to apply well-
established laws from the theory of electricity to action potentials. They interpreted
the lipid bilayer of the membrane as a capacitor, ion channels as conductors,
electrochemical gradients driving the flow of ions as batteries, and ion pumps as
current sources. To refresh our memory, let us revisit the basic terms of electricity.
• The most fundamental, subatomic property that describes relative amounts of
electrons and protons in atoms and molecules is the electrical charge q.
Of particular importance here are ions, which have either an excess or a shortage
of electrons compared with neutral atoms. The electrical charge of a macro-

scopic entity is equal to the sum of the charges of all its particles. Thus, in excit-
able cells, the electrical charge is the net positive or negative charge carried by
the ions in the system. A charge creates an electrical field in the surrounding
space, which exerts a force on other electrically charged objects.
• An electrical conductor is a material that contains movable electrical charges. In
electrical circuits, conductors typically consist of metals such as copper, while
the conductor in our case consists of watery solutions containing ions and the
channels that allow ions to cross the lipid bilayer.
• Electric current is the actual movement or flow of electric charge. It is typically
denoted by I and, because it describes the change in charge at a given point, it is
given as
I = q . (12.10)
• The electrical potential is the potential energy per unit of charge in an electric
field. The difference between the electrical potentials at two points is called volt-
age (denoted by V) or voltage drop and quantifies the ability to move electrical
charge though a resistance. Sometimes this potential difference is called the
electromotive force, or EMF.
• Capacitance, denoted by C, is a measure of the amount of stored electrical charge
in relation to an electric potential. It is related to charge and voltage as
q
C= . (12.11)
V
• Electrical conductance and resistance are properties of materials and are inverse
to each other. Resistance R describes how much the flow of charged particles,
such as ions, is impeded, while conductance g measures how easy it is for charge
to flow through some material. It is the inverse of resistance:
1
g= . (12.12)
R
Famous physicists such as Maxwell, Faraday, Ohm, Coulomb, and Kirchhoff

discovered laws relating the various quantities in electrical systems. Of particu-
lar relevance here are three.
• Ohm’s law states that the current through a conductor is directly proportional to
the voltage across the two points, and inversely proportional to the resistance
between them. Thus, for time-dependent V and I, we have
I
V = IR = . (12.13)
g
• Kirchhoff’s two most important laws address the conservation of charge and
energy at any point in an electrical circuit (with the exception of a capacitor).
Thus, if there are k currents flowing toward or away from a point, their sum is
zero. In mathematical terms,
∑I
k
k = 0.
(12.14)
• Similarly, the sum of all k voltages around any closed electrical circuit is zero,
if the positive and negative signs are assigned correctly. Thus,
∑V = 0.
k
k
(12.15)
Because an electric current describes the flow of electric charge, the total current
equals the sum of all currents in the system. In other words, the currents in a system
with k resistances are additive in the sense that
V 
I= ∑ I = ∑  R  = V ∑ g .
k
k
k
k
k
k (12.16)
These relationships look rather static, but they also apply in a system whose
electrical state changes over time. Of particular importance here, dynamic changes
in voltages and currents across the membrane of a cardiac cell can lead to the gen-
eration of a time-dependent action potential. To describe this action potential
mathematically, we acknowledge explicitly that voltage and currents are time-
dependent by defining V(t) as the difference between internal and external voltages:
V(t) = Vi(t) − Ve(t). V(t) is a time-dependent variable, which we may differentiate
with respect to time. Using (12.10), we obtain
1 I
V = q = . (12.17)
C C
According to the laws of electricity, the sum of the capacitive current I and the
ionic current Iion at the membrane must be zero at every time point t. Substituting
this conservation law into (12.17) and rearranging yields
CV + I ion = 0. (12.18)
This is our fundamental equation. However, in order to account for the various ion
fluxes, which sometimes run in opposite directions, it is now necessary to split the
voltage and ionic current into partial voltages and currents that are associated with
Na+, K+, and a pool L, which accounts for all other ions, such as calcium, chloride, and
magnesium. The symbol L was chosen to suggest leakage. Thus, we define VNa, VK, and
VL, representing contributions to the resting potential, along with the corresponding
currents INa = gNa(V − VNa), IK = gK(V − VK), and IL = gL(V − VL). Recalling the relation-
ship (12.13) between current, conductance, and voltage, we can rewrite (12.18) as
CV = − g eff (V − Veq ). (12.19)
Here, the effective conductance geff = gNa + gK + gL is composed of the conductances

related to Na+, K+, and the collective pool L for other ions. The resting (or equilibrium)
potential is composed of individual contributions of conductances and voltages as
g NaVNa + g KVK + g LVL

Veq = . (12.20)
g eff
These terms are substituted into (12.19), and one may furthermore account for the
membrane current, which is 0 at equilibrium or is equal to Iapp applied to the system
in typical experiments. It is customary in this field not to express relationships in
absolute potentials or voltages but instead to consider a relative potential v that
describes the deviation from the resting potential:
v = V − Veq . (12.21)
With this convention, we obtain an ODE describing dynamic changes in voltage as
Cv = − g Na (v − vNa ) − g K (v − vK ) − g L (v − v L ) + I app . (12.22)
Note that the capacitance C and the different conductances g are properties of the
material and are therefore constant over time in most electrical circuits. In the
case of the heart, the capacitance is related to the thickness of the membrane, which
is not entirely constant in a cardiac cell, but C is usually considered constant,
which is probably a reasonable approximation over the short period of time of an
action potential. Importantly, the conductances correspond to channels in

membranes, and some of these are ion-gated or voltage-gated and are therefore
functions of the electrochemical milieu of the cell. In other words, both the voltage
and the conductances are functions of time and far from constant.
Hodgkin and Huxley formulated the dependences of conductances on voltage
essentially as black-box models that were inspired by experimental data, but do not
have a direct interpretation in terms of biological mechanisms. In a stroke of genius,
they defined
g K = g Kn 4 ,
(12.23)
g Na = g Nam 3h ,
and considered gL as constant, because it was seen as negligible in comparison with

gNa and gK. The terms g K and g Na were defined as constant conductivity parameters,
while the quantities n, m, and h, which may be interpreted as characteristics of the
potassium and sodium gates, were given as the solutions of differential equations
that also depend on voltage:
n = α n (v )(1 − n) − β n (v )n ,
= α m (v )(1 − m) − βm (v)m ,
m (12.24) 100
h = α h (v)(1 − h) − β h (v )h.
voltage (mV)
The terms n4 and m3h are proportional to the fractions of open ion channels or the 45
probability that these channels are open. If these quantities are 1 or 0, then all
corresponding channels are open or closed, respectively. The power of n was chosen to
reflect four subunits of the potassium channel, which at the time was a postulate.
–10
The dependences on voltage in the Hodgkin–Huxley model are typically formulated as 0 10 20
time (ms)
α n (v ) = 0.01(10 − v )(e (10−v )/10 − 1)−1 , β n (v ) = 0.125e −v /80 ,
Figure 12.14 Simulated action potential
α m (v ) = 0.1(25 − v)(e (25−v )/10 − 1)−1 , βm (v ) = 4.0e −v /18 , (12.25) in a nerve cell. The action potential was
α h (v ) = 0.07e −v /20 , β h (v ) = (e (30−v )/10 + 1)−1 , computed from the Hodgkin–Huxley model
(12.22)–(12.25). The shape is different than in
cardiomyocytes, but an adjusted Hodgkin–
with initial values specified as v(0) = 10.5, n(0) = 0.33, m(0) = 0.05, h(0) = 0.6, and Huxley model can generate cardiac action
C = 1, vNa = 115, vK = −12, vL = 10.6, g Na = 120, g K = 36, and gL = 3 (see, for example, potentials as in Figure 12.12B.
[7, 59]). The apparently arbitrary form of the equations for n, m, and h comes from
arguments describing how these quantities decay over time in a first-order fashion.
With these definitions and settings, the dynamics at the membrane is described
as a system of four ODEs. The first represents the action potential itself, (12.22),
while the remaining three are equations describing time-dependent changes in
conductances and are defined as in (12.24) and the associated functional depen-
dences (12.25). Putting the four equations into a numerical solver and specifying the
initial values, we quickly obtain the fruit of our labor (Figure 12.14): an action
potential in a nerve cell that shows rapid depolarization and repolarization.
The numerical solution also shows the dynamics of the auxiliary variables n, m, and
h, even though they do not really have a direct meaning, except insofar as they 1.0
n
represent features of the different ion channels (Figure 12.15). m
The appearance of this action potential reflects that of nerve cells. It is much h
faster than in cardiomyocytes, and the pronounced plateau during repolarization in
0.5
phase 2 (see Figure 12.12B) is entirely missing. Furthermore, the nerve model exhib-
its a slight undershoot before returning to the resting potential. Nonetheless,
variations in Hodgkin and Huxley’s equations, such as adaptations in the represen-
tation of the potassium current, were shown to lead to longer action potentials and 0
repolarization plateaus as we find them in cardiomyocytes [8, 9, 56–58]. 0 10 20
time (ms)
Before we continue, let us summarize the key points of this section. The electro-
chemical changes during contraction and relaxation in a nerve cell—and by extension
Figure 12.15 Temporal responses of
in a cardiac cell—can be modeled in analogy to an electrical circuit, and Hodgkin gating variables in the Hodgkin–Huxley
and Huxley showed that the physical laws describing such a circuit can be applied to model (12.24). Close inspection shows that
this physiological system. The most significant difference between a standard electri- the variables do not return to their starting
cal circuit and a cell is that some of the channels in the cell membrane are gated, that values after the action potential.
is, controlled by voltage or ions, whereas conductance in a typical electrical circuit is

constant. As a consequence, there is a logical cycle of influences where ion concen-
trations affect voltage and voltage affects the movement of ions and thereby the
chemical and electrical balances in the system. Hodgkin and Huxley modeled the
voltage dependence of the channels based on mathematical assumptions regarding
the functional dependences and by fitting experimental data. The amazing outcome
is that their model works very well under a number of conditions.
12.11 Repeated Heartbeats

We now have an action potential model at our disposal, (12.22)–(12.24), but it fires 100
only once. If we simply run the simulation longer, nothing happens. The system has
approached a steady state and stays there. How does it lead to pacemaker activity and
a heartbeat? The vast majority of the heart cells receive electrical signals from the SA
voltage (mV)
node, through the AV node and the Purkinje fibers, which are mediated through gap
45
junctions. Thus, their triggers are changes in the action potential of their membrane.
The pacemaker cells, in turn, autonomously fire in a process of depolarization that
involves influxes and effluxes of ions, and we have seen before that depolarization
begins with funny currents, but really proceeds when Ca2+ rushes into the cytosol.
Let us recall that n, m, and h in the Hodgkin–Huxley model describe voltage- –10
0 40 80
dependent gates and that their conductance is affected by changes in voltage and
time (ms)
thus by changes in ions such as calcium. If we look more closely at the dynamics of
n in the action potential model, we see that its final value is about 0.415, while its Figure 12.16 Creating repeated
initial value is about 0.33 (see Figure 12.14). What happens if we artificially reset the heartbeats. Artificial resetting the value of
value of n to the initial value after the action potential has fired? This is a simple n in (12.24) to the initial values at times 20,
simulation experiment and, voilà, the action potential fires again. In Figure 12.16, 40, and 60 results in a beating pattern of
we repeatedly reset n at times t = 20, 40, and 60. We can see that the artificially repeated action potentials.
induced potentials are a little lower than the first potential, but that the shape of the
subsequent beats is identical. Would we receive the same peaks in all cases if we not
only reset n, but also m and/or h?
Because a four-variable system is difficult to study mathematically, Fitzhugh and
Nagumo developed a simpler variant of the Hodgkin–Huxley model, in which they
separated the variables into fast and slow variables [60, 61]. The result had only two
variables, and it turns out that their model is a variant of the van der Pol oscillator
that we discussed at the beginning of this chapter and in Chapter 4! We have come
full circle and in the process given some biological justification to van der Pol’s
expectation [34] that his oscillator might describe heart rhythms and maybe even
serve as a pacemaker. Many researchers have performed very elegant analyses of
the mathematical structure of the Fitzhugh–Nagumo equations and their bifurca-
tion structure; the interested reader is referred to [7, 59].
To generate the repeated action potential, we cheated a little bit by manually
resetting n at times 20, 40, and 60. Obviously, the real SA node does not need our
input for every beat. So we should ask whether it is possible to replace this artificial 120
repeated triggering with a more natural pacemaker. One candidate could be our cal-
cium oscillation model, because we can easily argue that the calcium oscillations are
associated with depolarization and repolarization. The strategy is therefore simple:
voltage (mV)
we set up a simulation program that contains the Hodgkin–Huxley model (12.22)– v

50
(12.25) and also the calcium oscillation model (12.5) and then couple the two sys-
tems. For instance, let us simulate a situation where the calcium concentration in
the cytosol, y, affects the conductance of h. The simplest, but not the only, way of
achieving the coupling is simply to add y to the equation for h. Revisiting the output y
of the calcium model, we see that the value of y is mostly low, close to 0, but spikes –20
regularly up to about 0.4. These low values make sense, because the calcium level in 0 50 100
time (ms)
the cytosol is about 10,000 times lower than calcium levels in the extracellular space.
The simulation indeed shows the desired effect: without artificial intervention, the
Figure 12.17 Creating self-sustained
action potential spikes and depolarizes in a repeated manner (Figure 12.17), as we
heartbeats. The combined Hodgkin–
see it in an ECG. Of course, our success is no proof that the stitched-together model Huxley/Somogyi–Stucki model generates
is an appropriate model of a beating heart, but it demonstrates how complex prob- repeated action potentials without external
lems might be approached in a modular fashion, at least as a proof of principle and cues. Note that y is shown with twenty-fold
a means for investigating the internal logic of a system. magnification.
ISSuES OF A FAILING HEART 355
ISSuES OF A FAILING HEART

Although the heart is incredibly resistant to fatigue and faithfully keeps on pumping,
age and disease eventually take a toll. Uncounted things in the complex systems that
work within the heart with unparalleled precision for most of our lives can go wrong,
and the result is often heart failure, which presently affects about 5 million individu-
als in the United States, of whom about 20% die within 1 year of diagnosis. Indeed,
heart failure is the leading cause of death in the Western world.
It is very clearly beyond the scope of this chapter to elaborate on all, or even the
most prevalent, physiological changes in the heart that might result in heart failure.
Instead, as a glimpse into multiscale modeling, the following sections discuss just
one thread that connects the intracellular level with failure at the organ level. In line
with our emphasis throughout the chapter, we focus on action potentials and the
dynamics of calcium.
The first class of heart problems is related to abnormalities in the pacemaker
system, which can lead to different arrhythmias. For instance, in sick sinus syndrome,
which is often related to senescence of the SA node, the heart rate is too slow, which
can lead to dizziness, sluggishness, and shortness of breath. The slow heartbeat
(bradycardia; bradys is Greek for slow) may alternate with fast atrial arrhythmia
(tachyarrhythmia; tachys is Greek for fast), leading to brady–tachy syndrome.
A delay in conduction between the SA node and the AV node can lead to a first-
degree heart block, which often does not lead to noticeable symptoms, because the
AV node can serve as a sort of back-up pacemaker. By contrast, a block below the
AV node, namely in the bundle of His and the Purkinje system, may be fatal if
untreated. An implanted pacemaker may alleviate such a block and prevent cardiac
arrest.
A fast heartbeat sometimes originates in one of the ventricles, leading to a life-
threatening arrhythmia called ventricular tachycardia. A very common cause of this
condition is scarring of the heart muscle due to an earlier heart attack. The scar does
not conduct electrical signals, so that the signal propagates on both sides around the
damaged tissue. Earlier, we mentioned that three-dimensional modeling with finite
element methods has become sophisticated enough to mimic this disease on the
computer, and the result of such an analysis was shown in Figure 12.3.
A seemingly minute, but indeed significant, change in the proper functioning of
the heart during aging is a slowing down of the calcium fluxes between the sarco-
plasmic reticulum (SR) and the cytosol [62]. If the calcium dynamics in individual
myocytes becomes less efficient, the consequences are felt in the contractions of the
heart: it does not fill with blood and empty as quickly as it used to. The slower
calcium dynamics often goes hand in hand with an age-related stiffening of the
arteries, and together they can lead to shortness of breath, especially during strenu-
ous activities. These systemic connections can be explained by the mechanical
features of the heart: perfusion is reduced, and this reduction leads to an accumula-
tion of fluid beneath the skin, called edema, and to the entire spectrum of problems
associated with congestive heart failure.
Another change in cardiomyocyte activity that is associated with slower calcium
dynamics is a lengthening of the action potential during repolarization (see
Figure 12.12C). The longer action potential keeps the ion gates in the cell membrane
open for a slightly longer period of time, thereby allowing more calcium to enter and
exit the cytosol between beats and giving the weakened SR more time to release or
take up calcium. The downside is a higher risk of unbalanced calcium flow and a
slower response to increased demands on heart activity. Age also affects the con-
tractile elements in the myocytes, and in particular the heavy chains of myosin,
which are responsible for forming cross-bridges with actin (see Figure 12.4).
Some of these changes can be detected in the expression of genes, which is different
in young and old individuals. Finally, as with nearly all cells, cardiomyocytes are
susceptible to unstable oxygen molecules, called free radicals, that create oxidative
stress within the cells and may impair energy metabolism and other functionalities
by damaging proteins, DNA, membranes, and other complex structures in the cell.
In cardiomyocytes, free radicals are doubly problematic, because of the high energy
demands of these cells and because free radicals can damage the calcium pumps in
the membrane of the SR, which connects this aging mechanism again to reduced
calcium dynamics. During aging, many cardiomyocytes die, and even in healthy
septuagenarians, the number of cardiac cells may be decreased by 30%.
To the detriment of many elderly people, some of the effects of aging on the heart
form a vicious cycle that gradually diminishes efficiency. When the calcium dynam-
ics slows down and the arteries stiffen, the heart has to work harder to compensate.
As in other muscles, the harder work leads to enlargement of the cardiomyocytes,
which in turn leads to a thickening of the heart walls, as described by the Frank–
Starling law. This normal thickening is exacerbated by diseases such as coronary
heart disease and high blood pressure. Together, the metabolic and anatomic
changes make the heart more susceptible to cardiovascular conditions such as left
ventricular hypertrophy and atrial fibrillation. The latter is a common cardiac
arrhythmia that involves the two atria: the heart muscle quivers instead of contract-
ing in a coordinated fashion. Strong fibrillations may lead to heart palpitations and
congestive heart failure.
12.12 Modeling Heart Function and Failure Based

on Molecular Events
The models that we have developed in this chapter are snapshots of heart function
at different levels. Considering these snapshots collectively will give an impression
of how difficult it is to connect heart function and failure to specific molecular
events through comprehensive, true multiscale models, as they are envisioned in
the Physiome [13] and Virtual Physiological Rat [63] Projects. Nonetheless, the
following will discuss in broad strokes how our models can inform models at higher
levels in order to connect intracellular calcium-associated events with macroscopic
heart problems at the organ level. This line of reasoning with models was proposed
by Raimond Winslow and his collaborators over a span of many years. Most of the
technical details will have to be skipped here since they involve a lot of mathemat-
ics, but the interested reader may start diving into this field with reviews such as
[64–66].
We begin our journey at the molecular scale, and specifically at the cell
membrane of the cardiomyocyte, the sarcolemma. The sarcolemma contains
invaginations into the cytosol that effectively increase its surface area. The invagina-
tions, called T tubules, contain calcium channels that functionally connect the
sarcolemma and the membrane of the tubular network of the SR (Figure 12.18).
This functional connection occurs in spatially restricted microdomains of about
T-tubule sarcolemma
dyad
Ca2+
Ca2+ Figure 12.18 Simplified illustration of a
dyad and its role in the cardiac cycle [67].
During the action potential, calcium from the
sarcoplasmic
extracellular space enters the dyad through
reticulum L-type Ca2+ channels (LCCs) in the T tubules.
LLCs The influx activates arrays of ryanodine
receptors (RyRs), which are located on
the opposite side of the dyad on the
sarcoplasmic reticulum (SR). The opening of
Ca2+
RyRs the RyRs leads to a calcium flux from the SR
into the dyad, from where it diffuses into the
cytosol, increasing its calcium concentration
greatly. The cytosolic calcium binds to
cytosol
myofilaments, which contract and shorten
the cell. (Data from Tanskanen AJ, Greenstein
extracellular space JL, Chen A, et al. Biophys. J. 92 [2007]
3379–3396.)
(A) 10 (B) 10 (C) 10 Figure 12.19 Three-dimensional structure

8 8 of a ryanodine receptor. (A) View from the
9 6
5
7 9 6 7 cytoplasm to the lumen of the sarcoplasmic
4 3 4 reticulum (SR). (B) View in the opposite
3 3 direction. (C) Side view. The numbers identify
2 1 9
5 different globular domains and TA denotes
TA 10 TA
the transmembrane assembly. (From
6
8 Zissimopoulos S & Lai FA. Ryanodine receptor
structure, function and pathophysiology. In
Calcium: A Matter of Life or Death [J Krebs &
M Michalak, eds], pp 287–342. Elsevier, 2007.
10 nm width, called dyads. Each cardiomyocyte contains roughly 12,000 such

dyads. From the lumen of the T tubules, the dyads can receive calcium through
specific L-type Ca2+ channels (LCCs) that peek through the sarcolemma. They can
also receive calcium from the SR through ryanodine receptor (RyR) proteins
(Figure 12.19). The role of this structural set-up is to facilitate the calcium-induced
calcium release (CICR) that we discussed before. In response to membrane
depolarization, the LCCs open and thereby produce a small flux of Ca2+ ions from
the T tubule into the dyad. The Ca2+ ions cause the RyRs on the opposite site of the
dyad to open, and the result is a strong calcium flow from the SR into the cytoplasm
that ultimately leads to cardiomyocyte contraction. This LCC–RyR excitation–
contraction mechanism is very fast and occurs in a range of micro- and milliseconds
[14]. With a length in the nanometer range and a volume of only about 10−19 liter, the
dyads may be very small, but they are “at the heart of the heart.” To allow relaxation
of the filament complex, intracellular calcium is returned to the SR by SERCA
(SR Ca2+ ATPase) pumps and also pumped out of the cell through sodium–calcium
exchangers or plasma membrane Ca2+ ATPase. The reduction in calcium concentra-
tion causes the bonds between the myosin and actin filaments to loosen.
Experiments have shown that not every RyR is associated with an LCC, but that
there is a stochastic distribution of one LCC among about five RyRs (Figure 12.20).
A careful computational analysis [64] showed that the particular shape of the RyR
protein restricts the movement of Ca2+ ions and funnels them to their target binding
sites. In contrast to a random process such as Brownian motion, this guiding action
enormously increases the efficiency with which an external excitation by calcium
ultimately leads to myocyte contraction.
(A) (B) (C)
Ryr
LCC
200 nm
Figure 12.20 Spatial distribution of RyRs on the surface of the sarcoplasmic reticulum and of fewer LCCs on the adjacent membrane of the
T-tubule, which is part of the sarcolemma. (A) Electron micrograph of Ca2+ release units in myocytes of a chicken heart. (B) Organization of cardiac RyR
units into ordered arrays. (C) Schematic representation of RyR and LCC units on opposing sides of a dyad. ((B) from Yin C-C, D’Cruz LG & Lai FA. Trends
Cell Biol 18 [2008] 149–156. With permission from Elsevier. (C) adapted from Tanskanen AJ, Greenstein JL, Chen A, et al. Biophys. J. 92 [2007] 3379–3396
and from Winslow RL, Tanskanen AJ, Chen M & Greenstein JL. Ann. NY Acad. Sci. 1080 [2006] 362–375. With permission from Elsevier. With permission from
John Wiley & Sons.)
The number of calcium molecules needed for mediating the signal from LCC to
RyR is small. Even at the peak of the calcium flux, only about 10–100 free Ca2+ ions
are found in each dyad. This low number has two consequences for any model
representation: the process is stochastic and the noise level is high—one may find
an average number of 50 ions in a dyad, but the number in any given dyad could also
be 20 or 120. A situation leading to such a random occupancy is typically represented
as a stochastic process [68, 69]. We briefly discussed stochastic processes in Chapters
2 and 4, comparing and contrasting them with deterministic dynamic processes.
The key feature of stochastic processes is that random numbers affect their progres-
sion. The simplest stochastic processes are probably flipping a coin and rolling a die.
Typical models for stochastic processes are Poisson processes and Markov models.
The defining feature of the dyad is the channeling of ions, and the location of all
Ca2+ ions in the vicinity of dyads is therefore of great importance. As we discussed in
Chapters 2 and 4, the consideration of spatial aspects in a dynamical process typi-
cally calls for a representation in the form of a partial differential equation (PDE).
Thus, to determine whether a Ca2+ ion will enter a dyad, a detailed model should
quantify the probability P(x, t) that a Ca2+ ion is close enough to and properly aligned
with a dyad position x at time t. From a mathematical point of view, the combination
of free motion and stochastic binding events would suggest as the default model a
representation of the movement of individual calcium molecules in the form of a
stochastic PDE, which is very complicated. Indeed, such detailed focus on each and
every dyad, or maybe even every Ca2+ ion, would be extremely involved, because of
the roughly 12,000 dyads per cell and the billions of Ca2+ ions flowing between the
extracellular space, the cytosol, and the SR. As a consequence, simulations would
take weeks to complete, even on a high-performance computer, and computation
time would become prohibitive even for modeling one single cardiomyocyte,
let alone the entire heart.
It may be infeasible to model processes with such resolution and detail, but dis-
cussion of the intricate molecular details and biologically observed mechanisms is
important because it indicates how a stochastic model design would proceed in
principle and also because it allows us to assess where one might make valid simpli-
fying assumptions without compromising realism. To the rescue in our specific case
comes the very fact that makes a direct simulation so complicated, namely, the large
number of dyads and the fact that they are more or less randomly distributed along
the T tubules. A reasonable first simplification in such a situation is to ignore spatial
considerations, with the argument that one may study a local area of the cell
membrane, a single T tubule, or even a single dyad and subsequently scale-up the
results statistically to the thousands of dyads in each cell. The important advantage
is that this simplification step circumvents the need for PDEs to describe the detailed
mechanisms in favor of a model consisting of ordinary differential equations (ODEs)
that use average numbers of ions in dyads instead of a specific number in each dyad.
Before we make this step, we must convince ourselves that we do not lose too much
resolution. A strict mathematical justification would require setting up a detailed
model and comparing it with the simplifying approximation. Indeed, Winslow’s
group performed a careful analysis demonstrating that a simplified model based on
a single dyad retained important features of the CICR mechanism.
We cannot present the details of this analysis here, but we can ask ourselves
which aspects are really important to us. In the end, we need to know how many
calcium ions enter the collection of dyads. Does the large population of dyads buffer
the stochastic fluctuations among individual dyads and make their collective behav-
ior predictable? The affirmative answer comes from a fundamental law of statistics,
known as the central limit theorem. This theorem states that the sum of many ran-
dom events of similar character follows a normal distribution and that the standard
deviation of this distribution becomes smaller if more and more events are consid-
ered. In other words, the more items are part of the system, the more the population
of items will behave predictably like their average. The theorem is very difficult to
prove, but Exercise 12.17 suggests some easy computational experiments that might
convince you that it is correct.
Unfortunately, even the simplified model of one dyad is still too complicated as
a module in a multiscale model of a whole cardiomyocyte. A second opportunity for
simplification arises from a technique called separation of timescales. This generic
technique is useful if a system involves different types of processes, some of which

are very fast and others comparatively slow. The argument is that on the timescale of
the fast processes, the slow processes are essentially constant and their dynamics
can simply be ignored. On the timescale of the slow processes, it is often reasonable
to expect that much faster processes assume their steady-state level essentially
instantaneously and are therefore almost always constant. The result of this separa-
tion is a set of two or more simpler models, and one only analyzes the one whose
timescale is of particular interest in a given situation. In the case of calcium trans-
port associated with the dyads, the processes at LCCs and RyRs are comparable in
speed, but much faster than the release of large quantities of calcium from the SR.
A possible strategy is therefore to focus only on the LCC and RyR as a coupled
system that can be represented as a low-dimensional system of ODEs [66, 70].
Experimental work has shown that opening and closing of each LCC depend on
the local calcium concentration. Specifically, the data suggest that Ca2+ binding
induces the channel to switch from an active mode (mode normal) to an inactive
mode (mode Ca) where transitions to the open state are very rare. A first model
describing the opening and closing dynamics assumed that four subunits could
independently close each LCC [71, 72]. Thus, five active states (C0, …, C4) were
defined, with the subscript representing the number of permissive subunits. Simi-
larly, five inactive states were defined. As a result, any given LCC could be in one of
12 states, namely, 10 closed states (5 active and 5 inactive) and 2 open states (active
or inactive). In line with experimental results, the transitions among active states
and among inactive states were made voltage-dependent, while transitions between
closed and open states were made voltage-independent. The transitions from active
to inactive states were formulated to depend on the Ca2+ concentration, with an
increase in calcium causing the LCC to shift to a gating mode permitting only
infrequent opening.
Rigorous analysis of this model justified a simplification from five to two pairs of
active and inactive states. It also showed that the transition probability from an
inactive closed state to an open state was negligibly small. Thus, the result was a
simplified LCC model consisting of five states: two closed active states C1 and C2,
one open state O, which is accessible from C2, and two closed (Ca2+ inactivated)
states I1 and I2 accessible from C1 and C2, respectively (Figure 12.21A).
The rates of transitions between the two active states and between the two
inactivated states are fast (the rates denoted by f in Figure 12.21A). By contrast,
inactivation by Ca2+ and activation are slower (the rates denoted by s). This differ-
ence in rates suggested a separation of timescales and thus a possible further
simplification to only three states (Figure 12.21B), using the argument that the fast
transitions become less relevant on the timescale of the slow transitions.
It is straightforward to set up mass-action equations for the two systems. Imple-
menting both in the same program allows us to judge the loss in resolution due to
(A) (B)
f1 s1 s1
C1 C2 O C OO
f–1 s–1 s–1
s–2 s2 s–3 s3 s– s+
f2
I1 I2 I
f–2
Figure 12.21 Original and simplified models of calcium dynamics in dyads. Computational validation experiments are needed to assess the degree
to which the simpler three-state model in (B) reflects the responses of the more complex five-state model in (A). (Adapted from Hinch R, Greenstein
JL & Winslow RL. Prog. Biophys. Mol. Biol. 90 [2006] 136–150. With permission from Elsevier.)
model reduction. The two models, inspired by Hinch and co-workers [66, 70],
have the following forms:
Five-state model
C1 = 50C 2 + 0.5I1 − 40C1 − C1 , C1 = 0.60,

C 2 = 40C1 + I 2 − 50C 2 − 2C 2 + 0.5O − 2C 2 , C 2 = 0.49,
I1 = 40 I 2 + C1 − 60 I1 − 0.5I1 , I1 = 0.79, (12.26)
I2 = 60 I1 + 2C 2 − 40 I 2 − I 2 , I 2 = 1.18,
O = 2C 2 − 0.5O + perturb , O = 1.95;
Three-state model
C = 1.1I − 2C + 1.1OO − 2C , C = 1.09,

I = 2C − 1.1I , I = 1.97, (12.27)
= 2C − 1.1OO + perturb ,
OO OO = 1.95.
The two systems have parameters that lead to similar steady states for C1 + C2 and C,
I1 + I2 and I, as well as O and OO. The simulation begins close to these states. At time
t = 2, we perturb the open state (O, OO) by setting the parameter perturb = 1. At
t = 4, we send the signal perturb = −1, and at t = 6, we return to normal by setting
perturb = 0. The result is shown in Figure 12.22A. We see that the two systems
respond quite similarly.
For the second part of the simulation, we multiply all fast (f ) parameters by 0.05,
which makes them similar in magnitude to the slow (s) parameters, or by 0.005,
which makes them even slower. Figure 12.22B and C demonstrate that the systems
return to similar, although not exactly the same, steady states as before, but that the
transient dynamics diverge a little more than before (Figure 12.22A). This one set of
simulations suggests that the three-state model is doing quite well. In reality, one
would execute many more and different types of perturbations with actual parameter
values and compare the transients of the five- and three-state systems. If the
differences in results can be deemed negligible for all relevant scenarios, the three-
state model is a sufficient representation of the larger, five-state system.
The more realistic model of Hinch and colleagues [66] is substantially more
complicated than this illustration, because the voltage and calcium dependences of
the transitions are explicitly taken into account, parameters are measurable
quantities, obtainable from specific experiments, and the channel current is
represented with the Goldman–Hodgkin–Katz equation (12.9) that we discussed
before. Even in this more complex implementation, validation studies have attested
that the reduction to the three-state model was of acceptable accuracy.
Both experiments and mechanistic models indicate that RyR inactivation is the pri-
mary mechanism of termination of calcium release from the SR. This inactivation pro-
cess has characteristics similar to the LCC activation process, and model construction
begins again with five variables for different states of the RyR, which can be reduced to
three states, with similar arguments of timescale separation as for the LCC model.
(A) (B) (C)

4 ts = 1 ts = 0.05 ts = 0.005
concentration
Figure 12.22 Comparison of responses

of the five-state and three-state models
(12.26) and (12.27). The accuracy of the
approximation generated with the simpler
0 model depends on the parameter settings
0 7.5 15 0 7.5 15 0 7.5 15
time and, in particular, on the magnitude of the
timescale separation ts. See the text for
Ctot Itot O C I OO
details.
OuTLOOk FOR PHYSIOLOGICAL MuLTISCALE MODELING 361
The next step toward a more comprehensive model is the merging of the LCC
and RyR models. A minimal representation for each such calcium release unit
(CaRU) consists of one LCC, a closely apposed RyR, and the dyadic space. One LCC
opening may activate four to six RyRs (see Figure 12.20). The Ca2+ flux from the
dyadic space to the cytoplasm is assumed to be governed by simple diffusion. Based
on this assumption, it is possible to establish an approximate relationship between
the concentration of calcium in the dyadic space and that in the cytoplasm. This
relationship, which also depends on the currents through the LCC and RyR, is very
important because it greatly reduces the number of differential equations in the
model: assuming that the relationship is valid, it is no longer necessary to compute
the calcium concentration for each dyad. The CaRU model can have nine different
states, resulting from combinations of the three LCC states and the three RyR states,
but can again be simplified to three states under the assumption that a separation of
timescales is legitimate.
Hinch and collaborators showed that all the different pieces (component
models) can be merged into a whole-cell Ca2+ regulation model that describes the
electrical, chemical, and transport processes in a cardiomyocyte very well [66]. This
model contains the CaRU system and the different pumps and ion channels that we
discussed before, as well as mechanisms of energy balance. Needless to say, this
model is rather complicated. Because of the strategy of simplifying details, followed
by validating each simplification, the model is not 100% realistic, but it seems suffi-
ciently accurate, as well as amenable to solution and analysis in a reasonable
amount of computer time.
Indeed, this model is accurate enough to allow the representation of some heart
diseases. Of particular importance is sudden cardiac death, which may have a wide
variety of root causes. Here, the models permit one chain of causes and effects, lead-
ing from changes in LLCs and RyRs to a prolongation of the action potential (see
Figure 12.12C), reduced Ca2+ flux, increased Ca2+ relaxation time, and subsequent
arrhythmia (cf. [73]). Interestingly, some of the molecular events can be seen in the
expression of ion channels and of genes that code for transporters and Ca2+ han-
dling proteins [29]. Such changes in expression can be represented as correspond-
ing changes in model parameters (see Chapter 11), and the model does indeed lead
to a prolonged action potential.
The models thus connect genetic, molecular, physiological, and clinical phe-
nomena in a stepwise, causative fashion. The same models can furthermore be
extended into a whole-heart model, which demonstrates how molecular aberrations
at the cellular level may lead to macroscopic changes, such as arrhythmias, ventricu-
lar tachycardia, and fibrillation. For such a whole-heart model, one must account for
the shape of the heart and the complex orientation of fibers within the heart muscle,
as we discussed early in the chapter. At this point, it is no longer possible to avoid or
simplify the PDEs or finite element representations that permit a detailed descrip-
tion of how electric impulses move along the surface of the heart (see Figure 12.3).
Nevertheless, because of the simplified, yet validated, submodels, in particular for
the LCC and RyR, it is actually feasible to construct detailed whole-heart models that
capture the entire chain of causes and effects. In the chain of events discussed above,
changes in channels, transporters, and genes affecting calcium dynamics within
individual cardiomyocytes lead to alterations in the electrical conductance in the
heart tissue, which may lead to arrhythmia and myocardial infarction.
OuTLOOk FOR PHYSIOLOGICAL MuLTISCALE

MODELING
This chapter has served two purposes. First, of course, it has shed light on the com-
plexity of the heart and on diverse modeling approaches. It is clear that no single
modeling strategy is superior to all the others, and that model structure and com-
plexity are directly related to the questions to be asked. For a full understanding, the
vastly differing aspects of the organ must all be understood and integrated. These
include the electrochemical processes that we have focused on in this chapter, but
also the structural, mechanical, physical, and morphological phenomena that are
crucial for the conductance of electrical signals and the beating and pumping action
of the heart. They should also allow investigations of the energy needs and of
changes in the heart due to exercise, diet, aging, and disease.
A tool for such a comprehensive task might ultimately be a template-and-anchor
model [6] that uses the high-level blood pumping function of the heart as the template
and embeds into this template electrophysiological, biochemical, and biophysical
submodels of much finer granularity as anchors. More likely will be a set of comple-
mentary models that focus on one or two levels and inform models at other levels, as
has been proposed in the Physiome and Virtual Physiological Rat Projects [10, 13, 63].
As a second purpose of this chapter, the discussion has shown how it is possible to
concatenate different models so that one may causally connect morphological, physi-
ological, and clinical changes in an organ to altered aspects in molecular structures
and events, and to variations in the expression state of the genome [64–66]. Such a
chain of models requires that large-scale systems be divided into functional modules,
which are analyzed by focusing on events at their specific spatial, temporal, and orga-
nizational scales. Within the human body, such systems of systems are abundant. The
liver consists of more or less homogeneous lobes, which themselves consist of hepa-
tocytes that produce over 1000 important enzymes and are essential for the produc-
tion of proteins and other compounds needed for digestion and detoxification. Ini-
tially ignoring the requirements and means of metabolite transport, much of the
function of the liver can be studied on the level of individual hepatocytes. However,
moving to higher levels of functionality, one must address the overarching functions
of the tissues and the whole organ, such as the regenerative potential of the liver with
all its spatial and functional implications. The Virtual Liver Network approaches this
multiscale topic with different types of computational models [16, 74].
Like the liver, other organs consist primarily of large numbers of similar mod-
ules, which suggests analogous modeling approaches [17, 63, 75]: the kidney is com-
posed of nephrons, and the lung consists of bronchi, bronchioles, and alveoli. While
this conceptual and computational dissection of large systems into smaller modules
is of enormous help for modeling purposes, caution is necessary, because the dis-
section sometimes eliminates crucial interactions between components. A prime
example is the brain, which consists of billions of neurons, whose individual func-
tion we understand quite well, but also of highly regulated brain circuits and other
physiological systems, which are often ill characterized. We have a good grasp of the
signal transduction processes occurring at synapses, where neurons communicate
with each other, but the sheer number of neurons, their biochemical and physiolog-
ical dynamics, and the complex distribution of synapses throughout the brain lead
to emerging properties such as adaptation, learning, and memory that we are still
unable to grasp (see Chapter 15). Beyond organs, we need to be recognizant of the
fact that problems in one system easily spread to other systems. For instance, a full
understanding of cancer without consideration of the immune system is without
doubt incomplete. One thing is clear: micro- and macro-physiology will continue to
present us with grand challenges, but also with a glimpse of where we might want to
target our research efforts in the years to come.
EXERCISES
12.1. As a nerdy Valentine’s gift, produce two- and (b) What happens if both equations in
three-dimensional hearts, using one of the many (4.76) are multiplied by the same positive
available heart curves [76–78], such as number? What happens if the multiplier is
negative? What happens if the two equa-
(1) ( x 2 + y 2 − 1)3 − x 2 y 3 = 0;
tions are multiplied by different
(2) x = sin t cos t ln | t |, y = | t | cos t , − 1 ≤ t ≤ 1; numbers?
(3) ( x 2 + 94 y 2 + z 2 − 1)3 − x 2 z 3 (c) Study the effects of larger perturbations on v.
Study the effects of perturbations on w; interpret
− 809 y 2 z 3 = 0, −3 ≤ x , y , z ≤ 3. what such perturbations mean.
Customize your hearts by changing or adding (d) Study simultaneous perturbations in both
parameters of the models. v and w.
12.2. (a) Explore the role of the parameter k in the van (e) Study the effects of regularly iterated perturba-
der Pol oscillator in Chapter 4; see (4.73)–(4.76). tions (given to the system at regular intervals by
Report your findings. an outside pacemaker).
EXERCISES 363
12.3. (a) Initiate the stable oscillator System B in (12.1) of this equation. What would it mean (biologically
with different initial values. Begin with X1 = 0.05, and mathematically) if rEC = pCE = 0, rEC − pCE = 0,
X2 = 2.5, and X1 = 0.75, X2 = 2.5. Try other combi- and rEC − pCE ≠ 0?
nations. Report and interpret your findings. 12.8. Confirm numerically that the oscillation in the
(b) Modulate the oscillator by slightly changing system (12.5) is stable. Is (12.5) a closed system?
some of the exponents or by multiplying the entire Explain.
systems or some terms by constant (positive or
12.9. Mine the literature or the Internet for three-
negative) factors. Report your findings.
dimensional molecular structures of channel
(c) Multiply both equations in the oscillator proteins in cardiomyocytes.
System A in (12.1) by the same X1 or X2, raised to
some positive or negative power, such as X 2−4 . 12.10. The Na+ concentration is usually much higher
Visualize the results with time courses and outside the cell (about 150 mM) than inside the
phase-plane plots in which X2 is plotted against X1. cell (between 5 and 20 mM). Furthermore, while
Describe in a report how the shape of the oscilla- there is a lot of calcium in cells, much of it is
tion changes. bound, and the free calcium concentrations are
much lower than those of Na+, with values of
12.4. (a) In the series of examples of spurious sine [Ca2+]i = 10−4 mM and [Ca2+]o between 1 and
function impulses, we changed the frequency. 3 mM. Compute the equilibrium potentials for
As a consequence, the oscillations became Na+ and Ca2+ and discuss their contributions to
chaotic, and eventually the entire heartbeat the membrane potential.
stopped. Explore what happens if the frequency
is retained (for instance, a = 10, a = 1, a = 0.5, or 12.11. The form of the equations for n, m, and h in the
a = 0.0001) but the amplitude of the oscillation is Hodgkin–Huxley model derives from the following
altered. argument. Let x be one of the three quantities; then
we can write its differential equation in simplified
(b) Explain why it is not clinically sufficient that a
form as
pacemaker generates a regular impulse pattern of,
say, 70 beats per minute. What else is needed? x = α x (1 − x ) − β x x .
12.5. Explore further the degree to which a highly Show that this equation has the solution
irregular heartbeat can be made regular by a
simplistic pacemaker (see the discussion of x (t ) = x ∞ + ( x 0 − x ∞ )e −t /τ x ,
Figure 12.9). Multiply all equations of the chaotic where x(0) = x0, x∞ = αx/(αx + βx), and τx = 1/
Lotka–Volterra (LV) oscillator in Chapter 10 by 50, αx + βx). Interpret x∞.
which should result in a rather irregular heartbeat 12.12. Study the effects of initial conditions (for v, n, m,
with about 60–70 beats per minute. Formulate a and h) on the shape of the action potential. Use as
sine oscillator (pacemaker) as a pair of ODEs. settings for other parameters vNa = 120 mV,
( s = qc, c = −qs). Add to each equation in the LV vK = −12 mV, vL = 10.6 mV, C = 1, and Iapp = 1.
system a term of the form p(s + 1). Confirm with
p = 0 the chaotic arrhythmia. Perform three series 12.13. Compare the results of the Hodgkin–Huxley
of experiments, with the goal of regularizing the model in the text with two other representations
arrhythmia. In the first series, fix q = 6, increase p for the gating functions n, m, and h. Use in both
from 0 in small steps, and study the resulting cases the parameters v(0) = 12, n = 0.33,
heartbeat. In the second series, set p at some m = 0.05, h = 0.6, C = 1, vNa = 115, vK = −12, and
small value and vary q. In the third series, vary vL = 10.6. Study the voltage and gating functions
both. Finally, select a combination of p and q that and discuss repeated beats.
seems to work well, and then alter the demand of (a) Explore the following functions (adapted
the body for more or fewer heartbeats by multi- from [79]) with g Na = 120, g K = 3.6, and gL = 0.3:
plying all LV equations with the same value from
the interval [0.5, 1.5]. 0.01(v + 50)
α n (v ) = ,
12.6. Describe what the setting of [Ca2+]ext as an 1 − e −(v+50)/10
independent variable in Section 12.7 entails. β n (v ) = 0.125e −(v+60)/80 ,
What is the importance of its numerical value? 0.1(v + 35)
What assumptions have to be made and justified α m (v ) = ,
1 − e −(v+35)/10
to assign the status of an independent rather than
a dependent variable? What would have to be βm (v ) = 4.0e −(v+60)/18 ,
changed if [Ca2+]ext were considered a depen- α h (v ) = 0.07e −(v+60)/20 ,
dent variable? Explain biologically and
1
mathematically. β h (v ) = − (v +30 )/10
.
1+ e
12.7. Adding the two equations in (12.4) together yields
the much simpler equation x + y = rEC − pCEy. (b) Use the following functions with g Na = 120,
Explain the biological and mathematical meaning g K = 3.6, and gL = 0.3:
0.01(50 − v ) the sum of numbers rolled, divided by the

α n (v ) = , number of dice. Compare and interpret the
1 − e (v−50)/10
results of these experiments and the previous
β n (v ) = 0.125e (v−60)/80 , experiment with one die. Create a histogram
0.1(35 − v ) of the results.
α m (v ) = ,
1 − e (v−35)/10 (3) Write a computer program to do a series of
βm (v ) = 4.0e (v−60)/18 , analogous experiments, where you assume
that each of 12,000 dyads contains a random
α h (v ) = 0.07e (v−60)/20 ,
number of Ca2+ ions that is between 10 and
1 100. Let the program graph the sum of ions,
β h (v ) = .
1 + e (v−30)/10 divided by the number of dyads, for each step
in the experiment. Interpret the results
12.14. Perform simulations assessing the effects of
mathematically and biologically. Create a
resetting v, n, m, and h, and combinations thereof,
histogram of the results.
in the Hodgkin–Huxley model. Report and
interpret your results. 12.18. Analyze the accuracy (loss of accuracy) due to
the separation of timescales with a simple
12.15. Add to the equation for n in (12.24) an artificial
linear pathway model with feedback that is
leakage term of the type −pn, where p is a small
affected by noise. The pathway has the structure
number. Test with simulations the effects of
shown in Figure 12.23 and the describing
different values of p on the dynamics of voltage v.
model is given by
If the system starts to beat, test whether the
pattern is a limit cycle. A strict mathematical test is x1 = px 5−0.3 − x10.5 , x1 = 2, p = 1,
difficult for these types of systems, leading
x 2 = x10.5 − x 20.75 , x 2 = 1,
Fitzhugh and others to study a reduced system in
which n and h were assumed to be constant. This x 3 = x 20.75 − x 30.4 , x 3 = 1,
assumption allowed them to study the bifurcation x 4 = x 30.4 − x 40.5 , x 4 = 1,
and limit-cycle behavior in a two-variable system
x 5 = x 40.5 − x 50.4 , x 5 = 1.
that was much easier to handle. If you have
sufficient mathematical background, read the The noise is modeled as a sine wave with small
elegant analysis in the original papers [60, 61] or in amplitude, namely
textbooks in the field [7, 59] and summarize the
highlights of the analysis. s = (1 − c ) ⋅ (ts ), s = 1, ts = 1,
12.16. Test with simulations whether the combined c = (s − 1) ⋅ (ts ), c = 0.8.
Hodgkin–Huxley/Somogyi–Stucki model truly
exhibits a stable limit cycle. The parameter ts represents the timescale. To
12.17. In the context of calcium dynamics in the dyad, we familiarize yourself with the two components,
evoked the central limit theorem. Explore this implement them separately in a differential
theorem in three steps: equation solver and explore the effects of
changing the initial value of x1, the value of the
(1) Roll one die 50 times and report the outcome parameter p, and the initial values of s and c.
on a graph showing on the horizontal axis the Once you have a feel for these components,
number of the roll (1–50) and on the vertical connect the two models by putting them
axis the result (the number rolled, between together in your solver, replacing p with s in
1 and 6). Explain to which aspect of the CICR the first equation and setting ts = 0. What do
system this experiment corresponds. Create a these settings mean? Now the actual series of
histogram of the results. experiments begins. For each experiment, set ts
(2) Repeat the experiment with 2, 3, and 4 dice to a different value between 0 and 1000.
simultaneously and report on the vertical axis Report and interpret your findings.
–
x1 x2 x3 x4 x5
noise
Figure 12.23 Simple pathway with feedback and exposure to noise. The effect of the noise depends
on the timescales of the pathway and the input noise.
REFERENCES 365
12.19. (a) Introduce more and different perturbations to different diseases or temporary problems with the
one or more variables of the systems in (12.26) heart. Similarly search for proteins. Pinpoint as
and (12.27). Explain the results. closely as possible where and how alterations in
(b) In the perturbations in (12.26) and (12.27), these genes or proteins affect heart function and
we set the parameter perturb first to 1, then to where they would enter any of the models we
−1, and then to 0. At the end, the systems are discussed.
back at their original steady states. Explain why 12.21. Mine the literature and the Internet for finite
the system does not return to its original steady element methods in heart modeling. Select one
state if, for instance, the second study using these methods and summarize it in a
perturbation is −2. two-page report.
12.20. Search the literature and the Internet for genes 12.22. Write a two-page report about spiral waves in
whose altered expression is associated with cardiovascular disease.
REFERENCES
[1] Klabunde RE. Cardiovascular Physiology Concepts, 2nd ed. [21] Allen D & Kentish J. The cellular basis of the length–tension
Lippincott Williams & Wilkins, 2011. relation in cardiac muscle. J. Mol. Cell. Cardiol. 9 (1985)
[2] Willis-Knighton Cardiology. Heart Anatomy. http://www. 821–840.
wkcardiology.com/Patient-Information/Heart-Anatomy/. [22] Shiels HA & White E. The Frank–Starling mechanism in
[3] Cherry E & Fenton F. The Virtual Heart. http://thevirtualheart. vertebrate cardiac myocytes. J. Exp. Biol. 211 (2008) 2005-13.
org/HHindex.html. [23] Smith L, Tainter C, Regnier M & Martyn DA. Cooperative cross-
[4] Cleveland Clinic. Your Heart and Blood Vessels. http:// bridge activation of thin filaments contributes to the Frank–
my.clevelandclinic.org/heart/heart-blood-vessels/default. Starling mechanism in cardiac muscle. Biophys. J. 96 (2009)
aspx. 3692–3702
[5] National Institute on Aging, US National Institutes of Health. [24] Smith NP, Mulquiney PJ, Nash MP, et al. Mathematical
Aging Hearts and Arteries: A Scientific Quest, Chapter 3: modeling of the heart: cell to organ. Chaos Solutions Fractals
Cellular Clues. https://www.nia.nih.gov/health/publication/ 13 (2002) 1613–1621.
aging-hearts-and-arteries/chapter-3-cellular-clues. [25] 3D model of blood flow by supercomputer predicts heart
[6] Full RJ & Koditschek DE. Templates and anchors: attacks. EPFL, École Polytechnique Fédérale de Lausanne,
neuromechanical hypotheses of legged locomotion on land. Press Release, 20 May 2010. https://www.sciencedaily.com/
J. Exp. Biol. 202 (1999) 3325–3332. releases/2010/05/100520102913.htm.
[7] Keener J & Sneyd J. Mathematical Physiology. II: Systems [26] Cherry EM & Fenton FH. Visualization of spiral and scroll waves
Physiology, 2nd ed. Springer, 2009. in simulated and experimental cardiac tissue. New J. Phys.
[8] Noble D. Cardiac action and pacemaker potentials based on 10 (2008) 125016.
the Hodgkin–Huxley equations. Nature 188 (1960) 495–497. [27] Joyner RW, Wilders R & Wagner MB. Propagation of pacemaker
[9] Noble D. A modification of the Hodgkin–Huxley equations activity. Med. Biol. Eng. Comput. 45 (2007) 177–187.
applicable to Purkinje fibre action and pace-maker potentials. [28] DiFrancesco D & Noble D. A model of cardiac electrical activity
J. Physiol. 160 (1962) 317–352. incorporating ionic pumps and concentration changes. Philos.
[10] Bassingthwaighte J, Hunter P & Noble D. The Cardiac Trans. R. Soc. Lond. B Biol. Sci. 10 (1985) 353–398.
Physiome: perspectives for the future. Exp. Physiol. 94 (2009) [29] Nerbonne JM & Kass RS. Molecular physiology of cardiac
597–605. repolarization. Physiol. Rev. 85 (2005) 1205–1253.
[11] Crampin EJ, Halstead M, Hunter P, et al. Computational [30] Nattel S. New ideas about atrial fibrillation 50 years on. Nature
physiology and the Physiome Project. Exp. Physiol. 89 415 (2002) 219–226.
(2004) 1–26. [31] Fink M, Niederer SA, Cherry EM, et al. Cardiac cell modelling:
[12] Hunter PJ & Borg TK. Integration from proteins to organs: the observations from the heart of the Cardiac Physiome Project.
Physiome Project. Nat. Rev. Mol. Cell Biol. 4 (2003) 237–243. Prog. Biophys. Mol. Biol. 104 (2011) 2–21.
[13] Physiome Project. http://www.physiome.org/Links/websites. [32] Nyhan D & Blanck TJJ. Cardiac physiology. In Foundations of
html. Anesthesia: Basic Sciences for Clinical Practice, 2nd ed. (HC
[14] Greenstein JL & Winslow RL. Integrative systems models Hemmings & PM Hopkins, eds), pp 473–484. Elsevier, 2006.
of cardiac excitation–contraction coupling. Circ. Res. 108 [33] van der Pol B & van der Mark J. Frequency demultiplication.
(2011) 70–84. Nature 120 (1927) 363–364.
[15] Waters SL, Alastruey J, Beard DA, et al. Theoretical models for [34] van der Pol B & van der Mark J.The heart beat considered
coronary vascular biomechanics: progress and challenges. as a relaxation oscillation. Philos. Mag. (7th series) 6 (1928)
Prog. Biophys. Mol. Biol. 104 (2011) 49–76. 763–775.
[16] Virtual Liver Network. http://www.virtual-liver.de/. [35] Yin W & Voit EO. Construction and customization of stable
[17] Tawhai MH, Hoffman EA & Lin CL. The Lung Physiome: oscillation models in biology. J. Biol. Syst. 16 (2008) 463–478.
merging imaging-based measures with predictive [36] Beard RW, Filshie GM, Knight CA & Roberts GM. The
computational models. Wiley Interdiscip. Rev. Syst. Biol. Med. significance of the changes in the continuous fetal heart rate
1 (2009) 61–72. in the first stage of labour. Br. J. Obstet. Gynaecol. 78(10) (1971)
[18] Blue Brain Project. http://bluebrain.epfl.ch/. 865–881.
[19] CellML. http://www.cellml.org. [37] Williams KP & Galerneau F. Intrapartum fetal heart rate
[20] SBML. The Systems Biology Markup Language. http://sbml. patterns in the prediction of neonatal acidemia. Am. J. Obstet.
org/Main_Page. Gynecol. 188 (2003) 820–823.
[38] Ferrario M, Signorini MG, Magenes G & Cerutti S. Comparison [59] Izhikevich EM. Dynamical Systems in Neuroscience: The
of entropy-based regularity estimators: Application to the Geometry of Excitability and Bursting. MIT Press, 2007.
fetal heart rate signal for the identification of fetal distress. [60] Fitzhugh R. Impulses and physiological states in
IEEE Trans. Biomed. Eng. 33 (2006) 119–125. theoretical models of nerve membrane. Biophys.
[39] Lipsitz LA. Physiological complexity, aging, and the path to J. 1 (1961) 445–466.
frailty. Sci. Aging Knowledge Environ. 2004 (2004) pe16. [61] Nagumo JA & Yoshizawa S. An active pulse transmission
[40] Lipsitz LA & Goldberger AL. Loss of “complexity” and aging. line simulating nerve axon. Proc. Inst. Radio Eng. 50 (1962)
Potential applications of fractals and chaos theory to 2061–2070.
senescence. JAMA 267 (1992) 1806–1809. [62] Mubagwa K, Kaplan P, Shivalkar B, et al. Calcium uptake by the
[41] Glass L & Mackey MC. From Clocks to Chaos: The Rhythms of sarcoplasmic reticulum, high energy content and histological
Life. Princeton University Press, 1988. changes in ischemic cardiomyopathy. Cardiovasc. Res. 37
[42] Thompson JMT & Stewart HB. Nonlinear Dynamics and Chaos, (1998) 515–523.
2nd ed. Wiley, 2002. [63] The Virtual Physiological Rat Project. http://virtualrat.org/.
[43] Endo M. Calcium release from the sarcoplasmic reticulum. [64] Winslow RL, Tanskanen A, Chen M & Greenstein JL. Multiscale
Physiol. Rev. 57 (1977) 71–108. modeling of calcium signaling in the cardiac dyad. Ann.
[44] Vinogradova TM, Brochet DX, Sirenko S, et al. Sarcoplasmic NY Acad. Sci. 1080 (2006) 362–375.
reticulum Ca2+ pumping kinetics regulates timing of [65] Cortassa S, Aon MA, Marbán E, et al. An integrated model
local Ca2+ releases and spontaneous beating rate of of cardiac mitochondrial energy metabolism and calcium
rabbit sinoatrial node pacemaker cells. Circ. Res. 107 dynamics. Biophys. J. 84 (2003) 2734–2755.
(2010) 767–775. [66] Hinch R, Greenstein JL & Winslow RL. Multi-scale models
[45] Somogyi R & Stucki J. Hormone-induced calcium oscillations of local control of calcium induced calcium release. Prog.
in liver cells can be explained by a simple one pool model. Biophys. Mol. Biol. 90 (2006) 136–150.
J. Biol. Chem. 266 (1991) 11068–11077. [67] Tanskanen AJ, Greenstein JL, Chen A, et al. Protein geometry
[46] Stucki J & Somogyi R. A dialogue on Ca2+ oscillations: an and placement in the cardiac dyad influence macroscopic
attempt to understand the essentials of mechanisms leading properties of calcium-induced calcium release. Biophys.
to hormone-induced intracellular Ca2+ oscillations in various J. 92 (2007) 3379–3396.
kinds of cell on a theoretical level. Biochim. Biophys. Acta 1183 [68] Xing J, Wang H & Oster G. From continuum Fokker–Planck
(1994) 453–472. models to discrete kinetic models. Biophys. J. 89 (2005)
[47] Bertram R, Sherman A & Satin LS. Electrical bursting, calcium 1551–1563.
oscillations, and synchronization of pancreatic islets. Adv. [69] Smith GD. Modeling the stochastic gating of ion channels.
Exp. Med. Biol. 654 (2010) 261–279. In Computational Cell Biology (CP Fall, ES Marland, JM Wagner
[48] Alberts B, Johnson A, Lewis J, et al. Molecular Biology of the & JJ Tyson, eds), pp 285–319. Springer, 2002.
Cell, 6th ed. Garland Science, 2015. [70] Hinch R, Greenstein JL, Tanskanen AJ, et al. A simplified
[49] Sipido KR, Bito V, Antoons G, et al. Na/Ca exchange and local control model of calcium-induced calcium release
cardiac ventricular arrhythmias. Ann. NY Acad. Sci. 1099 (2007) in cardiac ventricular myocytes. Biophys. J. 87 (2004)
339–348. 3723–3736.
[50] Ojetti V, Migneco A, Bononi F, et al. Calcium channel blockers, [71] Imredy JP & Yue DT. Mechanism of Ca2+-sensitive inactivation
beta-blockers and digitalis poisoning: management in the of L-type Ca2+ channels. Neuron 12 (1994) 1301–1318.
emergency room. Eur. Rev. Med. Pharmacol. Sci. 9 (2005) [72] Jafri MS, Rice JJ & Winslow RL. Cardiac Ca2+ dynamics: the roles
241–246. of ryanodine receptor adaptation and sarcoplasmic reticulum
[51] Sherwood L. Human Physiology: From Cells to Systems, load. Biophys. J. 74 (1998) 1149–1168.
9th ed. Cengage Learning, 2016. [73] Phillips RM, Narayan P, Gomez AM, et al. Sarcoplasmic
[52] DiFrancesco D. The role of the funny current in pacemaker reticulum in heart failure: central player or bystander?
activity. Circ. Res. 106 (2010) 434–446. Cardiovasc. Res. 37 (1998) 346–351.
[53] Lakatta EG, Maltsev VA & Vinogradova TM. A coupled system [74] Hoehme S, Brulport M, Bauer A, et al. Prediction and
of intracellular Ca2+ clocks and surface membrane voltage validation of cell alignment along microvessels as order
clocks controls the timekeeping mechanism of the heart’s principle to restore tissue architecture in liver regeneration.
pacemaker. Circ. Res. 106 (2010) 659–673. Proc. Natl Acad. Sci. USA 107 (2010) 10371–10376.
[54] Winslow RL, Rice J, Jafri S, et al. Mechanisms of altered [75] Physiome Project: Urinary system and Kidney; http://
excitation–contraction coupling in canine tachycardia-induced physiomeproject.org/research/urinary-kidney
heart failure, II: Model studies. Circ. Res. 84 (1999) 571–586. [76] Beutel E. Algebraische Kurven. Göschen, 1909–11.
[55] Hodgkin A & Huxley A. A quantitative description of [77] Heart Curve. Mathematische Basteleien. http://www.
membrane current and its application to conduction and mathematische-basteleien.de/heart.htm.
excitation in nerve. J. Physiol. 117 (1952) 500–544. [78] Taubin G. An accurate algorithm for rasterizing algebraic
[56] Noble D. From the Hodgkin–Huxley axon to the virtual heart. curves. In Proceedings of Second ACM/IEEE Symposium on
J. Physiol. 580 (2007) 15–22. Solid Modeling and Applications, Montreal, 1993, pp 221–230.
[57] Brady AJ & Woodbury JW. The sodium–potassium hypothesis ACM, 1993.
as the basis of electrical activity in frog ventricle. J. Physiol. [79] Peterson J. A simple sodium–potassium gate model.
154 (1960) 385–407. Chapter 2: The basic Hodgkin–Huxley model. cecas.clemson.
[58] Fitzhugh R. Thresholds and plateaus in the Hodgkin–Huxley edu/~petersj/Courses/M390/Project/Project.pdf.
nerve equations. J. Gen. Physiol. 43 (1960) 867–896.
FuRTHER READING
Alberts B, Johnson A, Lewis J, et al. Molecular Biology of the Cell, Bassingthwaighte J, Hunter P & Noble D. The Cardiac Physiome:
6th ed. Garland Science, 2015. perspectives for the future. Exp. Physiol. 94 (2009) 597–605.
FuRTHER READING 367
Crampin EJ, Halstead M, Hunter P, et al. Computational Nerbonne JM & Kass RS. Molecular physiology of cardiac
physiology and the Physiome Project. Exp. Physiol. repolarization. Physiol. Rev. 85 (2005) 1205–1253.
89 (2004) 1–26. Noble D. From the Hodgkin–Huxley axon to the virtual heart.
Keener J & Sneyd J. Mathematical Physiology. II: Systems J. Physiol. 580 (2007) 15–22.
Physiology, 2nd ed. Springer, 2009. Winslow RL, Tanskanen A, Chen M & Greenstein JL. Multiscale
Klabunde RE. Cardiovascular Physiology Concepts, 2nd ed. modeling of calcium signaling in the cardiac dyad. Ann.
Lippincott Williams & Wilkins, 2011. NY Acad. Sci. 1080 (2006) 362–375.
13
Systems Biology in
Medicine and Drug
Development
• Identify molecular causes of human variation
• Understand disease as a dynamic deviation in parameter space
• Describe the concepts of transforming an average disease model into a
personalized disease model
• Understand the drug development process
• Construct and analyze antibody–target binding models
• Construct and analyze pharmacokinetic models
• Analyze differential equation models of disease models
ARE YOU UNIQUE?

13.1 Biological Variability and Disease
If you and I and everybody else were exactly the same, life would be rather boring,
but medicine and drug development would be much easier. In a simplistic analogy,
if one knows how to exchange a windshield wiper motor in one 2005 Ford Mustang,
one knows how to do it in another 2005 Ford Mustang. Not so with complex dis-
eases. Two people may present with exactly the same symptoms, but react in a
distinctly different manner to the same dose of the same drug. As a consequence,
and in contrast to car repair, it is not sufficient to test a drug regimen on only one
person—instead it is necessary to perform large, complex, and expensive clinical
trials that often require many years of investigation. Then again, in spite of the clear
differences, we humans cannot be all that different from each other. After all, medi-
cine has become quite successful, and most prescription drugs actually work as
advertised.
So, how different are we really? Let us start with our genes. As we discussed in
Chapter 6, each of us has roughly three billion nucleotide pairs. For most of them,
there are no differences between two people and even among the entire human
population. If indeed there is a genetic difference, say in base 5,000,000, between
any two people on Earth, then this base is the location of a SNP (pronounced ‘snip’
370 Chapter 13: Systems Biology in Medicine and Drug Development
and meaning single nucleotide polymorphism, which roughly translates into

“many different forms in a given base of the DNA”). According to the best current
estimates, SNPs occur within the entire human population about every 300 bases,
which implies that you and I differ genetically by far, far less than 1% [1]. This does
not sound like much, but SNPs account for about 90% of all genetic variation, with
other changes occurring in deletions, multiplications of DNA, translocations of
DNA from one chromosome to another, and other rather rare alterations. Just con-
sidering SNPs, is there a sufficient repertoire of differences to make 6 or 7 billion
people unique? A quick and very rough computation leaves no doubt that we will
not run out of unique people any time soon. If every 300th nucleotide is changeable,
and if we simplistically assume that the probability of a single nucleotide (A, T, C, or
G) is the same at each SNP, then there are 10 million locations for SNPs, with a
choice of four nucleotides at each location. Try to compute 410,000,000. It is a truly
enormous number that dwarfs the number of all humans who have ever lived. It
even dwarfs by far the number of all atoms in the observable universe, which astron-
omers believe to be about 1080, give or take a few zeros. Adding to this genetic inter-
personal diversity, SNPs are not the only determinants of our individuality. While
genes are generally considered the blueprint of an organism, we know that even
organisms with exactly the same genetic material are not identical. First, there are
identical twins, who may look indistinguishable, but may have rather different traits
and personalities. Also, we have learned from the cloning of animals such as Dolly
the Sheep and CopyCat the Cat that the “offspring” differs from the mother, for
instance, in the coloration pattern.
If genes are not telling the whole story, what else is there? One relatively new
field of biology focuses on epigenetics, which studies differences in the activation
and deactivation patterns of genes among otherwise similar cells or organisms (see
Chapter 6). These differences are often the consequence of modifications at the
gene level, which, however, are not encoded in the DNA sequence. Instead, the DNA
is methylated or otherwise modulated with residues at certain locations, which
tends to affect the rate of gene transcription. Another possibility is a change in his-
tone proteins, around which the DNA is wrapped, and which may again affect the
rate of gene activation or deactivation. Other variations in an organism are, of
course, related to the environment, including exposure to harmful chemicals, diet,
and all our lifestyle choices.
All in all, we are different, and no two people are the same. This simple fact
makes medicine and the development and application of prescription drugs chal-
lenging. Fortunately, the situation is not entirely hopeless. First, not every point
mutation or SNP has a major effect. Much of our DNA does not code for proteins,
and mutations in coding or noncoding sections may not always be critical. As a
result of a mutation, an amino acid may be different in the protein for which the
gene codes, but the incorrect amino acid may be located outside the active site of
the protein, where no great harm is done. It is also possible that the protein is slightly
impaired but that the many control mechanisms in the body offer effective compen-
sation. Maybe most importantly, if we concentrate on one particular disease, very
many features and processes elsewhere in the body become irrelevant. Because
these facts come to our rescue, we are usually successful treating a child’s ear infec-
tion, irrespective of her size, weight, race, or hair color. Nonetheless, many diseases
are so severe, and prescription drugs may have such worrisome side effects, that a
new focus in the medical field is on personalized “precision” medicine that seeks to
develop treatments customized to each individual patient.
13.2 Modeling Variability and Disease

The situation of “heterogeneity in medicine” can be readily illustrated with an
example from metabolism, where, in a typical case, an inter-personal difference will
be manifest in the activity of some enzyme or modulator. It is easy to set up small
dynamic models to get a feel for what such a difference may involve (see Chapters 2
and 4). As an example, let us study a generic branched pathway with some regula-
tory signals (Figure 13.1). The mathematical form is not really important for this
type of illustration, and we translate the diagram into a mixed Michaelis–Menten/
power-law model:
ARE YOU UNIQUE? 371
X4 Figure 13.1 Example pathway for

illustrating variability and disease. The
generic pathway consists of five metabolites,
two feedback signals (red), and one
+
activation signal (green). The fluxes to and
X1 X2 X3
from X4 (gray) are very small.
– –
X5
X6
Vmax 1 X 1
X 1 = Input − ,
K M 1 (1 + X 6 K I 6 ) + X 1
Vmax 1 X 1 V X
X 2 = − max 2 2 ,
K M 1 (1 + X 6 K I 6 ) + X 1 K M 2 + X 2
V X
X 3 = max 2 2 − β31 X 3 331 X 5 351 − β32 X 3 332 X 5h 352 ,
h h h
KM2 + X2 (13.1)
X 4 = β31 X 3 331 X 5h 351 − β 4 X 4h 44 ,
h
V X
X 5 = β32 X 3h 332 X 5 352 − max 5 5 ,
h
KM5 + X5
V X
X 6 = max 5 5 h
− β6 X 6 66 .
KM5 + X5
Parameter values for the model were chosen almost arbitrarily and are given in
Table 13.1. We begin our analysis by computing the steady state, which, with three
significant digits, is (X1, …, X6) = (2.57, 4.00, 3.23, 3.70, 3.30, 3.13). For a typical sim-
ulation, we initiate the system at this state and double the input during the time
period [2, 3]. A plot of the results is shown in Figure 13.2. As is common for this type
of pathway structure, the variables rise and fall in order of their position and then
return to the original steady state. Notably, X4 responds very slowly, because the flux
going through this pool is very small in comparison with the others.
It is difficult to predict intuitively how such a system responds to perturbations
in parameter values, but it is easy to check computationally. For instance, we may
perform a series of simulations in which we raise one parameter at a time by 20%
and study the effect on the steady state. Interestingly, we find that most perturba-
tions are very well buffered. Usually, only one or two of the steady-state values devi-
ate by more than 10% from their original values, and deviations of between 20% and
30% only occur in the variables directly affected by the alteration. As a specific
example, if Vmax5 is raised by 20%, then X3 and X5 decrease by about 21%, but all
other variables are essentially unchanged. Exercises 13.1–13.4 explore this example
further. In reality, a deviation of 20% often goes unnoticed and does not lead to
disease or even symptoms.
While quite simplistic, this cursory analysis implies that regulated systems have
a strong buffering capacity. Indeed, a Japanese group made the same observation,
when they exposed Escherichia coli cells to various genetic and environmental per-
turbations [4]. While the gene activities and enzymes close to the perturbation were
often strongly affected, the remaining metabolic profile was by and large unaffected,
owing to a compensatory rerouting of fluxes in the cells. In this light, the reader
TABLE 13.1 PARAMETER VALUES FOR THE PATHWAY IN FIGURE 13.1 AND (13.1)
Input Vmax1 KM1 KI6 Vmax2 KM2 β31 h331 h351
2 4 1 2 5 6 0.005 0.8 0.2
β32 h332 h352 β4 h44 Vmax5 KM5 β6 h66
2 0.4 −0.4 0.005 0.9 8 10 1 0.6
6 Figure 13.2 Typical scenario simulation of

the pathway in Figure 13.1. The simulation
X1 used the equations in (13.1) and the
X2 parameter values in Table 13.1. From time
t = 10 to time t = 12, the input to the system
is doubled from 2 to 4.
X3
concentration
4
X4
X5
X6
2
0 25 50
time
should revisit the discussions regarding robustness of small-world networks (see

Chapter 3).
The analysis of perturbations and their ultimate effects provides several insights.
First, most steady-state values are not much affected by changes in any of the
parameters. Second, an enzyme alteration at the beginning or center of the pathway
may affect immediately associated metabolites, but the end products are often only
mildly influenced. A direct implication is that, in many disease-related cases,
changes in the intermediate metabolites may not be of relevance. Third, the dynamic
response to a change in input or enzyme activity may consist of a transient meta-
bolic deviation, which, however, may not be of long-term pertinence with respect to
disease. What is not evident from the example is that larger and more complex sys-
tems are usually less prone to local perturbations, such as an altered enzyme activ-
ity, than smaller, less regulated systems.
PERSONALIZED MEDICINE AND PREDICTIVE HEALTH

Even if it seems that biomedical systems can compensate for altered components
and processes quite well, we know of course that there are many diseases that cause
long-term pain and suffering. In a relatively small number of cases, these diseases
are caused by a single variation at the genome level. Perhaps the best example is
sickle cell anemia, which is caused by a point mutation in the hemoglobin gene
HBB. While there are many variants, the most common is very well understood: it
consists of the substitution of the hydrophilic glutamic acid with the hydrophobic
amino acid valine at the sixth amino acid position of the HBB polypeptide chain. So,
we know exactly what is wrong, but we cannot do much about it.
In contrast to single-gene diseases like sickle cell anemia and cystic fibrosis,
most disease pathologies result from a combination of genetic predisposition (as
manifest in several gene alterations) and epigenetics, as well as environmental and
possibly other factors, which nowadays is collectively called the exposome [5]. These
combinations make it difficult to answer questions such as: How can it be that some
nonsmokers develop lung cancer while some long-term chain-smokers do not?
Why do people respond differently to the same treatment? These types of questions
are considered to be at the core of two new and related approaches to medicine:
personalized or precision medicine and predictive health.
The key goal of personalized medicine is the prevention and treatment of disease
in a fashion that is custom-tailored to the individual. For instance, instead of deter-
mining the dose of a prescription drug simply based on body weight, and using
average information obtained from large populations to assess the efficacy of the
drug per kilogram of body weight, the dose is computed based on the personal char-
acteristics of the patient. These include obvious information such as body weight,
PERSONALIZED MEDICINE AND PREDICTIVE HEALTH 373
sex, and age, but also body mass index, gene markers, relevant enzyme profiles,
metabolic rate, and a host of other biomarkers that could be of relevance to the
disease and its treatment. Ideally, such a personalized, precise treatment in the
future will maximize efficacy and minimize undesired side effects.
Predictive health is a conceptual continuation of personalized medicine. Its key
goal is to maintain health rather than to treat disease. Thus, predictive health aims
to predict disease in an individual before it manifests itself. The main targets are
chronic diseases such as diabetes, cancer, and neurodegeneration. These diseases
often begin with a long, silent phase, where a person does not even perceive symp-
toms. However, this early phase constitutes a measurable deviation from the norm,
and although the person is apparently healthy, her or his biochemistry, physiology,
gene expression, or protein and metabolite profiles already contain early warning
signs indicating that the person is on an undesirable trajectory toward a chronic
disease. If we had a reliable and comprehensive catalog of such warning signs, and
if healthy people received early screening and treatment, while they were still appar-
ently healthy, then chronic diseases could potentially be prevented or managed
from their early stages on. There is plenty of indication that such a treatment option
would be much cheaper than treatments of chronic disease that are typical in today’s
world, and efforts are underway to promote predictive health [6].
One intriguing challenge for any type of personalized medicine and predictive
health is that it is not at all easy to define what exactly health or disease means. We
have seen above that there can be a great deal of variability among healthy individu-
als, but we also know that the transition step from health to disease does not have to
be large. A straightforward distinction between health and disease may be a different
set of important parameters, such as blood pressure, cholesterol, or some genetic
predisposition (see above and [7]). An extension of this view is that such changes col-
lectively may disturb the normal state of homeostasis and lead to allostasis, which
in the language of physiological modeling corresponds to a different, and in this case
undesirable, steady state [8]. A more recent view is that disease may be the trade-off
with the physiological robustness of the human body or the loss of complexity in the
dynamics of a physiological subsystem (see [9, 10] and Chapter 15).
13.3 Data Needs and Biomarkers

Two components are required to realize ideas of personalized medicine and predic-
tive health: first, lots of data; second, computational methods that are capable of
integrating and analyzing these data. The data must cover many physiological
aspects of individuals when they are healthy, and follow them throughout their
healthy life and possibly into disease. The data from a single person will not shed
sufficient light on complex disease development processes, but a growing collection
of many personal trajectories, some staying within the healthy range and some
diverting toward disease, will reveal an increasingly comprehensive picture of how
disease develops and may suggest which early and very early signs (biomarkers)
should be considered in regular screening examinations of healthy individuals. In
more general terms, these data and their analysis will permit new and sharper defi-
nitions of what it means to be healthy, premorbid (with early signs of disease but
without too severe symptoms), or diseased [3].
The tracing back of a disease to disease predictors, late biomarkers, and ultimately
early biomarkers (Figure 13.3) may seem to lie far in the future. However, in some
instances, it has already become standard. A good example is an elevated blood sugar
level, which in itself does not constitute a disease but is often the precursor of type 2
diabetes. Another example is amniocentesis, a means of diagnosing a fetus within the
womb by testing a small amount of amniotic fluid for genetic abnormalities and infec-
tions. Even though a baby might appear healthy in an ultrasound examination, the
amniocentesis might foreshadow diseases later in life, based on genetic markers.
Biomarkers can also be indicators of whether a particular drug treatment might
be effective, before the drug is actually administered [11]. In the new field of phar-
macogenomics, analyses of expressed genes, especially those from the class encod-
ing cytochrome P450 enzymes, are used to predict the sensitivity of individuals to
specific drugs. A beautiful example is a prognostic genome analysis in children with
acute lymphoblastic leukemia (ALL). It had been known that some children with
Figure 13.3 Hierarchical biomarker

network. Disease is usually preceded by
disease predictors. These are most likely
preceded by late and early biomarkers.
disease
early biomarkers late biomarkers disease predictors
ALL responded very well to standard drug treatments, while others developed
severe side effects that greatly compromised the efficacy and applicability of the
drugs. Scientists at St. Jude Children’s Research Hospital in Memphis, Tennessee,
discovered that a microarray test was able to classify young ALL patients before
treatment into responders and nonresponders [12]. In this particular case, 45 differ-
ently expressed genes were found to be associated with success or failure of all four
drug treatments tested, and 139 genes were identified in children responsive to one
drug, but not the others.
As in the example of ALL, the first step toward rational predictions of disease and
treatment outcome will be a statistical association analysis, where certain patterns
in the data are predominantly observed in cases of disease or drug resistance, while
other patterns are more often associated with health and a successful drug treat-
ment. As an extension of standard association analyses of statistics, it is possible to
train an artificial neural network (ANN) or some other machine learning tech-
nique to consider many markers and to classify individuals as healthy or at-risk.
Because ANNs can be large, they have the potential of identifying at-risk patients
earlier and more accurately than standard association methods of statistics. An
example is the computer-assisted ProstAsure test for prostate cancer [13].
The identification of associations between biomarkers and disease is invaluable,
but not the ultimate goal of predictive health. Building upon association data, the
next step is the construction of causal models, which not only associate certain bio-
markers with disease, but also demonstrate that a particular disease is actually
caused by a variation in these biomarkers. Once this causality has been established,
the biomarker is no longer just a diagnostic or prognostic symptom, but becomes a
true drug target. The following sections indicate how modeling and systems biology
of the future might contribute to our understanding of personalized health, disease,
and treatment.
13.4 Personalizing Mathematical Models

Let us begin by painting a picture, in very broad strokes, of how today’s medicine
works. Figure 13.4 shows a simplified diagram. The left side depicts, in reddish col-
ors, the typical paths of current knowledge discovery. The two main sources of input
are epidemiological studies, which study large populations in order to associate dis-
eases with risk factors, and modern biology, which characterizes biochemical and
physiological mechanisms that might lead to disease if they are altered and
quantifies these mechanisms with pertinent parameters. The results of these two
approaches ultimately enter clinical trials, which lead to prescriptions of treatments
that are, by the nature of the process, based on population averages. Process descrip-
tions and parameters have also been the basis for the design of disease models that,
owing to the nature of the underlying data, again address average individuals. These
models can be diagnosed and analyzed, as we discussed in Chapter 4, and may lead
to better characterizations of a disease, as well as to a better understanding of
averaged treatment options.
Systems biology enters the field with two components. Experimental systems
biology offers new and very powerful means of identifying physiological mecha-
nisms and their parameters. It will also be the foundation of information gathering
in personalized precision medicine and predictive health, as discussed before.
MOLECULAR BIOLOGY
EPIDEMIOLOGY BIOCHEMISTRY EXPERIMENTAL
PHYSIOLOGY SYSTEMS BIOLOGY
hypothesized
risk-factor/disease MODEL DESIGN
associations
“averaged”
model
physiological
mechanism
PERSONALIZED
PERTURBATION SIMULATION
SIMULATION
process
parameters NUMERICAL
SOLUTION sensitivity, health–disease
robustness classification
CLINICAL
TRIALS personalized personalized
risk profile health model
suggested
prevention
personalized personalized
“averaged” health prediction treatment
treatment
COMPUTATIONAL SYSTEMS BIOLOGY
Figure 13.4 Traditional and personalized medicine. The simplified diagram shows in reddish colors the current paradigm of medicine and in blue
and green colors an emerging paradigm involving systems biology. (Adapted from Voit EO & Brigham KL. Open Pathol. J. 2 [2008] 68–70.)
Computational systems biology will use averaged disease models, personalize

them, based on individualized data, yield personal risk profiles, and suggest possi-
ble prevention strategies.
The computational aspect of personalized medicine and predictive health
requires the personalization of models [7, 14]. Conceptually, this step is fairly simple,
and companies such as Entelos [15] have achieved successes in simulating virtual
patients. In generic terms, one begins with a disease model that had been con-
structed and parameterized with the best available data and information currently
available. This model reflects population averages, rather than one single individual.
Suppose now that many biomarkers had been measured for a specific person and
that they do not wildly fluctuate over time. A comprehensive model should directly
or indirectly permit these biomarkers to be entered into the model. For instance, an
altered enzyme activity might be translated into the KM value of some enzymatic
step. A deactivated gene could, at least in a first approximation, be translated into
the decreased activity of a protein, which would in turn correspond to a decreased
transport rate, altered catalytic conversion, a muffled signal transduction process,
or some other effect on the model. Because the biomarkers were measured in a spe-
cific person, their values are most likely different from the average over an entire
population. Thus, the original average model is customized by changing each aver-
age biomarker value to the measured, personal biomarker value, and every person
can be seen as a unique point in a huge, dynamically changing parameter space.
In the simple example at the beginning of the chapter, we modified parameters
by increasing them by 20%. In some sense, this exercise corresponded to an abstract
personalization. Here, we simultaneously change many parameters, namely, all

those for which personalized biomarker measurements are available. Presumably,
this personal biomarker profile is not complete, which requires us to retain average
values for the remaining parameters, in the hope that the individual is more or less
normal in aspects that were not specifically measured. If possible, missing measure-
ments can later be substituted with personal values, or one might perform a
simulation study, for instance, using Monte Carlo simulation (see Chapter 2), to
determine how influential the average parameters really are. If the system is not sensi-
tive to these parameters, it might not even be necessary to obtain personalized values.
The personalization of a model opens new avenues for various prevention and
treatment options. First, one will probably simulate the model from the current “ini-
tial” state forward. Such a simulation shows whether the individual converges to
some healthy steady state or whether the individual biomarker profile leads to the
slow, but unrelenting accumulation of molecules such as cholesterol (see Exercises
13.2 and 13.3). In addition to the steady state, the model also indicates the speed of
change and shows whether any transients during the scenario simulation could be
dangerous. This baseline simulation is expanded into scenario simulations, where
one attempts to intercept undesired trajectories, such as the accumulation of harm-
ful substances. It is expected that a reliable model will identify successful strategies.
It may also help screen for interventions that at first seem a good idea, but in fact do
not really work, for one unexpected reason or another.
The collection of many personalized health and disease trajectories will begin to
shape a deeper understanding of causes, symptoms, and progressions of disease
processes. It will show that some abnormal biomarkers alone are not dangerous, but
that they may forecast disease if they are combined with other altered biomarkers.
Ultimately, a comprehensive collection of trajectories will help us determine effec-
tive early biomarkers that precede late biomarkers and precursors of disease, as
shown in Figure 13.3.
Looking at the complexity of metabolic and physiological systems, one might
come to the conclusion that personalized medicine is a dream without much of a
chance of realization. Three arguments counter this pessimistic view.
First, physicians have actually been pursuing the same strategy since the begin-
nings of medicine: they gained their knowledge from population averages and
learned to customize it toward a patient. Maybe the biomarkers were initially sim-
pler, consisting of blood pressure, temperature, and past personal trends such as
weight loss. However, in more recent times, physicians have been using many more
and specific biomarkers, and the only true difference from a computerized system
of personalized medicine is that doctors perform the integration of data in their
heads, aided by their knowledge and experience. Looking into the future, this inte-
gration will become increasingly more difficult, because the number of measurable
biomarkers will continue to climb rapidly, eventually overwhelming even the most
astute human mind. It seems that the only remedy against this overload will be
computer-aided data integration. One should add that computer predictions will
eventually become so complex that the physician is no longer able to explain how a
suggestion from a computer was obtained.
The second argument is the success of computational expert systems in select
areas of medicine, such as the identification and severity of poisonings from symp-
toms [16] and emergency medicine [17]. While these systems are based on methods
of artificial intelligence, which are quite different from systems modeling with dif-
ferential equations, their successes suggest that computer-assisted diagnostics and
treatment are not out of the question in medicine. In the much simpler case of car
mechanics, computerized diagnostics has become an accepted standard.
Third, while most of our computational models of molecular systems are pres-
ently too coarse and insufficiently comprehensive to facilitate personalized medi-
cine, specific numerical, model-based predictions are already possible in other
medical fields. For instance, before cosmetic surgeries are performed, they are
often simulated first in a virtual environment that permits personalization.
Similarly, a group at Emory University and the Georgia Institute of Technology have
developed an image-based software system for blood flow reconstruction surgery
on the heart that is custom-tailored to a specific patient [18]. The computational
platform allows surgeons to perform, in a simulation, alternative surgical scenarios,
(A) (B) Figure 13.5 Health is related to

biomarker 2 biomarkers within their normal ranges.
(A) For a single biomarker, a normal range
is relatively easy to define. (B) For several
biomarkers, the healthy domain may be
different from a simple combination of
normal ranges.
normal
normal biomarker normal biomarker 1
and the system computes the postoperative hemodynamic flow properties of each
resulting reconstruction.
It is clear that the complexity of physiological systems in health and disease will
render it very challenging to design and implement precise models in specific indi-
viduals. Nonetheless, in most cases, it is not necessary to target the entire human phys-
iology at once. Instead, it will be feasible to move gradually from what we have already
accomplished in medicine to increasingly detailed systems of disease processes.
The starting point is a consideration of what health and disease really mean.
Precise definitions are difficult, but it is sufficient here to begin with the pragmatic
definition of a normal range for a biomarker. We all know that a blood pressure of
220 over 120 is probably not so good in the long run, and that the body core tem-
perature has to remain within a rather tight range for us to function normally.
Generically, one may define a normal range for a biomarker as shown in Figure
13.5A, which acknowledges that the boundary between normal and abnormal is
fuzzy. The situation becomes more complicated when we consider two biomarkers
simultaneously (Figure 13.5B). In many cases, combinations of extreme values of
both biomarkers are no longer healthy, so that the green domain is not a rectangle
but an ellipse or some other shape with truncated corners (Figure 13.6A). The con-
sequence is that the normal domain of a biomarker is no longer absolute. In Figure
13.6B, the orange and blue “patients” have exactly the same value for biomarker 1,
but blue is healthy, while orange is not. The purple patient has (borderline) healthy
biomarkers 1 and 2, but the combination is no longer in the healthy domain. The
turquoise patient, by contrast, shows an unhealthy reading for biomarker 1, but is
still considered healthy. Expressed differently, the outcome of only a single bio-
marker test can lead to false-positive or false-negative outcomes. In the first case,
the test falsely suggests disease, whereas in the second case, the test falsely
(A) (B)
biomarker 2 biomarker 2
normal
normal
Figure 13.6 Delineation of health and

disease. For simplicity, the healthy domain
in (A) is defined by the dotted hexagon.
(B) shows cases where the consideration of
a single biomarker leads to “false positives”
or “false negatives.” The issue is much more
pronounced when many more biomarkers
normal biomarker 1 normal biomarker 1 are involved.
indicates health. Although simplified, this discussion indicates that individual bio-
markers are seldom absolute predictors of health or disease and that their values
must be seen within the context of a larger health system where biomarkers are
interdependent on one another.
Today’s physicians know the normal ranges for many biomarkers, as well as
some of the more important combinations of a few biomarkers. With this knowl-
edge, it is possible to develop models that take account of information from many
sources and integrate it in a meaningful manner. Over time, such models will
improve in quality and reliability, and eventually will be able to take hundreds of
biomarkers and classify whether their combinations are healthy, premorbid, or
indicative of disease. Furthermore, model-based integration of repeated measure-
ments of biomarker combinations over the lifetime of an individual will show
whether this individual is on a healthy trajectory or whether he or she is moving
dangerously close to the realm of disease. Analyzing and interpreting thousands of
such trajectories together will create a rational basis for predictive health [3].
THE DRUG DEVELOPMENT PROCESS

Between the dawn and dusk of the twentieth century, the average human life expec-
tancy in the United States went up from about 45 to 77 years [19, 20]. Much of this
enormous extension of life can be credited to vast improvements in medicine and
the availability of drugs and biomedical devices, along with better hygiene, plenty of
food, and a deeper understanding of physiology, molecular biology, and microbiol-
ogy. While the maximal lifespan for humans is unknown, current estimates lie
between 120 and 130 years, and as more and more individuals are becoming cente-
narians, the emphasis of extending life is shifting toward good health at an advanced
age. Realizing this emphasis will only be possible with continued developments and
improvements in medicine and drug development.
Unfortunately, the rate at which new drugs are introduced has been slowing down
considerably in recent years. It is not the case that the pharmaceutical industry has run
out of drug targets. In fact, high-throughput biology has been suggesting new drug
targets galore. Rather, the main reason seems to be that “Big Pharma” is leery of the
length and cost of the drug development process, which may take 10–20 years and
require an investment of between one and two billion dollars. Patents protect new
drugs for only about 15– 20 years, depending on the country, before cheaper generic
drugs become competitors. In areas such as HIV/AIDS, the investment in a drug must
be recouped much faster, because new formulations and cocktails may begin to chal-
lenge a drug after just a few years. Adding to this situation is the enormous attrition
rate of new chemical entities (NCEs), that is, of active ingredients, during the drug
development process [21, 22]. Out of 100 NCEs, only one to five make it into preclinical
trials, and only a few percent of these survivors will ultimately be approved. Obviously,
these numbers depend on many aspects of the disease and the NCE, but indicate that
the a priori probability of success of an NCE is low.
Reasons for the low success rate are plentiful and include the fact that the phar-
maceutical industry has begun to attack very complex diseases, often even with
more resources for research and development than in the past, and that the stan-
dards set forth by the Food and Drug Administration (FDA) in the United States and
similar health agencies elsewhere are becoming more and more stringent. The rea-
sons for dismissal of an NCE are manifold as well, but two of them stand out. First,
30–60% of the dismissed drugs are abandoned owing to issues of efficacy and bio-
availability. In other words, the human body does not accept the drug in sufficient
quantities to permit a therapeutic effect. Another 30% have to be eliminated because
of toxicity and serious side effects, while 7% are not commercially viable [19]. Thus,
the landscape of drug discovery, development, and marketing is treacherous and
requires significant resources.
So, what exactly are the problems, and are there ways of solving them? Certainly,
nobody intends to cut corners or to simplify the testing procedures. Instead, the
challenge to be solved may be posed as (1) increasing the overall success rate of the
drug development pipeline and (2) accelerating the elimination of an NCE from the
pipeline. The former is rather obvious. The latter is important because the later
THE DRUG DEVELOPMENT PROCESS 379
discovery preclinical development clinical development postclinical development Figure 13.7 The drug development
pipeline. The pipeline consists of distinct
target ID lead optimization clinical phase I clinical phase III launch phases, which any drug has to survive
before it can be used for treatment. The
process is very costly and time-consuming.
Without compromising safety, costs and
time can only be saved if drug candidates
with undesirable properties are screened
out early.
hit ID, lead ID development clinical phase II FDA approval process
of drug candidate
phases of the pipeline, and in particular the phases of clinical trials, may cost hun-
dreds of millions of dollars. Thus, if a problematic new NCE can be eliminated early,
substantial amounts of resources are saved. The drug pipeline is depicted in
Figure 13.7. It consists of four major phases that are readily subdivided into further
segments. The outcome of testing at each segment of each phase may suggest aban-
doning further consideration of the NCE.
The first phase is concerned with early-stage research and discovery, target iden-
tification, the identification of small-molecule hits, and validation. At the beginning,
a specific biological target is identified as being associated with a particular disease.
The target could be a receptor, a signaling molecule, an enzyme, or some other pro-
tein that plays a crucial role in the disease. The hit one is hoping for is a molecule that
interacts with the target in such a fashion that the disease process is interrupted. In
the case of a receptor, a hit molecule may be an artificial ligand that attaches to the
receptor and blocks it from accepting the disease-associated ligand. In the case of an
enzyme, the hit might be an inhibitor. It is not surprising that many possible hits
must be investigated. Often, these hits come from extensive combinatorial molecu-
lar libraries that a pharmaceutical company has assembled. In most cases, only one
or a few of the hit molecules will continue in the drug development pipeline as leads
for further testing. Hit and lead identifications are nowadays executed in a high-
throughput fashion that utilizes many of the methods of modern molecular biology
and experimental systems biology [23]. In addition to the initial identification, the
target, hit, and lead must be independently validated. This step provides proof of
concept that the target may be worth pursuing with the identified leads. For instance,
if the target is a receptor, then binding affinities, length of receptor occupancy, and
accessibility of the receptor to the lead compound must be investigated.
Validated hits with favorable characteristics that show potential for an improved
medication become leads in the next phase of preclinical drug development. The
first component of this phase is lead optimization. The lead is subjected to further
tests of medicinal chemistry and, in particular, to an optimization process that
investigates small molecular variations, which may affect receptor binding or the
strength of inhibition with respect to a target enzyme. An important guideline is the
exploration of quantitative structure–activity relationships (QSARs), which
attempt to assign specific features to chemical structures that might be used as side
groups of the lead molecule [24, 25]. In the case of proteins, a similar approach is the
creation of peptoids, which are artificial peptide-like compounds that can be
designed to work as drugs with desirable properties [26]. Such drugs could be
intriguing because of the preeminent role of specific peptide stretches in protein–
protein interactions, molecular recognition, and signaling.
The second component of lead optimization is the assessment of dose, efficacy,
and acute toxicology in animals. Compounds that are found to be effective in ani-
mals and do not lead to acute toxicity are moved forward in the drug development
pipeline as drug candidates. The further investigation of drug candidates constitutes
the second step of preclinical drug development. Here, toxicity over longer time
horizons is assessed, along with pharmacokinetic tests that reveal how long the
compound resides within some tissue and the bloodstream. The half-lives in differ-
ent organs may differ drastically, because, for instance, metabolism in fat is typically
slow, while it is the role of the liver to detoxify the body as quickly as possible. In the
context of pharmacokinetics, one sometimes speaks of investigating ADME: absorp-
tion, distribution, metabolism, and excretion of a drug. Drug candidates surviving
the muster of preclinical assessment are moved forward to clinical testing [27].
The clinical phase may easily consume half of all costs of the entire drug devel-
opment process and half of the time. It is commonly subdivided into three subse-
quent types of clinical trials. The first, phase I, is primarily concerned with safety
issues. The compound, which will already have been found to be safe in animal
studies, is administered to a few dozen healthy human volunteers over a time span
of 1–2 years. If toxicity or other undesired side effects are detected, the testing is
stopped and the drug is not permitted to move forward in its current form. Through-
out the clinical phase, studies also investigate different formulations and dosing
regimens. In phase II, both efficacy and safety are tested on a relatively small cohort
of 100–300 patients. This phase again lasts for 1–2 years. Finally, in phase III, the
drug candidate is used in a randomized clinical trial involving a few thousand
patients. The foci of this 2- to 3-year-long study are again efficacy and safety. Because
volunteers, patients, and hospitals are involved in clinical trials, this phase of the
drug development pipeline is very costly. Roughly 20% of drug candidates survive
the three phases of clinical trials.
Once all preclinical and clinical testing has yielded satisfactory results, the drug
candidate must be approved by the FDA in the United States and/or corresponding
health agencies in other countries. Drug applications to the FDA require vast
amounts of documentation about the positive and negative results of all former tests
and trials, and approval typically takes 1–2 years. The final, post-approval phase of
the drug pipeline contains the launch. It consists primarily of the sharing of infor-
mation with physicians, hospitals, and trade magazines, as well as direct advertising
and marketing. This phase can cost $100 million or more [28].
THE ROLE OF SYSTEMS BIOLOGY IN DRUG

DEVELOPMENT
Results of deliberations by think tanks have suggested that academia and the phar-
maceutical industry should form partnerships, in which academicians would pri-
marily focus on drug discovery and possibly some of the preclinical studies, while
the pharmaceutical industry would take the lead in the expensive clinical trials.
However the roles of all potential contributors to the process are split up, it is more
important here to ask how systems biology might be able to help improve drug dis-
covery and development. Several roles of systems biology are already emerging, and
the industry has been embracing them to various degrees for quite a while. Indeed,
the potential of systems biology in drug discovery has been recognized for some
time (see, for example, [29–31]).
Looking at the drug development pipeline shown in Figure 13.7 from a systems
point of view, two distinct types of improvements would obviously contribute to
desirable solutions. First, it would be desirable to increase the number of promising
hits. And second, it would be highly desirable to find out as early as possible whether
particular hits or leads would later be doomed. This activity would not reduce over-
all attrition, but it would screen out hits and leads before their testing became very
expensive, which happens particularly in the clinical development phase.
Regarding increased numbers of hits, there is no doubt that methods of experi-
mental systems biology and of high-throughput data acquisition and analysis are
very promising tools for the separate or combined identification of targets, hits, and
leads, as well as for global toxicity studies. All major pharmaceutical industries are
already heavily engaged in parallelized “omics,” robotic screening, and bioinformat-
ics, and some smaller companies have sprung up that concentrate on genomic anal-
yses that are made available to larger companies; examples include Selventa [32]
and GNS Healthcare [33]. Proteomic studies are also being pursued, especially
within the realm of signaling systems, where sophisticated methods are beginning
to permit crisp assessments of the phosphorylation states of specific signaling pro-
teins [34, 35].
In addition to searching for NCEs, the pharmaceutical industry has begun to
invest heavily in the exploration of biologics [36, 37]. In contrast to synthetically
produced, small-molecule NCEs, these biologics are derived from naturally
occurring biological materials. They often have a favorable safety profile, because
THE ROLE OF SYSTEMS BIOLOGY IN DRUG DEVELOPMENT 381
they are less likely to elicit an immune response, and they break down into natural
molecules such as amino acids. In fact, it is sometimes even therapeutically ben-
eficial to supply an unchanged biomolecule to a patient who shows a correspond-
ing deficiency. For example, insulin is given to patients with type 1 diabetes and
factor VIII is prescribed for hemophilia A. In other cases, natural compounds are
used as templates that are chemically modified, in order, for instance, to improve
the binding to a receptor. Because it is possible to attach a wide variety of residues
to the natural base compounds, the number of potentially efficacious biologics is
enormous. As a consequence, the biologics sector of the pharmaceutical industry
has in recent years been growing at an annual rate of 20% and now is a multi-
billion-dollar market. Monoclonal antibodies constitute the best-known category
of biologics. Other examples include engineered proteases, for instance, from the
blood clotting cascade, and aptamers [38]. Aptamers are artificial oligonucle-
otides that do not code for a peptide or interact with DNA but are designed to bind
and thereby deactivate a target, such as the active domain of a disease-linked pro-
tein. Modern methods of experimental systems biology make it relatively easy
and cheap to produce aptamers in vast variations. Overall, the role of experimen-
tal systems biology in drug discovery seems quite evident, and we will not discuss
it further.
13.5 Computational Target and Lead Identification

Different branches of computational systems biology have also begun to be helpful
for target and lead identification (see, for example, [39]). For instance, methods of
molecular modeling and combinatorial chemistry permit the computational assess-
ment of molecules and active groups and are becoming increasingly reliable predic-
tors of the efficacy and toxicity of permutated drugs and their side groups, as well as
of biologics. These predictions are, in some sense, based on a new generation of
sophisticated QSARs that have been in use for many years. Systems biology is also
used increasingly for receptor binding and dosing studies with newly created mole-
cules. Not to be overlooked, computational methods of intelligent data and text
mining are rapidly moving to the forefront. With thousands of biomedical journals
being published on a regular basis, automated alerts and sophisticated searches
have become a necessity. In addition to simple key words or phrases, modern search
engines are rapidly becoming better in semantic searches that may even include
images.
Many aspects of systems biology look promising for reducing the cost of drug
discovery through effective early screening and hit elimination. The key to mov-
ing the elimination of NCEs and biologics closer to the front of the drug pipeline
is a solid understanding of the mechanisms of action of a hit or lead with respect
to a particular target. Because disease processes are complex, such enhanced
understanding greatly benefits from computational models. Beautiful examples
of very comprehensive disease models, for instance for diabetes and rheuma-
toid arthritis, have been developed by Entelos [15] and utilized by larger phar-
maceutical companies. Entelos and other companies of this type construct these
models with methods discussed in earlier chapters of this book, by mining the
literature exhaustively, curating the results very carefully, and converting all
pertinent information into processes and customized parameters for simula-
tions of virtual patients. These examples are real-world extensions of the
sandbox analysis that we used in the beginning of the chapter (see (13.1)) for
illustrative purposes.
While comprehensive models show clear potential, even simple models can be
of great value. For instance, it is not always necessary to model a signal transduction
pathway in great detail, and it may suffice instead to construct a black-box model
that is used as a surrogate for a comprehensive systems model [36]. In the following,
we discuss three examples. The first pertains to receptor dynamics, which plays a
major role for very many drugs. The second example discusses compartment mod-
eling and pharmacokinetics, which have already been mentioned in Chapter 4.
Finally, we will show how dynamic pathway models may help weed out NCEs and
contribute to an understanding of safety and toxicity.
Figure 13.8 Many drugs work by binding

to proteins. Here, the FDA-approved drug
indinavir docks into the cavity of an HIV
protease in a lock-and-key mechanism
(Protein Data Bank: PDB 10DW and 1HSG).
(Courtesy of Juan Cui and Ying Xu, University
of Georgia.)
13.6 Receptor Dynamics

Essentially all prescription drugs need to be taken up by cells, and very many of
them are associated with membrane-bound receptors (Figure 13.8). It is therefore
of great interest to study binding characteristics, and even relatively small mathe-
matical models can yield valuable insights. As an example, let us consider two sce-
narios involving therapeutic antibodies. In the first, very much simplified, case,
adapted from a study by the pharmaceutical industry [37], a monoclonal antibody
M is injected intravenously and in the bloodstream binds to a target T, which could
be a signaling peptide or a hormone. On encountering one another, M and T form a
complex C, which prevents T from exerting a negative effect on the system (Figure
13.9). The goal of the treatment is to reduce T to a low level, for example, 20%.
Ignoring all spatial aspects, it is easy to set up a diagram and a dynamic model of the
system (Figure 13.10; see also Exercises 13.15–13.17). It shows that the target is syn-
thesized and also cleared by the body. The antibody is also cleared, although with a
different rate. The complex in this simple example is assumed to be relatively stable
and, owing to its size, not cleared by the body within the time horizon of interest
[36]. The default assumption for setting up a model of the system is mass action
kinetics, which leads to the following representation:
T ′ = kT+ − kT−T − konTM + koff C ,

M ′ = koff C − konTM − kM− M , (13.2)
C ′ = konTM − koff C .
Note that the injection of antibody is not modeled explicitly. Instead, at time
t0, the value of M is set equal to a value corresponding to the antibody kM–
M
inject
kon
koff
C
kT+
T
kT–
Figure 13.10 Diagram of a monoclonal

antibody M binding to a target T. The
Figure 13.9 Treatment of a disease with biologics. Monoclonal antibodies binding yields a complex C that renders
(green circles) are injected into the bloodstream, where they bind reversibly the target dysfunctional. Rate constants are
to their target molecules. defined in (13.2) and the text.
(A) (B)
100 100
T0 = 10 T0 = 10
T0 = 5
T0 = 5
T0 = 2
M (%)
T (%)
50 50
T0 = 2
T0 = 1
T0 = 1
0 0
0 20 40 0 20 40
time (days) time (days)
Figure 13.11 Simulation results with the model (13.2) of antibody binding. For a fixed initial monoclonal antibody concentration M0, both (A) the
rebounding dynamics of the free target T as a percentage of the initial target concentration T0 and (B) the free antibody concentration M as a percentage
of M0 depend strongly on T0. Note that time is given in days, rather than hours.
concentration in the bloodstream. We simulate the system with initial values (T(t0),
M(t0), C(t0)) = (T0, M0, C0) = (different values, 100, 0.001) and parameter values
kT+ = T0kT− , kT− = ln(2)/10, kM− = ln(2)/(21⋅ 24), kon = 0.036, and koff = 0.00072, given for a
timescale of hours, as suggested in [37, 40]. Let us check the rationale for these val-
ues. The synthesis rate of T, kT+ , is set such that the system without any antibodies or
complexes is in a steady state. This assumption is often (although not always) true.
Setting all terms containing M or C to zero and equating the right-hand side of the
first equation in (13.2) to zero results in the given definition of the synthesis rate.
The degradation rate of T, kT− , corresponds to a half-life of 10 hours. The clearance
rate of M corresponds to a half-life of 3 weeks. The rates of complex formation and
dissociation are typical for these types of complexes.
It is now easy to run simulations over a typical period of say 40 days, with differ-
ent values for T(t0). Some results are shown in Figure 13.11. They show that the
concentration of the target drops immediately and then recovers after some while.
The higher the initial target concentration, the faster T recovers. The figure also
shows the disappearance of free antibody, which is due to binding to T and also to
clearance.
The second example is similar in spirit, but it is now assumed that the system
consists of a receptor R, a ligand molecule L that naturally binds to the receptor,
thereby contributing to the disease process, and a therapeutic antibody A. An
important question for drug design is whether the antibody should be developed to
bind to the receptor or to the ligand. In either case, the association of ligand
and receptor would be impeded, so does it really matter? A sketch of the situation
(Figure 13.12) accounts for both scenarios; in reality, binding occurs only to the
ligand or only to the receptor. The structural diagram, again accounting for both
Figure 13.12 Alternative options for antibody treatment. The monoclonal antibody (green circles),
which is injected into the bloodstream, binds reversibly either to the receptor (left) or to the ligand
(right) associated with the target disease.
C3 Figure 13.13 Diagram of a monoclonal

antibody A binding to a target receptor
R or to a target ligand L. Three types of
complexes arise: between receptor and
ligand (C1), and either between receptor
and antibody (C2) or between ligand and
k3+
antibody (C3). Rate constants are defined in
inject
k3– kL+
(13.3) and the text.
A L
kA– kL–
k2– k1–
k1+
k2+
C2 R
kR+ kR– C1
situations, is given in Figure 13.13, and the equations describing the diagram are as
follows:
R ′ = kR+ − kR− R − k1+ RL + k1− C1 − k2+ RA + k2− C 2 ,

L  kL  kL L  k1 RL  k1 C1  k3 LA  k3 C 3 ,
A  k A A  k2 RA  k2C 2  k3 LA  k3C3 ,
(13.3)
C1  k1 RL  k1 C1 ,
C 2  k2 RA  k2 C 2 ,
C 3  k1 LA  k3C 3 .
For the case of receptor binding by the antibody, the blue boxes are zero, whereas
the pink boxes are zero for ligand binding. Parameter values are given in Table 13.2;
they are inspired by actual measurements [37, 40]. The initial values R0, L0, and A0
are at first set to 4, 100, and 100, respectively, but are varied in different simulations.
As an illustration, we explore the system for the case where the antibody binds to
the ligand. Thus, R0 = 4, A0 = 100, and L0 is varied for different scenarios. Represen-
tative results (Figure 13.14) suggest the following. For low numbers of ligands, the
antibody reduces their concentration effectively, and the ligand concentration
remains below 30% of the initial value for about 2 days. By contrast, if the initial
ligand concentration is at the order of 10, the antibodies reduce it initially only to
about 40%, and within 2 days the concentration has reached 60%. For even higher
concentrations, for example, L0 = 100, the ligand concentration is almost back to
100% within 2 days. In addition, there is a strong accumulation of ligand–antibody
complex, which might cause undesirable side effects of the treatment. These results
suggest that designing an antibody to a ligand can be effective if the ligand concen-
tration is relatively low. If the concentration is high, the antibody is no longer
efficacious. Exercise 13.18 explores whether an antibody against the receptor is to
be preferred in this case.
In addition to changing R0 or L0, we can also easily explore different doses of
antibody. For instance, one might surmise that a higher dose might counteract the
decreasing efficacy of the treatment for higher ligand concentrations. To test this
hypothesis, we consider the case L0 = 10, and quadruple the dose of antibodies to
400. In comparison with the lower dose, the ligand concentration is better
TABLE 13.2: PARAMETER VALUES FOR THE SYSTEM (13.3)

kR+ kR− kL+ kL− k A+ k1+ k1− k2+ k2− k3+ k3− Ci
− −
R0 k R 0.0693 L0 k L 0.1155 0.0014 4 1 0.036 0.036 0.0018 0.036 0.001
(A) (B)
100 100
L0 = 100
L0 = 100
L0 = 10
L (%)
50 50
C3
L0 = 0.01
L0 = 1
L0 = 10
L0 = 0.01
L0 = 1
0 0
0 20 40 0 20 40
time (days) time (days)
Figure 13.14 Simulation results with a model of different types of antibody binding (13.3). For fixed initial monoclonal antibody and receptor
concentrations A0 and R0, both (A) the dynamics of free-ligand rebounding (L, as a percentage of the initial target concentration L0) and (B) the
concentration of the complex, C3, depend critically on L0.
controlled, but the concentration of the complex is again high (Figure 13.15). An
alternative to a higher dose is repeated injection of antibodies. Exercise 13.19 asks
you to explore variations on this theme.
13.7 Pharmacokinetic Modeling

Computational systems biology has been playing a significant role in drug develop-
ment for several decades by supporting pharmacokinetic and pharmacodynamic
assessments of the levels of therapeutic agents in different tissues and organs of the
body (see, for example, [41]). Two classes of approaches use either relatively simple
compartment models [42] or a particularly powerful expansion of these toward
physiologically based pharmacokinetic (PBPK) models [2, 43].
The base concepts for pharmacokinetic analyses and simulations are relatively
simple, at least conceptually. The body is represented as a set of homogeneous com-
partments, and a drug that was administered orally or through an injection migrates
among these compartments. Some compartments may excrete the drug (lungs
through exhalation, kidneys through urine, the gastrointestinal (GI) tract through
feces), and the liver, in particular, may break it down through enzymatic activity. The
key question is how long the drug resides in these different compartments and in
what concentrations, because this information is important for developing an effec-
tive dosing regimen.
To obtain an impression of the principles of this type of approach, let us consider
a simple compartment model that consists of just two compartments: the blood-
stream (B) and the liver (L) (Figure 13.16). In this simplified system, B receives the
drug through injection and later also from other organs, which, however, are not
modeled here explicitly; the input rate is kOB. B loses some of the metabolite to other
organs (with rate kBO) and exchanges the metabolite with the liver, with rates kBL and
100
C3
L (%)
50
Figure 13.15 Simulation result with the

model (13.3) for increased antibody loads.
If the initial antibody concentration A0 is
increased to 400, the free ligand rebounds
0 more slowly (compare with the curve L0 = 10
0 10 20 in Figure 13.14), but the concentration of the
time (days) complex, C3, is much higher.
Figure 13.16 A two-compartment

model for the exchange of a metabolite
between blood (B) and liver (L). Several
processes govern this exchange. They can be
formulated as a mathematical model such
as (13.4) in a straightforward manner (see
kOB Chapter 2). The parameters are explained in
kLB the text.
kLO
L B IN
kBL
kBO
kLB, respectively. The liver delivers some of the metabolite to the intestines or gall-
bladder with rate kLO.
Calling the concentration of the metabolite in the liver L and that in the blood B,
a linear model with possible additional input IN reads
B = kOB + kLB L − (kBO + kBL )B + IN ,

(13.4)
L = kBL B − (kLO + kLB )L.
Mathematically, kOB and IN could be merged, but we keep them separate to distin-
guish normal physiology, where B is continually augmented with material from the
rest of the body, and with additional input from an intravenous injection.
Let us suppose for a moment that no metabolite is transported from the blood-
stream into the liver (kBL = 0). The equation for L then has the form L = −kL, with
rate k = kLO + kLB, which immediately signals that L is losing material in an exponen-
tial fashion (see Chapters 2 and 4). As a second situation, imagine no influx through
kOB but a bolus injection IN at some time point. In other words, after the brief injec-
tion of IN, there is no further input and the only active processes are the exchanges
between the two compartments and efflux from both. After a short initial phase of
equilibration between L and B, the dynamics now consists merely of material shut-
tling between or leaving the pools, and these losses occur in an exponential fashion,
which corresponds to straight lines in a logarithmic plot (Figure 13.17A). This solu-
tion has the explicit form
B(t ) = c11e λ1t + c12e λ2t ,

(13.5)
L(t ) = c21e λ1t + c22e λ2t ,
where λ1 and λ2 are the eigenvalues of the system matrix A and c11, …, c22 are con-
stants that depend on the numerical specifications of the model (see Chapter 4).
Because the eigenvalues might be real or imaginary numbers, the functions in (13.5)
could in principle decrease toward zero or oscillate.
The most interesting situation occurs if none of the rate constants are zero and
the system is in a steady state, but then receives an injection IN. For instance, if
kOB = 1.2, kLB = 0.5, kBO = 0.9, kBL = 0.8, kLO = 0.3, and IN = 0, the system happens to
have the steady state B = L = 1. Suppose that we inject 5 units of IN within the time
interval 3 ≤ t ≤ 4. Since B is the primary recipient, it increases, but not in a linear
fashion, because some of the new material immediately starts moving into L, which
increases subsequently. After some while, IN has been cleared from the system,
which returns to its steady state (Figure 13.17B). Many scenarios can easily be tested
in this fashion (see Exercises 13.20–13.23).
(A) (B)
2.0 4
log L
log(concentration)
concentration
–1.5 2 L
log B
–5 0
0 5 10 0 5 10
time time
Figure 13.17 Responses of the two-compartment model (13.4) for the exchange of a metabolite between blood (B, with concentration B)
and liver (L, with concentration L). (A) After a one-time input, the material equilibrates and afterwards is lost in an exponential fashion (note the
logarithmic vertical axis). (B) The system operates at a steady state with constant influx, effluxes, and exchanges between B and L. Between times
3 and 4, B receives an injection with additional material, which moves between compartments and is ultimately cleared from the system.
If we realistically assume that the liver enzymatically degrades the metabo-

lite, the equation of L is augmented with additional degradation terms, maybe in
the form of Michaelis–Menten, Hill, or power-law functions, which make the
equation nonlinear (see Exercise 13.24). The nonlinearities make the model
more realistic but preclude some strictly mathematical analyses; however, they
are not a hindrance for all kinds of simulations. As so often, the true challenge of
this expansion lies in an accurate determination of suitable functions and param-
eter values.
More sophisticated and realistic than simple compartment models are PBPK
models that consist of multiple compartments representing organs or tissues,
which are functionally connected through the circulatory system (Figure 13.18). blood
The compartments are quantitatively represented with measurable characteristics
such as volume and perfusion rate. A drug enters the body by mouth, inhalation,
or injection and is transported to and among the various tissues via the blood- brain
stream. Organs such as the liver metabolize some of the drug. The rest cycles
through the body and is eventually excreted through feces and urine or, in some
cases, by exhalation or sweating, which causes the concentration eventually to lung
decrease to insignificantly low levels. Typical questions are: How long will it take
before a drug reaches a target tissue? How long will it remain in this tissue in ther-
apeutically relevant doses? What is the best route and dosing regimen for drug fat
administration? These questions are not easy to answer intuitively, because some
tissues are very strongly perfused (for example, the brain and liver), while others
are not (for example, fat). Also, many prescription drugs are lipophilic, so that they
kidney
have a much higher affinity to fat than, for instance, muscle. Furthermore, the
drug usually cycles through the system several times, and its residence time in the
various organs depends critically on its lipid content and other physicochemical liver
properties. PBPK models can be used not only to study the dynamics of the origi-
nal drug, but also the accumulation of toxic substances, including breakdown
products of the drug.
Probably the most important feature of PBPK models is that they can be utilized
as a mechanistic tool for scaling drug scenarios from animal experiments with mice,
rats, or dogs to humans. The same models are set up for the different species, but
Figure 13.18 Diagram of a generic
with parameter values that are adjusted, for instance, to the volumes of organs,
physiologically based pharmacokinetic
tested for one or two species, and then used to make predictions about the drug (PBPK) model. Blood circulates among
dynamics in humans. A typical case study is presented in Box 13.1. It is clear that the organs and tissues, carrying a drug of
success rate of the clinical drug development phase rises substantially if one can interest. The liver metabolizes some of the
rely on PBPK models that were set up for animals as reliable predictors of efficacy, drug, and the lung, kidney, and liver “lose”
safety, and toxicity in humans. drug by exhalation or excretion.
BOX 13.1: PHARMACOKINETICS OF METHOTREXATE
Methotrexate (MTX) is a powerful prescription drug that The movement through the bile duct is modeled with the
inhibits folate metabolism, which is crucial for the production equations
and repair of DNA and therefore the growth of cell populations.
MTX is used to treat various cancers and autoimmune diseases Bi = τ −1(Bi −1 − Bi ) for i = 1, 2, 3. (3)
and induces medical abortions. It is on the World Health
Organization’s List of Essential Medicines, which contains the Here, τ is the holding time in each section of the bile duct, and B0
most important medications needed in a basic health system is the Michaelis–Menten term in (2). The dynamics in the gut tissue
[44]. Being a very powerful medication, it is not surprising is represented as
that MTX can have serious, even life-threatening, side effects 4
 G 1  kG Gi 
in the liver, kidneys, lungs, and numerous other tissues. Any
treatment with MTX therefore requires a well-balanced dosing
VG G = qG  P −  +
 rG  4 ∑  K
i =1
G + G i
+ bGi  .

(4)
regimen.
Bischoff and colleagues [45, 46] studied the pharmacokinetics The gut tissue receives material from the four compartments
of MTX with a PBPK model that was originally implemented for G1, …, G4 of the gut lumen (see Figure 1). These transport
mice, but then also tested in rats and dogs. Model predictions processes are represented with a Michaelis–Menten process
were subsequently compared with plasma concentrations in that accounts for saturation as well as a nonsaturable effect.
humans, where they reflected the actual observations quite well. The concentrations of MTX in these compartments are modeled
Bischoff’s papers are quite old, but illustrate the principles of individually and for the entire gut lumen as
pharmacokinetic modeling very well. The following analysis is
directly adapted from this work [46]. VGL 1 k G 
G1 = B3 − kf VGLG1 −  G 1 + bG1 ,
The specific model diagram is shown in Figure 1. It is 4 4  K G + G1 
fundamentally the same as Figure 13.16, but the authors VGL 1 k G 
found it necessary to account for multistep processes of bile Gi = kf VGL (Gi −1 − Gi ) −  G i + bGi  , i = 2, 3, 4 , (5)
4 4  K G + Gi 
secretion (B1, …, B3) and transport through the gut lumen
4
(G1, …, G4).
∑G .
)= 1
The model is set up as explained in the text. In fact, the (GL i
4 i =1
equations for plasma, muscle and kidney are identical to
(13.6)–(13.8). The equation for liver is very similar, but additionally
By entering the equations into a solver like PLAS or MatLab and
contains terms for input from the gastrointestinal (GI) tract and
using the parameter values in Table 1, it is now possible to test
also for biliary secretion, which is modeled as the Michaelis–
the model against actual measurements. As a first example, Figure
Menten process B0. Thus, it has the form
2 shows the results of a simulation where 3 mg MTX per gram
of body weight were injected intravenously into a mouse. The
 L  G L
VL L = (qL − qG )  P −  + qG  −  − B0 , (1) injection may be modeled with an input (IN) to the equation for
 rL   rG rL  plasma or, more simply, by setting the initial value of P as 66 (3 mg
times body weight divided by the volume of plasma) and all other
where initial values close to zero (for example, 0.001). Note that results
of PBPK analyses are traditionally displayed with a logarithmic
 L
kL   scale on the y-axis. Similar to the simple two-compartment model
 rL  (Figure 13.17), one observes that all concentration profiles, after
B0 = . (2)
L the initial phase, become more or less linear, which corresponds
KL +
rL to exponential decay in all tissues. Exercises 13.25 and 13.26
plasma
qL– qG
qG
liver GI tract
biliary gut
secretion B0 absorption
B1 B2 B3 G1 G2 G3 G4
τ τ τ
bile duct gut lumen

feces
qK
kidney Figure 1 Diagram for the movement of methotrexate among
different compartments within the body. Note that secretion
from the liver is subdivided into three compartments of the bile
urine duct and that the gut lumen is subdivided into four sections.
qM
(Adapted from Bischoff KB, Dedrick RL, Zaharko DS & Longstreth
muscle JA. Methotrexate pharmacokinetics. J. Pharmacol. Sci. 60 [1971]
1128–1133.)
experiment, even though the concentration values are of course

TABLE 1: MODEL PARAMETERS FOR A different. Exercise 13.27 analyzes the MTX distribution among
METHOTREXATE CASE STUDY FOR THREE SPECIES* tissues in humans.
Model
Parameter Mouse Rat Human 100
component
Body weight (g) 22 200 70,000
methotrexate concentration (µg/g)

Volume (mL) 10
Plasma VP 1 9.0 3000
GL
Muscle VM 10.0 100 35,000
Kidney VK 0.34 1.9 280 1 L
Liver VL 1.3 8.3 1350
K
Gut tissue VG 1.5 11.0 2100
0.1 P
Gut lumen VGL 1.5 11.0 2100
Plasma flow rate (mL/min)
M
Muscle qM 0.5 3.0 420
0.01
Kidney qK 0.8 5.0 700 0 60 120 180 240
Liver qL 1.1 6.5 800 time (min)
Gut tissue qG 0.9 5.3 700

Figure 2 Methotrexate distribution among mouse tissues over
Partition coefficient time. Comparison of model predictions (lines) and experimental data
Muscle rM 0.18 0.11 0.15 (dots) for intravenous injection of 3 mg methotrexate per gram of
body weight in a mouse. GL, gut lumen; K, kidney; L, liver; M, muscle;
Kidney rK 3.0 3.3 3.0
P, plasma. (Data from Bischoff KB, Dedrick RL, Zaharko DS & Longstreth
Liver rL 7.0 3.3 3.0 JA. Methotrexate pharmacokinetics. J. Pharmacol. Sci. 60 [1971]
Gut tissue rG 1.2 1.0 1.0 1128–1133.)
Kidney clearance
kK 0.2 1.1 190
(mL/min) 100
Bile secretion parameters
methotrexate concentration (µg/g)
Clearance (mL/min) kL/KL 0.5 3.0 200

10 GL
Residence time (min) τ 2.0 2.0 10
Gut lumen parameters
Transit time (min) 100 100 1000
1 L
Reciprocal transit K
kf 0.01 0.01 0.001
time (min–1)
P
Maximal velocity kG 0.2 20 1900
0.1
Michaelis constant KG 6.0 200 200
Rate constant, M
b 0.001 0.0 0.0
absorption (mL/min)
0.01
*Adapted from Bischoff KB, Dedrick RL, Zaharko DS & Longstreth 0 60 120 180 240
JA. Methotrexate pharmacokinetics. J. Pharmacol. Sci. 60 (1971) time (min)
1128–1133.
Figure 3 Methotrexate distribution among rat tissues over time.
address a simulation of 300 mg MTX per gram of body weight and Comparison of model predictions (lines) and experimental data
a variant modeling strategy where the injection is represented as (dots) for intravenous injection of 6 mg methotrexate per gram of
a bolus. body weight in a rat. GL, gut lumen; K = kidney; L, liver; M, muscle; P,
As a second example, let us consider an injection of 6 mg plasma. (Data from Bischoff KB, Dedrick RL, Zaharko DS & Longstreth
MTX per gram of body weight into a rat. The results are shown in JA. Methotrexate pharmacokinetics. J. Pharmacol. Sci. 60 [1971]
Figure 3. The overall appearance is similar to that of the mouse 1128–1133.)
One somewhat tricky technical issue with PBPK models is that the typical mea-
surement unit in biochemistry, physiology, and pharmaceutical science is a con-
centration, but the various organs and tissues have very different volumes, so the
same amount of a drug moving, say, from the bloodstream to the lung tissue all of a
sudden corresponds to a different concentration. PBPK models deal with this issue
by multiplying concentrations by volumes, thereby computing the drug dynamics in
absolute amounts. Outside this feature, most PBPK models use standard mass-
action kinetics and Michaelis–Menten functions, but they could of course also be
formulated with other rate functions, such as power laws.
As an important example, the dynamics of a drug in the blood plasma is typically
written as
q q q
VP P = IN + L L + K K + M M − (qL + qK + qM )P . (13.6)
rL rK rM
The left-hand side shows the time derivative of the concentration P of the drug in
plasma, multiplied by the volume VP of plasma, and is thus the time derivative of the
actual amount of drug. On the right-hand side, we have amounts of drug entering
(positive signs) and leaving (negative signs) the bloodstream. Beyond the input
(IN), typically an intravenous injection, we see terms with two new types of param-
eters: flow rates between compartments, traditionally denoted by q and a subscript
identifying the compartment, and so-called partition coefficients, which are
denoted by subscripted r parameters and represent the fact that only portions of a
tissue actively take up or release the drug. The quantities denoted by capital letters
are concentrations of the drug in liver (L), kidney (K) and muscle (M). Depending
on the application, other compartments, such as brain, lung, and fatty tissue, could
also be considered.
The equations for other organs are constructed similarly, but are usually simpler.
For instance, the drug concentration (or amount) in muscle is formulated as
= qM  P − M  .
VM M (13.7)
 rM 
Similarly, the typical equation for the kidneys is written as
 K K
VK K = qK  P −  − kK . (13.8)
 rK  rK
It contains an extra term representing the excretion of the drug through urine. The
case study in Box 13.1 provides further details and a few case-specific twists.
In principle, PBPK models may account for arbitrarily many compartments, and
there is no real limit to their complexity. Typically, all transport steps are modeled as
first-order processes, and metabolism within organs is not considered, with the
possible exception of the liver, so that standard PBPK models consist of systems of
linear differential equations, for which there are many methods of analysis
(see Chapter 4). Of course, this is an approximation.
All in all, PBPK models have been very successful in the pharmaceutical industry
because they permit predictions of dynamically changing drug distributions among
organs and tissues based on animal experiments and on individual human-specific
parameters like organ volumes, partition coefficients, and flow rates that can be
estimated for humans without exposing them to the new drug.
13.8 Pathway Screening with Dynamic Models

Most dynamic pathway models do not account explicitly for different spatial loca-
tions such as organs. Instead, they attempt to capture quantitatively how a drug
alters the movement of metabolites through a diseased pathway system and hope-
fully leads to a more desirable steady-state metabolic profile. The specific goal is
typically an increase or decrease in a specific target metabolite. In addition to rep-
resenting the primary pathway metabolites, dynamic models permit analyses of
side effects within the pathway, such as the accumulation of toxic substances or the
depletion of vital metabolites. Sufficiently accurate pathway models are thus very
versatile and can play a role in efficacy, safety, toxicity, and dosing assessments
[47]. Furthermore, similar to pharmacokinetic models, they have the potential to
aid extrapolations from animal to human test results. Sophisticated models in the
future will be capable of capturing large portions of human physiology and metab-
olism in sufficient detail. They will become reliable enough to screen out seemingly
promising but ultimately ineffective leads early on in the drug development pipe-
line and thereby save valuable resources. Pathway models for drug discovery are
reviewed in [48].
As an illustration of the role of pathway models in target screening, consider a
very simplified model of purine metabolism (Figure 13.19). This biosynthetic path-
way is responsible for generating sufficient amounts of purines, which are compo-
nents of DNA and RNA and can also be converted into a number of other crucial
metabolites. Here, we assume that the cells are not growing, so that hardly any new
DNA is needed, and that the RNA pool is in equilibrium, where as much is produced
as degraded.
For our simplistic illustration, we focus on a disorder called hyperuricemia,
which is an elevation of the uric acid concentration in the blood and can be the
result of various perturbations within the purine pathway. A common disease asso-
ciated with hyperuricemia is gout. The model corresponding to the pathway in Fig-
ure 13.19 was inspired by earlier modeling efforts [14, 49–51], and, although it is
much simpler than the original model, it allows us to illustrate some salient points.
The mathematical representation in power-law form is as follows:
P = vin − v PG − v PI − v PHI − v PH ,
I = v PI + v PHI + vGI − v IG − v IH ,
G = v PG + v IG − vGI − vGX ,
(13.9)
H = v PH + v IH − v PHI − v HX − v H out ,
X = vGX + v HX − v XU ,
U = v XU − vU out ,
where the fluxes, in self-explanatory notation, take the form
Vin
VPG P VPH
VPI
VPHI
I
VIG
VIH
VGI
G H
VH out
VGX VHX
Figure 13.19 Simplified representation

of purine metabolism. The diagram, and
X
the corresponding model (13.6), (13.7),
consist of metabolites (boxes), enzymatic
VXU conversions and transport steps (blue
arrows), and inhibitory signals (red arrows).
P, phosphoribosyl pyrophosphate; I,
U
inosine monophosphate; G, guanylates; H,
VU out hypoxanthine and inosine; X, xanthine; U,
uric acid.
vin = 5,
v PG = 320 P 1.2G −1.2 , vGX = 0.01G 0.5 ,
v PI = 0.5P 2 I −0.6 , v IH = 0.1I 0.8 ,
v PH = 0.3P 0.5 , v HX = 0.75H 0.6 , (13.10)
v PHI = 1.2 PI −0.5 H 0.5 , v H out = 0.004 H ,
v IG = 2 I 0.2G −0.2 , v XU = 1.4 X 0.5 ,
vGI = 12G 0.7 I −1.2 , vU out = 0.031U.
With these functions and numerical settings, the steady state is approximately (P, I,
G, H, X, U) = (5, 100, 400, 10, 5, 100) (cf. [51]).
We discuss two perturbations within this pathway that lead to hyperuricemia.
The first is hyperactivity of the enzyme phosphoribosylpyrophosphate synthase
(PRPPS), which catalyzes the input step vin. The second perturbation is due to a
severe deficiency in the enzyme hypoxanthine–guanine phosphosribosyltransfer-
ase (HGPRT), which in the simplified model affects vPG and vPHI. In both cases, the
concentration of uric acid (U) is significantly increased, among other problems. It is
easy to simulate these conditions with the model in (13.9) and (13.10).
To represent PRPPS hyperactivity, we simply multiply vin by an appropriate fac-
tor, such as 2. In response, the steady-state profile changes to about (P, I, G, H, X,
U) = (13, 260, 1327, 23, 14, 170). All metabolites are elevated, including U, which is
not surprising, because more material is flowing into the system. For the simulation
of HGPRT deficiency, we multiply vPG and vPHI by a number between 0 and 1, where
0 would correspond to total absence of enzyme activity and 1 is the normal case. If
the enzyme has only two-thirds of its normal activity, the situation is actually not so
bad: the steady-state profile in this case is about (P, I, G, H, X, U) = (6, 103, 370, 11, 6,
109), which is different from the normal profile, but not by all that much. The situa-
tion deteriorates critically if the activity drops much further, leading to a disease that
is known as Lesch–Nyhan syndrome and is accompanied by severely compromised
brain function. If we reduce the activity in the model to 1%, the metabolite profile
becomes (P, I, G, H, X, U) = (11, 97, 168, 21, 12, 157), where the guanylate levels are
low and U is high.
The pertinent question here is whether it is possible to interfere effectively with
any of the enzymes in order to compensate the changes in the metabolite profile
that are caused by a disorder like PRPPS hyperactivity. The most straightforward
strategy would be to inhibit PRPPS, but that is currently not possible. One active
ingredient on the market is allopurinol, which is structurally similar to hypoxan-
thine and inhibits the enzyme xanthine oxidase, which oxidizes both hypoxanthine
to xanthine and xanthine into uric acid. Multiplying vHX and vXU by 0.33, reflecting
33% remaining enzyme activity, does indeed control the uric acid level. The result-
ing profile is (P, I, G, H, X, U) = (10, 259, 1124, 66, 54, 109). However, as is also
observed in clinical studies, hypoxanthine and xanthine accumulate, along with the
other metabolites in the system.
It is rather evident that an effective remedy should remove disease-associated
excess material from the system. Uricase is an enzyme that catalyzes the oxidation
of uric acid to 5-hydroxyisourate and has been formulated as a drug for hyperurice-
mia. To test its efficacy, we multiply vU out by a factor representing the efficacy of the
drug. If the activity of vU out is increased by 50%, the uric acid concentration indeed
comes within 10% of the normal value, and the metabolite profile is (P, I, G, H, X,
U) = (13, 252, 1215, 22, 13, 110). Alas, as in the previous scenario, P, I, and G accu-
mulate quite significantly.
The enzyme information system BRENDA [52] indicates that the key initial
enzyme of purine biosynthesis, amidophosphoribosyltransferase (EC 2.4.2.14),
has a variety of inhibitors. Would it be useful to design a drug based on one of
these? We can easily explore a hypothetical inhibitor affecting the corresponding
process vPI. For example, reducing the activity to 5% leads to the metabolite profile
(P, I, G, H, X, U) = (22, 235, 1475, 11, 7, 118). As in the previous case, the uric
acid is more or less under control. Also as before, P, I, and G are greatly increased.
In contrast to the allopurinol treatment, hypoxanthine and xanthine are now

close to normal.
Are there other options? The diagram of the pathway contains two exit routes:
one through U and one through H. Increasing vH out by 100 does indeed control U,
along with X and H, but P, G, and I still accumulate. Could it potentially be effective
to reroute excess P so that intermediates do not accumulate? Indeed, if we increase
vPH by a factor of 8 in addition to increasing vH out by 100, essentially all excess mate-
rial is funneled out of the system and the resulting profile is almost normal, namely,
(P, I, G, H, X, U) = (5, 110, 426, 11, 6, 111). The model obviously cannot tell whether
such a combined strategy can be developed as a useful pharmaceutical option or if
it leads to complications elsewhere in the body. However, it demonstrates that
modeling can be a powerful tool for screening hypotheses (see Exercises 13.28 and
13.29).
Dynamic models in pharmaceutical research are not limited to metabolism and
transport between organs, the two systems chosen here because of their relative
simplicity. Chemical, hormonal, electrical, and mechanical signaling, as well as
aspects of higher-level systems physiology, will play increasingly important roles as
diseases of the brain and of the cardiovascular, endocrine, and immune systems
become sufficiently well understood for mechanical investigations at the molecular
level. An example of such efforts is the Physiome Project [53], which has as its cen-
tral theme the creation of integrated computational models of complete organ sys-
tems and, eventually, organisms.
13.9 Emerging Roles of Systems Biology in Drug

Development
The examples shown in this chapter indicate that the role of computational systems
biology in drug development has already begun to shift from “potential” to “actively
contributing.” In many cases, the contributions are still modest, but they do suggest
the growing contribution of systems biology to this field. In addition to molecular
and QSAR modeling, systems biology should be expected to play an increasingly
important role in exploring the following [41]:
• the feasibility of an envisioned disease-disrupting drug treatment

• the choice of a target (ligand, receptor, complex; upstream or downstream)
• the efficacy of a therapeutic molecule
• the balance between safety and efficacy
• the toxicity of NCEs and biologics
• the consequences of properties of a therapeutic molecule (half-life, affinity,
attainable concentration, etc.)
• the best therapeutic modality (antibodies, peptide mimics, engineered enzymes,
RNA interference (RNAi), etc.)
• the route of administration
• the optimal dosing regimen
• the need to account for diffusion and transport (within tumors, through mem-
branes, across the blood–brain barrier, etc.)
• the effects of clearance
• the extrapolation of test results between animals and from animals to humans
Considering the fact that the young field of systems biology is already involved in
many areas of drug discovery, it is fair to expect that systems biology will move into
the mainstream of pharmaceutical research and development in the foreseeable
future. At the same time, considerations of health will shift more from treating to
preventing disease, and the application of prescription drugs will become increas-
ingly personalized. Systems biology will provide powerful tools to assist with these
shifts.
EXERCISES
13.1. Using the example in (13.1), perform the same type 13.13. Again using the simple model (13.1), study the
of analysis by decreasing one enzyme activity at a long-term transient effects of a decreased clearance
time. Are the results symmetric in a sense that they of X4. For instance, if a concentration of X4 above 12
mirror the results of increasing activities? or 40 (roughly 3 or 10 times the normal level,
13.2. In the example in (13.1), the flux through X4 is very respectively) is deleterious, how long does it take
small. Suppose that this flux represents clearance of before the patient should be considered sick?
an undesired metabolite by the kidneys, and that a 13.14. Search the literature for data on the success or
patient with kidney disease can eliminate X4 at only attrition of NCEs and biologics in different disease
10% capacity or not at all. Implement these two areas (inflammation, cardiovascular, oncology, etc.).
scenarios in the model and report the results. Identify at which stages of the drug development
13.3. Using the example in (13.1), study the effects of a process the failure rate is the highest and the lowest.
decreased efflux from X6. Compare the results with 13.15. Expand the simple antibody–target model (13.2)
those of decreasing the clearance of X4. to include clearance of the complex C. Perform
13.4. Using the example in (13.1), determine the most simulations and compare the results with
influential combination of two parameters with corresponding results of the model without
respect to X6. Develop a strategy for testing the clearance of C. Explore different half-lives for
effects of three or four simultaneous perturbations the clearance process.
(between 5% and 50%) in enzyme activities in 13.16. Expand the simple antibody–target model (13.2)
the model. Look in the modeling chapters for to allow for multiple doses of antibody injection.
alternatives to exhaustive analyses, which become Perform simulations with different dosing regi-
infeasible owing to combinatorial explosion. mens. Determine an optimal dosing schedule that
13.5. Define health and disease without expressing one keeps the target under control, while minimizing
in terms of the other. the total dose and adhering to a patient-friendly
schedule.
13.6. In a model of a metabolic health and disease
system, are biomarkers to be represented by 13.17. Chabot and Gomes [36] state that the half-life of a
parameters, independent or dependent variables, collection of different clearance processes can be
two of the three, or all three? Explain your answer computed as
and provide examples where applicable.
1 1 1
13.7. Search the literature for examples of biomarkers = +
t1/2; overall t1/2; proteolysis t1/2; kidney
that are not genomic or related to enzymes. Discuss
how they would enter mathematical models of 1 1
+ + .
some health and disease system under t1/2; endocytosis t1/2; other
investigation.
13.8. Discuss whether false-positive or false-negative Prove this statement mathematically or design an
interpretations of biomarker readings, such as ordinary differential equation model to test it
those corresponding to the pink and green points computationally. Compare the collective half-life
in Figure 13.6, become more of a problem or less of with the half-lives of the individual clearance
a problem in a larger system of biomarkers. processes.
13.9. Search the literature for reports on drug treatments 13.18. Use the receptor–ligand–antibody model (13.3)
whose efficacy can be predicted from genome to study the dynamics of ligands if the antibody
analyses. binds to the receptor. Compare your results with
those in the text.
13.10. Search the literature for nongenomic biomarkers
that are predictive of disease or the success of some 13.19. Use the receptor–ligand–antibody model (13.3) to
specific medication. allow for multiple doses of antibody injection.
Perform simulations with different dosing
13.11. Discuss the role of cytochrome P450 enzymes as
regimens.
possible biomarkers for disease, disease resistance,
or the efficacy of drugs. Formulate your findings as 13.20. Simulate constant input IN in the compartment
a report. model (13.4).
13.12. Using the simple model (13.1) at the beginning of 13.21. Simulate three equal daily doses of a drug that
the chapter, define health and disease states and enters the compartment system (13.4) as input IN.
analyze parameter settings within the disease Can you set the doses such that the drug concen-
domain. Is it possible to simulate situations where tration in blood is constant? Or that it at least stays
the analysis of a single biomarker suggests a within some predefined range?
healthy state, whereas the consideration of several 13.22. What happens if the efflux parameters kBO and kLO
biomarkers suggests disease? Either show examples in the compartment model (13.4) are zero? Make a
or prove that such a situation is not possible. prediction before you simulate.
REFERENCES 395
13.23. Develop a compartment model like (13.4) that input bolus IN, which is set at 66 for one time unit,
contains fatty tissue in addition to blood and liver. 33 for two time units, or 22 for three time units.
Set the parameters such that they reflect very low 13.27. Using the information from Box 13.1, simulate the
perfusion in fat. Perform representative dynamic MTX distribution among tissues in
simulations. humans, given an injection of 1 mg per kg of body
13.24. Develop a compartment model like (13.4) that weight. Compare the results with the measure-
accounts for degradation of the toxic substance in ments in Table 13.4. Note that measurements for
the liver. Compare the results for degradation other tissues are not available.
processes in the form of a Michaelis–Menten rate 13.28. Explore other possible treatments of hyperuricemia
law and a power-law function. with the simplified purine model (13.9).
13.25. Simulate an injection of 300 mg MTX per gram of 13.29. Using the simplified purine model (13.9), compare
body weight and compare the results with the data the efficacy of combined treatments (at least at two
in Table 13.3. target sites) with inhibition or activation of individual
13.26. Using the information from Box 13.1, explore the steps. Is it possible that drugs are synergistic in a sense
differences between modeling the drug injection that their combination is stronger than the sum of the
either as a change in the initial value of P or as an two individual treatments?
TABLE 13.3: MEASURED CONCENTRATIONS (µg/g) OF MTX IN RESPONSE TO AN

INJECTION OF 300 mg MTX PER GRAM BODY WEIGHT IN MOUSE*
Time after injection (min) 20 50 70 80 150 180 200
Plasma 100 25.6 18.2 10 2.2 1.5 4.3
Gut lumen 1292 1668 1978 2154 2346 426 256
*From Bischoff KB, Dedrick RL, Zaharko DS & Longstreth JA. Methotrexate pharmacokinetics.
J. Pharmacol. Sci. 60 (1971) 1128–1133.
TABLE 13.4: MEASURED CONCENTRATIONS (µg/g) OF MTX IN PLASMA

FOLLOWING AN INJECTION OF 1 mg MTX PER GRAM BODY WEIGHT IN TWO
HUMANS*,†
Time after injection (min) 15 45 60 120 150 240 300 360
Experiment 1 1.10 0.821 0.693 0.453 — 0.284 — 0.211
Experiment 2 1.91 1.31 1.06 — 0.560 — 0.23 —
*From Bischoff KB, Dedrick RL, Zaharko DS & Longstreth JA. Methotrexate pharmacokinetics.
J. Pharmacol. Sci. 60 (1971) 1128–1133.
†
The data were obtained in two experiments.
REFERENCES
[1] Human Genome Project Information. SNP Fact Sheet. https:// (J Schulkin, ed.), pp 17–64. Cambridge University Press,
ghr.nlm.nih.gov/primer/genomicresearch/snp. 2004.
[2] Reddy MB, Yang RS, Clewell HJ III & Andersen ME (eds). [9] Kitano H, Oda K, Kimura T, et al. Metabolic syndrome and
Physiologically Based Pharmacokinetic Modeling: Science robustness tradeoffs. Diabetes 53 (Suppl 3) (2004) S6–S15.
and Applications. Wiley, 2005. [10] Lipsitz LA. Physiological complexity, aging, and the path to
[3] Voit EO. A systems-theoretical framework for health frailty. Sci. Aging Knowledge Environ. 2004 (2004) pe16.
and disease. Abstract modeling of inflammation and [11] van’t Veer LJ & Bernards R. Enabling personalized cancer
preconditioning. Math. Biosci. 217 (2008) 11–18. medicine through analysis of gene-expression patterns.
[4] Ishii N, Nakahigashi K, Baba T, et al. Multiple high-throughput Nature 452 (2008) 564–570.
analyses monitor the response of E. coli to perturbations. [12] Foubister V. Genes predict childhood leukemia outcome.
Science 316 (2007) 593–597. Drug Discov. Today 10 (2005) 812.
[5] Miller GW. The Exposome: A Primer. Academic Press, 2014. [13] Babaian RJ, Fritsche HA, Zhang Z, et al. Evaluation of
[6] Emory/Georgia Tech Predictive Health Institute. http:// ProstAsure index in the detection of prostate cancer: a
predictivehealth.emory.edu/. preliminary report. Urology 51 (1998) 132–136.
[7] Voit EO & Brigham KL. The role of systems biology in [14] Voit EO. Computational Analysis of Biochemical Systems:
predictive health and personalized medicine. Open Pathol. J. A Practical Guide for Biochemists and Molecular Biologists.
2 (2008) 68–70. Cambridge University Press, 2000.
[8] Sterling P. Principles of allostasis: optimal design, predictive [15] Entelos, Inc. http://www.Entelos.com.
regulation, pathophysiology and rational therapeutics. [16] Darmoni SJ, Massari P, Droy J-M, et al. SETH: an expert system
In Allostasis, Homeostasis, and the Costs of Adaptation for the management on acute drug poisoning in adults.
Comput. Methods Programs Biomed. 43 (1994) 171–176.
[17] Luciani D, Cavuto S, Antiga L, et al. Bayes pulmonary [36] Chabot JR & Gomes B. Modeling efficacy and safety of
embolism assisted diagnosis: a new expert system for clinical engineered biologics. In Drug Efficacy, Safety, and Biologics
use. Emerg. Med. J. 24 (2007) 157–164. Discovery: Emerging Technologies and Tools (S Ekins & JJ Xu,
[18] Sundareswaran KS, de Zelicourt D, Sharma S, et al. Correction eds), pp 301–326. Wiley, 2009.
of pulmonary arteriovenous malformation using image- [37] Mayawala K & Gomes B. Prediction of therapeutic index
based surgical planning. JACC Cardiovasc. Imaging 2 (2009) of antibody-based therapeutics: mathematical modeling
1024–1030. approaches. In Predictive Toxicology in Drug Safety (JJ Xu & L
[19] Kola I & Landis J. Can the pharmaceutical industry reduce Urban, eds), pp 330–343. Cambridge University Press, 2010.
attrition rates? Nat. Rev. Drug Discov. 3 (2004) 711–715. [38] Nimjee SM, Rusconi CP & Sullenger BA. Aptamers:
[20] Central Intelligence Agency. CIA World Factbook: Life an emerging class of therapeutics. Annu. Rev. Med.
Expectancy at Birth. https://www.cia.gov/library/ 56 (2005) 555–583.
publications/the-world-factbook/fields/2102.html#us. [39] Arakaki AK, Mezencev R, Bowen N, et al. Identification of
[21] Research and development costs for new drugs by metabolites with anticancer properties by computational
therapeutic category. A study of the US pharmaceutical metabolomics. Mol. Cancer 7 (2008) 57.
industry. DiMasi JA, Hansen RW, Grabowski HG, Lasagna L. [40] Gomes B. Industrial systems biology. Presentation at
Pharmacoeconomics (1995) 152–169. PMID: 10155302. Foundations of Systems Biology in Engineering, FOSBE’09,
[22] An analysis of the attrition of drug candidates from four Denver, CO, August 2009.
major pharmaceutical companies. Waring MJ, Arrowsmith [41] Bonate PL. Pharmacokinetic–Pharmacodynamic Modeling
J, Leach AR, et al. Nat. Rev. Drug Discov. 14 (2015) 475–486. and Simulation, 2nd ed. Springer, 2011.
PMID: 26091267. [42] Jacquez JA. Compartmental Analysis in Biology and Medicine,
[23] Goodnow RAJ. Hit and lead identification: integrated 3rd ed. BioMedware, 1996.
technology-based approaches. Drug Discov. Today: [43] Lipscomb JC, Haddad S, Poet T, Krishnan K. Physiologically-
Technologies 3 (2006) 367–375. based pharmacokinetic (PBPK) models in toxicity testing
[24] Andrade CH, Pasqualoto KF, Ferreira EI & Hopfinger AJ. and risk assessment. In New Technologies for Toxicity Testing (M
4D-QSAR: perspectives in drug design. Molecules 15 (2010) Balls, RD Combes & N Bhogal N, eds), pp 76–95. Landes
3281–3294. Bioscience and Springer Science + Business Media, 2012.
[25] Funatsu K, Miyao T & Arakawa M. Systematic generation of [44] World Health Organization (WHO). WHO Model List of
chemical structures for rational drug design based on QSAR Essential Medicines. http://www.who.int/medicines/
models. Curr. Comput. Aided Drug Des. 7 (2011) 1–9. publications/essentialmedicines/en/.
[26] Zuckermann RN & Kodadek T. Peptoids as potential [45] Bischoff KB, Dedrick RL & Zaharko DS. Preliminary model for
therapeutics. Curr. Opin. Mol. Ther. 11 (2009) 299–307. methotrexate pharmacokinetics. J. Pharmacol. Sci. 59 (1970)
[27] Wishart DS. Improving early drug discovery through ADME 149–154.
modelling: an overview. Drugs R D 8 (2007) 349–362. [46] Bischoff KB, Dedrick RL, Zaharko DS & Longstreth JA.
[28] DiMasi JA, Hansen RW & Grabowski HG. The price of Methotrexate pharmacokinetics. J. Pharmacol. Sci. 60 (1971)
innovation: new estimates of drug development costs. 1128–1133.
J. Health Econ. 835 (2003) 1–35. [47] Kell DB. Systems biology, metabolic modelling and
[29] Kitano H. Computational systems biology. Nature 420 (2002) metabolomics in drug discovery and development. Drug
206–210. Discov. Today 11 (2006) 1085–1092.
[30] Hood L. Systems biology: integrating technology, biology, [48] Yuryev A (ed.). Pathway Analysis for Drug Discovery:
and computation. Mech. Ageing Dev. 124 (2003) 9–16. Computational Infrastructure and Applications. Wiley, 2008.
[31] Voit EO. Metabolic modeling: a tool of drug discovery in the [49] Curto R, Voit EO & Cascante M. Analysis of abnormalities
post-genomic era. Drug Discov. Today 7 (2002) 621–628. in purine metabolism leading to gout and to neurological
[32] Selventa. http://www.selventa.com/. dysfunctions in man. Biochem. J. 329 (1998) 477–487.
[33] GNS Healthcare. https://www.gnshealthcare.com/. [50] Curto R, Voit EO, Sorribas A & Cascante M. Validation and
[34] Huang PH, Mukasa A, Bonavia R, et al. Quantitative analysis of steady-state analysis of a power-law model of purine
EGFRvIII cellular signaling networks reveals a combinatorial metabolism in man. Biochem. J. 324 (1997) 761–775.
therapeutic strategy for glioblastoma. Proc. Natl Acad. Sci. [51] Curto R, Voit EO, Sorribas A & Cascante M. Mathematical
USA 104 (2007) 12867–12872. models of purine metabolism in man. Math. Biosci.
[35] Sachs K, Perez O, Pe’er S, et al. Causal protein-signaling 151 (1998) 1–49.
networks derived from multiparameter single-cell data. [52] BRENDA: The Comprehensive Enzyme Information System.
Science 308 (2005) 523–529. http://www.brenda-enzymes.org/.
[53] IUPS Physiome Project. http://www.physiome.org.nz/.
FURTHER READING
Auffray C, Chen Z & Hood L. Systems medicine: the future of Kell DB. Systems biology, metabolic modelling and metabolomics
medical genomics and healthcare. Genome Med. 1 (2009) 2. in drug discovery and development. Drug Discov. Today 11
Bassingthwaighte J, Hunter P & Noble D. The Cardiac Physiome: (2006) 1085–1092.
perspectives for the future. Exp. Physiol. 94 (2009) 597–605. Kitano H. Computational systems biology. Nature 420 (2002) 206–210.
Bonate PL. Pharmacokinetic–Pharmacodynamic Modeling and Kitano H, Oda K, Kimura T, et al. Metabolic syndrome and
Simulation, 2nd ed. Springer, 2011. robustness tradeoffs. Diabetes 53 (Suppl 3) (2004) S6–S15.
Gonzalez-Angulo AM, Hennessy BT & Mills GB. Future of Lipsitz LA. Physiological complexity, aging, and the path to frailty.
personalized medicine in oncology: a systems biology Sci. Aging Knowledge Environ. 2004 (2004) pe16.
approach. J. Clin. Oncol. 28 (2010) 2777–2783. Reddy MB, Yang RS, Clewell HJ III & Andersen ME (eds).
McDonald JF. Integrated cancer systems biology: current progress Physiologically Based Pharmacokinetic Modeling: Science
and future promise. Future Oncol. 7 (2011) 599–601. and Applications. Wiley, 2005.
FURTHER READING 397
Voit EO. A systems-theoretical framework for health and disease. West GB & Bergman A. Toward a systems biology framework for
Abstract modeling of inflammation and preconditioning. understanding aging and health span. J. Gerontol. A Biol.
Math. Biosci. 217 (2008) 11–18. Sci. Med. Sci. 64 (2009) 205–208.
Voit EO & Brigham KL. The role of systems biology in predictive health Weston AD & Hood L. Systems biology, proteomics, and the
and personalized medicine. Open Pathol. J. 2 (2008) 68–70. future of health care: toward predictive, preventative,
Voit EO. The Inner Workings of Life. Vignettes in Systems Biology. and personalized medicine. J. Proteome Res. 3 (2004)
Design of Biological
Systems 14
• Recognize the importance of studying design principles
• Identify concepts and basic tools of design analysis
• Discuss different types of motifs in biological networks and systems
• Set up static and dynamic models for analyzing design principles
• Describe methods for manipulating biological systems toward a goal
• Understand the concepts and basic tools of synthetic biology
• Characterize the role of systems and synthetic biology in metabolic
engineering
This chapter intentionally carries a potentially ambiguous title, because it discusses

two topics that fit under the same heading and are closely related to each other, but
are also genuinely distinct. The first aspect addresses questions of the natural
designs of biological systems, as we observe them: Are there structural patterns or
motifs that we find time and again in biological networks? If so, what are their roles
and advantages? Are there design principles that are prevalent in dynamic biologi-
cal systems? Why are some genes regulated by an inducer and some by a repressor?
Does it really matter? What is the advantage of a feedback signal inhibiting the first
step of a linear pathway rather than the second or third step?
The second aspect covered in this chapter addresses how we can alter the exist-
ing design of a biological system in a goal-oriented fashion or even create new
biological systems from scratch. This artificial design serves two purposes. First, it
adheres to Richard Feynman’s famous quote “What I cannot create, I do not under-
stand.” In other words, by creating first small and then larger biological modules
that function as predicted, we can demonstrate that we comprehend the principles
that underlie the organization and function of natural systems. Or, if the artificial
system behaves differently than expected, it indicates that there are aspects we do
not understand correctly. Second, the targeted alteration of existing biological
systems or the de novo creation of novel systems offers the potential of using biology
for new purposes. This second aspect of design is nowadays called synthetic biology
and can be seen as applied systems biology or a true engineering approach to
biology. Presently, the most prevalent application is metabolic engineering, which
addresses the microbial production of valuable compounds such as human insulin
or the large-scale microbial production of bulk materials such as ethanol or citric
acid. A different type of application is the artificial creation of microbial systems as
very sensitive sensors for environmental toxins or as drivers of environmental
remediation for soils containing metals or radionuclides.
400 Chapter 14: Design of Biological Systems
Of course, the natural and artificial sides of biological design studies have quite
a degree of overlap. First, they both deal with the structure, function, and control of
biological systems, with one being a manifestation of evolution and the other the
result of clever molecular biology and engineering. Second, both directly or
indirectly involve optimization. Natural designs are often assumed to be optimized
by the forces of selection, although this assumption cannot really be proven. It seems
indeed to be the case that natural designs are typically (always?) superior to compa-
rable alternatives, but who is to say that evolution has stopped at the beginning of
the third millennium of our era? Should one really doubt that organisms will evolve
further and that they will have out-competed today’s species in a few million years?
You could even ask yourself: am I optimal?
Optimization is certainly center-stage in the design of synthetic biological
systems. In terms of de novo designs, we may initially be happy if we succeed with
the creation of a biological module that functions at all, but it is our human nature
immediately to start improving and trying to optimize such a creation. Even more
pronounced is the role of optimization in the manipulation of existing systems,
for instance, in the cheap production of biofuel. Is it possible to increase the yield
in yeast-based alcohol fermentation? What are the optimal strain, medium, and
mode of gene expression? Can we make the production cheaper by using organ-
isms such as Escherichia coli instead of yeast? Is it possible to use autotrophic
organisms like plants or algae that are able to convert freely available sunlight and
carbon dioxide into sugars, and to manipulate them into converting the sugars
into ethanol, butanol, or harvestable fatty acids? How could we maximize yield?
No matter whether the investigation addresses natural or artificial systems, the
exploration and exploitation of biological design principles requires a mix of biology,
computation, and engineering, and thus of systems biology. So, let us embark on the
study of biological design, be it natural or created artificially, exploring optimality or
at least superiority with respect to reasonable alternatives.
NATURAL DESIGN OF BIOLOGICAL SYSTEMS

If every two biological systems had entirely different structures and modes of opera-
tion, the study of biology would have a lot of similarity with casual stargazing, where
we may find patterns remotely resembling people, animals, or objects, but where
there is no rhyme or reason for why such patterns happen to be there. As humans
have done for thousands of years, we simply describe and name them. We could
study biological systems one at a time and classify them by some features, but we
would not gain much insight into the rationale behind these features and their spe-
cific roles. Thus, when we talk about investigating the design of a biological system,
we are really trying to address design patterns, design principles, or motifs that show
up time and again with relatively minor variations. Instead of comparing biology
with stargazing, we should therefore compare biological designs with technological
and engineered designs. A good watchmaker can explain the role of every spring or
cogwheel in a particular location of a mechanical watch, even if he or she has never
seen this particular watch before. The shapes of cars are far from coincidental, and
style and uniqueness are extremely constrained by functionality, aerodynamics,
desirable features, comfort, and price. After all, why do almost all passenger cars
have four wheels, while most large trucks have more? Why are brake lights red?
Systems biology must eventually come to a point where we understand in detail why
particular genes are turned on and off in particular sequences, why cells and organ-
isms are structured in the way they are, and what is behind the details of the function
and operation of organs, tissues, cells, and intracellular components in particular
situations.
14.1 The Search for Structural Patterns

The investigation of design principles is an interesting line of research that began
several decades ago [1–4], but has taken off only recently (see, for example, [5–8]).
It consists of two parts. First, we need to discover designs that are used over and over
NATURAL DESIGN OF BIOLOGICAL SYSTEMS 401
again in biology, and second, we have to figure out why these designs are apparently
preferred over other imaginable designs.
In some cases, design patterns are easy to detect, especially if they are associated
with macroscopic features. Birds have wings, fish have fins, spiders have eight legs,
and most mammals have hair. In other cases, such discoveries are much more diffi-
cult. It took quite a bit of molecular biology to figure out the almost ubiquitous struc-
ture of lipid bilayers or the genetic code that seems to be universal among essentially
all earthly organisms. In yet other cases, the design patterns are so intricate that they
easily escape casual observation. As a specific example, we have discovered with
sophisticated computational analyses that genomes are not arbitrarily arranged in
gene sequences, but have much more structure. Bacterial genes are arranged in
operons, regulons, and über-operons [9], and it appears that operons involved in
multiple pathways tend to be kept in nearby genome locations [10] (Figure 14.1).
Species show their own codon bias, and bacterial genomes consist of almost unique
nucleotide compositions and combinations that resemble molecular fingerprints or
barcodes [11] (see Chapter 6). These types of patterns were only discovered with the
help of very large datasets and custom-tailored computer algorithms.
It has been known for some while that the expression of a gene is correlated with
the expression of other genes. But are there patterns? Microarray studies, taken at
several time points after some stimulus, suggest that groups of genes may be co-
expressed, while the expression of other genes is inversely correlated or correlated
with some time delay. In many cases, the clusters of co-expressed genes can be
explained with the coordinated function of the proteins for which they code, but in
other cases they remain a puzzle (see, for example, [12]).
Like gene sequences, metabolic and signaling pathways show commonalities
among different cells and organisms, some of which have been known for a long
time. Feedback inhibition by the end product is usually exerted on the first step of a
linear pathway, and one can show—with mathematical rigor—that this is indeed a
design that is superior to other modes of inhibition, such as inhibition of the last
step, or of several steps, in a reaction sequence (cf. [1, 13]). Signaling is often accom-
plished with cascades of proteins that are phosphorylated and dephosphorylated at
three levels (Chapter 9). Why three? We have recently learned to look for generic
patterns in the connectivity patterns of natural and artificial networks (see Chapter
3). For instance, most metabolites are associated with only a few reactions, and
relatively few metabolites are associated with many different production and
degradation steps [14]. This same pattern of only a few highly connected hubs within
a large network shows up time and again, in biology, the physical world, and society.
Just study the flight routes of a major airline!
Interestingly, but not really surprisingly, many networks we find in biology are
very robust. In a beautiful demonstration of this feature, Isalan and collaborators
F C
G H A B species I
E D
A B Figure 14.1 Conceptual representation
E F C D of operons and über-operons. Operons
are contiguous, co-transcribed, and co-
G H regulated sets of genes. While they are
usually poorly conserved during evolution,
E F C D B G H A species II the rearrangement of genes often maintains
A B C D individual genes in specific functional and
E F regulatory contexts, called über-operons.
The ancestral genome (left) is assumed to
G H A B C D G contain three clusters of similarly regulated
species III and functionally related genes, A–H, in
E F H 1
three separate genomic locations. During
evolution, the clusters are rearranged
A B C D G H into new clusters. Further genome
rearrangements may occur, but some
E F clusters and individual genes are retained in
A B C H the same neighborhoods. (From Lathe WC III,
species IV Snel B & Bork P. Trends Biochem. Sci. 25 [2000]
E F D G
randomly rewired the transcriptional network of E. coli and found that only about
5% of these scrambled networks showed growth defects [15]. A similar change in the
connectivity of gene transcription networks seems to have occurred naturally in the
evolution of yeast, and it could be that such flexibility is a general feature of gene
transcription systems. It is intriguing to ponder whether it might be possible to use
rewiring as a precisely targeted tool in synthetic biology [16].
Another important aspect of biological systems is that they employ similar
components time and again in a modular and hierarchical manner. This use of the
same motifs found within different networks and at different organizational levels
renders complex biological networks simpler than they might otherwise have been,
and indeed offers hope that we might eventually understand critical aspects of the
complexity of biological systems [17].
The identification and rigorous characterization of design principles requires
methods of formalizing complex systems and their organization. This task may be
accomplished in different ways. One school of thought has been focusing on network
graphs and their properties, while another has been looking for design principles in
fully regulated dynamic systems and the corresponding differential equation mod-
els. Interestingly, both approaches ultimately employ a similar strategy, namely
comparisons between the observed system structures and reasonable alternatives.
In the case of graphs, the properties of an observed network are compared with the
corresponding properties of a randomly connected network. For instance, one
might look for the number of nodes that have more links than one would statistically
expect in a random network (see Chapter 3). In the case of dynamic systems, an
observed design is compared with a very similar dynamic system structure that dif-
fers in just one particular feature of interest, such as a regulatory signal.
14.2 Network Motifs

The first step in an analysis of static network motifs is the abstraction of the network (A) (B)
as a graph. As Chapter 3 discussed in detail, a typical analysis of such a graph reveals
global network features, such as the small-world property, which is related to the – +
clustering coefficient, the degree distribution, and the average shortest pathway X X
length between any two nodes [18]. Beyond connectivity patterns, the search for
design principles in graph models focuses on network motifs, which are specific,
small subgraphs that are encountered more often than probability theory would Figure 14.2 Network motifs can be very
simple. Shown here are negative (A) and
predict in a randomly connected graph [5, 19].
positive (B) auto-regulation motifs. The
The simplest network motif, containing just one node, is auto-regulation, which regulation arrows may encompass several
may be negative or positive (Figure 14.2). It should be noted that this motif allows steps.
for more than a single node: the arrow feeding back may consist of a chain of pro-
cesses. For instance, negative auto-regulation is frequently found in transcription
factors that repress their own transcription. While the transcription factor is the only
node explicitly considered, the system in reality also contains at least one gene,
mRNA, ribosomes, and other molecular components. Negative auto-regulatory
loops are very stable and reduce fluctuations in inputs. However, if the feedback is
delayed, the system may also oscillate and eventually lose stability. Positive auto- S1 S2
regulation permits a cell to assume multiple internal states, and, in particular,
toggling between two stable states. We discussed one example in Chapter 9. Other
examples have been reported for the genetic switch between lytic and lysogenic
responses in the bacteriophage λ as well as in signaling cascades and cell cycle
phenomena (see, for example, [20, 21]). We discuss a toggle switch in more detail in
the case study in Section 14.9.
A slightly more complicated, quite prevalent network motif is the bi-fan, which
consists of two source nodes sending signals to two target nodes (Figure 14.3) [19].
This design is interesting, because it allows temporal regulation [22]. In the case of
T1 T2
signal transduction, it can also sort, filter, de-noise, and synchronize signals. The
interactions between the two signals arriving at a target may be of different types. In
the case of an and gate, the signal is only transduced to T1 (or T2) if both sources, S1
and S2, fire, while in the case of an or gate, a single source, S1 or S2, is sufficient. An Figure 14.3 The bi-fan is a very prevalent
or gate confers reliability onto the system, because one source is sufficient, while motif. In this type of small network, two
the and gate permits the detection of signal coincidence and can be employed to source nodes (S1 and S2) send signals to two
prevent a target from being turned on spuriously when only one signal is present. target nodes (T1 and T2).
It is not surprising that small motifs can be combined, both in nature and in the-
ory, to form larger regulatory structures. These larger structures can be implemented
in parallel or by stacking smaller motifs. They can also be complemented with
additional components and links, thereby creating a potentially huge repertoire of
possible responses. A well-known example of moderate complexity is the cross-talk
between signaling cascades, which can create sophisticated response patterns such
as in-band filters, where the system responds only to a signal within a strictly
bounded range, or an exclusive or function, where the system responds only if exactly
one of two sources is active, but not both [23].
While they are very important and instructive, small motifs often do not reveal
characteristic topological features of larger networks [24]. At the same time, search-
ing for all possible motifs of even moderate size within a large network is not a trivial
matter; in fact, it is currently not computationally feasible to identify all motifs of 10
or more nodes within realistically sized networks of a few thousand nodes and links.
A strategy for addressing the situation is to reduce the scope of the search.
For instance, Ma’ayan and collaborators successfully searched large networks exclu-
sively for cycles consisting of between 3 and 15 directional, nondirectional, or bidi-
rectional links. Particularly important among their findings was that biological and
engineered networks contain only small numbers of feedback loops of the types
shown in Figure 14.4. Among them, all-pass-through feedback loops that are coher-
ent in a sense that all arrows point in the same direction pose the risk of destabiliz-
ing the system, while anti-coherent cycles that contain sinks and sources are stable.
Lipshtat and collaborators showed that feedback loops in signal transduction sys-
tems can exhibit bistable behavior, oscillations, and delays [22]. Consistent with the
argument of instability in coherent feedback loops, they found that hubs in natural
networks are mostly either source or sink nodes, but not pass-through nodes. They
also observed that networks with fewer nodes involved in feedback loops are more
stable than corresponding networks with more nodes involved in feedback loops.
They concluded that the prudent placement of feedback loops might be a general
design principle of complex biological networks.
Similar to some of the feedback loops are feedforward loop motifs that contain
a direct and an indirect path (Figure 14.5). Each arrow may be activating or inhibi-
tory, which for the case of three nodes leads to eight combinations with different
response patterns [5]. Four of these are coherent, in the sense that the indirect path
overall sends the same activating or inhibitory signal as the direct path, while the
other four combinations are incoherent, because they contain opposite signals. As
an example, the incoherent motif in Figure 14.5C indicates inhibition along the
direct path, but double inhibition, and thus activation, along the indirect path. Both
coherent and incoherent patterns have been observed in natural systems, and the
motif with all positive signals (Figure 14.5A) is most prevalent among them.
An interesting example of a feedforward loop is a mechanism for sensing stress
signals in E. coli and many other bacteria [17]. When the bacterium senses the stress,
the feedforward loop triggers the production of flagellum proteins (Chapter 7),
which continues for about an hour, even if the signal disappears. This time period is
(A) (B)
X1 X1
X6 X2 X6 X2
Figure 14.4 Subtle differences in

network structure can have significant
consequences on the stability of a
X5 X3 X5 X3
network. The coherent all-pass-through
feedback loop in (A) poses the risk of
destabilization, while the anti-coherent
X4 X4 network in (B) is stable, with two source
nodes (X1 and X3), two sink nodes (X2 and X6),
and two pass-through nodes (X4 and X5).
(A) (B) (C) Figure 14.5 Examples of feedforward

loop motifs that contain a direct and an
X1 X1 X1 indirect path. The motifs in (A) and (B) are
coherent, because the signals arriving at X3
are both activating (A) or both inhibiting (B),
while the motif in (C) is incoherent: the direct
path signals inhibition, while the indirect
path signals double inhibition and thus
activation.
X2 X2 X2
X3 X3 X3
sufficient for the bacterium to assemble the flagellum and swim away. Like other
motifs, the feedforward loop in this case creates functional robustness by allowing
the system to operate continuously, even if it is exposed to random fluctuations.
14.3 Design Principles

The discovery of design principles in dynamic systems is based on comparisons
between similar systems, but it is slightly more complicated than in static graphs.
Michael Savageau initiated this discovery effort several decades ago [1, 4]. He asked
whether there were objective, quantifiable reasons for particular designs that had
been observed in nature many times. As a representative example for the way one
might think about design principles, we revisit an early analysis by Irvine and Sav-
ageau of a cascaded system involved in immune responses (Figure 14.6) [3, 25].
Here a source antigen Ag0 (= X0) describes a potentially infectious microorganism in
the environment and Ag (= X1) refers to microorganisms in the body that disrupt the
host’s physiological functioning. The variable E0 (= X2) represents immature effector
cells, while E (= X3) represents mature effector lymphocytes. S0 (= X4) and S (= X5) are
precursors and mature suppressor lymphocytes, respectively. In Irvine and
Savageau’s set-up, the variables X0, X2, and X4 are constant inputs to the system and
therefore independent; that is, they do not require their own differential equations
(see Chapter 2).
The question of primary interest was why suppressor lymphocytes inhibit the
maturation of effector cells, as suggested by experimental observations. Such a
question is very difficult to answer with wet experiments, because one cannot easily
eliminate this inhibition signal in a living organism without changing all kinds of
other features of the system. Irvine and Savageau therefore developed and used a
mathematical approach, called the method of mathematically controlled compari-
son (MMCC), which is particularly powerful for canonical models (see Chapter 4),
where each parameter has a clearly defined role and meaning. In this method, two
system models are analyzed side by side. One model corresponds to the observed
Figure 14.6 Simplified immune cascade.

An external source antigen Ag0 stimulates
an increase in the internal antigen level
Ag0 Ag Ag. Ag triggers the maturation of effector
lymphocytes from precursors (E0) to fully
functional cells (E). E in turn triggers the
maturation of precursor suppressor cells (S0)
E0 E into functional suppressor lymphocytes S.
In the observed system (Model A), S inhibits
the maturation of E0 (red arrow), while this
feedback is missing in the hypothetical
alternative system (Model B). See the text for
S0 S
further explanations.
design, while the other, alternative, model is different in the one feature of primary
interest, which in this case is the inhibition signal. The system is shown in
Figure 14.6.
According to the requirement of internal equivalence within MMCC, the param-
eters that reflect features common to both models receive the same values. The dif-
ference between the models is thus manifest in one distinguishing feature, which is
typically reflected in one or two terms of the model and its parameters. While many
analyses within MMCC can actually be performed symbolically [3], we use for our
demonstration a numerical system implementation of an S-system model (see [25]
and Chapter 4):
X 1 = 2.02 X 0 X 10.9 − 2 X 1 X 30.5 ,

X 3 = α X 1g 31 X 2g 32 X 5g 35 − X 3 , (14.1)
X 5 = X 0.5
3 X 4 − X 5.
Note that the parameters in the production term for X3 are not yet numerically spec-
ified. We do know that in the system with feedback (which we will refer to as Model
A), the kinetic order g35 must have a negative value (here g35 = −0.5), because it rep-
resents inhibition, while it is zero in the case without feedback (which we call Model
B). This numerical difference changes the value of the production term of X3 and
thereby the steady state of the entire system. To counteract this undesirable side
effect, MMCC suggests enforcing external equivalence, which means that the two
systems should appear to the outside observer as similar as possible. For instance,
adjusting in Model B the rate constant α, which has a value of 1 in Model A, could be
used to ensure that both systems have the same steady state. In the present case, we
have further degrees of freedom, because the distinguishing term also contains the
parameters g31 and g32, which in Model A have values of 0.5 and 1, respectively.
In order to achieve maximal external equivalence, it turns out to be sufficient to set
g31 in Model B to 0.4. Thus, for Model A, the production term for X3 is X 10.5 X 2 X 5−0.5 ,
while it is X 10.4 X 2 in Model B [25]. With these settings, both systems have not only the
same steady state but also the same expenditure in material and energy, which are
reflected in the same antigen and effector gains.
We now have two systems that to the uninitiated observer appear to be essen-
tially identical, but differ internally in their suppression mechanism, with which
suppressor lymphocytes inhibit the maturation of effector lymphocytes. What is the
relevance of this observed mechanism? MMCC answers this question by formulat-
ing objective criteria for desirable system responses. In general, such criteria for
functional effectiveness depend on the nature of the phenomenon and may include
the response time with which the system reacts to a perturbation; characteristic
measures of stability, sensitivity, and robustness; the accumulation of intermediate
components; or the extent to which the transients between stimulus and the return
to normalcy are affected by external perturbations. These criteria provide a set of
metrics against which the performance of the alternative models is measured.
In the given case of the immune response cascade, the criteria for functional
effectiveness were defined as follows: the base levels of systemic antigen and effector
lymphocytes should be as low as feasible; the antigen gain and the effector gain,
which are the increases in the steady-state levels of systemic antigen and effector
lymphocytes following a source antigen challenge, should be minimal; the peak
levels of systemic antigen and effector lymphocytes during the response should be
as low as possible; and the entire cascade should be relatively insensitive to
fluctuations in the environment.
The base levels and gains in Models A and B are identical, because they were
equated to maximize external equivalence, so they cannot be used to investigate the
superiority of either design. By contrast, the dynamic responses were not con-
strained and are easy to evaluate with simulations. As an example, one may initiate
the two systems at their common steady state (X1S = 1.033724, X3S = 1.013356,
X5S = 1.0066656) and raise the source antigen X0 from 1 to 4 at time t = 1. Indeed, the
two models respond to the “infection” with different transients in antigens and
effector lymphocytes (Figure 14.7). The model with suppression (Model A) is
deemed superior, because the antigen and effector concentrations show smaller
200 10 Figure 14.7 Comparison of time courses of

two designs of immune cascades. Dynamic
B responses of the observed system design
effector concentration
antigen concentration
B (Model A) and a hypothetical, alternative,

design (Model B), which lacks inhibition
A
100 5 of effector lymphocyte maturation, to a
A
fourfold source antigen challenge at time
t = 1.
0 0
0 6 12 0 6 12
time time
overshoots and the effector concentration responds more quickly. For other source
antigen levels, the numerical features are different, but the superiority of the
observed design (Model A) persists. Thus, lower peak levels are inherent functional
characteristics of suppression. Algebraic analysis furthermore indicates that Model
A is less sensitive to potentially harmful fluctuations in the environment [25]. Over-
all, the observed design has preferable features over the a priori also reasonable
design without suppression.
Since its inception, MMCC has been applied to a variety of systems, primarily in
the realm of metabolism and signaling, and variations have been proposed to study
entire design classes at once (for example, [13]), characterize design spaces [26, 27],
or establish superiority of a design in a statistical sense [28]. A particularly intriguing
example is Demand Theory [2, 29, 30], which rationalizes and predicts the mode of
gene regulation in bacteria, based on the microbes’ natural environments, and thus
reveals strong natural design principles (see Chapter 6). Another recent example is
the two-component sensing system in bacteria, which we have already discussed in
Chapter 9. A careful analysis of this system [31, 32] led to a deeper understanding of
its natural design and provided clues for how to manipulate a system with methods
of synthetic biology. For instance, Skerker and collaborators [33] explored the large
amount of available sequence data available for E. coli’s envZ–OmpR two-
component sensing system and identified individual amino acid residues that
always varied together as a pair. They speculated that these pairs were critical for
sensor specificity. Subsequently, they substituted the specificity-determining resi-
dues of one sensor with those of a different sensor, which indeed allowed the
sensor’s response to be triggered by the “wrong” signal.
14.4 Operating Principles

As a variation on the theme of design principles, one may also ask questions
regarding operating principles for systems with given designs. An example is the
controlled comparative analysis of different regulation patterns in a branched path-
way (Figure 14.8), where the goal might be to increase the yield v40 by altering the
ratios of fluxes flowing into different branches, v10/v12 and/or v23/v24 [34]. It turns out
? v10
–
X1
v01
X3
v12 v30
–
v23
Figure 14.8 Optimal operating strategies
? X2
depend on both structural and regulatory
v24 pathway features. Depending on the
existence of no, one, or two feedback
inhibition signals (indicated by question
marks), the optimal strategy for increasing
X4 the flux v40 by manipulation of the flux ratios
v40 v10/v12 and v23/v24 is different.
GOAL-ORIENTED MANIPULATIONS AND SYNThETIC DESIGN OF BIOLOGICAL SYSTEMS 407
that a variety of manipulations exhibit equal yields if the pathway is not regulated,
but that some manipulations are significantly more effective than others if some of
the intermediates exert feedback inhibition.
The discovery and analysis of design and operating principles is an ongoing
effort. A different aspect than we have discussed so far is the search for rules related
to spatial phenomena. For instance, if a protein is to be localized in the cell nucleus,
it often contains a high proportion of hydrophobic surface regions and a charge that
is complementary to that of the nuclear pore complexes, which control transport
between the nucleus and cytoplasm [35]. It may also be that some intracellular
gradients and heterogeneous concentration patterns are associated with the spatial
design issue. A special issue of Mathematical Biosciences [8] contains a collection of
papers representing the state of the art.
GOAL-ORIENTED MANIPULATIONS AND SYNThETIC

DESIGN OF BIOLOGICAL SYSTEMS
14.5 Metabolic Engineering
Between the discovery and analysis of natural design principles on the one hand
and the creation of novel biological systems from component parts (that is, syn-
thetic biology) on the other lies a large field of activities that attempt to manipulate,
and subsequently optimize, existing systems toward a desired goal. The most well-
developed line of research in this context is the optimization of metabolic pathway
systems in microbes, which falls under the umbrella of metabolic engineering.
There is no clear dividing line between manipulating existing systems and changing
them to such an extent that they can be considered new, and so both metabolic
engineering and synthetic biology flow into each other and can be seen as modern
endeavors in a long-term effort to alter and control biological systems that reaches
back about as far as recorded history and the first humans dabbling in agriculture.
After all, the breeding of plants and animals is a slow form of genetic and metabolic
engineering and thus of pushing biological systems toward a desired goal.
While excruciatingly slow by modern standards, the success of breeding efforts
over hundreds of years has been stunning, and one can only be amazed on compar-
ing modern corn or carrots with the original wild types from which they were
derived. The selection of the most desirable plant and animal species proceeded
throughout human history at a modest pace until the second half of the twentieth
century, when it started speeding up tremendously. This huge jump in the pace of
development and success occurred when we became knowledgeable enough to
explain the intuitive concepts of Mendelian genetics on the basis of a theoretical
foundation, and when we began to decipher the structure and function of genes,
and later the control of their expression.
Once the role of genes became clearer, effective strategies to generate higher
yield emerged. For instance, by exploiting nature’s own method of improving
design, the method of directed or adaptive evolution was invented to increase the
efficient production of proteins (see, for example, [36]). In its natural form, evolu-
tion is slow, but methods of bioinformatics overcame this problem, because it
became feasible with efficient algorithms to compare pertinent gene sequences
from many different organisms or strains that code for the same target protein. Such
a comparison shows which amino acids vary a lot and which are essentially unal-
tered among different candidate sequences. Considering that most mutations are
deleterious, the immutable amino acids are more likely to have an important role,
whereas other amino acids can probably be changed without much consequence.
By focusing on the important locations, it is possible to engineer bottleneck enzymes
in target pathways and to improve productivity significantly. A beautiful example
for this strategy is the alteration of an isoprenoid pathway in E. coli, where produc-
tivity increased roughly 1000-fold within a few years [37]. Another instance is the
optimization of a biosynthetic pathway by multiplex automated genome engineer-
ing, which allowed the creation of over four billion combinatorial genomic variants
per day and the isolation of variants with over fivefold increased yields within just
three days [38].
In relatively straightforward cases, the enzyme that is to be changed is clear.

However, this is not necessarily so in more complex systems. For instance, citric acid
production by microbes has been gradually improving for almost one hundred
years [39]. As a consequence, simple changes in one or two genes or enzymes are no
longer able to improve yield further, because they have already been detected during
the stepwise improvement of strains and media. It has therefore been proposed to
set up stoichiometric or dynamic models of pathway systems and to optimize them
toward a desired goal, while taking into account physiological constraints on metab-
olite concentrations, fluxes, and metabolic burden [40–42]. The last of these means
that it is costly for the organism to maintain mRNAs, proteins, and metabolic
products, and if these are expendable, the bacterium has an advantage if it gets rid
of them. Very many articles have been written about this approach, and we will not
pursue it further here, because its goals and concepts are intuitive, but the
techniques of analysis and implementation are rather difficult. The case study in
Section 14.7 addresses a related issue.
14.6 Synthetic Biology

While genetic and metabolic engineering have made great strides, industrially rele-
vant applications often take years to complete. Several research groups have there-
fore embarked on the development of synthetic biology as a potentially efficient strat-
egy for manipulating microorganisms. Current methods of synthetic biology primarily
manipulate genes and their expression in new ways, and the hope is that microorgan-
isms can be coerced into producing organic compounds that they have never encoun-
tered before and that they do not need. If the early successes are any indication, we
have every reason to expect that the field will explode over the next decades as we
develop faster and more finely tuned means of inserting foreign DNA into other
organisms, and even for moving DNA between prokaryotes and eukaryotes [43–45].
In the process, we will learn to master the control of the timing and extent of translat-
ing this DNA into RNA, proteins, and ultimately metabolic and signaling functions.
The main attraction of any gene-based strategy is that it allows rather precise
alterations in the activities of key enzymes, which themselves are specific and very
effective. Even complex metabolic conversions, which would require a long
sequence of steps in conventional synthetic chemistry, can sometimes be catalyzed
in a single enzymatic reaction [46]. Furthermore, it is becoming possible to combine
several enzymatic reactions, thereby eliminating the need for intermediate purifica-
tion steps. Finally, if microorganisms can be coaxed into producing complex organic
compounds, the products are usually enantiomerically pure and not, for example,
in racemic mixtures of l and d forms, which often result from chemical synthesis
and are very expensive to separate [47].
While many of the current methods might be outdated and replaced in a few
years, it is important to give a brief account of the methods and techniques of gene
manipulation that are currently used to improve the yield of desired organic
compounds. The targets are mostly microorganisms, and this focus is unlikely to
change, owing to the relative ease with which microorganisms can be cultured and
manipulated, even in very large batches. Good reviews include [16, 48–50].
The overall challenge of altering the metabolic responses of microorganisms
can be approached in two ways:
1. Increasing or rerouting an indigenous metabolic flux toward a desired goal.
2. Inserting into the organism entirely foreign metabolic reactions or even whole
pathways.
An example of the first approach is the increased production of lysine in Corynebac-
terium glutamicum, which does not secrete this compound in the wild type, but has
been modified so that new strains produce hundreds of thousands of tons per year
now (see, for example, [51]). An example of the second approach is Genentech’s
production of human insulin in the bacterium E. coli, which had never encountered
this protein in its evolutionary history and has no use for it [52].
Equipping microorganisms with new DNA requires two steps: the injection of
foreign genes into the organism, and the targeted and controlled expression of
GOAL-ORIENTED MANIPULATIONS AND SYNThETIC DESIGN OF BIOLOGICAL SYSTEMS 409
these genes. Adding foreign genes to a microorganism is typically accomplished

with plasmids, which are essentially naked strands of DNA (see Chapter 6). In
genetic engineering, they are often called vectors. At first glance, this operation is
not all that difficult, because bacteria take up plasmids quite easily, either directly
through the cell wall in a process called transformation, through viruses called bac-
teriophages, or through conjugation, a process in which two bacterial cells mate.
To make sure that all bacteria in the culture eventually contain the vector, the plas-
mid is constructed to contain not only the desired genes, but also genes that make
the bacterium resistant to some specific antibiotic. Once the transformation has
taken place, the culture is exposed to the antibiotic, and essentially all cells without
the plasmid die.
The complications with transformation in genetic engineering come from sev-
eral innate response mechanisms. In particular, many bacteria possess mechanisms
that lead to the expulsion of foreign DNA. One suspected reason for this elimination
is the bacterium’s attempt to minimize its metabolic burden, as discussed above,
because any increased burden may cause the growth rate to drop so much that the
bacterium is no longer competitive. One important step is therefore the removal of
innate resistance genes and thereby the creation of strains that retain foreign DNA
over sufficiently long time periods [53]. Another challenge is the transformation of
the microorganism in such a fashion that it does not require a large amount of
expensive supplements in the growth medium. Finally, to be commercially relevant,
the cultures must be robust and permit scaling from the bench to large industrial
bioreactors.
Once the DNA has been inserted into the bacterial host, it acts like the host’s own
genetic material. The challenge for metabolic engineers is now to turn on the desired
genes at the right times and to the right extents. In other words, one must be able to
affect the degree of transcriptional activity and regulation, which, among other
effectors, depend on the number and interactions of transcription factors and the
locations of their binding sites [16]. The main tools for this purpose are inducible
promoters, which are DNA regions that facilitate the transcription of specific genes
[49, 54] (see also Chapter 6). These promoters may be native, and if the foreign gene
is inserted into the right location, the promoter will turn it on. The promoter may
also be included in the vector and turned on or off through the presence or absence
of a chemical, physical, or physiological inducer. Examples of chemical inducers are
steroids and antibiotics such as tetracycline, physical inducers include light and
temperature, while a physiological inducer is starvation for a required nutrient such
as phosphate.
Unfortunately, most of the commonly used inducible promoters are binary: they
are turned either completely on or completely off, but cannot be turned on at 30%,
50%, or 80% levels. It seems that this binary outcome might in many cases be a mat-
ter of inducer transport into the cell. New developments are beginning to amelio-
rate this problem. For instance, it is becoming possible to uncouple the expression
of the transporter from the control by the inducer, use an inducer that can diffuse
into the cell without a transporter, or use a natural inducer that does not depend on
transport. It is also possible, at least in principle, to create constitutive promoters of
different strengths by manipulating their nucleotide sequences. Because of their
importance, efforts are steadily increasing to create extensive promoter libraries
(see, for example, [55]).
If entire metabolic pathways are to be added to the organism, it is not sufficient
just to turn a gene on at the right time—it is also necessary to coordinate the
expression of several genes [49]. The most natural approach to this task is to group
genes into operons under the control of the same promoter. If necessary, the rela-
tive expression of different genes in the same operon can be controlled via mRNA
stability or post-translationally. A further potential challenge is that intermediates
in the new pathway may be toxic to the microorganism, which means that it must
be possible to turn off production when it is not needed and especially during
growth.
The more we learn about the natural control of genes and pathways, the more
targets for manipulation we will discover for the advancement of genetic and
metabolic engineering. Barely known until rather recently, small interfering or
silencing RNAs, microRNAs and other noncoding RNAs (see Chapter 6), as well as
aptamers and riboswitches, promise a slew of new tools for controlling transcrip-
tion and translation [16] (see also Box 14.1), and inversion recombination
elements such as fim and hin are even able to rearrange DNA sequences within the
genome [56].
It is also becoming possible to interact with the function of cells at the protein
level, for instance, by specifically controlling their degradation with proteases from
E. coli [57]. Furthermore, controlling the intracellular scaffolds permits the co-
location of enzymes, which tends to increase flux, reduce leakage, and improve
specificity [58]. Ideally, such a control scaffold should be available on demand. Like
other methods at the protein level, it will provide a tool for spatiotemporal control of
in vivo processes.
As biological engineering techniques become more mature, it will be useful, and
indeed mandatory, to develop experimental protocols and standards, as well as to
share information and materials through registries of biological components [59]
and standardized formats called datasheets, which succinctly characterize the
features of building blocks for synthetic biology, such as sensors, actuators, and
other logical elements [60]. Important features to be documented include the pur-
pose of the device module, along with its operational contexts, interfaces, and
compatibility with other devices; quantitative measures of its functions, such as the
characteristics of its input–output transfer function; properties of its dynamic
behavior; and indicators of its reliability [61]. Standards are important because they
characterize components in sufficient detail to permit predictions for larger systems
composed of these components. Standards also allow abstracted representations,
such as we are familiar with in electric wiring diagrams.
Along with design standards and the creation of standardized modular parts
[59], future synthetic biology will use methods of computer-aided design (CAD) that
have been tremendously successful in various engineering disciplines for a long
time. For instance, platform-based design (PBD), a special case of CAD, has proven
very useful in the design of embedded electronic systems [62], and one might expect
that PBD methods will support the de novo creation of robust, reliable, and reusable
biological systems within given design constraints and will usher in procedures
for automatically optimized designs, given the specifications of the intended
system [61, 63].
BOX 14.1: RIBOSWITCHES
Riboswitches are gene regulatory elements that are found

as sections of 5′ untranslated regions of mRNAs. They are
broadly distributed across bacteria and, for instance, account
for the regulation of more than 2% of all genes in Bacillus
subtilis, underscoring their importance in the control of cellular
metabolism. The 5′ untranslated region of many mRNAs of genes
involved in the metabolism and transport of purines contains
a guanine-responsive riboswitch that directly binds guanine,
hypoxanthine, or xanthine to terminate transcription. Figure 1
shows the crystal structure of the purine-binding domain of
the guanine riboswitch from the xpt-pbuX operon of B. subtilis
bound to hypoxanthine, a prevalent metabolite in the bacterial
purine salvage pathway. The structure reveals a complex RNA
fold containing several phylogenetically conserved nucleotides
that create a binding pocket that almost completely envelops
the ligand. Hypoxanthine functions to stabilize this structure
and to promote the formation of a downstream transcriptional
terminator element, thereby providing a mechanism for directly
repressing gene expression in response to an increase in
intracellular concentrations of metabolite.
Figure 1 Structure of a riboswitch. The natural guanine-responsive
riboswitch in B. subtilis is shown here in complex with the metabolite
hypoxanthine (mRNA_metabolite_1u8d). (Courtesy of Ying Xu and
Huiling Chen, University of Georgia.)
CASE STUDIES OF SYNThETIC BIOLOGICAL SYSTEMS DESIGNS 411
CASE STUDIES OF SYNThETIC BIOLOGICAL SYSTEMS

DESIGNS
14.7 Elementary Mode Analysis in Metabolic Engineering
Many laboratories around the world have become involved in new strategies for
improving the yield of biofuels and, in particular, the creation of microbial strains
that convert non-edible plant biomass into ethanol, butanol, fatty acids, or hydrogen.
Friedrich Srienc and his laboratory set out to develop a high-efficiency E. coli strain
that would convert hexoses and pentoses into ethanol [64]. Instead of over-expressing
promising genes or knocking out genes that were assumed unimportant, as had
been done in the past, they used a computationally based, systematic top-down
approach that started with a wild-type strain and eventually led to a highly efficient
production strain. The guiding concept was that cells specialized for a purpose such
as ethanol production could otherwise afford minimized functionality when living
under controlled, favorable conditions. The goal was to create a strain that would,
under optimized conditions, grow, replicate, and produce economically interesting
amounts of ethanol, but not necessarily do much else. While obviously intriguing,
this goal was certainly not trivial to accomplish. For instance, how could one predict
which reaction steps are truly necessary for survival? The investigators approached
and solved the challenge with a combined computational and molecular strategy.
The first task was to identify redundancies within the metabolic pathway struc-
ture. These redundancies are beneficial for survival in uncertain environments, but
are presumably not needed under ideal laboratory conditions. In other words, if the
bacterium had two ways of producing the same metabolite, one pathway could
possibly be eliminated, thereby reducing the organism’s physiological machinery
and decreasing its metabolic burden. The tool for this identification of absolutely
necessary pathways was elementary mode analysis (EMA) [65]. Elementary modes
(EMs) are thermodynamically feasible pathways that form an essentially unique,
minimal set of irreversible reactions taking place at steady state. The set is minimal
in a sense that the removal of any reaction will disrupt the functioning of the entire
pathway system. Because of this feature, elementary modes can be seen as the
inherent building blocks of a metabolic pathway system. Expressed mathematically,
any steady-state flux distribution can be decomposed into a linear combination of
elementary modes, with, say, 20% EM1, 50% EM4, 30% EM5 and 0% of all other EMs.
The concept of EMA was introduced by Schuster and Hilgetag [66] as an alternative
to simple stoichiometric network analysis or flux balance analysis (see Chapter 3),
with the specific goal of extracting minimal sets of most relevant pathways from
complicated metabolic networks. Liao and co-workers [67] were the first to
demonstrate the applicability of the method by optimizing the biosynthesis of
3-deoxy-d-arabino-heptulosonate-7-phosphate (DAHP) in E. coli.
EMA begins with a stoichiometric system at steady state, which can be
represented as
NR = 0, (14.2)
where N is the stoichiometric matrix that describes which metabolite is involved in
which reaction and the vector R contains the fluxes between metabolite pools and
the outside of the system (see Chapters 3 and 8). Equation (14.2) usually has infi-
nitely many solutions, because realistic systems tend to have more fluxes than
metabolites. Within this space of solutions, an elementary mode is a subset of reac-
tions that satisfy (1) the steady-state equation in (14.2) with nonzero fluxes for
reactions within the subset and zero fluxes for reactions outside the subset; (2) the
necessary thermodynamic constraints of all reactions such that the signs of the
nonzero reactions are consistent with the directions of the arrows in the network
diagram; and (3) a nondecomposability constraint, which means that no smaller
subset of reactions in an elementary node can itself be an elementary mode.
The latter condition ensures that no elementary mode can be removed or validly
split into simpler modes without rendering some connections between two nodes
impossible.
The first step of an EMA consists of the construction of a network describing the
pathway system of interest and its stoichiometric matrix. In the second step, one
screens the reactions and their directionalities and identifies all nondecomposable
pathways within this network. Mathematically speaking, one determines all flux
vectors that contain a minimum of active reactions and otherwise contain as many
flux values of zero as possible, while still satisfying (14.2). Under different experi-
mental conditions, cells may use any combination of these elementary modes.
The next step is EM selection, where one retains only those pathways that connect a
desired input with a desired output. Other pathways are subject to deletion, for
instance, through knockouts, which tends to increase the efficiency of the organism
with respect to the desired pathway. In an optional final step, the chosen pathway
may be further subjected to targeted evolution. It is easy to imagine that a realisti-
cally sized pathway system contains very many elementary modes. Fortunately,
algorithms have been developed to automate this analysis [68–71].
As an illustration, consider the metabolic network shown in Figure 14.9A. It is
adapted from [65] and is characterized by the following stoichiometric matrix:
r1 r2 r3 r4 r5 r6r r7 r8r r9
A 1 −1 0 0 −1 0 0 0 0
B 0 0 0 0 1 −1 −1 −1 0
C 0 1 −1 0 0 1 0 0 0
D 0 0 1 −1 0 0 2 0 0
E 0 0 1 0 0 0 0 0 −1
Two comments are in order here. First, reaction r7, entering node D, has a coefficient
of 2, but a coefficient of −1 with respect to B. The reason is as follows: Each molecule
of type B can be converted into C, which then splits into two molecules of types D
and E. Naively speaking, B and C molecules are twice as big as D or E molecules. At
the same time, r7 converts a B molecule directly into type D. To be compatible with
the path through C, each B molecule must result in two D molecules. Second, two
reactions are shown as reversible. For simplicity, the stoichiometric matrix only
shows the dominant reaction. For r6r, the main direction is from B to C, while the
minus sign associated with r8r in row B indicates a flux out of the system. Nonethe-
less, r8r may also be an influx, in which case the corresponding flux value is set to −1,
and r6r may be used from C to B, again by setting the corresponding flux value to −1.
Either by hand, or using an algorithm like those in [68, 69] or [70, 71], one now
enumerates all nondecomposable connections between an input and an output.
For example, starting at Ae, a mode may run through r1, r2, and r6r (in reverse) and
leave the system through r8r toward Be. The corresponding mode (flux vector) is
(1 1 0 0 0 −1 0 1 0)T
(where T indicates the matrix transpose); it is shown as EM1 in Figure 14.9B. As a

second example, consider EM2, which starts at Be, and runs through r8r and r6r to C,
where r3 splits toward D and E, from where material leaves the system through r4
toward De and r9 toward Ee.
Analysis of this type shows that the network contains eight elementary modes,
which are shown in Figure 14.9B. If the goal is to convert Ae into De, some modes are
not useful and should be directly eliminated from further consideration. In this
case, EM1 and EM3 are excluded, because they produce Be instead of De. Moreover,
EM2 and EM5 are not pursued further, because they use Be as substrate. The most
efficient pathways are EM6 and EM7. EM4 and EM8 are also admissible, but are
suboptimal because they yield Ee in addition to De and thereby convert valuable
substrate into an undesired by-product.
In the actual case of ethanol production, the investigators constructed and
experimentally validated a comprehensive model of the intermediary metabolism
of E. coli, which included reactions involved in the dynamics of pentoses and hex-
oses, including glucose, galactose, mannose, xylose, and arabinose. Wild-type
E. coli does not possess the enzyme pyruvate decarboxylase, which converts
(A)
Be
r8r
r5 r7
r1 r6r r4
Ae A D De
r2 r3
r9
E Ee
(B)
EM1 EM2 EM3 EM4
EM5 EM6 EM7 EM8
Figure 14.9 Principles of elementary mode analysis (EMA). (A) Illustrative example of a pathway system with four external pools (red, subscript e), five
internal metabolite pools (green), and several irreversible and two reversible reactions (r6 and r8). (B) The eight elementary modes are shown in purple.
The most efficient pathways for producing De from Ae are EM6 and EM7. EM4 and EM8 are suboptimal, because they also yield Ee. (Adapted from Trinh CT,
Wlaschin A & Srienc F. Appl. Microbiol. Biotechnol. 81 [2009] 813–826. With permission from Springer Science and Business Media.)
pyruvate to acetaldehyde, from which ethanol is produced, and the enzyme was
therefore provided through a plasmid. This process generated an ethanol produc-
tion system similar to that in the bacterium Zymomonas mobilis and the yeast
Saccharomyces cerevisiae.
The next step was the calculation of all elementary modes (EMs) with a computer
algorithm that analyzed all reactions and their directionalities within this network
and identified all unique pathways [65]. While the details are too complicated for
this case description, the results are in principle the same as those shown in
Figure 14.9. In the actual case of ethanol production in E. coli, the investigators
identified over 15,000 EMs with which the bacterium can metabolize xylose and
arabinose, the most prevalent pentoses found in biomass.
The next step was the selection of desirable EMs and the elimination of inefficient
EMs. Among the 15,000 EMs identified, about 1000 consumed xylose and arabinose
under the desired anaerobic conditions. Amazingly, the deletion of just seven
reactions reduced this number to only 12 ethanol- and biomass-generating EMs.

The seven reactions for disruption were identified by application of basic EMA rules.
Namely, the consequences of eliminating a reaction were studied with respect to
the remaining EMs. Then, the maximal yields in biomass and ethanol were assessed
for the remaining pathways. Finally, the reaction with the least number of EMs for
ethanol and biomass production was eliminated. The 12 EMs were further reduced
to six, because the other six EMs made use of the pyruvate dehydrogenase complex,
which is not active under anaerobic conditions. Only two of the remaining six EMs
co-produced a sufficient amount of ethanol and biomass.
The investigators used the same rationale and methods for ethanol production
based on hexose consumption. Furthermore, it was possible to study the simultane-
ous consumption of both pentoses and hexoses. In this case, almost 100,000 EMs
were identified, of which only 18 remained for anaerobic conditions. Of these, 12
utilized either glucose or xylose individually, and six consumed both. The set of EMs
using hexoses and pentoses could be further reduced by blocking glucose transport
into the cell or glucose phosphorylation within the cell.
Intriguingly, all inferences above were based on purely theoretical consider-
ations. However, they were subsequently validated by the actual genetic engineer-
ing of a deletion strain as suggested by the computational analysis, and the
agreement with the prediction was excellent, with a yield of about 0.4–0.5 gram of
ethanol per gram of sugars.
14.8 Drug Development

The laboratory of Jay Keasling at Berkeley has been at the forefront of synthetic biol-
ogy since the inception of the field. One of the hallmark projects of the laboratory
has been the production of an antimalarial drug, artemisinin, at a cost that allows
developing countries to buy the drug for their malaria patients [72, 73]. Malaria has
been a persistent problem for a long time. Hippocrates already described the disease
about 2500 years ago, and yet even today nearly 500 million people in the tropics
and subtropics become infected every year and nearly three million patients die [74].
Artemisinin is a natural secondary metabolite of the Chinese sweet wormwood
Artemisia annua and has been used in Chinese treatments of malaria for a long
time. It kills Plasmodium falciparum, one of the predominant malaria parasites, by
releasing large doses of reactive oxygen species inside a patient’s red blood cells.
Annually, about 100 million treatment doses are being sold, but harvesting the drug
from A. annua is not feasible at a cost that developing countries can afford [75].
Even in microbes, the total synthesis of artemisinin would be expensive and diffi-
cult, and the Keasling laboratory therefore set out to re-engineer the baker’s yeast
Saccharomyces cerevisiae to produce the immediate precursor of artemisinin, arte-
misinic acid, in sufficient quantities and at an acceptable price.
The biochemical context for this precursor is given by terpenoids, sesquiterpe-
noids, and isoprenoids. Sequencing of A. annua clones indicated that enzymes
from the mevalonate and farnesyl pyrophosphate (FPP) pathways are involved in
artemisinin biosynthesis [76]. FPP is found in many fungi and other organisms and
can be converted into about 300 different structures. One of these, amorpha-4,11-
diene, is the precursor of artemisinin.
Thus, the task became to re-engineer yeast toward three goals (Figure 14.10).
First, it was necessary to increase the amount of FPP. This was accomplished by
enhancing the activities of enzymes and thereby increasing the flux into the
endogenous mevalonate pathway. The initial substrate of this pathway is acetyl-CoA,
which is readily produced from simple sugars. The second goal was to introduce a
gene coding for amorphadiene synthase, which converts FPP into amorpha-4,11-
diene. Finally, it was necessary to recreate in yeast the three-step conversion of
amorpha-4,11-diene to artemisinic acid, which in A. annua is accomplished with a
cytochrome P450 mono-oxygenase and the corresponding redox partner NADPH:
cytochrome P450 oxidoreductase. As mentioned earlier, FPP is the starting point for
many other products, and in order to channel FPP toward amorpha-4,11-diene,
it was furthermore necessary to down-regulate the initial step of ergosterol biosyn-
thesis from FPP, which was accomplished through reduced transcription of squalene
synthetase, which converts FPP into squalene. The down-regulation was achieved
P450
ADS
A-4, 11-D artemisinic acid
A-CoA FPP
squalene ergosterol
Figure 14.10 Simplified representation of the steps toward generating artemisinic acid in yeast. The task involves enhancing the flux of the
mevalonate pathway from acetyl-CoA (A-CoA) to farnesyl pyrophosphate (FPP) (green arrow), down-regulating the flux from FPP toward squalene and
ergosterol (red arrow), and introducing two new enzyme systems (pale orange): amorphadiene synthase (ADS), which converts FPP into amorpha-4,11-
diene (A-4,11-D), and the complex (P450) of a specific cytochrome P450 mono-oxygenase and its NADPH-oxidoreductase partner.
through reduction of the expression of the corresponding gene ERG9 by placing it

under the control of the methionine-repressible PMET3 promoter and including a
gain-of-function mutation in the gene UCP2, which regulates sterol production.
It turned out that yeast transports the synthesized artemisinic acid out of the cell,
which made harvesting and purification relatively easy.
With concerted efforts, the engineered yeast eventually became capable of
producing artemisinic acid at a significantly higher specific productivity than
A. annua. In fact, the yield of amorpha-4,11-diene was increased roughly 10,000
times, and it seems feasible to improve yield by another order of magnitude or more.
It remains necessary to scale the processes up to an industrial scale and to optimize
yield further, in order to reduce the cost of artemisinin sufficiently. Unfortunately,
recent reports indicate that the Plasmodium parasite is already beginning to show
resistance, for instance in Cambodia [77]. However, it is hoped that anti-malarial
drug cocktails, as recommended by the World Health Organization, rather than
monotherapy, will delay or even prevent resistance and permit the effective use of
artemisinin in the future.
Similar to the development of drugs against diseases like malaria, geneticists
have been using synthetic biology to create transgenic foods with added benefits.
Specifically, almost 50% of the world population suffers from vitamin deficiency.
As is to be expected, this problem occurs mostly in developing countries, where the
typical cereal-based diet does not change from day to day. An early and very famous
example of a transgenic crop was Golden Rice, a rice variety that was genetically
engineered by the Swiss research group of Ingo Potrykus to synthesize β-carotene,
which is a precursor of vitamin A [78]. The name of this transgenic strain reflects its
golden color, which is due to the synthesized carotene. Just 5 years later, its successor,
Golden Rice 2, generated over 20 times the amount of the original transgenic crop
[79]. In a similar vein, the group of Paul Christou set out to create transgenic staple
crops with increased vitamin content. They have succeeded in altering a South
African corn variety containing 169 times the normal amount of β-carotene, six
times the normal amount of vitamin C, and twice the normal amount of the
B vitamin folate [80]. The Golden Rice Project [81] is dedicated to research efforts of
this type.
14.9 Gene Circuits

Arguably the best-studied systems in synthetic biology today are artificial gene
circuits [82]. These circuits are created by inserting foreign genes into microorgan-
isms and changing their transcription and/or translation in a controlled fashion.
One of the first working examples was a genetic toggle switch, which was con-
structed such that the circuit permitted bistability and memory [83]. Specifically,
the system consisted of two genes that each coded for a transcriptional repressor of
the other (Figure 14.11). An external inducer allowed the inhibition of one repres-
sor, thereby moving the system to a stable state where one gene was being tran-
scribed and the other repressed. The stability of this state also led to hysteresis
(see Chapter 9), where the system remained in the particular on–off state without
Figure 14.11 Schematic representation

of a toggle switch [76]. Synthetic biology
has advanced far enough to create gene
promoter A repressor A circuits that exhibit responses as predicted
by mathematical models.
inducer A
inducer B
promoter B repressor B GFP
further exposure to the external inducer, until another perturbation forced the
system into the second steady state.
Gardner and collaborators [83] represented the experimental system with a
simple mathematical model. As we discussed in Chapter 9, the creation of bistability
usually involves a sigmoidal production function that intersects with a linear degra-
dation process. Gardner and co-workers did indeed use this set-up. Inspired by
generic mathematical representations for gene expression processes, they defined
dimensionless equations for the toggle switch:
du α
= 1 − u,
dt 1 + v β
(14.3)
dv α2
= − v,
dt 1 + u γ
where u and v are the concentrations of repressors 1 and 2, respectively, α1 and α2

are their effective rates of synthesis, and β and γ are cooperativity parameters with
respect to promoters 2 and 1, respectively. While we could easily simulate the equa-
tions with a numerical solver, the most interesting aspect of the system is whether it
might possess bistability. Using methods we discussed in Chapters 2, 4, 9, and 10,
we can determine the stable and unstable steady states by setting the right-hand
sides of both equations in (14.3) equal to zero, which leads to two relationships
between u and v:
α1
u= (14.4)
1+ vβ
and
1/γ
α 
u =  2 − 1 . (14.5)
 v 
These two nullclines describe all points for which the time derivative of u or v,
respectively, is zero. Points where these two functions intersect are therefore steady
states. The number of such intersection points depends on the parameter values.
For instance, the two functions intersect three times in Figure 14.12A, and further
(A) (B)
8 8
u= 6 u= 6
1 + v3 1 + v2
6 6 Figure 14.12 Slight alterations in
( ) ( )
3 1/3 3 1/3
u = – –1 u = – –1 numerical features can lead to different
v v
qualitative system behaviors. The
u 4 u 4
parameter values in (14.3)–(14.5) determine
whether the system exhibits bistability (A)
2 2 or not (B). The values here differ in only one
of the cooperativity coefficients, namely, β,
which is equal to 3 and 2, respectively, in (A)
0 0
0 1 2 3 4 0 1 2 3 4 and (B). Other parameter values are α1 = 6,
v v α2 = 3, and γ = 3.
analysis shows that the two points shown in orange are stable steady states, while
the brown point is unstable. By contrast, the system in Figure 14.12B has only one
stable point.
At the same time as Gardner and colleagues proposed the toggle switch, Elowitz
and Leibler [84] created in E. coli a low-copy-number plasmid encoding an
oscillating system, called the repressilator. This system contained a compatible
higher-copy-number reporter plasmid with a tetracycline-repressible promoter
fused to a gene coding for green fluorescent protein (GFP). The system was stimu-
lated with a transient pulse of IPTG, which is known to interfere with the repression
by LacI.
The repressilator circuit thus contained three gene modules. The first gene pro-
duced the repressor protein LacI, which inhibits the transcription of the second
repressor gene, tetR, which was obtained from a tetracycline-resistance transposon
and whose product inhibits the expression of the λ-phage gene cI. The product, CI,
inhibits lacI expression, thereby creating a cyclical system (Figure 14.13). The oscil-
lations were visible in single cells under a fluorescence microscope. They were sus-
tained for a few cycles, but not synchronized, and varied quite a bit in frequency and
amplitude, even between sibling cells.
Combining properties of the toggle switch and the original oscillator, Atkinson
and collaborators [85] improved on the simple repressilator with a more complex
circuit design that led to predictable oscillation patters that were robust to perturba-
tions and even synchronized in large populations. The design contained an activa-
tor and a repressor module (Figure 14.14). The activator module consisted of a
LacI-repressible glnA promoter, which controlled the expression of the activator
NRI. Because phosphorylated NRI in turn activates the glnA promoter, this module
is auto-activated. The repressor module consisted of a phosphorylated NRI-
activated glnK promoter fused to the gene coding for LacI expression. Thus, the out-
put of the activator module was connected to the input of the repressor module. The
researchers demonstrated that this circuitry reproducibly generated synchronous
damped oscillations. More importantly, the behavior of the oscillator was consistent
with sophisticated theoretical predictions. For instance, as predicted by the model
analysis, the oscillation changed to a toggle switch when the enhancer-binding site
of the repressor system was removed and the production of LacI thereby became
constitutive.
At a higher functional level, stable oscillations are at the core of circadian
rhythms. While these have been studied for a long time, both experimentally and
with mathematical modeling, synthetic biology offers new avenues for investigating
these intriguing systems [86, 87].
The sky is the limit when it comes to creating novel gene circuits, and one should
expect a steady increase in the number of ingenious inventions. For instance, James
stimulus
promoter A repressor A
promoter B repressor B B GFP

of the repressilator. This interesting gene
circuit consists of three genes that inhibit
each other’s expression in a circular manner.
promoter C repressor C The result is the ability to exhibit oscillatory
expression patterns [77].

of a gene circuit that exhibits sustained,
NRI NRI~P synchronized oscillations. Lac-O is a LacI
operator site, and E-B-S A and E-B-S B are
Lac-O E-B-S A promoter A Lac-O enhancer-binding sites. (Adapted from
Atkinson MR, Savageau MA, Myers JT & Ninfa
AJ. Cell 113 [2003] 597–607. With permission
from Elsevier.)
E-B-S B promoter B
LacI
Collins’s group created an E. coli strain that can count [88]. Christopher Voigt’s and
Andrew Ellington’s laboratories engineered bacteria that became able to see light,
detect edges between light and dark, and essentially act like a photographic film
[89]. The genetic constructs consisted of a combination of a chimeric light-sensitive
protein with components from the freshwater cyanobacterium (blue–green alga)
Synechocystis and the bacterium E. coli, the LuxI biosynthetic enzyme from the
marine bacterium Vibrio fischeri, which produces a cell–cell communication signal,
a transcriptional repressor from bacteriophage λ, and a Lac system producing black
pigment. The implementation of the edge-detection algorithm consisted of three
plasmid backbones carrying about 10,000 DNA base pairs. A good how-to review of
strategies for engineering bacterial sensor systems and synthetic gene circuits can
be found in [90].
In a different application, Voigt’s laboratory developed a method for designing
synthetic ribosome binding sites, which allowed them to control protein expression
in a rational, predictable, and accurate fashion [91]. This type of control could have
enormous implications for connecting gene circuits and for the targeted production
of organic compounds in metabolic engineering.
In nature, bacteria communicate with each other and often exhibit coordinated,
population-based behaviors through a process called quorum sensing (QS)
(see Chapter 9). In this process, bacteria produce and release species-specific or
universal chemical signaling molecules. Other bacteria of the same or another spe-
cies sense these molecules and correspondingly regulate their gene expression in a
manner that depends on the amount of chemical signal and thus on the cell density
in the environment. The result is a population response that occurs only when suf-
ficiently many bacteria are on board. Natural examples of QS can be seen in the
formation of biofilms, which exist everywhere from human teeth to sewage pipes
(see Chapter 15), in bacterial pathogenicity, and in virulence. It seems possible, at
least in principle, to use a detailed understanding of the QS process as a weapon
against some of these undesired bacterial phenomena [92]. It also seems that QS
could have great potential for the creation of novel microbial systems that react
collectively as a population to stimuli of interest. For instance, bacterial population
sensors might be used for the detection of environmental chemicals or pathogens or
for the creation of new materials [93].
Many species of bacteria employ only one QS circuit, while others use two or
more circuits for communication. As we learn more about the details of these
circuits, more targets for possible disruption at the genetic or signaling level
emerge. The reward for such interruption could be significant, because biofilms
are often undesirable and many pathogenic species develop drug resistance by
means of QS processes. A promising approach to developing interruption strate-
gies and directing multicellular functioning is the re-creation of computational
and molecular QS systems with methods of synthetic biology [93, 94]. One way to
accomplish this task is by rewiring of the input–output relationships of a signal-
ing pathway. For instance, exchanging the sensors for different signaling path-
ways permits targeted engineering of the input to a microbial system, while alter-
ing the molecule carrying the signal from the receptor to downstream effectors
affects its output [16].
ExERCISES 419
ThE FUTURE hAS BEGUN

Understanding and reliably altering the design of biological systems is in some sense
the ultimate goal of systems, synthetic, and applied biology. While there is much value
in modeling specific systems for specific purposes, discovering design principles
brings us closer to a rational approach to—and ultimately a theory of—biology (see
Chapter 15). It will also reveal new avenues for using biology in medical treatment,
metabolic engineering, the discovery of toxins in the environment, and probably a
slew of other activities that we cannot even foresee at this point. Working on the design
of biological systems requires multipronged approaches of molecular biology, engi-
neering, and modeling. Early successes have demonstrated the utility of linking
genome information with metabolic and signaling pathways and with cell physiology
and systems-based models. In metabolic engineering, the traditional trial-and-error
approaches are being successfully supplanted with rational methods for altering
existing genomic and metabolic designs and with the creation of new pathways and
their control. On a small scale, we have developed the base capabilities to design func-
tional circuits, tested them with mathematical models, and implemented them in
actual microbial systems [50]. In the present first stage of synthetic biology, we are
primarily mimicking natural systems. Soon, we will be able to create novel, function-
ing systems from basic components [16, 59]. Already, we have at our disposal a variety
of means for controlling microbial function at the genomic level [48, 49], and we are
beginning to hone our capabilities of fine-tuning expression at the RNA and protein
levels [91, 95]. What would have seemed to be a fantasy only a few decades ago has
become reality. But it is only the beginning. There is no doubt that we have before us a
long and exciting journey whose destination we cannot even imagine.
ExERCISES
14.1. Discuss methods and challenges of introducing represent positive effects. Choose simple parameter
foreign genes into an organism and having them values and perform simulations of the behaviors of
expressed. the two loops. In particular, investigate the stability
14.2. Yao and co-workers [96] studied the restriction of the two loops.
point within the mammalian cell cycle, where the 14.5. Set up differential equations for the two loop
cell either commits to entering S-phase or remains structures in Figure 14.4, where some of the arrows
in G1. To analyze this bistable behavior, they represent positive effects and the remaining arrows
constructed a simplified model, whose core represent negative (inhibitory) effects. Choose
consisted of the self-activation of the E2F simple parameter values and perform simulations
transcription factor, which was inhibited by the of the behaviors of the two loops. Pay special
retinoblastoma tumor suppressor protein (Rb). attention to the stability of the loops. Document
Read the original article and implement the your findings in a report.
differential equation system that is given in the 14.6. Draw diagrams of all three-node feedforward loop
article. Analyze which components are essential motifs that contain a direct and an indirect path.
for bistability. Summarize your findings in a Report whether they are coherent or not.
report. Note that the equations in the article
contain several typos. That’s life, and a good 14.7. Perform other simulations with the immune
opportunity to learn troubleshooting. Compare suppression model and test whether Model A is still
the equations with the table of processes and superior to Model B.
the diagram of the system to detect and 14.8. Implement the system in Figure 14.8 as a power-law
reconcile inconsistencies. system and test the effect of different flux split
14.3. Add to the diagram in Figure 14.3 inputs that feed ratios (v10/v12 and/or v23/v24) on the output v40.
S1 and S2. Then write ordinary differential equations Study the influence of the two potential feedback
for the bi-fan motif and test different combinations signals (question marks in the figure) on the results.
of constant inputs as well as positive and negative 14.9. Search the literature for information on metabolic
signals between the source and target nodes. Study burden. What exactly is it, why is it important in
and and or gates when combining the two signals. metabolic engineering, and how can one
Is the structure of the model description unique for circumvent problems associated with it?
a given combination of positive and negative 14.10. Determine the elementary modes in the simple
signals? reaction system in Figure 14.15A, either by hand or
14.4. Set up differential equations for the two loop with software like that in [68–71]. Assume that
structures in Figure 14.4, assuming that all arrows the nodes inside the shaded area are internal and
(A) (B)
r1 r2 r3 r1 r2 r3
Ae A B Be Ae A B Be
r4 r5 r4 r5
C C
r6 r6
Ce Ce
(C)
r1 r2 r3
Ae A B Be
r4 r5
C
r6
Ce
Figure 14.15 Generic metabolic reaction system. In (A) the pathway contains two inputs and two outputs. (B) and (C) represent slight variations.
(Adapted from Palsson BØ. Systems Biology: Properties of Reconstructed Networks. Cambridge University Press, 2006. With permission from Cambridge
University Press.)
that the dominant direction for r1 is toward A. Formulate the stoichiometric matrix. Determine
Define the stoichiometric matrix. What is/are the elementary modes of the network, either by
the optimal elementary mode(s) if the goal is hand or with the software in [68–71].
optimal production of Be? 14.14. The text states that elementary modes are
14.11. Formulate the stoichiometric matrix of the simple pathways that form an essentially unique, minimal
reaction network in Figure 14.15B. Determine the set of irreversible reactions taking place at steady
elementary modes of the network. state. The inclusion of the word “essentially”
14.12. Formulate the stoichiometric matrix of the simple means that the sets are unique except for simple
reaction network in Figure 14.15C. Determine the scaling (that is, multiplication by the same
elementary modes of the network. constant). Argue why this type of scalability is
always present in stoichiometric, flux balance, and
14.13. Identify the abbreviations in the simplified yeast
elementary mode analyses. Is the same true for
fermentation pathway in Figure 14.16. Assume
dynamic simulations?
that reactions outside the shaded area are external.
14.15. Explore examples of drug development through
metabolic engineering and synthetic biology.
Tre For instance, investigate the engineered
r10
GAP Pyr Ace EtOH production of the cancer drug paclitaxel or of
r2 r8 r9
r1 r5 r11 polyketides, a family of compounds containing
Glu G6P FBP r6
r3 r4 antibiotics like erythromycin. Describe your
DHAP
r7
Gly Suc findings in a report.
Glc 14.16. Explore the specific techniques of metabolic
engineering and synthetic biology that were used in
Figure 14.16 Simplified diagram of the yeast fermentation system.
the creation of Golden Rice, which has helped
The shaded area is considered inside the system, while EtOH, Glu, Glc,
Gly, Suc, and Tre, are external. The dashed arrow indicates a sequence of
address vitamin A deficiency in developing
several reactions that are not explicitly shown. (Adapted from Schwartz countries whose people rely on rice as
J-M & Kanehisa M. BMC Bioinformatics 7 [2006] 186. With permission from the primary staple. Describe your findings in a
Bio Med Central Ltd.) report.
REFERENCES 421
14.17. Implement the system of equations (14.3)–(14.5) in 14.21. Analyze what happens when both production
a numerical solver and demonstrate bistability or terms and degradation terms in the differential
lack thereof with simulations. equations in (14.3) are interchanged. Begin by
14.18. Review what it means when a system shows using the parameter values in Figure 14.12.
hysteresis. Demonstrate hysteresis or the lack 14.22. Implement the repressilator of Elowitz and Leibler
thereof in the system (14.3)–(14.5). [84] and perform simulations.
14.19. Change the parameters in (14.3)–(14.5) such that 14.23. Study the bistability and hysteresis in the
the system has only one steady state for low u and repressilator of Elowitz and Leibler [84]. Describe
high v. your findings in a report.
14.20. Is it possible to change the parameters in (14.3)– 14.24. Search the literature for a synthetic biological
(14.5) such that the system has no steady state? system not discussed in the text. Describe in a
If your answer is “yes,” show an example. If it is report the molecular and computational
“no,” explain why it is impossible. techniques that were used.
REFERENCES
[1] Savageau MA. Biochemical Systems Analysis: A [18] Barabási A-L. Linked: The New Science of Networks.
Study of Function and Design in Molecular Biology. Perseus, 2002.
Addison-Wesley, 1976. [19] Milo R, Shen-Orr S, Itzkovitz S, et al. Network motifs: simple
[2] Savageau MA. Design of molecular control mechanisms and building blocks of complex networks. Science 298 (2002)
the demand for gene expression. Proc. Natl Acad. Sci. USA 824–827.
74 (1977) 5647–5651. [20] Ptashne M. A Genetic Switch: Phage Lambda Revisited, 3rd ed.
[3] Irvine DH & Savageau MA. Network regulation of the Cold Spring Harbor Laboratory Press, 2004.
immune response: alternative control points for suppressor [21] Weitz JS, Mileyko Y, Joh RI & Voit EO. Collective decision
modulation of effector lymphocytes. J. Immunol. 134 (1985) making in bacterial viruses. Biophys. J. 95 (2008)
2100–2116. 2673–2680.
[4] Savageau MA. A theory of alternative designs for [22] Lipshtat A, Purushothaman SP, Iyengar R & Ma’ayan A.
biochemical control systems. Biomed. Biochim. Acta 44 (1985) Functions of bifans in context of multiple regulatory motifs in
875–880. signaling networks. Biophys. J. 94 (2008) 2566–2579.
[5] Alon U. An Introduction to Systems Biology: Design Principles [23] Schwacke JH & Voit EO. The potential for signal integration
of Biological Circuits. Chapman & Hall/CRC Press, 2006. and processing in interacting MAP kinase cascades. J. Theor.
[6] Ma’ayan A, Blitzer RD & Iyengar R. Toward predictive models Biol. 246 (2007) 604–620.
of mammalian cells. Annu. Rev. Biophys. Biomol. Struct. 34 [24] Ma’ayan A, Cecchi GA, Wagner J, et al. Ordered cyclic
(2005) 319–349. motifs contribute to dynamic stability in biological
[7] Shen-Orr SS, Milo R, Mangan S & Alon U. Network motifs in and engineered networks. Proc. Nat. Acad. Sci.
the transcriptional regulation network of Escherichia coli. Nat. USA 105 (2008) 19235–19249.
Genet. 31 (2002) 64–68. [25] Irvine DH. The method of controlled mathematical
[8] Alves R & Sorribas A (eds). Special issue on biological design comparison. In Canonical Nonlinear Modeling (EO Voit, ed.),
principles. Math. Biosci. 231 (2011) 1–98. pp 90–109. Van Nostrand Reinhold, 1991.
[9] Che D, Li G, Mao F, et al. Detecting uber-operons in [26] Savageau MA, Coelho PM, Fasani RA, et al. Phenotypes and
prokaryotic genomes. Nucleic Acids Res. 34 (2006) 2418–2427. tolerances in the design space of biochemical systems.
[10] Yin Y, Zhang H, Olman V & Xu Y. Genomic arrangement of Proc. Natl Acad. Sci. USA 106 (2009) 6435–6440.
bacterial operons is constrained by biological pathways [27] Lomnitz JG & Savageau MA. Elucidating the genotype–
encoded in the genome. Proc. Natl Acad. Sci. USA 107 (2010) phenotype map by automatic enumeration and analysis
6310–6315. of the phenotypic repertoire. Nature Partner J. Syst.
[11] Zhou F, Olman V & Xu Y. Barcodes for genomes and Biol. Appl. 1 (2015) 15003.
applications. BMC Bioinformatics 9 (2008) 546. [28] Schwacke JH & Voit EO. Improved methods for the
[12] Yus E, Maier T, Michalodimitrakis K, et al. Impact of genome mathematically controlled comparison of biochemical
reduction on bacterial metabolism and its regulation. systems. Theor. Biol. Med. Model. 1 (2004) 1.
Science 326 (2009) 1263–1268. [29] Savageau MA. Demand theory of gene regulation.
[13] Alves R & Savageau MA. Effect of overall feedback I. Quantitative development of the theory. Genetics
inhibition in unbranched biosynthetic pathways. Biophys. 149 (1998) 1665–1676.
J. 79 (2000) 2290–2304. [30] Savageau MA. Demand theory of gene regulation. II.
[14] Jeong H, Tombor B, Albert R, et al. The large-scale Quantitative application to the lactose and maltose operons
organization of metabolic networks. Nature 407 (2000) of Escherichia coli. Genetics 149 (1998) 1677–1691.
651–654. [31] Alves R & Savageau MA. Comparative analysis of prototype
[15] Isalan M, Lemerle C, Michalodimitrakis K, et al. Evolvability two-component systems with either bifunctional or
and hierarchy in rewired bacterial gene networks. Nature monofunctional sensors: differences in molecular
452 (2008) 840–845. structure and physiological function. Mol. Microbiol.
[16] Mukherji S & van Oudenaarden A. Synthetic biology: 48 (2003) 25–51.
understanding biological design from synthetic circuits. [32] Igoshin OA, Alves R & Savageau MA. Hysteretic and graded
Nat. Rev. Genet. 10 (2009) 859–871. responses in bacterial two-component signal transduction.
[17] Alon U. Simplicity in biology. Nature 446 (2007) 497. Mol. Microbiol. 68 (2008) 1196–1215.
[33] Skerker JM, Perchuk BS, Siryaporn A, et al. Rewiring the [58] Dueber JE, Wu GC, Malmirchegini GR, et al. Synthetic protein
specificity of two-component signal transduction systems. scaffolds provide modular control over metabolic flux. Nat.
Cell 133 (2008) 1043–1054. Biotechnol. 27 (2009) 753–759.
[34] Voit EO. Design principles and operating principles: the yin [59] Registry of Standard Biological Parts. http://partsregistry.org/.
and yang of optimal functioning. Math. Biosci. 182 (2003) [60] Arkin A. Setting the standard in synthetic biology. Nat.
81–92. Biotechnol. 26 (2008) 771–774.
[35] Colwell LJ, Brenner MP & Ribbeck K. Charge as a selection [61] Canton B, Labno A & Endy D. Refinement and standardization
criterion for translocation through the nuclear pore complex. of synthetic biological parts and devices. Nat. Biotechnol.
PLoS Comput. Biol. 6 (2010) e1000747. 26 (2008) 787–793.
[36] Bommarius AS & Riebel BR. Biocatalysis. Wiley-VCH, 2004. [62] Densmore D, Sangiovanni-Vincentelli A & Passerone R. A
[37] Yoshikuni Y, Dietrich JA, Nowroozi FF, et al. Redesigning platform-based taxonomy for ESL design. IEEE Des. Test
enzymes based on adaptive evolution for optimal function in Comput. 23 (2006) 359–374.
synthetic metabolic pathways. Chem. Biol. 15 (2008) 607–618. [63] Densmore D, Hsiau TH-C, Batten C, et al. Algorithms for
[38] Wang HH, Isaacs FJ, Carr PA, et al. Programming cells by automated DNA assembly. Nucleic Acids Res. 38 (2010)
multiplex genome engineering and accelerated evolution. 2607–2616.
Nature 460 (2009) 894–899. [64] Trinh CT, Unrean P & Srienc F. Minimal Escherichia coli cell for
[39] Currie JN. On the citric acid production of Aspergillus niger. the most efficient production of ethanol from hexoses and
Science 44 (1916) 215–216. pentoses. Appl. Env. Microbiol. 74 (2008) 3634–3643.
[40] Palsson BØ. Systems Biology: Properties of Reconstructed [65] Trinh CT, Wlaschin A & Srienc F. Elementary mode analysis:
Networks. Cambridge University Press, 2006. a useful metabolic pathway analysis tool for characterizing
[41] Stephanopoulos GN, Aristidou AA & Nielsen J. Metabolic cellular metabolism. Appl. Microbiol. Biotechnol. 81 (2009)
Engineering: Principles and Methodologies. Academic Press, 813–826.
1998. [66] Schuster S & Hilgetag S. On elementary flux modes in
[42] Torres NV & Voit EO. Pathway Analysis and Optimization in biochemical reaction systems at steady state. J. Biol. Syst.
Metabolic Engineering. Cambridge University Press, 2002. 2 (1994) 165–182.
[43] Lartigue C, Vashee S, Algire MA, et al. Creating bacterial strains [67] Liao JC, Hou SY & Chao YP. Pathway analysis, engineering,
from genomes that have been cloned and engineered in and physiological considerations for redirecting central
yeast. Science 325 (2009) 1693–1696. metabolism. Biotechnol. Bioeng. 52 (1996) 129–140.
[44] Mojica FJ & Montoliu L. On the origin of CRISPR-Cas [68] Poolman MG, Venkatesh KV, Pidcock MK & Fell DA. ScrumPy—
technology: from prokaryotes to mammals. Trends Microbiol. Metabolic Modelling in Python. http://mudshark.brookes.
24 (2016) 811–820. ac.uk/ScrumPy.
[45] Wang F & Qi LS. Applications of CRISPR genome engineering [69] Poolman MG, Venkatesh KV, Pidcock MK & Fell DA. A method
in cell biology. Trends Cell Biol. 16 (2016) 875–888. for the determination of flux in elementary modes, and its
[46] Koeller KM & Wong CH. Enzymes for chemical synthesis. application to Lactobacillus rhamnosus. Biotechnol. Bioeng.
Nat. Rev. Genet. 409 (2001) 232–240. 88 (2004) 601–612.
[47] Duncan TM & Reimer JA. Chemical Engineering Design and [70] von Kamp A & Schuster S. METATOOL 5.1 for GNU octave and
Analysis: An Introduction. Cambridge University Press, 1998. MATLAB. http://pinguin.biologie.uni-jena.de/bioinformatik/
[48] Drubin DA, Way JC & Silver PA. Designing biological systems. networks/metatool/metatool5.1/metatool5.1.html.
Genes Dev. 21 (2007) 242–254. [71] von Kamp A & Schuster S. Metatool 5.0: fast and flexible
[49] Keasling JD. Synthetic biology for synthetic chemistry. ACS elementary modes analysis. Bioinformatics 22 (2006)
Chem. Biol. 3 (2008) 64–76. 1930–1931.
[50] Marner WD 2nd. Practical application of synthetic biology [72] Martin VJJ, Pitera DJ, Withers ST, et al. Engineering a
principles. Biotechnol. J. 4 (2009) 1406–1419. mevalonate pathway in Escherichia coli for production of
[51] Tryfona T & Bustard MT. Fermentative production of lysine by terpenoids. Nat. Biotechnol. 21 (2003) 796–802.
Corynebacterium glutamicum: transmembrane transport and [73] Ro DK, Paradise EM, Ouellet M, et al. Production of the
metabolic flux analysis. Process Biochem. 40 (2005) 499–508. antimalarial drug precursor artemisinic acid in engineered
[52] First successful laboratory production of human yeast. Nature 440 (2006) 940–943.
insulin announced. Genentech Press Release, 6 [74] National Institute of Allergy and Infectious Diseases (NIAID).
September 1978. https://www.gene.com/media/press- Malaria. https://www.niaid.nih.gov/diseases-conditions/
releases/4160/1978-09-06/first-successful-laboratory- malaria.
production-o. [75] Graham IA, Besser K, Blumer S, et al. The genetic map of
[53] Sharma SS, Blattner FR & Harcum SW. Recombinant protein Artemisia annua L. identifies loci affecting yield of the
production in an Escherichia coli reduced genome strain. antimalarial drug artemisinin. Science 327 (2010) 328–331.
Metab. Eng. 9 (2007) 133–141. [76] Bertea CM, Voster A, Verstappen FW, et al. Isoprenoid
[54] Pátek M, Holátko J, Busche T, et al. Corynebacterium biosynthesis in Artemisia annua: cloning and heterologous
glutamicum promoters: a practical approach. Microb. expression of a germacrene A synthase from a glandular
Biotechnol. 6 (2013) 103–117. trichome cDNA library. Arch. Biochem. Biophys. 448
[55] Hammer K, Mijakovic I & Jensen PR. Synthetic promoter (2006) 3–12.
libraries—tuning of gene expression. Trends Biotechnol. [77] Dondorp AM, Nosten F, Yi P, et al. Artemisinin resistance in
24 (2006) 53–55. Plasmodium falciparum malaria. N. Engl. J. Med. 361 (2009)
[56] Ham TS, Lee SK, Keasling JD & Arkin A. Design and 455–467.
construction of a double inversion recombination switch [78] Ye X, Al-Babili S, Klöti A, et al. Engineering the provitamin A
for heritable sequential genetic memory. PLoS One 3 (2008) (beta-carotene) biosynthetic pathway into (carotenoid-free)
e2815. rice endosperm. Science 287 (2000) 303–305.
[57] Grilly C, Stricker J, Pang WL, et al. A synthetic gene network for [79] Paine JA, Shipton CA, Chaggar S, et al. Improving the
tuning protein degradation in Saccharomyces cerevisiae. Mol. nutritional value of Golden Rice through increased pro-
Syst. Biol. 3 (2007) 127. vitamin A content. Nat. Biotechnol. 23 (2005) 482–487.
FURThER READING 423
[80] Naqvi S, Zhu C, Farre G, et al. Transgenic multivitamin corn [88] Friedland AE, Lu TK, Wang X, et al. Synthetic gene networks
through biofortification of endosperm with three vitamins that count. Science 324 (2009) 1199–1202.
representing three distinct metabolic pathways. Proc. Natl [89] Tabor JJ, Salis HM, Simpson ZB, et al. A synthetic genetic edge
Acad. Sci. USA 106 (2009) 7762–7767. detection program. Cell 137 (2009) 1272–1281.
[81] Golden Rice Project. http://www.goldenrice.org/. [90] Salis H, Tamsir A & Voigt C. Engineering bacterial signals and
[82] Myers CJ. Engineering Genetic Circuits. Chapman & Hall/CRC sensors. Contrib. Microbiol. 16 (2009) 194–225.
Press, 2009. [91] Salis HM, Mirsky EA & Voigt CA. Automated design of
[83] Gardner TS, Cantor CR & Collins JJ. Construction of a synthetic ribosome binding sites to control protein
genetic toggle switch in Escherichia coli. Nature 403 expression. Nat. Biotechnol. 27 (2009) 946–950.
(2000) 339–342. [92] March JC & Bentley WE. Quorum sensing and bacterial
[84] Elowitz MB & Leibler S. A synthetic oscillatory network of cross-talk in biotechnology. Curr. Opin. Biotechnol. 15 (2004)
transcriptional regulators. Nature 403 (2000) 335–338. 495–502.
[85] Atkinson MR, Savageau MA, Myers JT & Ninfa AJ. [93] Hooshangi S & Bentley WE. From unicellular properties to
Development of genetic circuitry exhibiting toggle switch multicellular behavior: bacteria quorum sensing circuitry and
or oscillatory behavior in Escherichia coli. Cell 113 (2003) applications. Curr. Opin. Biotechnol. 19 (2008) 550–555.
597–607. [94] Li J, Wang L, Hashimoto Y, et al. A stochastic model of
[86] Rust MJ, Markson JS, Lane WS, et al. Ordered phosphorylation Escherichia coli AI-2 quorum signal circuit reveals alternative
governs oscillation of a three-protein circadian clock. Science synthesis pathways. Mol. Syst. Biol. 2 (2006) 67.
318 (2007) 809–812. [95] Beisel CL & Smolke CD. Design principles for riboswitch
[87] Cookson NA, Tsimring LS & Hasty J. The pedestrian function. PLoS Comput. Biol. 5 (2009) e1000363.
watchmaker: genetic clocks from engineered oscillators. [96] Yao G, Lee TJ, Mori S, et al. A bistable Rb–E2F switch underlies
FEBS Lett. 583 (2009) 3931–3937. the restriction point. Nat. Cell Biol. 10 (2008) 476–482.
FURThER READING
Alon U. An Introduction to Systems Biology: Design Principles of Jeong H, Tombor B, Albert R, et al. The large-scale organization of
Biological Circuits. Chapman & Hall/CRC Press, 2006. metabolic networks. Nature 407 (2000) 651–654.
Alon U. Simplicity in biology. Nature 446 (2007) 497. Keasling JD. Synthetic biology for synthetic chemistry. ACS Chem.
Arkin A. Setting the standard in synthetic biology. Nat. Biotechnol. Biol. 3 (2008) 64–76.
26 (2008) 771–774. Ma’ayan A, Blitzer RD & Iyengar R. Toward predictive models of
Baldwin G, Bayer T, Dickinson R, et al. Synthetic Biology—A mammalian cells. Annu. Rev. Biophys. Biomol. Struct. 34 (2005)
Primer, revised ed. Imperial College Press/World 319–349.
Scientific, 2016. Marner WD 2nd. Practical application of synthetic biology
Barabási A-L. Linked: The New Science of Networks. Perseus, 2002. principles. Biotechnol. J. 4 (2009) 1406–1419.
Canton B, Labno A & Endy D. Refinement and standardization Mukherji S & van Oudenaarden A. Synthetic biology: understanding
of synthetic biological parts and devices. Nat. Biotechnol. biological design from synthetic circuits. Nat. Rev. Genet.
26 (2008) 787–793. 10 (2009) 859–871.
Church GM & Regis E. Regenesis: How Synthetic Biology Will Myers CJ. Engineering Genetic Circuits. Chapman & Hall/CRC Press,
Reinvent Nature and Ourselves. Basic Books, 2012. 2009.
Covert MW. Fundamentals of Systems Biology: From Synthetic Salis H, Tamsir A & Voigt C. Engineering bacterial signals and
Circuits to Whole-Cell Models. CRC Press, 2015. sensors. Contrib. Microbiol. 16 (2009) 194–225.
Drubin DA, Way JC & Silver PA. Designing biological systems. Savageau MA. Demand theory of gene regulation. I. Quantitative
Genes Dev. 21 (2007) 242–254. development of the theory. Genetics 149 (1998) 1665–1676.
Friedland AE, Lu TK, Wang X, et al. Synthetic gene networks that Voit EO. The Inner Workings of Life. Vignettes in Systems Biology.
count. Science 324 (2009) 1199–1202. Cambridge University Press, 2016.
Emerging Topics in
Systems Biology 15
• Present an overview of the current state of the art in systems biology
• State some of the pressing challenges for systems biology in the near future
• Explain why some application areas are difficult to analyze and in need of
systems analysis
• Explain where current modeling techniques fall short and what could be done
to remedy them
• Discuss what a theory of biology would entail
This final chapter discusses where systems biology stands at present and where it
might go in the near future. It first presents some challenging application areas
awaiting systems analysis and then lists some of the foreseeable modeling needs
and theoretical developments.
Looking into the future is tricky business. Many a forecaster has met with ridicule
once the future arrived and turned out to be drastically different than even the wildest
imagination had thought possible. The world has used many, many more than five
computers, as allegedly predicted in the early 1950s by Thomas John Watson, the
President of IBM. We have not run out of copper, as predicted by the Club of Rome in
the 1970s, when the number of telephone lines was rapidly growing, apparently with-
out end, and before optical fibers had been invented [1]. People are not flying through
town with individual jet packs, as portrayed in so many old science fiction movies.
The future of systems biology is particularly difficult to predict, even for the rela-
tively short time period of the next 10 or 20 years, because it is still such a young disci-
pline that any type of extrapolation into the unknown should immediately trigger
warning signals. If we take molecular biology as an example and try to imagine what
forecasts might have looked like 50 years ago, we probably see very quickly how dra-
matically predictions can go astray. At a time when a doctoral degree was awarded for
sequencing a single gene, who could have imagined microarrays, spotted by robots?
Who could have foreseen the incredible advances in molecular and cellular imaging?
So, with respect to systems biology, it seems that we can only be certain that there is
not much reliability when it comes to predicting the face of systems biology in 2050 or
even in 2025. Not even the goals are crisp and clear. Some systems biologists have
declared as their ultimate goal a computational model of a complete cell or organism.
Others believe that the discovery of general design principles is of greater importance.
In spite of all this uncertainty, it might be useful to speculate about some application
areas and methodological trends that are appearing on the horizon.
426 Chapter 15: Emerging Topics in Systems Biology
EMERGING APPLICATIONS
15.1 From Neurons to Brains
Both experimental biology and biomathematics have long been fascinated with
nerves and brains [2, 3]. It was over half a century ago that Hodgkin and Huxley pro-
posed their pioneering modeling work on squid neurons, for which they received
the Nobel prize (see [4] and Chapter 12). Yet, in spite of a lot of research, we still do
not know how our memory works or what determines whether a person will be
afflicted with Alzheimer’s disease later in life.
By necessity, modeling approaches in neurobiology have mostly addressed
either small systems in detail or large systems in a much coarser fashion.
Uncounted variations on models for individual action potentials have been pro-
posed, and careful mathematical analyses have been able to explain the basis of
their functioning [5–7]. Using action potentials as input, biochemical models
have been developed that describe the dynamics of neurotransmitters in indi-
vidual neurons, as well as the transduction of signals across the synapse between
two neurons (see, for example, [8–12]). At a higher level of organization, models
of neuronal signal transduction have focused on the function and interactions of
different brain modules, but by doing so they have had to ignore most details
regarding the cellular and intracellular events that are the basis for neuronal sig-
naling [13].
Naturally, complex and detailed models are expected to permit more realistic
analyses than minimal models. Yet, very simplified models have their own appeal
[14, 15]. The extreme in simplification is a biologically inspired construct, called an
artificial neural network (ANN), in which neurons are coded exclusively by their
activation state as on or off, 1 or 0, and this state is transduced throughout the net-
work in a discrete fashion (see Chapters 4, 9, and 13). At each step, the 1’s and 0’s
reaching a neuron are multiplied by numerical weights, and the sum of the weighted
inputs determines whether this neuron will be on or off for the next step of the dis-
crete process. This drastic simplification of a realistic network of biological neurons
has the advantage that large arrays of thousands of artificial neurons can be studied
very effectively. Intriguingly, it has been shown that, over time, ANNs can be trained
to recognize patterns by changing the weights with which each neuron processes its
inputs [16, 17]. As a result, ANNs are capable of representing an unlimited spectrum
of responses, and extensive research has led to applications of ANNs to all kinds of
classification and optimization tasks that no longer have anything to do with neuro-
science. For instance, ANNs can be trained to find and read zip codes on envelopes
and distinguish a sloppily written 7 from a 1.
At the other end of the spectrum between simplicity and complexity, the cul-
mination of simulating realistic interaction networks of neurons is presently the
Blue Brain Project [18], for which a large group of scientists is collecting detailed
micro-anatomical, genetic, and electrical data of one specific human brain sec-
tion, the neocortical column, and integrating them into three-dimensional simu-
lation models that account for very large populations of over 200 types of neurons
(Figures 15.1 and 15.2). The sheer size of the project requires enormous super-
computing power, and the results are fascinating. A major issue with this approach
is that the simulation results have become so complex that their analysis is a sig-
nificant challenge in itself.
The different approaches to understanding the brain highlight distinct aspects of
neurons and neural networks and ignore others. The grand challenge for neuro-
systems biology is now the creation of conceptual and computational multiscale
models that integrate electrical, biochemical, genetic, anatomical, developmental,
and environmental aspects associated with the brain in a manner that offers deep
insights into fundamental functions such as learning, memory, and adaptation, as
well as the development and progression of neurodegenerative diseases and addic-
tion [19]. The first step will be to bridge the gap from single-neuron models to
detailed, realistically sized neural networks. This step is by no means trivial, since Figure 15.1 Image of a model
not only are neural networks in the brain very complex in structure, but they can reconstruction of the rat neocortical
dynamically change in activity, for instance, between wakefulness and sleep. column. (Courtesy of the Blue Brain Project,
Switches between these states are often rapid, and the rules guiding them are largely EPFL.)
EMERGING APPLICATIONS 427
Figure 15.2 Toward a wiring diagram

L1 of the brain. Simplified Blue Brain
g representation of the cortical laminar
p2/3 structure in the human brain. (From de
g
Garis H, Shuo C, Goertzel B & Ruiting L.
L2/3 Neurocomputing 74 [2010] 3–29. With
g
ss4(L2/3)
g
L4
ss4 (L4)
g p4
cortico-cortical
nonspecific
g
p5(L5/6)
p5(L2/3)
specific
L5
g
p6(L4) g
p6(L5/6)
L6 cortex
wm cortico-cortical to other cortical

areas
specific premotor
and
motor
centers
types of neurons types of synapses
pyramidal (p) cell local excitatory
spiny stellate (ss4) local inhibitory
cell of layer 4 global cortical reticular
g thalamic
basket (b) cortico-thalamic
interneuron nucleus
thalamo-cortical (RTN)
nonbasket (nb) sensory input
interneuron brainstem
thalamo-cortical modulation Tls
thalamus
(TC) relay cell g gap junctions

Tln
RTN neuron g g TCn
TCs
Figure 15.3 Accounting for all neurons.
specific nonspecific
thalamic Caenorhabditis elegans is a simple worm
interneuron sensory input with exactly 959 cells, 302 of which are
neurons. (Courtesy of kdfj under the Creative
Commons Attribution-Share Alike 3.0
Unported license.)
unknown [20]. Initial attempts in this direction are models that study bursting pat-
terns in networks of maybe a dozen to 100 neurons [21]. A complementary approach
is the creation of computer languages that are custom-tailored to analyses of how
anatomical features, synapses, ion channels, and electrochemical processes govern
the functionality of the brain [10].
A good starting point for studying realistic neural networks might be a relatively
simple model organism such as the fascinating worm Caenorhabditis elegans, which
consists of exactly 959 somatic cells, of which exactly 302 are neurons (Figure 15.3).
Researchers in Japan have even created a database of synaptic connectivities in this
organism [22]. Another candidate is the California sea slug Aplysia californica, which
has become a model for neurobiological studies, because it has only about 20,000
neurons, but exhibits interesting responses to stimulation, such as the ejection of red
ink (Figure 15.4). Eventually, models of these and more complex neural networks Figure 15.4 The sea slug Aplysia
will be extended and become reliable enough to permit targeted interrogations and californica has become a model organism
analyses of brains in higher organisms. They will lead to deeper understanding and for studying learning and memory. Here,
might help us explain neurodegenerative disease, addiction, epilepsy, and other the slug releases a cloud of ink to confuse
an attacker. (Courtesy of Dibberi under the
menacing diseases, as well as normal, physiological phenomena such as intelligence Creative Commons Attribution-Share Alike
and conscience (see, for example, [23, 24]). 3.0 Unported license.)
15.2 Complex Diseases, Inflammation, and Trauma

The brain is certainly fascinating, but it is not the only challenging system in medi-
cine. Many complex diseases, and even normal development and aging, are insuf-
ficiently understood, and systems biology may provide some of the tools to address
them in a holistic manner. This section discusses two representative cases.
Diabetes has relatively clear causes, outcomes, and symptoms, but is a complex,
systemic disease that may be exacerbated by physiological and pathological responses
of other organs and reactions by the immune system. An example is reduced renal
clearance, which directly or indirectly affects virtually all aspects of physiology or
pathology, beginning with a reduced removal of toxins and with changes in blood
pressure, volume, and flow, which may affect cardiovascular functioning, leading to
easy bruising and bleeding with a chance of infection, vomiting, headaches, and sei-
zures. Mathematical models of diabetes go back a long way [25] and, coming from
humble beginnings, they have reached a level where they are beginning to be predic-
tive. Indeed, simple diabetes models have shed light on the kinetics of insulin in vivo,
demonstrated the importance of pancreatic beta-cell compensation, and identified a
particular composite model parameter, the disposition index, as representative of the
ability of the pancreas to compensate for insulin resistance [26].
While diabetes modeling has made great progress [27, 28], there is still a lot to be
learned, and as we gather more biological and clinical information, models will
become increasingly important tools of integration. The beginnings of this type of
integration can be seen in recent physiologically based simulation models of diabe-
tes and metabolic syndrome [29, 30] that capture the function of regulatory path-
ways in these diseases and have been used to connect diabetes to the risk of heart
problems [31]. Given that type 2 diabetes afflicts roughly 20 million people in the
United States alone and is a primary cause of heart and kidney disease, as well as of
blindness, the rewards from developing a deeper understanding of the disease and
its sequelae are beyond any doubt.
Another disease-related and very intriguing modeling challenge is just now in the
initial phase of being addressed. It pertains to infection and inflammation, and to the
ensuing immune responses, which are usually beneficial, but can also become very
dangerous. Mammals, including humans, possess two systems that protect them
from infecting pathogens: the innate and the adaptive immune systems. According
to current understanding, the former recognizes unique molecular features of patho-
gens with specific receptors and responds very quickly by activating countermea-
sures [32]. If the innate response is unable to contain the infection, the adaptive
immune system is called up. This system responds on a timescale of days and ulti-
mately equips the organism with a memory of earlier pathogen infections and
immune responses. Although the adaptive system operates at a much slower speed,
it interacts directly with the innate system: T cells of the adaptive response and mac-
rophages of the innate immune system are part of a positive-feedforward loop, in
which they activate each other through signaling proteins called cytokines. The
innate immune cells are needed to trigger responses in the adaptive system, while
the activated adaptive immune cells effect the amplification of anti-pathogenic
responses in the innate cells [33]. Positive-feedback loops are potentially dangerous,
because they run the risk of spinning out of control. In the body, this risk is internally
controlled by regulatory T cells that normally suppress the innate response in order
to preclude excessive responses.
Thus, the innate and adaptive immune systems constitute an integrated defense
system, with each component depending on the other for amplification and regula-
tion. This fine-tuned system works very well, as long as all components are opera-
tional. However, if one system component is compromised, the system can derail. For
instance, if the CD4+ T-cell count is very low, the regulation of the innate immune
system is lacking, and it is possible that inflammatory cytokines will dramatically
accumulate into a cytokine storm that can be strong enough to kill the host through
tissue damage and septic shock (Figure 15.5). As an example, a cytokine storm in the
lung can become severe enough to block the airways through the accumulation of
fluids and immune cells. Under normal conditions, cytokines move through the cir-
culatory system in picomolar concentrations. However, these levels can increase
1000-fold during infection or trauma, especially in otherwise-healthy young
(A) Healthy host Figure 15.5 A brewing storm? (A)

In healthy hosts, T cells of the adaptive
T cell immune system interact with macrophages
of the innate immune system and probably
suppress the production of cytokines.
(B) In compromised hosts with insufficient
functional T cells, the production of
pathogen macrophage cytokines resolution cytokines is uncontrolled and can result
in a lethal cytokine storm. (Adapted from
Palm NW & Medzhitov R. Nat. Med. 13 [2007]
1142–1144. With permission from Macmillan
Publishers Limited, part of Springer Nature.)
(B) Compromised host
dysfunctional
T cell
pathogen macrophage cytokines cytokine storm
individuals. There is speculation that the high number of deaths among young adults
during the 1918 influenza pandemic and during the 2005 bird flu outbreak might have
been due to cytokine storms.
Of course, as so often in biology, the diagrams of inflammation in Figure 15.5 are
woefully over-simplified. In reality, the body produces numerous families of cyto-
kines, including interleukins, chemokines, and interferons. These families are
diverse and form complex interaction networks, and we are even beginning to rec-
ognize that the early distinction between “pro-inflammatory” and “anti-inflamma-
tory” cytokines no longer makes real sense in many instances, because a pro-
inflammatory cytokine may trigger the production of an anti-inflammatory cytokine,
so that the pro-inflammatory cytokine indirectly becomes anti-inflammatory. Cyto-
kine storms have been estimated to involve more than 150 different cytokines, coag-
ulation factors, and radical oxygen species, and the ultimate response to an altera-
tion of one cytokine species within such a complex network depends on much more
than the type of cytokine; in particular, it seems that the preconditioning history of
the network is of extreme importance [34]. Furthermore, cells of the innate immune
system initiate both innate and adaptive immune responses [33] through the release
of cytokines, which makes intuitive predictions of responses very difficult. To make
matters even more complicated, there is plenty of crosstalk among different cell
types and their signaling activities. These large numbers of different players and
interactions clearly suggest that methods of systems biology will be needed to eluci-
date immune responses and cytokine storms [34].
Infection is often the first trigger for calling up the immune system, which rem-
edies the situation. Unfortunately, the response of the immune system itself can
cause damage and inflammation in associated tissues. In fact, it is possible that
most of the infecting pathogens are killed but a vicious cycle continues, in which
tissue damage causes more inflammation, which in turn causes more damage.
Again, this positive feedback is driven by cytokines and immune cells [35].
Intriguingly, the point has been made that inflammation is a unifying factor
tying together diseases that at first appear to be very different, such as atherosclero-
sis, rheumatoid arthritis, asthma, cancer, AIDS, metabolic syndrome, Alzheimer’s
disease, chronically non-healing wounds, and sepsis [34, 35]. At the same time,
inflammation is associated with natural aging processes as well as the beneficial
effects of exercise and rehabilitation. When seen in such broad contexts, inflamma-
tion per se is not a detrimental process, but the reflection of a complex communica-
tion network, with which the organism signals potentially harmful perturbations.
Under healthy conditions, the system works remarkably well, but if it becomes too
unbalanced for an extended period of time, disease is the consequence.
While the inflammatory processes in their entirety may appear to be over-

whelmingly complex, they constitute a homeostatic control system whose basic
structure is ubiquitous and actually quite simple. Focusing on this simplicity per-
mits mathematical analyses that can yield interesting insights [34]. Specifically, the
inflammatory response may be characterized as a balance of processes (Figure
15.6). The simplified representation accommodates several entry points into the acute
inflammatory response and furthermore includes a damage signal that can feed
back in a detrimental manner. Under normal conditions, a pathogen evokes a
strong inflammatory response and simultaneously a weaker anti-inflammatory
response. The inflammation triggers an attack on the pathogen and secondarily
activates the anti-inflammatory machinery. At the same time, pro-inflammatory
agents, such as tissue necrosis factor (TNF), can cause tissue damage or dysfunction.
Anti-inflammatory agents, such as transforming growth factor β (TGF-β1), sup-
press inflammation and stimulate healing of damaged tissues. Tissue damage, if
not controlled, can lead to further inflammation.
Given a simplified base structure as in Figure 15.6, it is not too difficult to imagine
that the outcome of a pathogen attack is determined by the balance between the
different processes in the system. For instance, if the stimulation of the anti-
inflammatory machinery is sufficiently strong, inflammation will be suppressed and
tissues have a chance to heal: in Figure 15.6, the processes in green overcome the
red processes. By contrast, if the processes shown in red are too powerful, one can
see that not only is tissue damaged, but a vicious cycle continues between inflam-
mation and further tissue damage, even after the pathogen has been defeated, as we
discussed earlier. The diagram also shows that inflammation can be triggered by
tissue damage, which may be due to trauma or disease.
Recent years have seen enormous progress in this field, and we are beginning to
understand enough details of the inflammatory network to think about initiating for-
mal systems biological investigations. These investigations should ideally include pre-
clinical studies, clinical trials, in-hospital care, and the long-term use of treatment
[34]. The challenges faced when attempting to model inflammation are manifold. Not
only are the normal inflammatory processes and their regulation complicated, but
components of the inflammatory response system also interact with non-inflamma-
tory physiological systems, which makes it difficult to deduce the architecture and
control of the system from biological and clinical observations. Furthermore, individ-
ual inflammatory responses depend critically on the preconditioning state of the
organism, that is, on the current health status as well as the health history of the
affected organism [34].
Distinct approaches have been employed to model inflammation (for reviews,
see [34, 36]). They have included mechanistically based ordinary and partial dif-
ferential equation models and simulations, and different stochastic and hybrid for-
mulations, as well as rule- and agent-based models, which will be discussed in a
later section of this chapter. A few examples of inflammation analyses are pre-
sented in the next few paragraphs.
alarm/danger
signal
inflammation
damage,
pathogen
dysfunction
Figure 15.6 Inflammation as an
imbalance. This very simplified diagram
demonstrates the fine balance between
counteracting processes leading either
to inflammation or to recovery from an
anti- infection (red and green, respectively).
inflammation
(Adapted from Vodovotz Y, Constantine G,
Rubin J, et al. Math. Biosci. 217 [2009] 1–10.
A mechanistic mathematical model was constructed to elucidate processes of

bacterial infection, cytokine dynamics, the acute inflammatory response, global tis-
sue dysfunction, and a therapeutic intervention [37]. The model was subsequently
used to simulate a clinical trial designed to demonstrate the benefits of a therapy
with different doses of a neutralizing antibody directed against TNF in patients with
sepsis. The simulation predicted cases with favorable and unfavorable outcomes
and also revealed why an actual clinical trial had failed and how the trial should
have been designed to succeed.
An equation-based model was developed to investigate whether different shock
states following an infection could be explained by a single inflammatory response
system [38]. Supported by different sets of experimental mouse data, the model
demonstrated different trajectories of the inflammatory process to either resolution
or shock, which depended on the individual cellular and molecular state of each
mouse. This modeling effort demonstrated three aspects. First, it proved that the
complex biology of inflammation can indeed be captured by mechanistic models.
Second, it made clear that the shock response to different physiological challenges
depends critically on the preconditioning state of the organism. And finally, the
model presented trends in the progression of the inflammation process that were
personalized and predictive with respect to survival.
Another example of the usefulness of differential equations in the analysis of
inflammation is a model for the receptor-mediated inflammatory response in
humans exposed to toxins produced by bacteria inside the body [14, 39]. This
model is interesting because it integrates genomic information and extracellular
signaling processes into the bacterial infection process. By changing its parameter
settings, the single model allows for the characterization of different outcomes,
including healthy resolution responses, persistent infection responses, aseptic
inflammation, and desensitization.
An interesting example of agent-based inflammation modeling is a hierarchical
concatenation of interacting submodels that collectively integrate basic scientific
and clinical information on acute inflammation into a modular, multiscale architec-
ture [40]. The individual submodels address pertinent cellular response mecha-
nisms to infection and inflammation, the permeability of gut epithelia, which are
the ultimate barrier against microbial pathogens, the inflammatory response of the
gut to ischemia, and responses in the lung. The model successfully matched clinical
observations on the pathogenesis of multiple organ failure and the underlying
crosstalk between gut and lung.
Cancer has been a target of modeling for a long time, and some of the early
approaches still have relevance. The initial concepts and models considered the pro-
cess of carcinogenesis (cancer formation) as a multistage or multihit process (Figure
15.7). This process is thought to begin with normal cells which, as a result of some
initiation event, become premalignant. The premalignant cells can remain in this
stage of promotion for a long time. A transformation event converts one of the prema-
lignant cells into a malignant cell, which replicates during a progression phase into a
tumor. Radiation, exposure to toxins, and a number of other processes can function as
possible initiation and transformation events [41]. In this view of carcinogenesis, a
single cell moving through all phases is seen as sufficient to lead to a tumor. Early
mathematical models assessed the probabilities or rates of transitions within this
multistage process, and also allowed for cell replication and death in the various
phases [42, 43].
A more modern view of carcinogenesis still contains the same phases, but recog-
nizes that tumor formation is much more complicated. First, it has become clear that
cancer cell populations are heterogeneous and contain cells with different degrees of
carcinogenic potential [44]. Thus, killing most of the cancer is not necessarily the best
Figure 15.7 Conceptual phases and events

initiation transformation during carcinogenesis. Over an extended
period of time, a few of very many normal
cells become initiated, but still appear
normal premalignant promotion malignant progression normal. A transformation event turns one of
cell cell cell tumor these cells malignant, and unless the body
destroys this cell, it will eventually multiply
and grow into a tumor.
strategy if the most virulent cells survive the treatment and subsequently form an even
more aggressive tumor. Instead, a better strategy may call for controlling the tumor
rather than attempting to remove it [45]. Second, several interesting studies (see, for
example, [46]) have demonstrated that essentially all middle-aged and older adults
harbor microscopic tumors, most of which stay dormant for a long time, if not for the
person’s entire life. As one impressive example, less than 1% of the adult population is
ever diagnosed with thyroid cancer, but 99.9% of all individuals autopsied for other rea-
sons were found to have thyroid cancer lesions. Third, recent studies show that small
tumors can have distinctly different fates (the orange arrow in Figure 15.7). Some grow
into full-size tumors and maybe even metastasize, but others will grow and shrink in an
oscillatory pattern, some will regress on their own, and some may stay dormant at a
very small size that per se is not harmful [44, 47]. What determines these diverging tra-
jectories is not entirely understood, but it seems that a change in the balance between
stimulators and inhibitors may flip a switch during the progression phase, upon which
the tumor starts to grow and to recruit blood vessels. This recruitment process, called
angiogenesis, apparently constitutes at least one significant bottleneck in tumor for-
mation [48].
Future models of carcinogenesis will probably retain the backbone of the multi-
stage cancer model, but will incorporate it into systems models that take account of
what we learn at the genomic level about small-RNA regulation (see Chapter 6), of
immune responses to the emergence of cancer cells, of the recruitment of progenitor
stem cells, of the recruitment of blood vessels, and of the interactions between cancer
cells and healthy tissues. It is clear that these models must either be of an integrating
multiscale nature or address some aspect that can subsequently be embedded in a
larger systems model. The models will be useful for forecasting the dynamics of carci-
nogenesis and the efficacy of different treatment options. They may also reveal over-
arching principles that differentiate tumor development from normal physiology.
In more generic terms of modeling human development, health, and disease,
some research teams have begun to consider disease, aging, and frailty as the con-
sequence of a loss of complexity, arguing that some highly periodic processes
become diseased if they are too regular and do not exhibit the characteristics of frac-
tals or even chaos [49, 50]. An example is a very regular fetal heartbeat, which very
often is a cause for concern [51–53].
15.3 Organisms and their Interactions with the Environment

It is complicated enough to design models for systems within individual cells and
organisms, but the degree of complexity jumps again when it is our task to inves-
tigate organisms in their environments. Any comprehensive study of this type
immediately requires multiple scales in time, space, and organization; determin-
istic processes mingle with stochastic effects; and data and other pieces of infor-
mation are so diverse and heterogeneous that it often seems overwhelmingly dif-
ficult to merge and integrate them in our minds. It seems again that system
approaches will eventually be our best bet. This section discusses two examples in
which systems thinking and computational modeling show very high potential.
Metapopulations consist of several populations, each of a different species, that
live in the same environment and interact with each other. Although we tend to think
of populations in large expanses of the environment, metapopulations are similarly
intriguing in very small ecological niches, such as localized parcels of soil, the human
oral cavity, or even the gut of a termite. Whatever the context, microbial metapopula-
tions come in enormous sizes. It has been estimated that 5 × 1030 prokaryotic cells
inhabit the Earth [54]. A single gram of soil can contain enough microbial DNA to
stretch almost 1000 miles [55]. A lake may contain 20,000 different species of bacteria
[56]. Humans harbor 10 times as many bacteria as their own cells. Very many of these
live in the gut, where their density can reach 1012 cells per milliliter. This number is
enormous, but the bacteria are so small in size that the gut microbiome makes up only
between 1% and 2% of a human’s body mass. The human gut microbiome can easily
contain hundreds or even thousands of different species, and the totality of different
genes in this metapopulation probably exceeds our own genome size by several orders
of magnitude. Furthermore, the composition of the gut microbiome changes
dynamically and drastically in response to age, diet, diseases, and of course the inges-
tion of antibiotics [57]. Because the gut microbiome is directly or indirectly associated
with a number of diseases, the US National Institutes of Health in 2007 launched an
NIH Roadmap effort to use genomic technologies to explore the role of microbes in
human health and disease [58]. Similar projects have been launched elsewhere [59,
60].
Similarly important, ubiquitous metapopulations can be found in biofilms, which
cover moist surfaces from teeth to the insides of sewage pipes (Figure 15.8). Like the
gut microbiome, biofilms consists of hundreds or thousands of different species, and
waste products from one species are used as nutrients by others. Intriguingly, the dif-
ferent species in metapopulations, and the individual organisms within these species,
compete for the same physical space and the same resources, yet they depend on
each other and often demonstrate a strict division of labor. Thus, in contrast to the
typical competition in shared environments, different microbes assume distinctly dif-
ferent and complementary roles, to a point where a species cannot survive without
the others [57].
Understanding microbial metapopulations faces significant challenges, the most
restrictive of which might be the fact that the vast majority of species in these popula-
tions are unknown. They cannot even be cultured in the laboratory and are therefore
difficult to characterize. Even if a few representative species can be grown in the labo-
ratory, their artificial environment is so deprived in comparison with their natural
surroundings that the extrinsic validity of subsequent results often seems doubtful. Figure 15.8 Scanning electron microscope
An interesting strategy for overcoming this problem was the introduction of image of a microbial biofilm. This
metagenomics [61]. Here, samples of entire microbial ecosystems are taken directly metapopulation developed on the inside of
from their natural environment and, without any attempt to culture them, analyzed untreated water tubing within some dental
as if they came from a single species. While this procedure is now technically pos- equipment. Such water lines often contain
sible, the challenge shifts to extracting reliable information from the mixture of mil- bacteria, as well as protozoa and fungi. (From
lions of new genomic sequences that in some unknown fashion belong to thousands Walker JT & Marsh PD. J. Dentistry 35 [2007]
of species. Needless to say, metagenomics stretches current methods of bioinfor-
matics to the limit. As a consequence, a full assembly of genomes is presently only
possible for mixtures of a very few species [51].
In addition to metagenomic studies, techniques are now becoming available for
community proteomics, which has the potential of yielding unprecedented insights
into the composition and function of microbial metapopulations. These techniques
have already been applied to a variety of systems, including sludge, biofilm, soil, and
gut microbiomes [62], but, as for many other proteomics techniques, the experi-
mental and computational analysis is challenging.
Various application areas have witnessed a time trend from molecular biology to
bioinformatics and subsequently to systems biology. The same trend holds for
microbial metapopulations [63, 64], where it is clear that systemic analyses will be
very challenging, fascinating, and absolutely necessary.
At the other end of the size spectrum, the oceans are without doubt among the
most challenging multiscale systems that await more systematic analyses than have
been possible in the past. They provide most of the oxygen we breathe and provide
the ultimate storage for water and a lot of food. As with many subspecialties of biol-
ogy, many aspects of the dynamics of and in oceans have been addressed with
reductionist methods and models of specific details, yet it has so far not been pos-
sible to integrate the huge range of spatial, temporal, and organizational scales into
reliable computational models. Nonetheless, we can see the rudimentary begin-
nings of systemic analyses in many places.
A first step beyond observational and wet sciences is the rise of ecoinformatics,
which uses methods of bioinformatics to manage, analyze, and interpret the enor-
mous quantities of data that have relevance for ecological systems and, in particular,
the oceans [65]. The challenges are great, not merely because of the sheer volume of
information, but also owing to the highly heterogeneous nature of this information
as well as a reluctance to adopt worldwide data standards and experimental and ana-
lytical protocols. Such challenges are currently being addressed with the creation of
data warehouses, the collection and structuring of metadata, laboratory information
management systems (LIMS), data pipelines, customized mark-up languages, spe-
cialized ontologies, and semantic web tools, including computer-based reasoning
systems. However, efficient storage and access to data alone will not be sufficient.
Beyond direct data analyses and automated inferences, computational means of

information integration, simulation, and interpretation are needed. In other words,
ecoinformatics must be complemented with dynamical systems ecology.
Recognizing this need, the US Department of Energy sponsors a program, Sys-
tems Biology for Energy and Environment, that identifies the Earth’s living systems
as a potential source of capabilities that can be put to use to meet national chal-
lenges and has the ultimate scientific goal of understanding the principles underly-
ing the structural and functional design of living systems [66]. The program focuses
on energy, environmental remediation, and carbon cycling and sequestration.
Within the latter focus area, the role of the oceans clearly plays a crucial role,
because marine cyanobacteria (blue–green algae) are responsible for half of the
global carbon dioxide fixation. There is little doubt that the well-being of these
organisms depends directly on nutrients, pH, temperature, salinity, contamination,
and the surrounding multispecies communities, but the exact parameters of opti-
mal, suboptimal, and barely tolerable growth conditions are as yet unknown.
An example of how molecular and systems biology have begun to reach out to
investigations of global ocean processes is a study on marine cyanobacteria, which
are among the most abundant organisms on Earth (Figure 15.9). These small crea-
tures are able to convert solar energy into biomass and thereby account for 20–30%
of the Earth’s photosynthetic productivity. Furthermore, they recycle about 40% of
the entire carbon that cycles between the oceans, the atmosphere, the Earth’s crust,
soils, microorganisms, plants, and animals. Researchers at Harvard’s Department of
Systems Biology found that cyanobacteria of the genus Synechococcus fix carbon
and convert it into sugar within spatially well-placed internal structures, called car-
boxysomes [67]. A single protein is responsible for the optimal spatial organization
of these carboxysomes, and mutations in the corresponding gene lead to a photo-
synthetic productivity that is decreased to 50% of that of the wild-type production.
Thus, a systemic, causal connection has been elucidated between the action of a
specific gene, the protein it codes for, the functionality of a microorganism, and a
truly global phenomenon. This same causal connection could become the starting
point for the design of genetically modified algae or bacteria for the production of
hydrogen, using methods of synthetic biology.
A different starting point for systemic studies of oceans is the careful investiga-
tion of coexisting marine populations using methods described in Chapter 10
(Figure 15.10). More and more detailed data for such studies are being collected
worldwide for different marine systems. For instance, Burkepile and Hay [68]
measured the effects of single-species and mixed-species herbivorous fish popu-
lations on the colonization and progression of algal communities and their impact
on coral growth. In carefully controlled experiments with otherwise comparable
systems, they demonstrated convincingly that species diversity is critical for the
health of coral systems and the resilience of the oceans following disturbances.
Dam and colleagues [56] used Lotka–Volterra models to study dynamic interac-
tion patterns among lake metapopulations consisting of almost 20,000 species.
(A) (B) (C)
Figure 15.9 Different strains of

cyanobacteria (blue–green algae) may
have very diverse shapes. (A) Bloom-
forming, filamentous, and colony-forming
Trichodesmium thiebautii (scale bar 100 μm).
(B) Unicellular Synechococcus PCC 6301 (scale
(E)
bar 1 μm). (C) Filamentous Symploca PCC
8002 (scale bar 10 μm). (D) Filamentous,
H nonbranching, and heterocystous (H) Nostoc
PCC 7107 (scale bar 10 μm). (E) Filamentous,
heterocystous (H) and branching (arrow)
(D) Fischerella PCC 7521 (scale bar 15 μm).
H (From Cox PA, Banack SA, Murch SJ, et al.
Proc. Natl Acad. Sci. USA 102 [2005] 5074–8.
H
With permission from National Academy of
Sciences, USA.)
MODELING NEEDS 435
Figure 15.10 Corals and other

invertebrates compete for space. Data
characterizing interacting populations
are currently being collected and create
a starting point for systems-biological
model analyses. (From Burkepile DE & Hay
ME. Coral reefs. In Encyclopedia of Ecology
[SE Jorgensen & B Fath, eds], pp 784–796.
Elsevier, 2008. With permission from
Elsevier.)
MODELING NEEDS
To predict future modeling needs, it might be prudent to revisit the present state of
the art, consider how we got here, and then try to extrapolate trends into the coming
years. There is no doubt that computational modeling has come a long way. Thirty
or forty years ago, computing was done on mainframe computers that required
manual feeding of punch cards with instruction code. There was no Internet, no
parallel computing, no user-friendly software. Much has happened since those dark
days. Computational power and speed have doubled every 18 months since the
1950s, access to computing has increased millions of times, and today’s higher-level
languages such as MATLAB®, Mathematica®, and SBML have facilitated the develop-
ment and implementation of computational ideas among users most of whom
would certainly have been intimidated by the need to master computer languages
like Fortran, let alone assembly language, just a few decades ago.
While progress has been amazingly fast, it is clear that we still have a long way to
go. But the bottleneck is not just computational speed and convenience. In fact, out-
side certain areas, such as the prediction of protein structure and function and very
large-scale simulation and optimization studies, it is not necessarily the technical
aspects of computing that stand in the way of the next significant advance. New con-
cepts and radically new modeling techniques might be needed. Rainwater runs down
a pile of dirt in a very efficient fashion, yet capturing the process in a realistic mathe-
matical model requires enormous effort. Why is that? Spiders with rather small brains
are able to construct complex webs, and animals respond to situations never encoun-
tered before with appropriate and effective actions, but we are hardly capable of rep-
resenting such activities adequately in systems of differential equations. To realize
model-aided personalized medicine, it is not sufficient to characterize the direct
effects of a disease or treatment, but also their secondary and tertiary effects, along
with interactions between different treatments and the multitudinous personal char-
acteristics of an individual.
These challenges are indeed great, and it might well be possible that mathematical
modeling and systems biology are in need of a significant paradigm shift, comparable to
the one that physics experienced in the early 1900s when quantum theory opened the
door to concepts unimaginable just a few decades before. But before we throw our
hands in the air and resign ourselves to waiting for a “new math” to appear, before we
discard modeling as simplistic and powerless and instead suggest doing another set of
laboratory experiments, let us just see what the alternatives are. It is undisputed that
realistic systems consist of hundreds or thousands of important components and that
these make modeling very complicated. But the same hundreds or thousands of impor-
tant components make every other approach even more complicated. Yes, we may not
be able to keep track of hundreds of dynamically changing systems components in a

computer-aided model analysis right now, but we will never be able to manage them in
our unaided minds. Statistical association models may be sufficient for identifying aver-
ages, but they will not suffice to address specific cases, as required for instance in per-
sonalized medicine. Beyond these considerations of magnitude, many experiments
simply cannot be executed, because they are technically impossible, practically infea-
sible, or unethical. We cannot test a new drug on every human, and it is probably not a
good idea to introduce a new strain of algae in the oceans on a trial basis.
Thus, while there is a clear mismatch between the grand challenges to be solved
and the tools we have presently at our disposal, the magnitude of the tasks before us
should not be a reason to abandon systems biology but rather one to increase our
efforts—because we have no choice. Besides, it will be fascinating and fun to witness
the field tackle these issues. Step by step, sometimes evolutionary and sometimes
revolutionary, sometimes elegantly and sometimes with brute force, we will expand
and strengthen our arsenal of modeling methods, address small problems, then
larger problems, first connect two biological levels and later on more. As a warm-up,
we will study hundreds of isolated processes, then systems, and ultimately realistic
and relevant interacting systems of systems. The journey is open-ended, but it
invites us to take the next steps. And while we move along, new methods will allow
us to increase speed and efficiency, and we will learn which types of techniques
tend to work and which not. Ultimately, some smart student, uninhibited by tradi-
tion, may indeed design an entirely new mathematical approach. It is moot to spec-
ulate what this might look like, but it is possible to assemble a partial wish list of
topics that need to be addressed. Some of these appear to be high-priority
candidates.
Interestingly, our current models are often either too simple or not simple
enough. If our goal is to explore and manipulate a specific system, drastic simplifi-
cations may hurt us by making predictions unreliable. For instance, if the task is to
increase product yield in a population of microorganisms, accounting for as many
details as possible is more likely to lead to success than simplifications of the model.
Similarly, it seems that the consideration of a larger number of biomarkers will be
advantageous in personalized medicine. At the same time, as models become larger
and more complex, they become more difficult to analyze, understand, and predict,
while simpler models often offer insights that would be overwhelmed and over-
shadowed by too much detail. Thus, starting from the level of complexity of today’s
models, some models will have to increase and some decrease in scope and com-
plexity, depending on the purposes of the modeling effort. And so the circle closes:
the introduction to modeling in Chapter 2 made it clear that the goals and purposes
of a model dictate its structure and features, and this statement still holds. The fol-
lowing sections will discuss first extensions into greater complexity and then simpli-
fications that might lead us one day to one or more theories of biology.
15.4 Multiscale Modeling

The hallmark of systems biology is a global, integrated approach to biological chal-
lenges. Yet, if we look at our current modeling activities, we must admit that they
often seem like reductionism in a different form. Many models focus on one set of
components, whether these be genes, neurons, or different species within the same
environment, and on one or maybe two organizational levels within the multiscale
hierarchy of biology. We are generating models of gene expression and of physiol-
ogy, but it is rare to find models spanning the two levels, as well as all significant
aspects in between. For instance, imagine the task of predicting the effect on climate
of a new strain of marine algae with increased photosynthetic activity. One could
presumably estimate the total volume of additional carbon dioxide fixation, but how
would this change influence other components of the oceanic ecosystem, cloud for-
mation, and light absorption, and feed back on the new algae and their competi-
tors? We are far from making such predictions with any degree of reliability.
In addition to the challenge of designing comprehensive multiscale models, we
need effective methods of analyzing them and of exploring their conceptual bound-
aries. If we were to truly list each and every explicit and implicit assumption made
MODELING NEEDS 437
in a complex model, we would see how limited the scope of each model really is.
Pointing out these limitations is not a critique, but a means of taking an inventory
and of showing how much further we have to go.
In the context of multiscale analyses, we will need effective means of integrating
coarse- and fine-grained information. Models of phenomena such as complex dis-
eases will have to be modular and expandable. No models in the near future will cap-
ture realistic systems in their entirety, which means that we must develop techniques
that functionally and validly connect the well-defined core of a modeled system with
an ill-defined outside. We need to characterize local effects of a model on global
responses in its environment and capture how these global effects feed back to the
modeled system. We must also find ways to connect different timescales in an effective
manner. We must invent efficient techniques for merging deterministic and stochastic
features that change dynamically and in spatially heterogeneous environments.
15.5 A Data-Modeling Pipeline

In the past, there was a clear distinction between what biologists did and what mod-
elers did. While division of labor has been one of the greatest inventions, there is a
real risk of losing information and insight when systems biology is strictly separated
into experiments and model analyses. As a case in point, the biological literature
contains thousands of diagrams showing how one component activates or inhibits
another component or process in a system. As soon as a modeler picks up the dia-
gram and tries to convert it into a model, the biologist’s intuition and much undocu-
mented tangential information are lost, which is a real pity. An initial step toward
circumventing this problem would be a more systematic warehousing of data
together with contextual information, as we discussed in the context of ecoinfor-
matics. However, it is difficult to see how data management alone can solve the
problem, because intuition is hard to code. Thus, a data-modeling pipeline is
needed that loses as little information as possible. Two solutions come to mind.
First, the topic could be discussed in a team of biologists and modelers, which, how-
ever, would require that almost every biologist have access to a modeler. Second, the
modeling community could create tools that would allow the biologist to enter the
information embedded in the diagram, as well as other tangential and intuitive
information, into a pipeline that leads toward a computational model. Two rudi-
mentary approaches in this direction are agent-based modeling (ABM) and con-
cept-map modeling.
ABM approaches started to become popular not long ago [40, 69]. They originated in
the business world, but are also suitable for the simulation and analysis of any autono-
mous interacting agents, which could be individuals, machines, parts of cells, mole-
cules, or a variety of other entities. ABM is a very general stochastic modeling technique,
where the behavior of discrete, identifiable agents is governed by features and rules. The
features may be almost any quantifiable characteristics, and the rules may be simple
reactive decisions or the results of complex adaptive artificial intelligence [70]. The
agents may be of different types, operate on different length- and timescales, and follow
different sets of rules. All agents operate in a common virtual environment, where they
interact with each other according to rules and protocols, and the state of each agent is
updated in discrete time steps. In biology, typical agent-based models try to emulate
population behaviors in two- or three-dimensional space, such as schooling, flocking,
or swarming in fish, birds, and ants, where individuals are autonomous, yet influenced
by their neighbors to such a degree that the collective result has been termed swarm
intelligence [71]. The simplest models of agent-based behavior are probably the cellular
automata that we discussed in Chapter 9. Models of this type, as well as corresponding
partial differential equation models, have been used for many decades to describe the
formation of patterns, for instance, in developmental biology (see, for example, [72–74]).
Free software supporting ABM includes NetLogo [75].
We have already mentioned that ABM has found interesting applications in the
area of inflammation. Another typical application of ABM is tumor growth. An exam-
ple is a model in which cells are characterized by their gene expression and pheno-
type, which includes the presence of growth factor receptors [76]. The agents (cells)
are placed on a two- or three-dimensional grid and allowed to proliferate, migrate
according to chemotactic signals, turn quiescent, or undergo apoptosis. Simulations

with the model lead to experimentally testable hypotheses. For comprehensive refer-
ences on agent-based and cellular automata models in immunology and cancer
research, see [40, 77]. A specific example is illustrated in Figure 15.11.
While agent-based models are very appealing, especially for complex
phenomena occurring in spatially heterogeneous environments, they have signifi-
cant drawbacks compared with process-based models in the form of differential or
difference equations. First, the computational demands are often high, because
many rules have to be called up very frequently [76, 77]. A possible solution of this
problem is parallel implementation of the system code. Second, the estimation of
parameter values is often complicated [76]. Finally, it is very difficult to analyze
ABMs mathematically. While it is feasible to execute very large numbers of simula-
tions, general questions about the structure of the simulated system are hard to
answer. For instance, outside comprehensive trial and error, it is difficult to deter-
mine whether a system is able to exhibit stable limit-cycle oscillations.
In concept-map modeling, the components and processes within a diagram are
formulated as symbolic equations [78]. To some degree, this process can be auto-
mated if one uses canonical models in default representations (see Chapter 4). In
(A) TH-cell differentiation modulated by dendritic cells in the lymph node

mature naive
immature dendritic cell
dendritic cell T cell
regulatory
T cell
effector
T-helper 1 cell
effector
Patches: T-helper 2 cell
various types of
dendritic cells Agents:
various types of
T-helper cells
Simple rules:
Gradient: • Dendritic cells and effector T cells release
cytokines cytokines that influence T-cell polarization
• Based on co-stimulatory, TCR/MHC-peptide
World space: and cytokine interactions, naive T cells
lymph node differentiate into effectors
(B) An example agent-based model simulation of TH-cell differentiation
point of exit
(efferent
lymphatics)
agents (TH cells) world space

(lymph node)
point of cytokine gradients

entry (high
endothelial
cytokine venuel) Figure 15.11 Definition of an agent-
gradients
based model (ABM) for the differentiation
Example rule-set: of T-helper cells in a lymph node.
(if a naive
T cell falls on Differentiation is the result of interactions
patch A, it between naive T cells and dendritic cells.
differentiates The process also involves cytokines. (A) The
patch A (DC) into a T-helper
process in ABM terminology. (B) Result of an
1 cell)
patch B (DC) ABM simulation, using NetLogo [63]. (From
Chavali AK, Gianchandani EP, Tung KS, et al.
Trends Immunol. 29 [2008] 589–599. With
TOWARD A THEORY OF BIOLOGY ⋯ OR SEVERAL THEORIES? 439
(A) (B)
X2 X1
X3 X5
X4 X6
Figure 15.12 Concept-map modeling. (A) A typical diagram of a biological system is augmented by coarse, measured or expected dynamic
responses to a typical input (green arrow). (B) The result is a system of time courses (green lines) that can be formulated as symbolic equations.
Its parameter values can be estimated using inverse methods (see Chapter 5). Blue arrows indicate flow of material. Red and orange blunted lines
indicate known and hypothesized inhibition signals, respectively.
the second step, the biologist sketches the known or expected response of each sys-
tem component to a perturbation, such as an increase in input or the knock-down
of a gene (Figure 15.12). For instance, the biologist might expect some component
to increase slowly, then speed up, and finally reach a saturation level. This informa-
tion might have been measured or might reflect the biologist’s experience and intu-
ition. The measured or alleged responses are converted into numerical time series
that show how much and how fast each component is expected to change and what
the transient response might look like. Most experienced biologists will have good
intuition regarding such responses in systems that they have been investigating for
many years. The alleged time courses are now converted into parameter values of
the symbolic model that had been formulated from the components and the con-
nectivity of the original diagram. The method for this conversion is inverse param-
eter estimation, as discussed in Chapter 5. Additional information from other
sources can be added to the model as well. Ideally, the result is a numerical model
that captures the biologist’s intuition of the features of all system components.
Under opportune conditions, it is even possible to avoid parameters and execute
nonparametric studies directly based on data [79].
Whether parametric or nonparametric, it is now easy to perform model simula-
tions. These might try to assess the likelihood of a hypothesized mechanism (such
as the orange inhibitory signal in Figure 15.12) or simulate tested and yet untested
scenarios. The results are compared with the biologist’s expectations or with new
data. Confirmatory simulation results are of course welcome, but counterintuitive
results are at least as valuable, because they point to an incomplete understanding
of some detail. A thorough analysis of ill-fitting simulations provides concrete hints
of where the problems are most likely to lie. These problems might be structural and
due to missing or ill-placed processes or regulatory signals. They may also be
numerical in the sense that rate constants or other parameters are assigned very dif-
ferent values than what they should have. The analysis may also suggest the replace-
ment of the original canonical model formulation with more complicated functions.
Importantly, the process generates a computational structure that mimics the bio-
logical phenomenon and is incomparably easier to interrogate than the actual bio-
logical system. This interrogation will often lead to new hypotheses that are to be
tested in the laboratory. An example is [8].
TOWARD A THEORY OF BIOLOGY . . . OR SEVERAL

THEORIES?
According to the tenets of mathematical logic, a theory consists of a formal language
that permits the formulation of rules of inference and a set of sentences or theorems
that are assumed or proved to be true. The assumed sentences are often called axi-
oms, which form a set of common beliefs; an example is the axiom that every integer
has a successor. If these axioms are not accepted as true, then further inferences are
usually moot, because one cannot decide whether they are true or false. Valid appli-
cation of the rules of inference to axioms and proven theorems leads to new theo-
rems that can be proved to be true, as long as one accepts the axioms.
Faced with such rigor, does it even make sense to strive for a theory of biology?
What could serve as the language, what would be the rules of inference, what kinds
of axioms could one formulate? Are there absolute truths in biology? Would the
effort even pay off in the end if we could actually reach our goal? Indeed, it might
seem quite esoteric to pose a biological theory as a long-term goal, when so many
practical and pressing questions await answers. But, as the Prussian philosopher
and psychologist Kurt Lewin once said, “there is nothing so practical as a good the-
ory” ([80], p 288). To convince ourselves of the veracity of such a bold statement, just
look at mathematics and physics. The laws of mathematics and physics have helped
us understand the world, and the theories of physics are the foundations of engi-
neering, with all its enormous implications and applications. If we had similarly
strong laws in biology, or at least in certain subspecialties, we would no longer have
to question the logic of observed phenomena and could save plenty of time and
effort if the laws would allow us to make reliable inferences.
What might such biological axioms or laws look like? Some of them we actually
have already. A famous candidate for a biological axiom is the quote omnis cellula
e cellula (every cell comes from a cell) of Rudolf Virchow, a German cell physiolo-
gist of the nineteenth century. Or consider the genetic code that assigns amino
acids to certain codons, and which is essentially ubiquitous among all organisms
on Earth (see Chapter 7 for this code and a few exceptions). Without this “codon
law,” we would have to analyze every new gene sequence! Every time we discov-
ered a new organism, we would not only have to sequence its genome, but also
have to determine how this specific organism translated its gene sequences into
peptides and proteins. We know about the Central Dogma, and even though
details of this dogma have been challenged, for instance in the context of RNA
viruses, it is a very powerful tool that we have learned to take for granted within its
domain of applicability: if we knock out a gene, we do not expect the correspond-
ing protein to be present. Mendelian genetics is a good first approximation to the
processes of inheritance, and it is simply necessary to define the boundaries of
this “theory” a little bit more stringently than the Austrian monk Gregor Mendel
did. Many biologists consider evolution a mature theory. Within a more limited
scope, Michael Savageau’s demand theory (see [81] and Chapter 14) has the char-
acter of a theory. Hiroaki Kitano suggested a theory linking disease to the robust-
ness of the physiological system [29, 82]. A small group of scientists have begun to
create a theory of chemical reaction networks [83–85]. There are certainly laws
governing the development of an embryo, and we rely on these laws almost as
much as on physical laws. There are even laws at higher levels, such as the famous
phenomenon of behavioral conditioning that was studied by the Russian physi-
ologist and psychologist Ivan Petrovich Pavlov. Much effort from different angles
is currently being devoted to discovering fundamental laws that span several bio-
logical scales (see, for example, Chapter 14).
Up to now, the search for general principles has required us to boil down complex,
realistic systems and their models to essential features such as network motifs. This
reduction of complexity presently is an art that has eluded strict general guidelines,
which suggests that systematic methods of valid model reduction should be high on
our list of priorities. In a parallel line of thinking, efforts are underway to create
mathematical concepts and techniques of topological data analysis that facilitate the
detection and analysis of highly nonlinear and often hidden geometric structures in
massive datasets. Specific tasks to be addressed with these methods include how one
might infer such high-dimensional structures in data from low-dimensional represen-
tations and how discrete data points might be assembled into global structures [86].
In the near future, any budding theories of biological subdisciplines might be
modest in scope [81, 83, 87], but they will eventually span larger domains and later
lead to a comprehensive understanding of processes such as carcinogenesis and
the dynamics of metapopulations. We may also discover that the number of hard,
unbreakable laws in biology is limited, but that it is quite possible to identify fuzzy
or probabilistic laws. As we discussed in Chapter 14, a substantial number of
REFERENCES 441
researchers have begun to focus on the formulation of motifs, design principles,

and operating principles, under which many organisms operate. Many of these will
be encountered much more frequently than expected from a random design, but
not in all cases. Thus, some of these principles may hold almost always, whereas
other patterns may simply stand out more often than not, and only in certain con-
texts. These limits in scope will become parts of theories and theorems, as is com-
mon in mathematics, where certain statements are only true for certain sets of
objects.
A strong theory of subareas of biology, for instance in gene regulation, would
have tremendous implications. In analogy with the applications of physical the-
ory in engineering, a biological theory of gene regulation would enormously
streamline synthetic biology by permitting reliable predictions of which manipu-
lations would actually lead to the desired goals. In many cases, we are already
making these predictions, and they are often actually correct (Chapter 14). With
a theory and with knowledge of its boundaries, the degree of certainty would
come close to 100%.
Beyond all speculations regarding the future, a few trends are clear. First, we live
in an unprecedented and incredibly exciting time for biology, modeling, and sys-
tems biology. Second, we have barely scratched the surface, and a book like this may
cause a good chuckle 40 or 50 years from now. Hopefully by that time, our genera-
tion of systems biologists will have earned and deserved respect for how much we
accomplished with the little we had. Above all, there is a huge and fascinating world
out there that invites us to explore it.
REFERENCES
[1] Meadows DL & Meadows DH (eds). Toward Global [14] Voit EO. Mesoscopic modeling as a starting point for
Equilibrium: Collected Papers. Wright-Allen Press, 1973. computational analyses of cystic fibrosis as a systemic
[2] Izhikevich EM. Dynamical Systems in Neuroscience: The disease. Biochim. Biophys. Acta 1844 (2013) 258–270.
Geometry of Excitability and Bursting. MIT Press, 2007. [15] Qi Z, Yu G, Tretter F, et al. A heuristic model for working
[3] Keener J & Sneyd J. Mathematical Physiology. I: memory deficit in schizophrenia. Biochem. Biophys. Acta
Cellular Physiology, 2nd ed. Springer, 2009. 1860 (2016) 2696–2705.
[4] Hodgkin A & Huxley A. A quantitative description of [16] McCulloch W & Pitts W. A logical calculus of the ideas
membrane current and its application to conduction and immanent in nervous activity. Bull. Math. Biophys. 7 (1943)
excitation in nerve. J. Physiol. 117 (1952) 500–544. 115–133.
[5] Fitzhugh R. Impulses and physiological states in [17] Hornik K, Stinchcombe M & White H. Multilayer
theoretical models of nerve membrane. Biophys. feedforward networks are universal approximators. Neural
J. 1 (1961) 445–466. Netw. 2 (1989) 359–366.
[6] Herz AV, Gollisch T, Machens CK & Jaeger D. Modeling [18] Blue Brain Project: http://bluebrain.epfl.ch/.
single-neuron dynamics and computations: a balance of [19] Dhawale A & Bhalla US. The network and the synapse: 100
detail and abstraction. Science 314 (2006) 80–85. years after Cajal. HFSP J. 2 (2008) 12–16.
[7] Nagumo JA & Yoshizawa S. An active pulse transmission [20] Destexhe A & Contreras D. Neuronal computations with
line simulating nerve axon. Proc. IRE 50 (1962) 2061–2070. stochastic network states. Science 314 (2006) 85–90.
[8] Qi Z, Miller GW & Voit EO. Computational systems analysis [21] Rubin JE. Emergent bursting in small networks of model
of dopamine metabolism. PLoS One 3 (2008) e2444. conditional pacemakers in the pre-Botzinger complex. Adv.
[9] Qi Z, Miller GW & Voit EO. The internal state of medium Exp. Med. Biol. 605 (2008) 119–124.
spiny neurons varies in response to different input signals. [22] Oshio K, Iwasaki Y, Morita S, et al. Database of Synaptic
BMC Syst. Biol. 4 (2010) 26. Connectivity of C. elegans for Computation. http://ims.dse.
[10] Gleeson P, Croo S, Cannon RC, et al. NeuroML: a language ibaraki.ac.jp/ccep/.
for describing data driven models of neurons and networks [23] Tretter F, Gebicke-Haerter PJ, Albus M, et al. Systems
with a high degree of biological detail. PLoS Comput. Biol. biology and addiction. Pharmacopsychiatry 42 (Suppl. 1)
6 (2010) e1000815. (2009) S11–S31.
[11] Kikuchi S, Fujimoto K, Kitagawa N, et al. Kinetic simulation [24] Human Connectome Project. http://www.humanconnect
of signal transduction system in hippocampal long- omeproject.org/.
term potentiation with dynamic modeling of protein [25] Bergman RN & Refai ME. Dynamic control of hepatic
phosphatase 2A. Neural Netw. 16 (2003) 1389–1398. glucose metabolism: studies by experiment and computer
[12] Nakano T, Doi T, Yoshimoto J & Doya K. A kinetic model simulation. Ann. Biomed. Eng. 3 (1975) 411–432.
of dopamine- and calcium-dependent striatal synaptic [26] Bergman RN. Minimal model: perspective from 2005. Horm.
plasticity. PLoS Comput. Biol. 6 (2010) e1000670. Res. 64 (Suppl 3) (2005) 8–15.
[13] Tretter F. Mental illness, synapses and the brain— [27] Ali SF & Padhi R. Optimal blood glucose regulation of
behavioral disorders by a system of molecules with a diabetic patients using single network adaptive critics.
system of neurons? Pharmacopsychiatry 43 (2010) S9–S20. Optim. Control Appl. Methods 32 (2011) 196–214.
[28] De Gaetano A, Hardy T, Beck B, et al. Mathematical models [50] Lipsitz LA. Physiological complexity, aging, and the path to
of diabetes progression. Am. J. Physiol. Endocrinol. Metab. frailty. Sci. Aging Knowledge Environ. 2004 (2004) pe16.
295 (2008) E1462–E1479. [51] Beard RW, Filshie GM, Knight CA & Roberts GM. The
[29] Kitano H, Oda K, Kimura T, et al. Metabolic syndrome and significance of the changes in the continuous fetal heart
robustness tradeoffs. Diabetes 53 (Suppl 3) (2004) S6–S15. rate in the first stage of labour. Br. J. Obstet. Gynaecol.
[30] Tam J, Fukumura D & Jain RK. A mathematical model of 78 (1971) 865–881.
murine metabolic regulation by leptin: energy balance and [52] Williams KP & Galerneau F. Intrapartum fetal heart rate
defense of a stable body weight. Cell Metab. 9 (2005) 52–63. patterns in the prediction of neonatal acidemia. Am. J.
[31] Eddy D, Kahn R, Peskin B & Schiebinger R. The relationship Obstet. Gynecol. 188 (2003) 820–823.
between insulin resistance and related metabolic variables [53] Ferrario M, Signorini MG, Magenes G & Cerutti S.
to coronary artery disease: a mathematical analysis. Comparison of entropy-based regularity estimators:
Diabetes Care 32 (2009) 361–366. Application to the fetal heart rate signal for the
[32] Palm NW & Medzhitov R. Not so fast: adaptive suppression identification of fetal distress. IEEE Trans. Biomed. Eng.
of innate immunity. Nat. Med. 13 (2007) 1142–1144. 33 (2006) 119–125.
[33] Zhao J, Yang X, Auh SL, et al. Do adaptive immune cells [54] Wooley JC, Godzik A & Friedberg I. A primer on
suppress or activate innate immunity? Trends Immunol. metagenomics. PLoS Comput. Biol. 6 (2010) e1000667.
30 (2008) 8–12. [55] Trevors JT. One gram of soil: a microbial biochemical gene
[34] Vodovotz Y, Constantine G, Rubin J, et al. Mechanistic library. Antonie Van Leeuwenhoek 97 (2010) 99–106.
simulations of inflammation: current state and future [56] Dam P, Fonseca, LL, Konstantinidis, KT & Voit EO. Dynamic
prospects. Math. Biosci. 217 (2009) 1–10. models of the complex microbial metapopulation of Lake
[35] Vodovotz Y. Translational systems biology of inflammation Mendota. NPJ Syst. Biol. Appl. 2 (2016) 16007.
and healing. Wound Repair Regen. 18 (2010) 3–7. [57] Turroni F, Ribbera A, Foroni E, et al. Human gut microbiota
[36] Vodovotz Y, Constantine G, Faeder J, et al. Translational and bifidobacteria: from composition to functionality.
systems approaches to the biology of inflammation and Antonie Van Leeuwenhoek 94 (2008) 35–50.
healing. Immunopharmacol. Immunotoxicol. 32 (2010) [58] NIH launches Human Microbiome Project. US National
181–195. Institutes of Health Public Release, 19 December 2007.
[37] Clermont G, Bartels J, Kumar R, et al. In silico design of www.eurekalert.org/pub_releases/2007-12/
clinical trials: a method coming of age. Crit. Care Med. nhgr-nlh121907.php.
32 (2004) 2061–2070. [59] Canadian Microbiome Initiative. http://www.cihr-irsc.
[38] Chow CC, Clermont G, Kumar R, et al. The acute gc.ca/e/39939.html.
inflammatory response in diverse shock states. Shock [60] International Human Microbiome Consortium. http://www.
24 (2005) 74–84. human-microbiome.org/.
[39] Foteinou PT, Calvano SE, Lowry SF & Androulakis IP. Modeling [61] Rondon MR, August PR, Bettermann AD, et al. Cloning the
endotoxin-induced systemic inflammation using an indirect soil metagenome: a strategy for accessing the genetic and
response approach. Math. Biosci. 217 (2009) 27–42. functional diversity of uncultured microorganisms. Appl.
[40] An G. Introduction of an agent-based multi-scale modular Environ. Microbiol. 66 (2000) 2541–2547.
architecture for dynamic knowledge representation of [62] Wilmes P & Bond PL. Microbial community proteomics:
acute inflammation. Theor. Biol. Med. Model. 5 (2008) 11. elucidating the catalysts and metabolic mechanisms
[41] Feinendegen L, Hahnfeldt P, Schadt EE, et al. Systems that drive the Earth’s biogeochemical cycles. Curr. Opin.
biology and its potential role in radiobiology. Radiat. Microbiol. 12 (2009) 310–317.
Environ. Biophys. 47 (2008) 5–23. [63] Medina M & Sachs JL. Symbiont genomics, our new tangled
[42] Moolgavkar SH & Knudson AGJ. Mutation and cancer: bank. Genomics 95 (2010) 129–137.
a model for human carcinogenesis. J. Natl Cancer Inst. [64] Vieites JM, Guazzaroni ME, Beloqui A, et al. Metagenomics
66 (1981) 1037–1052. approaches in systems microbiology. FEMS Microbiol. Rev.
[43] Moolgavkar SH & Luebeck EG. Multistage carcinogenesis 33 (2009) 236–255.
and the incidence of human cancer. Genes Chromosomes [65] Jones MB, Schildhauer MP, Reichman OJ & Bowers S. The
Cancer (2003) 302–306. new bioinformatics: integrating ecological data from the
[44] Enderling H, Hlatky L & Hahnfeldt P. Migration rules: gene to the biosphere. Annu. Rev. Ecol. Evol. Syst. 37 (2006)
tumours are conglomerates of self-metastases. Br. J. Cancer 519–544.
100 (2009) 1917–1925. [66] Genomic Science Program: Systems Biology for Energy and
[45] Folkman J & Kalluri R. Concept cancer without disease. Environment. http://genomicscience.energy.gov/.
Nature 427 (2004) 787. [67] Savage DF, Afonso B, Chen A & Silver PA. Spatially ordered
[46] Black WC & Welch HG. Advances in diagnostic imaging and dynamics of the bacterial carbon fixation machinery.
overestimations of disease prevalence and the benefits of Science 327 (2010) 1258–1261.
therapy. N. Engl. J. Med. 328 (1993) 1237–1243. [68] Burkepile DE & Hay ME. Impact of herbivore identity on
[47] Zahl PH, Maehlen J & Welch HG. The natural history algal succession and coral growth on a Caribbean reef. PLoS
of invasive breast cancers detected by screening One 5 (2010) e8963.
mammography. Arch Intern. Med. 168 (2008) 2311–2316. [69] Bonabeau E. Agent-based modeling: methods and
[48] Abdollahi A, Schwager C, Kleeff J, et al. Transcriptional techniques for simulating human systems. Proc. Natl Acad.
network governing the angiogenic switch in human Sci. USA 14 (2002) 7280–7287.
pancreatic cancer. Proc. Natl Acad. Sci. USA 104 (2007) [70] Macal CM & North M. Tutorial on agent-based modeling
12890–12895. and simulation. Part 2: How to model with agents.
[49] Lipsitz LA & Goldberger AL. Loss of “complexity” and aging. In Proceedings of the Winter Simulation Conference,
Potential applications of fractals and chaos theory to December 2006 (LF Perrone, FP Wieland, J Liu, et al. eds),
senescence. JAMA 267 (1992) 1806–1809. pp 73–83. IEEE, Monterey, CA, 2006.
FuRTHER READING 443
[71] Robinson EJH, Ratnieks FLW & Holcombe M. An agent- [79] Faraji M & Voit EO. Nonparametric dynamic modeling.
based model to investigate the roles of attractive and Math. Biosci. (2016) http://dx.doi.org/10.1016/j.
repellent pheromones in ant decision making during mbs.2016.08.004.
foraging. J. Theor. Biol. 255 (2008) 250–258. [80] Lewin K. Resolving Social Conflicts and Field Theory
[72] Ermentrout GB & Edelstein-Keshet L. Cellular automata in Social Science. American Psychological Association,
approaches to biological modeling. J. Theor. Biol. 160 1997.
(1993) 97–133. [81] Savageau MA. Demand theory of gene regulation.
[73] Wolpert L. Pattern formation in biological development. I. Quantitative development of the theory. Genetics
Sci. Am. 239(4) (1978) 154–164. 149 (1998) 1665–1676.
[74] Gierer A & Meinhardt H. A theory of biological pattern [82] Kitano H. Grand challenges in systems physiology. Front.
formation. Kybernetik 12 (1972) 30–39. Physiol. 1 (2010) 3.
[75] Wilensky U. NetLogo. http://ccl.northwestern.edu/netlogo/. [83] Horn FJM & Jackson R. General mass action kinetics.
[76] Zhang L, Wang Z, Sagotsky JA & Deisboeck TS. Multiscale Archive Rat. Mech. Anal. 47 (1972) 81–116.
agent-based cancer modeling. J. Math. Biol. 58 (2009) [84] Feinberg M. Complex balancing in general kinetic systems.
545–559. Arch. Rat. Mech. Anal. 49 (1972) 187–194.
[77] Chavali AK, Gianchandani EP, Tung KS, et al. Characterizing [85] Arceo CP, Jose EC, Marin-Sanguino A, Mendoza ER.
emergent properties of immunological systems with Chemical reaction network approaches to biochemical
multi-cellular rule-based computational modeling. Trends systems theory. Math. Biosci. 269 (2015) 135–152.
Immunol. 29 (2008) 589–599. [86] Ghrist R. Barcodes: the persistent topology of data. Bull.
[78] Goel G, Chou IC & Voit EO. Biological systems modeling Am. Math. Soc. 45 (2008) 61–75.
and analysis: a biomolecular technique of the twenty-first [87] Wolkenhauer O, Mesarovic M & Wellstead P. A plea for
century. J. Biomol. Tech. 17 (2006) 252–269. more theory in molecular biology. In Systems Biology—
Applications and Perspectives (P Bringmann, EC Butcher,
G Parry, B Weiss eds), Springer, 2007.
FuRTHER READING
Bassingthwaighte J, Hunter P & Noble D. The Cardiac Physiome: Likic VA, McConville MJ, Lithgow T & Bacic A. Systems biology:
perspectives for the future. Exp. Physiol. 94 (2009) the next frontier for bioinformatics. Adv. Bioinformatics 2010
597–605. (2010) 268925.
Calvert J & Fujimura JH. Calculating life? Duelling discourses in Lucas M, Laplaze L & Bennett MJ. Plant systems biology: network
interdisciplinary systems biology. Stud. Hist. Philos. Biol. matters. Plant Cell Environ. 34 (2011) 535–553.
Biomed. Sci. 42 (2011) 155–163. McDonald JF. Integrated cancer systems biology: current progress
del Sol A, Balling R, Hood L & Galas D. Diseases as network and future promise. Future Oncol. 7 (2011) 599–601.
perturbations. Curr. Opin. Biotechnol. 21 (2010) 566–571. Noble D. The aims of systems biology: between molecules and
Gonzalez-Angulo AM, Hennessy BT & Mills GB. Future of organisms. Pharmacopsychiatry 44 (Suppl. 1) (2011) S9–S14.
personalized medicine in oncology: a systems biology Palsson B. Metabolic systems biology. FEBS Lett. 583 (2009)
approach. J. Clin. Oncol. 28 (2010) 2777–2783. 3900–3904.
Hester RL, Iliescu R, Summers R & Coleman TG. Systems biology Sperling SR. Systems biology approaches to heart development
and integrative physiological modelling. J. Physiol. and congenital heart disease. Cardiovasc. Res. 91 (2011)
589 (2011) 1053–1060. 269–278.
Ho RL & Lieu CA. Systems biology: an evolving approach in drug Tretter F. Mental illness, synapses and the brain—behavioral
discovery and development. Drugs R D 9 (2008) 203–216. disorders by a system of molecules with a system of
Hood, L. Systems biology and P4 medicine: Past, present, and neurons? Pharmacopsychiatry 43 (2010) S9–S20.
future. Rambam Maimonides Med. J. 4 (2013) e0012. Vodovotz Y. Translational systems biology of inflammation and
Keasling JD. Manufacturing molecules through metabolic healing. Wound Repair Regen. 18 (2010) 3–7.
engineering. Science 330 (2010) 1355–1358. Werner HMJ, Mills, GB, Ram PT. Cancer systems biology: a
Kinross JM, Darzi AW & Nicholson JK. Gut microbiome–host peak into the future of patient care? Nat. Rev. Clin. Oncol.
interactions in health and disease. Genome Med. 3 (2011) 14. 11 (2014) 167–176.
Kitano H. Grand challenges in systems physiology. Front. Physiol. West GB & Bergman A. Toward a systems biology framework for
1 (2010) 3. understanding aging and health span. J. Gerontol. A Biol.
Kitano H, Oda K, Kimura T, et al. Metabolic syndrome and Sci. Med. Sci. 64 (2009) 205–208.
robustness tradeoffs. Diabetes 53 (Suppl 3) (2004) S6–S15. Weston AD & Hood L. Systems biology, proteomics, and the
Kreeger PK & Lauffenburger DA. Cancer systems biology: future of health care: toward predictive, preventative,
a network modeling perspective. Carcinogenesis 31 and personalized medicine. J. Proteome Res.
(2010) 2–8. 3 (2004) 179–196.
Kriete A, Lechner M, Clearfield D & Bohmann D. Computational Wolkenhauer O, Mesarovic M & Wellstead P. A plea for more
systems biology of aging. Wiley Interdiscip. Rev. Syst. theory in molecular biology. Ernst Schering Res Found
Biol. Med. 3 (2010) 414–428. Workshop (2007) 117–137.
Glossary
The chapter numbers refer to the first substantial discussion of the topic.
Actin Apoptosis
A protein that is one unit of an actin filament in eukaryotic cells. A natural or induced cellular mechanism resulting in the self-
Thin actin filaments and thicker myosin filaments are key compo- destruction of a cell. Also called “programmed cell death.” (Chapter 11)
nents of the contractile machinery in skeletal and heart muscle cells.
(Chapter 12) Approximation
The replacement of a mathematical function with a simpler function.
Action potential Usually the two have the same value, slope, and possibly higher
A dynamic pattern in voltage exhibited by the membranes of excit- derivatives at one chosen point, the operating point. (Chapter 2)
able cells, including neurons and heart cells. Typical action potentials
consist of a fast depolarization and a slower repolarization phase. Arrhythmia
Excitable cells exhibit a resting potential between two action poten- An irregular heartbeat or contraction pattern. (Chapter 12)
tials. Action potentials are directly associated with neuronal signal Artificial neural network (ANN)
transduction, muscle contraction, and heartbeats. (Chapter 12) A machine learning algorithm, inspired by natural neural networks,
Ad hoc model that can be trained to classify data and model complex relationships
In contrast to a canonical model, a mathematical model that is con- between inputs and outputs. For instance, upon successful training,
structed for a specific purpose and without adhering to a general an ANN may be able to distinguish healthy and cancer cells, based on
modeling framework. (Chapter 4) images or biomarker profiles. (Chapter 13)
Adaptive Atrioventricular (AV) node

1. In immunology, a system of specialized cells, primarily B and T A relatively small group of cells with pacemaker potential in the wall
lymphocytes, that learn to recognize foreign antigens like patho- of the right atrium of the heart. Normally, the AV node receives elec-
gens, eliminate them or prevent their growth, and remember them trical signals from the sino-atrial node and directs them through the
in cases of re-infections. The system is activated by the innate bundle of His and the Purkinje fibers throughout the ventricles.
immune system. (Chapter 12)
2. In systems theory, a natural system or computational model Atrium (pl. Atria)
that changes its characteristics in response to outside stimuli. One of the two upper, smaller chambers of the heart. See also
(Chapter 9) Ventricle. (Chapter 12)
Affinity Attractor
A measure of the strength of an interaction between atoms or mole- In the field of dynamical systems, a point or set of points to which tra-
cules or for the tendency of an atom or molecule to enter into a chemi- jectories converge. The most typical attractors are stable steady-state
cal reaction with unlike atoms or molecules. (Chapter 5) points and stable limit cycles. In certain chaotic systems one talks of
Agent-based modeling (ABM) ‘strange attractors.’ The antonym is a repeller. (Chapter 4)
A computational modeling method in which each component (agent)
Automaticity
acts according to specific sets of rules. ABM is particularly useful for
Term for the ability of sino-atrial (SA) node cells in the heart to depo-
spatial modeling in heterogeneous environments. (Chapter 15)
larize spontaneously, thereby serving as pacemaker cells. (Chapter 12)
Algorithm
Autonomic nervous system
A series of computational instructions designed for solving a specific
Part of the brain that, largely without conscious input, controls the
task; usually implemented as computer code or software. (Chapter 5)
function of most inner organs. (Chapter 12)
Allostasis
Auto-regulation
An abnormal state of an organism that differs from normal
Regulation of an adaptive biological system by itself. An example is
homeostasis and is associated with inferior functioning or disease.
blood pressure regulation in the brain. (Chapter 14)
(Chapter 13)
Base pair (bp)
Allosteric (regulation)
A matching couple of two nucleotides on complementary DNA or
Regulation of enzyme activity through binding of a ligand to the
RNA strands, such as cytosine–guanine (DNA and RNA) and adenine–
enzyme outside its active site. (Chapters 7 and 8)
thymine (DNA) or adenine–uracil (RNA), which are connected
Amino acid through hydrogen bonds. (Chapter 6)
One of 20 molecular building blocks used to generate peptides and
Basin of Attraction
proteins according to the genetic code. (Chapter 7)
A subset of the state space of a model with the following property: If
Antagonism/Antagonistic a simulation starts at any point within this subset, it will converge to
Literally, “working against each other.” Nonlinear interaction be- the same attractor, which may be a steady-state point, limit cycle, or
tween two system components that lessens both components’ indi- chaotic attractor. (Chapter 9)
vidual effects. Antonym to synergism/synergistic. (Chapter 11)
Bayesian (network) inference
Antibody A statistical method for deducing from data the most probable con-
See Immunoglobulin. (Chapter 7) nectivity (although not causality) between nodes in a network or
446 GLOSSARY
graph that does not contain cycles. See also Directed acyclic graph Carcinogenesis
(DAG). (Chapter 3) The process of cancer development. (Chapter 15)
Behavior (of a system) Cardiac
A collective term for changes in the state of a system, often over time Related to the heart. (Chapter 12)
and in response to a perturbation or stimulus. (Chapter 2)
Cardiomyocyte
Bi-fan A contractible heart muscle cell. See also Excitable cell and
A network motif where two source nodes send signals to each of two Pacemaker cell. (Chapter 12)
target nodes. (Chapter 14)
Carrier protein
Bifurcation (in a dynamical system) A protein transporting specific ions and other small molecules
A critical value of a parameter in a system (usually of differential between the inside and outside of a cell. The required energy is of-
equations), where the behavior of the system changes qualitatively, ten supplied in the form of ATP. Also called a pump or permease.
for instance, from a stable steady state to an unstable state surround- (Chapter 12)
ed by a stable limit cycle oscillation. (Chapter 4)
Carrying capacity
Biochemical systems theory (BST) The number of individuals an environment can sustainably accom-
A canonical modeling framework that represents biological systems modate. See also crowding and logistic growth. (Chapter 10)
with ordinary differential equations, in which all processes are
described as products of power-law functions. See also Generalized Causality analysis
mass action (GMA) system and S-system. (Chapter 4) A statistical assessment of the likelihood that one component (or
condition) of a system directly controls or causes changes in another
Biofilm component (or condition). (Chapter 3)
A microbial metapopulation consisting of many species and collec-
tively forming a thin layer. (Chapter 15) cDNA
Complementary DNA; generated from mRNA by the action of the
Bioinformatics enzymes reverse transcriptase and DNA polymerase. (Chapter 6)
A field of study at the intersection of biology and computer science
that organizes, analyzes, and interprets datasets and provides compu- Cell cycle
tational support for -omics studies. (Chapter 3) The sequence of intracellular events that ultimately lead to the divi-
sion and duplication of a cell. (Chapter 3)
Biologics
Pharmaceutically active molecules, such as antibodies, that are Central Dogma (of Molecular Biology)
derived from naturally occurring biological materials. (Chapter 13) The general (and somewhat simplistic) concept underlying the uni-
directional information transfer from DNA to mRNA (transcription)
Biomarker and then to protein (translation). (Chapter 6)
A biological or clinical measurement that is associated with, or can be
used to predict, a specific event (such as a disease). Some biomarkers Central limit theorem
are causative, while others are merely symptomatic. (Chapter 13) A fundamental law of statistics stating that the sum of many random
variables typically approaches a normal (Gaussian) distribution
Bistability (bell curve). (Chapter 12)
The property of a system to possess two stable steady states (and at
least one unstable steady state). (Chapter 9) Channel protein
A protein spanning a cell membrane and forming a hydrophilic pore
Boolean model of a system that permits very fast travel of specific ions across the membrane.
Typically a discrete model, in which the state of a multivariate sys- (Chapter 12)
tem at time t + 1 is determined by a logical function of the states of the
system components at time t. (Chapter 3) Chaos/chaotic
Without order or predictability. In the terminology of systems theory, de-
Bootstrap
terministic chaos refers to a phenomenon of some nonlinear dynamic
A statistical method for determining desired features of a dataset of
systems that are completely deterministic and numerically character-
size n. It is based on constructing and analyzing many samples of
ized, yet exhibit erratic, unpredictable behaviors. Two simulations of
size n, whose data are randomly chosen one-by-one from the data-
these systems with different, but arbitrarily close, initial values eventu-
set and replaced before the next datum is chosen. See also Jackknife.
ally show no more similarities in their responses than two simulations
(Chapter 5)
with very different initial values. (Chapter 4)
Branch-and-bound method
Chaperone
One of several sophisticated search algorithms that find global
A protein that assists in the correct folding of proteins and the assem-
optima by systematically and iteratively subdividing the param-
bly of macromolecular structures. Chaperones also protect the three-
eter space and discarding regions that cannot contain the solution.
dimensional conformation of proteins in conditions of environmen-
(Chapter 5)
tal stress. Heat-shock proteins are chaperones that protect proteins
Bundle of His against misfolding due to high temperatures. (Chapter 7)
A collection of heart cells specialized for the rapid conduction of elec-
trical signals, functionally connecting the atrioventricular node and Chemical master equation
the Purkinje fibers. (Chapter 12) A detailed ordinary differential equation description of a chemical
reaction that accounts for the stochastic nature of every molecular
Calcium-induced calcium release (CICR) event associated with this reaction. (Chapter 8)
An amplification mechanism in cardiomyocytes, in which a small
amount of calcium input to the cell leads to a large efflux of calcium Cis-regulatory DNA
from the sarcoplasmic reticulum into the cytosol, where it leads to A region of DNA that regulates the expression of a gene located on the
contraction. (Chapter 12) same DNA molecule, for instance, by coding for a transcription fac-
tor, or serving as operator. (Chapter 6)
Canonical model
A mathematical modeling format that follows strict construction Clique
rules. Examples are linear models, Lotka–Volterra models, and mod- A part of an undirected network or graph where all nodes are neigh-
els within biochemical systems theory. (Chapter 4) bors, that is, directly connected to each other. (Chapter 3)
GLOSSARY 447
Closed system Coronaries

In contrast to an open system, a mathematical model without input Arteries and veins supplying the heart muscle with blood circulation.
or loss of material. (Chapter 2) (Chapter 12)
Clustering (of genes) Correlative model
The sorting of genes by some feature, such as a similar sequence or A model that relates one quantity to another, without explaining
co-expression, usually by means of a machine learning algorithm. what causes the relationship or what the relationship means. See also
(Chapter 11) Explanatory model. (Chapter 2)
Clustering coefficient Crosstalk

A quantitative measure for the connectivity of a network, either Information exchange between two processes or pathways; typically
throughout the entire network or within the neighborhood of a given in signal transduction. (Chapter 9)
node. (Chapter 3)
Crowding
Codon A collective term for impeding interactions within a growing popula-
A triplet of nucleotides that, when translated, represent an amino tion, such as the competition for resources that lead to a slowing down
acid. (Chapter 6) of population growth, often until the carrying capacity is reached.
(Chapter 10)
Coefficient
A (typically constant) multiplicative factor (number) in some term of Curated (database)
an equation or a matrix. See also Linear. (Chapter 4) Referring to the fact that information (in the database) was checked
by an expert. (Chapter 8)
Compartment
A defined space that is separated from its surroundings, for instance, Cytokine
by a membrane. See also Compartment model. (Chapter 4) A signaling protein involved in responses to infections and inflamma-
tion. (Chapters 7 and 15)
Compartment model
Usually a linear ordinary differential equation model of the Damped oscillation
dynamics of material moving between compartments, such as the An oscillation that subsides over time. See also Limit cycle and
organs and the bloodstream in an organism. See also Pharmacokinetic Overshoot. (Chapter 4)
model. (Chapter 4)
Data mining
Complexity The process of extracting patterns from large datasets and forging
An ill-defined conglomerate of features of a system that includes its them into useful information, by combining methods from machine
number of components, processes, and regulatory signals, as well as the learning, artificial intelligence, statistics, and database management.
degree of nonlinearity and unpredictability of the system. (Chapter 4) See also Text mining. (Chapter 8)
Concept-map modeling
Degree distribution
A strategy for converting biological diagrams and coarse input-
Statistical distribution of the number of edges per node in a graph or
response information into dynamic models. (Chapter 15)
network. (Chapter 3)
Conductance (electrical)
A measure of how easily electricity can move through some material. Denature
(Chapter 12) To change or destroy the structure of a macromolecule, typically a
protein or nucleic acid, by applying extreme physical or chemical
Conformation stress. Denatured proteins are often removed by the proteasome.
The three-dimensional arrangement of atoms in a large molecule like (Chapter 11)
a protein. The conformation can change with the molecule’s chemical
and physical environment. See also Tertiary structure. (Chapter 7) Dependent variable
A system variable that is potentially affected by the dynamics of the
Constraint (optimization) system. See also Independent variable. (Chapter 2)
Any type of threshold that a component, subject to optimization, is
not permitted to cross; a condition that variables must satisfy in an Depolarization
optimization problem. (Chapter 3) A rapid change, from negative to positive voltage, in the membrane
potential of neurons, muscle cells, and heart cells during an action
Continuous potential; it is followed by repolarization. (Chapter 12)
Antonym to discrete. The feature of a function or set not to have holes
or interruptions; in systems biology usually with respect to time, and Design
sometimes with respect to space. (Chapter 4) 1. Of a natural system. Similar to a network motif, the structural and
Continuum approximation regulatory composition of a dynamic system as it is observed.
An approximation of the mechanical features of an array of many The observed design is sometimes analyzed in comparison with
similar objects with the mechanics of a continuous material; used, for hypothetical alternatives, in order to reveal design and operating
instance, in modeling heart tissue. (Chapter 12) principles.
2. Of an artificial biological system. The core subject of synthetic
Control coefficient biology.
A quantitative measure of steady-state sensitivity in metabolic con- 3. Of an experiment. A set of statistical techniques assuring that
trol analysis. (Chapter 3) results from the experiment will likely be statistically significant.
(Chapter 14)
Controlled comparison
A mathematical technique for revealing design principles, based on Design principle
comparing the designs of two similar systems in a minimally biased A structural or dynamic feature of a system that, due to its functional
fashion. (Chapter 14) superiority, is found more often than alternative designs. See also
Cooperativity Motif and Operating principle. (Chapter 14)
The phenomenon that enzymes or receptors may have multiple bind-
Deterministic (model)
ing sites, which may increase (positive cooperativity) or decrease (nega-
A model that does not allow or account for random events. See also
tive cooperativity) their affinity. Hill models are often used to represent
Stochastic model and deterministic Chaos. (Chapter 2)
this phenomenon. See also Allosteric (regulation). (Chapter 7)
448 GLOSSARY
Diastole Elementary mode analysis (EMA)

Greek term for dilation and used to describe the relaxed state of the A method for determining independent reaction paths within a
heart. See also Systole. (Chapter 12) stoichiometric network. (Chapter 14)
Difference equation Emergent property

A discrete function in which the state of a system is defined as a func- A property of a system that cannot be assigned to the features of a
tion of its state at one or more previous time points. See also Iteration single component, but only becomes possible from nonlinear inter-
and Boolean model. (Chapter 4) actions between the system components. See also Superposition.
(Chapter 1)
Differential equation
An equation or system of equations containing functions and Environment (of a model, graph or system)
their derivatives. Most dynamical systems are represented as Components and processes outside the explicit definition of a model,
differential equations. See also Ordinary differential equation graph or system. (Chapter 2)
and Partial differential equation. (Chapter 4) Enzyme
Differentiation A protein that catalyzes (facilitates) a biochemical reaction. (Chapter 7)
1. In developmental biology, the natural process of converting a cell Epidemic
with many possible roles or fates, such as a fertilized egg or stem An infectious disease that sickens large percentages of a population.
cell, into a cell with one or a few specific functions. (Chapter 11) An extremely widespread epidemic is called a pandemic. (Chapter 10)
2. In mathematics, the process of computing a derivative (slope) of a
function. (Chapter 4) Epigenetics
The relatively new field of studying heritable features that are not
Directed acyclic graph (DAG) due to changes in the nucleotide sequence of an organism’s DNA.
A graph in which all edges have a unique direction, and which does (Chapter 6)
not allow feedback or the return to any node, once this node had
been left. (Chapter 3) Error (in modeling and statistics)
See Residual. (Chapter 5)
Discrete
Antonym to continuous. The feature of a function or set to exhibit in- Evolutionary algorithm
terruptions. For instance, a discrete timescale is defined for distinct One of several machine learning methods that, inspired by natural
time points t1, t2, t3, …, but not in between, like the display of a digital evolution, search for optimal and near-optimal solutions. See also
watch. (Chapter 2) Genetic algorithm and Optimization. (Chapter 5)
DNA chip Excitable (cell)
Similar to a microarray, a tool for comparing gene expression in dif- A neuron or muscle cell that responds to electrochemical stimula-
ferent cell types or under different conditions. In contrast to a micro- tion. See also Action potential. (Chapter 12)
array, a DNA chip contains artificially constructed, short nucleotide
Excitation–contraction (EC) coupling
probes. (Chapter 6)
A fundamental process in muscle physiology, where an electrical
Dyad stimulus (usually an action potential) is converted into an increase
A small space between the cell membrane and the sarcoplasmic in calcium concentration within the cytosol of a muscle cell, which
reticulum in cardiomyocytes, where calcium-induced calcium causes a sliding motion between actin and myosin filaments and
release is controlled. (Chapter 12) leads to contraction of the cell. (Chapter 12)
Dynamical system or model Exon
A system, or model of a system, whose state and behavior changes A segment of a gene sequence that is transcribed into mRNA and sub-
over time. (Chapter 2) sequently translated into protein. See also Intron. (Chapter 6)
Dynamics (of a model) Explanatory model
The collection of changes in model features over time. (Chapter 2) A conceptual or mathematical model that explains a phenomenon
(for example, development) based on events at a lower level of orga-
Edge (in a graph) nization (for example, gene expression). (Chapter 2)
The directed or undirected connection between two nodes in a graph
or between a node and the environment of the graph. (Chapter 3) Exponential growth
A model of growth, or growth law, that can be described by an ex-
Eigenvalue ponential function with a positive growth rate. The corresponding
One of a group of complex numbers that characterize various aspects process with a negative rate is called exponential decay. (Chapter 10)
of a linear or linearized system, such as its stability at a steady state
and the different speeds with which different variables evolve over Expressed sequence tag (EST)
time. (Chapter 4) A gene fragment of about 300 nucleotides that has been sequenced
from cDNA. (Chapter 6)
Elasticity
A quantitative measure of the sensitivity of an enzyme or reaction to Extrapolation
changes in the local environment, for instance, induced by changes in The prediction of model responses outside the range for which actual
substrate; used in metabolic control analysis. Equivalent to the con- data are available. (Chapter 2)
cept of a kinetic order in biochemical systems theory. (Chapter 3)
False negative
Electrocardiogram (ECG, EKG) The result of a test indicating a negative outcome even though the
A reading from several pairs of electrodes, attached to the torso and outcome is in actuality positive. For instance, a test may suggest the
limbs, of the electrical activity of the heart. (Chapter 12) absence of disease even though the person has the disease. See also
False positive. (Chapter 13)
Electrochemical
Related to electrical events associated with ions. (Chapter 12) False positive
The result of a test indicating a positive outcome even though the out-
Electrophysiology come is in actuality negative. For instance, a test may suggest cancer
Study of the normal functioning of electrical events in the body. even though the person does not have cancer. See also False negative.
(Chapter 12) (Chapter 13)
GLOSSARY 449
Feature selection Gene annotation

A technique in statistics and machine learning for reducing the num- The assignment of a protein or function to a gene sequence. See also
ber of variables to a subset of most influential variables. (Chapter 5) Gene ontology. (Chapter 6)
Feedback Gene ontology (GO)
A signal affecting a preceding reaction step or process in a generic A widely accepted, controlled vocabulary for gene annotation.
pathway. (Chapter 2) (Chapter 6)
Feedforward Generalized mass action (GMA) system
A signal affecting a downstream reaction step or process within a A modeling format within biochemical systems theory, based
generic pathway. (Chapter 14) on ordinary differential equations, in which every process is
represented with a product of power-law functions. See also
Filament (muscle cell)
S-system. (Chapter 4)
A microfiber composed of proteins like actin and myosin, which
endows muscle cells with contractility. (Chapter 12) Genetic algorithm
A machine learning method that, inspired by the process of genet-
Finite element approach ic inheritance, recombination, and mutation, searches for optimal
A set of methods for assessing the dynamics of fluids, materials, or sol- and near-optimal solutions. See also Evolutionary algorithm and
id bodies, based on subdividing these into very small units (elements) Optimization. (Chapter 5)
and studying the behavior of and between these elements. Often used
to analyze partial differential equations as well as changes within Genetic code
bodies that have complicated shapes. (Chapter 12) The assignment of each natural amino acid to one or more codons of
nucleotides in DNA or RNA. Because of the near-universality of the
Fitness genetic code, amino acid sequences in peptides and proteins can reli-
1. In genetics, a quantitative measure of superiority of a gene, its ably be predicted from their coding genes. (Chapter 6)
gene product, or the organism possessing the gene.
2. In Optimization, a quantitative measure of superiority of one Genetic engineering
solution over another. (Chapter 5) The targeted manipulation of genes toward a goal of (industrial)
interest. See also Metabolic engineering and Synthetic biology.
Fitting (of a model) (Chapter 14)
The automatic or manual adjustment of parameter values so that a
model matches observation data as closely as possible. See also param- Genome
eter estimation, optimization, residual, and validation. (Chapter 5) The totality of DNA in an organism. In addition to genes coding for
proteins, much of the DNA is used for other purposes, including
Fluid mechanics information regarding regulatory RNA. See also Proteome and -ome.
The study of the movement of particles within liquids, often us- (Chapter 6)
ing partial differential equations and finite element approaches.
(Chapter 12) GFP
Green fluorescent protein. A protein whose DNA sequence can be
Flux
inserted within the protein-coding region of a target gene, permit-
The amount of material (number of molecules) flowing through a
ting the localization of gene expression through a fluorescence assay.
pool (variable), pathway, or system during a defined time period.
Fluorescent proteins with many other colors are available now.
(Chapter 3)
(Chapter 6)
Flux balance analysis (FBA)
Global optimum
A modeling approach for (large) metabolic networks, which at a
The best possible point (with respect to some defined criterion or
steady state balances metabolic fluxes entering and leaving nodes,
metric). Finding the global optimum for a nonlinear system is a
according to the stoichiometry of each reaction, and which assumes
difficult optimization problem of computer science. See also Local
that the cell or organism optimizes some feature, such as its growth
optimum and Branch-and-bound method. (Chapter 5)
rate. See also Metabolic flux analysis. (Chapter 3)
Frank–Starling law Goldman–Hodgkin–Katz equation
A relationship between blood volume, pressures, forces, and wall An equation describing the equilibrium potential of a cell membrane
tensions in the heart. (Chapter 12) as a function of internal and external ion concentrations and perme-
ability constants. (Chapter 12)
Funny current
Relatively slow, mixed sodium–potassium current, entering a cell of G-protein
the sino-atrial node and triggering spontaneous depolarization. One of several specific proteins on the inside of the cell membrane
(Chapter 12) that interact with G-protein-coupled receptors in numerous signal
transduction processes. (Chapter 9)
Fuzzy (logic)
A set of logical rules that are applied in a probabilistic (rather than a G-protein-coupled receptor (GPCR)
deterministic) fashion. (Chapter 15) A class of membrane-associated receptors that are involved in
uncounted signal transduction and disease processes. (Chapter 9)
Gain
The sensitivity of a system response with respect to a small change in Graph
an independent variable. (Chapter 4) A visual or mathematical representation of a network consisting of
nodes (vertices) and edges connecting some of the nodes with each
Gap junction other. (Chapter 3)
A channel that connects the cytosols of two neighboring cells. The
channel consists of two protein complexes that form half-channels Graph theory
spanning the cell membranes. The two half-channels meet in A branch of mathematics concerned with graphs. Graph theory has
the extracellular space between the neighboring cells. See also found rich applications in biological network analysis. (Chapter 3)
Transmembrane protein. (Chapter 7)
Growth law
Gate (logic) A typically simple function describing changes in the size of a cell, organ-
A logical conjunction between two signals. In the and gate, the signal ism or population over time. The best-known examples are exponential
is only propagated if both incoming signals are active. (Chapter 14) and logistic growth, but many others exist. (Chapter 10)
450 GLOSSARY
Growth rate Initial value

Parameter in a growth law characterizing the speed of growth. A numerical start value that needs to be defined for every dependent
(Chapter 10) variable in a system of differential equations, before the system can
be solved numerically. (Chapter 4)
Half-life
The time it takes for the amount of some material to degrade from Innate (immune system)
100% to 50%, under the assumption of exponential decay. (Chapter 4) A generic defense system in a variety of multicellular organisms that
protects the host from infection by pathogens. In vertebrates, cells of
Heat shock protein the innate immune system, such as macrophages and neutrophils,
One of a class of specific proteins that are activated under heat stress and rush to the site of infection and destroy invaders. The innate immune
protect other proteins from unfolding or denaturing. (Chapter 11) system activates the adaptive immune system, which learns to rec-
Hill function ognize and remember pathogens. (Chapter 15)
A type of kinetic rate law that is well suited for modeling allosteric Intron
processes and cooperativity involving n binding sites of an enzyme. A segment of a gene sequence that is transcribed into RNA, but later
The number n is called the Hill coefficient; n > 1 means positive co- spliced out from the mature RNA. In protein coding genes, introns are
operativity, n < 1 negative cooperativity. For n = 1, the Hill function not translated into protein. See also Exon. (Chapter 6)
becomes a Michaelis–Menten rate law. Often used to describe sig-
moidal behavior. (Chapter 8) Ion
An atom or molecule that carries an electrical charge due to a differ-
Hit (drug development) ence between its numbers of electrons and protons. (Chapter 12)
A natural or artificially created molecule, investigated during drug
discovery and development, that interacts with a drug target and Irreversible (reaction)
interrupts a disease process. (Chapter 13) Feature of a (biochemical) reaction not permitting a back-reaction.
Absolutely irreversible reactions are rare, but the reverse direc-
Hodgkin–Huxley model tion is often negligible because of unfavorable energetics. See also
The first mathematical model describing electrical activity, trans- Reversible reaction. (Chapter 8)
membrane currents, and action potentials in neurons and other
excitable cells. (Chapter 12) Isozyme
One of several different enzymes that catalyze the same biochemi-
Homeostasis cal reaction(s), sometimes with different efficiency and sensitivity to
Maintenance of the normal (steady) state of an organism or physio- inhibitors or modulators. (Chapter 11)
logical system, which is achieved through the action of often complex
internal control systems. See also Allostasis. (Chapter 13) Iteration
A recurring step or sequence of steps in an algorithm or simulation,
Homogeneous
usually with slightly changed settings. (Chapter 2)
Possessing the same features throughout. (Chapter 10)
Jackknife
Hub
Similar to the bootstrap method, the jackknife is a statistical method
A particularly well connected, and usually important, node within a
for determining desired features of a dataset. It is based on the repeat-
graph, network, or system. (Chapter 3)
ed estimation of the desired feature when one or more data points are
Hybridization randomly left out. (Chapter 5)
A core technique in numerous experimental methods for the identifica-
Jacobian
tion of identical or similar DNA or RNA nucleotide sequences; based on
A matrix, consisting of partial derivatives, that approximately repre-
the fact that nucleotide strands specifically bind to their complementary
sents the dynamics of a nonlinear system close to some operating
sequence through base pairs, such as cytosine and guanine. (Chapter 6)
point. (Chapter 4)
Hysteresis
A property of nonlinear systems with two or more stable steady Kinetic order
states where the transient toward one of these states depends on the 1. In elemental chemical kinetics, the number of molecules of the
recent history of the dynamics of the system. (Chapter 9) same species undergoing a reaction.
2. In biochemical systems theory, the exponent of a variable in a
Identifiability power-law function, quantifying the effect of the variable on the
A property of a system and a dataset indicating whether the system process represented by the function; equivalent to the concept of
structure or parameters can be uniquely determined from the data- elasticity in metabolic control analysis. (Chapter 4)
set. See also Sloppiness. (Chapter 5)
Kinetic rate law
Immunoglobulin A function describing how the dynamics of a biochemical reaction is
Also called an antibody. A protein produced by white blood cells affected by substrates, enzyme, and modulators. See also Michaelis–
that is either attached to a B cell of the immune system or circulates Menten kinetics and Hill function. (Chapter 8)
through the blood and other bodily fluids. Immunoglobulins detect
and neutralize invaders such as bacteria, viruses, and foreign pro- Kinetics (in biochemistry)
teins. (Chapter 7) Pertaining to the dynamics of a biochemical reaction. (Chapter 8)
Immunoprecipitation Latin square or hypercube sampling
A technique for concentrating, isolating, and purifying proteins A statistical design and sampling method that circumvents large num-
from a cell extract or protein mixture by means of specific anti- bers of analyses that would be necessary if every combination of num-
bodies. The protein–antibody complex is extracted from the solu- bers in a complete grid were used. Instead, only one number per row
tion, the antibodies are removed, and the protein can be analyzed. and per column of the grid is selected. In higher dimensions, the grid is
(Chapter 7) replaced with a cube or hypercube, where the latter is the analog to a
cube in more than three dimensions. (Chapter 5)
Independent variable
A system variable that is not affected by the dynamics of the system Lead (drug development)
itself, but potentially by the environment or the experimenter. In A potential drug molecule, such as a hit with improved potency, that
many cases, an independent variable is constant during a given interrupts a disease process and has desirable properties related to
experiment. See also Dependent variable. (Chapter 2) efficacy, pharmacokinetics, and side effects. (Chapter 13)
GLOSSARY 451
Least-squares method Matrix (in linear algebra; pl. Matrices)

A statistical regression technique that optimally matches a model to An array of numbers that typically correspond to the coefficients in a
a given dataset. Optimality is defined based on the minimum of the set of linear equations; see also Vector. Many aspects within the fields
sum of squared differences between each data point and the corre- of linear algebra, multivariate calculus, systems analysis, and statis-
sponding model value. See also Residual. (Chapter 5) tics are based on vectors and matrices. (Chapter 4)
Limit cycle Metabolic burden

An attractor or repeller in the form of an oscillation in a dynamical The cost to the organism for maintaining certain levels of mRNAs,
system whose amplitude does not change over time. In the case of a proteins, and metabolic products, especially if they are expendable.
stable (unstable) limit cycle, nearby oscillations eventually converge (Chapter 14)
to (diverge away from) the attractor (repeller). (Chapter 4)
Metabolic control analysis (MCA)
Linear A framework of theorems and methods, based on sensitivity analy-
The mathematical property of a function or system to be describable sis, for the assessment of control in metabolic, genetic, and signaling
with sums of variables that are multiplied with constant coefficients systems at steady state. (Chapter 3)
(numbers). If f is a linear function of x, and x1 and x2 are any values of
Metabolic engineering
x, then the superposition principle f(x1 + x2) = f(x1) + f(x2) holds. The
A set of techniques from molecular biology, genetics, and engineering
antonym is a nonlinear function or system, where the above relation-
for manipulating a culture of microorganisms toward a desired goal,
ships do not hold. (Chapter 2)
such as the production of a valuable amino acid. (Chapter 14)
Linearization Metabolic flux analysis
The linear approximation of a system at a chosen operating point. A collection of experimental (labeling) and computational (flux bal-
The linearization is characterized by the Jacobian matrix. (Chapter 4) ance analysis-like) techniques for the assessment of flux distribu-
Local minimum (optimum) tions in metabolic networks. (Chapter 8)
A state of a system that is better (with respect to some defined cri- Metabolic pathway
terion or metric) than all states close by. However, farther away, the Any prominent sequence of (typically enzyme-catalyzed) reaction
system may have even better states; see also Global optimum. Local steps, involving the synthesis or degradation of organic molecules.
optima often prevent computer algorithms from finding the global Metabolic pathways are often used to organize the complexity of met-
optimum. (Chapter 5) abolic networks. (Chapter 8)
Logistic growth Metabolic profile
A sigmoidal model of growth, or growth law, that is initially exponen- A comprehensive set of metabolite concentrations characterizing a
tial but ultimately approaches an upper value, called the carrying ca- metabolic pathway or organism at one time point or a series of time
pacity. The deviation from exponential growth is often attributed to points. (Chapter 8)
crowding. (Chapter 10)
Metabolomics
Lotka–Volterra model The simultaneous study of many or all metabolites in a system. See
A canonical model, based on ordinary differential equations and also -ome and -omics. (Chapter 8)
used primarily in ecological systems analysis, where all processes are
represented as products of two variables and a rate coefficient. Each Metagenomics
equation j may also contain just the jth variable with a coefficient. Methods of genomics applied to a metapopulation. A collective term
(Chapter 4) for simultaneous genome analyses of many (coexisting) species in
environments such as soil, the oceans, sewage pipes, or the human
Machine learning gut. (Chapter 15)
A computational branch of the field of artificial intelligence, which
Metapopulation
trains algorithms to classify data according to given criteria and to
A collection of coexisting populations of different species. Commonly
reveal patterns within complex datasets. See also Parameter estima-
used in the context of microbial species. (Chapter 15)
tion, Data mining, and Text mining. (Chapter 5)
Michaelis–Menten mechanism (rate law)
MAPK A conceptual and mathematical model (rate law) of an enzyme-
Mitogen-activated protein kinase. One of a class of enzymes at the catalyzed reaction, in which substrate and enzyme reversibly first
core of many signaling cascades in eukaryotes. Typically, phosphor- form an intermediate complex, before the product is generated and
ylation events of different MAPKs occur at three hierarchical levels, the enzyme recycled. A suitable quasi-steady-state assumption con-
thereby transducing a signal, for instance, from the cell membrane to verts the describing system of mass-action differential equations
the nucleus. (Chapter 9) into a simple function that quantifies the rate of product formation in
Markov model terms of the substrate concentration. Its two parameters are Vmax, the
Usually a discrete, probabilistic mathematical model, in which the maximal rate of product formation, and the Michaelis constant KM,
state of a system is determined by the previous state of the system and which corresponds to the substrate concentration for which this rate is
a (fixed) set of transition probabilities for moving the system from any Vmax/2. A mathematical generalization is the Hill function. (Chapter 8)
one state to any other state. (Chapter 4) Microarray (technology)
A set of methods for comparing gene expression in different cell types
Mass spectrometry (MS)
or under different conditions. Each microarray consists of a solid sur-
An experimental technique used to characterize mixtures of
face that holds very small probes of known DNA (or sometimes other
molecules based on their mass and an applied electrical charge.
molecules). See also DNA chip and RNA-Seq. (Chapter 6)
(Chapter 7)
Microbiome
Mass-action (kinetics) A microbial metapopulation, usually living within the human body;
A simple quantitative description of the dynamics of chemical reac- for instance, in the oral cavity, vagina, or gut. (Chapter 15)
tions, consisting of the product of a rate constant and all substrates
involved in the reaction. If two substrate molecules of the same spe- MicroRNA
cies are needed for the reaction, the substrate enters the mass-action A short noncoding RNA of between 20 and 30 nucleotides that serves
function with a power of 2. See also Kinetics and Generalized mass a number of regulatory roles, including the silencing of genes. See also
action system. (Chapter 8) Small RNA. (Chapter 6)
452 GLOSSARY
Model (in biology) Network

1. In biological experimentation, the use of a particular species A collection of nodes (vertices) and connecting edges. (Chapter 3)
as a representative for a larger domain of organisms, such as a
“mouse model.” Network motif
2. Conceptual: a formalized idea of how components of a system See Motif. (Chapter 14)
might be related to and interact with each other. Neurotransmitter
3. Mathematical: a set of functions or (differential) equations that in A chemical such as dopamine or serotonin that is responsible for the
some manner represent a biological phenomenon. transmission of specific signals between neurons across a synapse.
4. Computational: the use of algorithms supporting the modeling of (Chapter 15)
a biological phenomenon. (Chapter 2)
New chemical entity (NCE)
Modeling (in biology)
A potential drug molecule, intensely investigated in the early phases
The conversion of a biological phenomenon into a mathematical or
of drug development, that had not been approved by the FDA in any
computational construct and the subsequent analysis of this construct.
prior application. (Chapter 13)
(Chapter 2)
NMR spectroscopy
Module
Abbreviation for nuclear magnetic resonance spectroscopy. A tech-
A more or less confined or autonomous (sub-)model that is part of a
nique that exploits the fact that atomic nuclei with an odd number
larger model. (Chapter 2)
of protons and neutrons spin on their own axes. This spin is assessed
Molecular dynamics with strong magnets and can be interpreted to yield the structure of
A computationally intensive simulation technique for exploring the molecules, such as proteins. The same method can be used for nonin-
location and motion of each atom in a large molecule such as a pro- vasive in vivo analyses of labeled metabolites. (Chapter 7)
tein. (Chapter 7)
Node
Monte Carlo simulation A variable or pool within a network or graph, which is connected by
A statistical technique for exploring possible, likely, and worst-case edges with other nodes in a graph, or with the environment of the
behaviors of complex systems that is based on thousands of simu- graph. Also called a vertex. (Chapter 3)
lations with different randomized combinations of model settings.
(Chapter 2) Noise
Summary expression for unexplained, apparently random errors in
Motif data. See also Residual. (Chapter 5)
A recurring structural feature of a network or system that is found
more often than one would expect in a corresponding, randomly Noncoding DNA/RNA
composed system. See also Design principle. (Chapter 14) DNA or RNA that is ultimately not translated into protein. (Chapter 6)
Multiscale modeling Nonlinear

The construction of models that simultaneously capture essential sys- Antonym of linear. The property of a curved appearance of a func-
tem features and behaviors at different temporal and organizational tion or a system. If f is a nonlinear function of x, and if x1 and x2 are
scales, such as gene expression, physiology, and aging. (Chapter 12) values of x, then f(x1 + x2) is generally not equal to f(x1) + f(x2). See also
Superposition and Synergism. (Chapter 2)
Multistage cancer model
One of several mathematical models formalizing the assumption that Nonlinear system
a normal cell experiences several discrete changes before becoming Any mathematical model that exhibits nonlinear features (possibly
a cancer cell. (Chapter 15) in addition to linear features). (Chapter 4)
Multivariate Northern blot

Simultaneously depending on several variables. (Chapter 2) A technique for measuring gene expression based on produced
mRNAs. See also RT-PCR and Western blot. (Chapter 6)
Myocardium
Nucleoside
Heart muscle. (Chapter 12)
A building block of DNA or RNA, consisting of a base and a sugar
Myocyte (deoxyribose or ribose, respectively). If a phosphate group is attached
Muscle cell. In the heart, it is called a cardiac myocyte or cardiomyo- as well, a nucleoside becomes a nucleotide. (Chapter 6)
cyte. See also Myofibril. (Chapter 12) Nucleotide
Myofibril A building block of DNA or RNA, consisting of a base, a sugar (de-
The basic contractile unit of a myocyte, consisting of actin, myosin, oxyribose or ribose, respectively), and a phosphate group. DNA bases
and other proteins. (Chapter 12) are adenine (A), cytosine (C), guanine (G), and thymine (T). In RNA,
thymine is replaced by uracil (U). (Chapter 6)
Myosin
A motor protein that forms filaments in eukaryotic cells. Together with Nullcline
actin filaments, myosin filaments are key components of the contrac- A set of points in a phase-plane plot containing all situations where a
tile machinery in skeletal and heart muscle cells. (Chapter 12) particular variable has a slope (“cline”) of zero (“null”) and therefore
does not change. Nullclines divide the space of system responses into
Na+–Ca2+ exchanger sections of qualitatively similar behaviors. (Chapter 10)
A transport protein for ions, which exchanges calcium inside an ex-
citable cell against sodium; in particular, after an action potential. Objective function
(Chapter 12) A function that is maximized (or minimized) in an optimization task.
In parameter estimation, the objective function is often the sum
Na+–K+ pump of squared errors between a model and the modeled data. See also
A transport protein that pumps Na+ out of an excitable cell, while K+ is Residual. (Chapter 5)
pumped in. The pump can also run in reverse. (Chapter 12)
-ome, -omics
Nernst equation Suffix added to words in order to indicate comprehensiveness or to-
An equation governing properties of electrochemical gradients. tality. Thus, the proteome consists of all proteins in a cell or organ-
It describes free energy in terms of differences in ion concentrations ism. One of the oldest -omes is the rhizome, the totality of all roots of
across a membrane. (Chapter 12) a plant. -ome is similar to the suffix -oma, which is used to describe
GLOSSARY 453
tumors. Some new word concoctions using -ome and -omics are quite represent time and one or more space coordinates. See also Differential
silly. See also Genome. (Chapter 6) equation and Ordinary differential equation. (Chapter 4)
Ontology Path length
A controlled vocabulary with generally accepted definitions of terms The number of consecutive edges connecting one node with another
within a limited domain of knowledge. A well-known example is gene node within a graph. Of interest is often the longest minimal path
ontology. See also Gene annotation. (Chapter 6) length in a graph. (Chapter 3)
Open system Peptide
In contrast to a closed system, a mathematical model with input and/ A chain of up to about 30 amino acids that is deemed too short to be
or output. (Chapter 2) called a protein, but otherwise has the same properties. (Chapter 7)
Operating point Peptide bond
The point at which a function and an approximation have the same The chemical (covalent) bond between amino acids in a peptide or
value, slope, and possibly higher derivatives. (Chapter 4) protein, in which the carboxyl group of one amino acid reacts with the
amino group of the neighboring amino acid. (Chapter 7)
Operating principle
The dynamic analog to a design principle or motif: a natural strategy Personalized (precision) medicine
for solving a task that is superior to alternative strategies. (Chapter 14) An emerging approach to health and disease that uses genom-
ic, proteomic, metabolic, and physiological biomarkers for the
Operon
custom-tailored prevention and treatment of an individual’s disease.
The clustered arrangement of (bacterial) genes under the control of
(Chapter 13)
the same promoter. Operons often lead to co-expression of function-
ally related proteins. (Chapter 6) Perturbation
Optimization A usually rapid shift in conditions applied either to a biological sys-
The mathematical or computational process of making a system tem or to a mathematical model. (Chapter 2)
behave as closely as possible to a given objective, for example, by Petri net
minimizing the error between some output of the model and the A specific formalism for modeling networks and systems, consisting
desired output. See also Objective function, Parameter estimation, of two distinct types of variables: pools and reactions. (Chapter 3)
and Residual. (Chapters 5 and 14)
Pharmacogenomics
Orbit The study of interactions between gene expression and drugs.
A closed trajectory. If a dynamical system starts at some point of an (Chapter 13)
orbit, it will reenter this point infinitely often again. An example is a
limit cycle. (Chapter 4) Pharmacokinetics/Pharmacokinetic model
The dynamics of a drug within an organ, body, or cell culture; a math-
Ordinary differential equation (ODE) ematical representation of this dynamics. A notable special case is a
A type of differential equation, or system of differential equations, in physiologically based pharmacokinetic (PBPK) model, which is a
which all derivatives are taken with respect to the same variable, which hybrid between a compartment model and a set of kinetic models
is usually time. See also Partial differential equation. (Chapter 4) within the compartments. (Chapter 13)
Oscillation Phase diagram
The behavior of a function (or dynamical system), where the value A rough sketch of domains in which a system exhibits different
of the function (or of a system variable) alternates between phases qualitative behaviors. For instance, the system may exhibit oscilla-
with positive and negative slopes. See also Damped oscillation, tions in one domain but not in another. See also Phase-plane plot.
Overshoot, and Limit cycle. (Chapter 4) (Chapter 10)
Overfitting Phase-plane analysis/plot
The fact that a model with (too) many parameters permits a good A graphical representation of a dynamical system, typically with two
representation of a training set of data, but not of other datasets. dependent variables, where one variable is plotted against the other,
(Chapter 5) rather than plotting both variables as functions of time. The plot is di-
Overshoot vided by nullclines into domains of qualitatively similar behaviors.
The dynamics of a function or system that temporarily exceeds its See also phase diagram. (Chapter 10)
ultimate final value. (Chapter 4) Phylogeny, Phylogenetics
Pacemaker cell Pertaining to studies of the relatedness of species throughout evolu-
Cells, primarily in the sino-atrial node of the heart, that depolarize tion. (Chapter 6)
spontaneously and therefore set the pace of the heartbeat. The sig- Physiologically based pharmacokinetic (PBPK) model
nal from the sino-atrial node eventually reaches all cardiomyocytes. A compartment model with biologically meaningful compart-
See also Atrioventricular node, Bundle of His, and Purkinje fiber. ments, such as organs, tissues, or the bloodstream. Often used to
(Chapter 12) assess the fate of drugs within a higher organism. The dynamics of
Parameter (of a model) drugs within compartments may be represented with kinetic models.
A numerical quantity in a mathematical model that is constant (Chapter 13)
throughout a given computational experiment or simulation. Its Polysaccharide
value may be altered in a different experiment. See also Parameter A linear or branched polymer consisting of many sugar (carbohy-
estimation. (Chapter 2) drate) molecules of the same or different type. (Chapter 11)
Parameter estimation Post-translational modification (PTM)
A diverse set of methods and techniques for identifying numerical Any chemical change to a peptide or protein after it has been trans-
values of model parameters that render the model representation of lated from mRNA and that is not encoded in the RNA sequence.
a given dataset acceptable. (Chapter 5) (Chapter 7)
Partial differential equation (PDE) Power function/Power-law function
An equation (or system of equations) that contains functions and de- A mathematical function consisting of the product of a (non-nega-
rivatives with respect to two or more variables, which, for instance, may tive) rate constant and one or more positive-valued variables that
454 GLOSSARY
are each raised to some real-valued exponent, which in biochemi- sumption permits the explicit formulation of Michaelis–Menten and
cal systems theory is called a kinetic order. See also Canonical Hill rate laws as functions, rather than systems of differential equa-
model. (Chapter 4) tions. (Chapter 8)
Power-law approximation Quorum sensing (QS)
Linear approximation of a function within a logarithmic coordinate A mechanism with which microbes collectively respond to their sens-
system, which always results in a (possibly multivariate) power-law ing that many other microbes in their environment send out chemical
function. (Chapter 4) signals. (Chapter 9)
Power-law distribution Random
Specific probability distribution that follows a power-law function Containing an aspect that is unpredictable, possibly because of lim-
with negative exponent. An example of recent prominence is the ited information. (Chapter 2)
distribution of nodes with certain numbers of edges in networks ex-
hibiting the small-world property. (Chapter 3) Random variable
Variable in a stochastic context, whose specific value is determined
Prediction (of a model) by a probability distribution. (Chapter 2)
A feature of a system, such as a numerical or approximate value of
a system variable, that has not been measured experimentally but is Rate constant
computed with the model. (Chapter 2) A positive or zero-valued quantity, usually in chemical and biochemi-
cal kinetics, representing the turnover rate of a reaction. See also
Predictive health Kinetic order, Elasticity, and Mass-action kinetics. (Chapter 4)
A novel paradigm of health care whose goal is maintenance of health
rather than treatment of disease. Predictive health heavily relies on Reaction (in biochemistry)
personalized assessments of biomarkers that begin when the indi- The conversion of a molecule into a different molecule, a process that
vidual is still apparently healthy and follows him or her into old age is often catalyzed by an enzyme. See also Reversible, Irreversible,
and, possibly, chronic disease. (Chapter 13) Kinetic, and Rate constant. (Chapter 8)
Probabilistic Receptor
Affected by random (stochastic) influences. (Chapter 2) A protein on the surface of (or spanning) a cell membrane that senses
external signals which, if strong enough, trigger a signal transduction
Probability distribution process. Prominent examples are G-protein-coupled receptors and
A succinct, collective characterization of all values that a random Toll-like receptors. See also Transmembrane protein. (Chapter 7)
variable may assume, along with their likelihoods. (Chapter 2)
Receptor tyrosine kinase (RTK)
Profile (of metabolites or proteins) A member of an important class of transmembrane proteins that
A collection of features characterizing many or all metabolites (or pro- transfer a phosphate group from a donor to a substrate and by do-
teins) of a biological system at one time point or at a series of time ing so transduce a signal from the outside of a cell to the inside.
points. See also Time-series data. (Chapter 8) (Chapter 7)
Proteasome Recursion
A cellular organelle consisting of a large protein complex and respon- In mathematics, a process characterized by a temporal sequence of
sible for the disassembly of proteins that had been tagged by ubiqui- states of a system that are defined in terms of the states of the system
tin. The cell reuses the peptides and amino acids generated in the at one or more earlier time points. (Chapter 10)
process. (Chapter 7)
Redox
Proteome Contraction of the terms reduction and oxidation. Used as an ad-
A collective term for all proteins in a cell, organism, or otherwise de- jective referring to the state, or change in state, of a chemical reac-
fined biological sample. See also Genome and -ome. (Chapter 7) tion system with respect to losses or gains of electrons. For example,
the physiological redox state is changed under oxidative stress.
Pump (in a cell membrane) (Chapter 11)
A carrier protein, also called a permease, that transports specific
ions across a cell membrane. (Chapter 12) Reductionism
The philosophy that the functioning of a system can be understood
Purkinje fibers by studying its parts and the parts of parts, until the ultimate build-
Modified heart muscle cells that rapidly conduct electrical ing blocks are revealed. Systems biology maintains that reductionism
impulses from the sino-atrial and atrioventricular nodes to the must be complemented with a reconstruction phase that integrates
excitable cardiomyocytes in the ventricles. See also Bundle of the building blocks into a functioning entity. (Chapter 1)
His. (Chapter 12)
Refractory period
Qualitative analysis A time period, typically toward the end of an action potential, during
A type of mathematical analysis that depends only minimally on which an excitable cell cannot be excited again. (Chapter 12)
specific numerical settings, such as parameter values, and therefore
tends to yield rather general results. (Chapter 10) Regression
A statistical technique for matching a model to experimental data in
Qualitative behavior (of a model) such a way that the residuals are minimized. (Chapter 5)
A model response that is not characterized by specific numerical fea-
tures, but rather as a member of a distinct class of features, such as Repeller
damped oscillations or limit cycles. (Chapter 2) In the field of dynamical systems, a set of points from which tra-
jectories diverge. The most typical repellers are unstable steady-
Quantitative structure–activity relationship (QSAR) state points and unstable limit cycles. The antonym is attractor.
A specific property of a molecule, such as drug activity, that is expe- (Chapter 4)
rientially or computationally inferred from its chemical composition.
(Chapter 13) Repolarization
As part of an action potential, the return in voltage across the mem-
Quasi-steady-state assumption (QSSA) brane of a neuron, muscle cell, or cardiomyocyte from a depolarized
The assumption that an intermediate complex between a substrate state to the relaxed state, which in most cells corresponds to the rest-
and an enzyme is essentially constant during a reaction. The as- ing potential. (Chapter 12)
GLOSSARY 455
Repressilator a cloud of dots. An example consists of math and physics scores of

A system of three genes (A, B, and C), where A represses B, B represses a class of high-school students. For each student, the math score
C, and C represses A. (Chapter 14) could be the x-coordinate and the physics score the y-coordinate of a
unique point that represents this student. (Chapter 5)
Residual (error)
The remaining difference (error) between a model and data, after the Search algorithm
model has been fitted. See also Regression and Least-squares meth- A set of computing instructions for finding an optimal solution, con-
od, and Parameter estimation (Chapter 4) sisting of iterations of searching and comparing with the so-far-best
solutions. See also Optimization and Parameter estimation (Chapter 4)
Resting potential
The electrical potential across the cell membrane of a relaxed, excit- Second messenger.
able cell, such as a neuron or cardiomyocyte. In neurons, the resting A small molecule, such as nitric oxide, calcium, cAMP, inositol trispho-
potential is about -70 mV, whereas it is about -90 mV in cardiomyo- sphate, or the sphingolipid ceramide, that is involved in signal trans-
cytes. See also Action potential. (Chapter 12) duction. (Chapter 9)
Retrovirus Secondary structure (RNA)
An RNA virus that needs to be reverse-transcribed by the host The folding and partial base matching of a single-stranded RNA with
into complementary DNA, before it can be integrated into the host itself, creating accurately matched regions, interrupted by loops and
genome. See also transcription. (Chapter 11) possibly ending in unmatched ends. (Chapter 6)
Reverse engineering Sensitivity analysis
A diverse set of methods for inferring structures, processes, or mech- A branch of general systems analysis that quantifies how much a sys-
anisms from data. An example is Bayesian network inference. tem feature changes if a parameter value is slightly altered. (Chapter 4)
(Chapter 3)
Separation of timescales
Reversible (reaction)
A modeling technique for systems containing processes that run at
A (biochemical) reaction that may occur in forward and backward
distinctly different speeds. By assuming that very slow processes are
direction, usually with a different rate. See also Irreversible reaction.
essentially constant and that very fast processes are essentially at
(Chapter 8)
steady state, one restricts a model to the one or two most relevant
Ribosome timescales. (Chapter 12)
A complex, consisting of specific RNAs and proteins, that facilitates
the translation of mRNA into protein. (Chapter 7) Signal
Any molecular, chemical or physical event that can be sensed and
Riboswitch transmitted by a cell or organism, usually through a receptor. (Chapter 9)
An RNA-based regulatory element of gene expression in bacteria.
(Chapter 14) Signal transduction (signaling)
A collective term for the management and integration of signals
Ribozyme received by the receptors of a cell. In many cases, signaling cascades
A ribonucleic acid (RNA) that catalyzes certain chemical reactions. The relay an external signal to the genome, where specific genes are up- or
word is composed from “ribonucleic acid” and enzyme. (Chapter 6) down-regulated in response. (Chapter 9)
RNA-Seq Signaling cascade
A relatively recent high-throughput method for measuring gene expres- Usually a hierarchically arranged set of proteins, which, by becoming
sion, based on cDNA sequencing. See also Microarray. (Chapter 6) phosphorylated or dephosphorylated, transduce a chemical or physi-
cal signal from the cell surface to the genome. (Chapter 9)
Robust
An often vaguely defined term for the tolerance of a system toward Silencing (of a gene)
perturbations or changes in parameters. See also Sensitivity and The ability of certain small RNAs to suppress gene expression.
Gain. (Chapter 4) (Chapter 6)
RT-PCR Simulated annealing
Abbreviation for: A parameter estimation technique inspired by a method of heat-
1. Reverse transcription followed by a quantitative polymerase chain ing and cooling metals in order to reduce impurities and defects.
reaction. A method for quantifying mRNA and thus gene expres- (Chapter 5)
sion. See also Northern blot.
2. Real-time polymerase chain reaction. A fluorescence-based Simulation
method for monitoring in real time how much DNA had been The process of using a mathematical model with different param-
polymerized since the beginning of the experiment. Also called eter and variable settings, with the goal of assessing its possible
RT-qPCR, where q means quantitative. (Chapter 6) range of responses, including best, worst, and most likely scenarios,
or of explaining scenarios observed in, or predicted for, a biological
Sarcoplasmic reticulum (SR) system. (Chapter 2)
A compartment in cardiomyocytes that can contain large quantities
of calcium. Controlled calcium movement between the sarcoplasmic Single nucleotide polymorphism (SNP)
reticulum and the cytosol is ultimately responsible for cell contrac- A point mutation in a specific nucleotide that distinguishes individu-
tion. (Chapter 12) als. Because multiple codons can be translated into the same amino
acid, the change is often neutral. (Chapter 13)
SBML (Systems Biology Markup Language)
A standardized computer language for modeling and sharing biologi- Sino-atrial (SA) node
cal systems and data. (Chapters 1 and 15) A relatively small group of pacemaker cells located between the atria
and ventricles of the heart, which depolarize spontaneously and
Scale-free thereby set the speed of a heartbeat. (Chapter 12)
Property of a graph or network to have similar features, independent
of the number of nodes. (Chapter 3) SIR model
An established prototype model for assessing the spread of commu-
Scatter plot nicable diseases. S, I, and R stand for susceptible, infective, and (from
Graphical representation of two (or three) features of some entity as the disease process) removed individuals. Thousands of variants have
coordinates of points in a coordinate system, which often looks like been proposed. (Chapter 2)
456 GLOSSARY
Sloppiness Stochastic simulation algorithm

The result of a parameter estimation process showing that many com- A specific algorithm for modeling kinetic reaction processes
binations of parameter values for a given model can fit the same datasets that involve small numbers of molecules and therefore do not
equally (or nearly equally) well. See also Identifiability. (Chapter 5) permit the use of methods of statistical mechanics and kinetics.
(Chapter 8)
Small RNA
Relatively short RNA sequences of 50–200 nucleotides that form Stoichiometry
specific loop structures and can bind to proteins or mRNAs, thereby The connectivity of a metabolic network, augmented with an account
affecting their function. See also MicroRNA. (Chapter 6) of the relative quantities of reactants and products in each reaction.
(Chapter 3)
Small-world (property of networks)
A special type of connectivity within networks, where a few hubs Strange Attractor
are heavily connected and the remaining nodes are more loosely The domain of a chaotic system to which trajectories converge and
connected. Thus, most nodes are not neighbors of one another, but that they do not leave once they are in the attractor. See also Attractor.
most nodes can be reached from any other node by a small number (Chapter 4)
of steps. The number of nodes with a given degree of connectivity fol- Stress
lows a power-law distribution with negative exponent. (Chapter 3) A situation that moves a cell or organism away from its normal
physiological state or range. See also Allostasis. (Chapter 11)
Smoothing
A mathematical technique for reducing noise in experimental Stress response element (STRE)
data and for bringing the data closer to a continuous trend line. A short sequence of (maybe five) nucleotides in the promoter region
(Chapter 5) of genes involved in responses to external stresses, such as heat,
osmotic pressure, strong acid, or the depletion of required chemicals
Sphingolipid
in a cell’s environment. (Chapter 11)
A member of a specific class of lipids that are crucial constituents of
membranes and also possess signaling function, for instance, during Subpopulation
differentiation, aging, apoptosis, and a number of diseases, includ- A subset of a population characterized by some common feature,
ing cancer. (Chapter 11) such as age, gender, race, or a particular disease status. (Chapter 10)
S-system Superposition (principle)
A modeling format within biochemical systems theory, in which A feature characteristic of linear, but not nonlinear, systems: the re-
every ordinary differential equation consists of exactly two prod- sponse of a system to two simultaneous stimuli is equal to the sum of
ucts of power-law functions (one of which may be zero): one col- the responses to the same two stimuli administered separately. If the
lectively representing all augmenting processes and one representing superposition principle holds, a system can be analyzed by perform-
all degrading processes. See also Biochemical systems theory and ing many stimulus–response experiments and adding the observed
Generalized mass action system. (Chapter 4) stimulus–response relationships. See also Synergism. (Chapter 4)
Stability Synergism/synergistic
Local stability refers to a steady state or limit cycle and indicates that Literally, “working together.” The phenomenon that the combined
a system will return to the steady state or limit cycle after small per- action of two system components (or people) has a stronger effect
turbations. Structural stability refers to the robustness of the system than the sum of their individual actions. Synergism is a feature of non-
behavior to slight, persistent changes in parameter values. See also linear systems and violates the superposition principle. Antonym of
Bifurcation. (Chapter 4) antagonism/antagonistic. (Chapter 1)
State (of a system or model) Synthetic biology
A comprehensive collection of numerical features characterizing a A diverse set of experimental and computational manipula-
system or model at a given point in time. (Chapter 2) tion techniques with the goal of creating new (usually microbial)
biological systems. (Chapter 14)
Static (model)
A model that describes a network or system without accounting for System
changes over time. (Chapter 3) An organized collection of dynamically interacting components.
Statistical Mechanics Systems biology
A body of concepts and theorems describing the mechanical and A branch of biology that uses experimental and computational tech-
energetic features of large assemblies of molecules or other particles niques to assess dynamic biological phenomena within larger con-
through appropriate averaging. Statistical mechanics relates the mac- texts. (Chapter 1)
roscopic bulk behavior of materials to the microscopic features of
Systole
molecules. (Chapter 8)
Greek term for contraction and used to describe the contracted state
Steady state of the heart. See also Diastole. (Chapter 12)
The state of a dynamical system where no dependent variable changes
Taylor approximation
in magnitude, although material may be flowing through the system. At a
A general type of approximation for functions of one or more vari-
steady state, the time derivatives in all differential equations describ-
ables around a specific operating point; resulting in a (possibly multi-
ing a system are equal to zero, and thus the equations become algebraic
variate) linear, polynomial or power-law representation. (Chapter 4)
equations. See also Homeostasis and Stability. (Chapter 2)
Tertiary structure (of a protein)
Stimulus The three-dimensional shape of a protein molecule. See also
An external perturbation to a system or model. Stimuli are often Conformation. (Chapter 7)
used to study possible behaviors of the system. (Chapter 2)
Text mining
Stochastic Use of a machine learning algorithm to extract information from
Containing probabilistic or random aspects; often considered the large amounts of text. See also Data mining. (Chapter 13)
antonym of deterministic. (Chapter 2)
Time-series data
Stochastic simulation Experimental or artificial data, consisting of values of variables at
A simulation with a model that is subject to random events. (many) subsequent time points. (Chapter 5)
GLOSSARY 457
Toggle switch (genetics) Two-dimensional gel electrophoresis

A phenomenon or model in which a gene is either fully expressed or An experimental method for separating proteins by their size and
silent, but does not allow intermediates. (Chapter 14) electrical charge. (Chapter 7)
Toll-like receptor (TLR) Ubiquitin
A member of a class of signaling proteins that recognize a variety of A small protein that, when attached to other proteins, tags them for
foreign macromolecules, especially those on the cell walls of bacteria, disassembly in the proteasome and later recycling of the resulting
and trigger the initiation of an immune response. See also Receptor peptides and amino acids. The tagging process is called ubiquitina-
and Transmembrane protein. (Chapter 7) tion. (Chapter 7)
Training set Uncertainty
A portion of a dataset used by a machine learning algorithm to dis- Degree to which an experimental result is not exactly known, owing to
tinguish different classes of outputs and to establish classification experimental inaccuracies or other challenges; see also Variability.
rules. See also Validation. (Chapter 5) Advanced measurement techniques might reduce uncertainty, but
Trajectory not variability. (Chapter 5)
The collective change in the state of a system over time; usually between Validation
a stimulus and some later state, such as a steady state. (Chapter 10) The process of testing (and confirming) the appropriateness of a mod-
Transcription el with data that had not been used for the construction of the model.
The process of creating a matching messenger RNA (mRNA) from a See also Training set, Identifiability, and Sloppiness. (Chapter 2)
DNA sequence. (Chapter 6) Variability
Transcription factor Degree to which an experimental result is affected by differences
A protein affecting the expression of a specific gene by binding to the among individual cells, organisms, or other items under investigation;
corresponding DNA. (Chapter 6) see also Uncertainty. Advanced measurement techniques might re-
duce uncertainty, but not variability. (Chapter 2)
Transcriptome
The totality of RNAs transcribed from the genome of an organism. See Variable
also -ome. (Chapter 6) A symbol representing a component of a system, which may or may
not change over time. See also Dependent and Independent vari-
Transfer RNA (tRNA) able. (Chapter 2)
One of several specific RNAs that facilitate the attachment of the
correct amino acid to a growing peptide or protein, while it is being Vector
translated from mRNA. This translation process occurs within a ribo- 1. In mathematics, a convenient arrangement of variables or numbers
some. (Chapter 6) in a one-dimensional array. Often used in conjunction with a matrix.
Many aspects within the fields of linear algebra, multivariate calcu-
Transient lus, and statistics are based on vectors and matrices. (Chapter 4)
A collective term for behaviors of a system between the time of a stim- 2. In genetic engineering, an artificially constructed virus or plasmid
ulus and the time the system reaches a steady state, a stable oscilla- used to introduce foreign DNA into a bacterium or host cell.
tion, or some other attractor or point of interest. (Chapter 4) (Chapter 14)
Translation Ventricle
The process of generating a protein or peptide whose amino acid One of the two larger chambers of the heart. See also Atrium. (Chapter 12)
sequence corresponds to the genetic code of an mRNA. The process
Vertex (pl. Vertices)
occurs within a ribosome and uses transfer RNA. See also Post-
Synonym of node. (Chapter 3)
translational modifications. (Chapter 6)
Voltage-gated
Transmembrane protein
Property of a membrane channel to open and close in response to the
A member of a large class of proteins that span the membrane of a
membrane potential. (Chapter 12)
cell and thereby allow communication between the outside and
inside of the cell. Transmembrane proteins are involved in signal Western blot
transduction and a large number of diseases. See also Gap junction A technique for measuring gene expression based on produced pro-
and G-protein-coupled receptor. (Chapter 7) tein. See also Northern blot. (Chapter 6)
Transposon XML
A stretch of DNA that can move from one location in a genome to Abbreviation for eXtensible Mark-up Language. Computer language
another. (Chapter 6) with a defined vocabulary that permits the structuring of data and
facilitates translation into different languages and software packages.
Trial (stochastic process)
See also Ontology and SBML. (Chapters 1 and 15)
One of many evaluations of a stochastic model. (Chapter 2)
X-ray crystallography
Two-component (signaling) (TCS) system A technique for determining the structure of a protein. The protein
While not exclusively limited to this definition, usually a specific type must first be crystallized. An X-ray shot through the crystal scatters
of signaling mechanism, with which microorganisms sense their in a characteristic fashion that allows the deduction of the molecular
environment. See also Receptor and Signal transduction. (Chapter 9) structure of the protein. (Chapter 7)
Index
Note: Page numbers followed by F refer to figures, those followed by T refer to tables, and those followed by B refer to information boxes
A continuum 336
linearized models 95–98, 96B, 97B, 97F
two-component signaling systems 266–269,
266F–269F, 268T
ABM (agent-based modeling) 110, 437–438, 438F power-law 108B, 109B, 239 vectors 409
gene regulation 191 Aptamers 410 “wild type” 16
ACO (ant colony optimization) 148 Aquaporins 217 Bacteriophages 409
Actin 217, 338, 338F Aquatic systems 3 Barcoding 175, 175F
Action potentials 337, 348–350 oceans 433–434, 434F, 435F Barrels 221, 222F
cardiomyocytes 337, 348–350, 349F, 350F Archaea 266 Base(s), DNA and RNA 171, 172F, 173F
lengthening during repolarization 355 Arginine (Arg) 203F, 204T Base pairs 171, 172F, 173F
models 350–354, 353F Arrhythmias 355 Basin of attraction 262
phases 349–350, 350F Artemisinic acid production 414–415, 415F Bassingthwaighte, James 332
prolongation 361 Artemisinin production 414–415, 415F Bayes’ formula 65
repeated 354, 354F Artificial neural network (ANN) 374, 426 Bayesian inference 64–69
SA and AV note cells 349 Ascorbate pathway 250 Bayesian reconstruction, interaction networks 63–69,
sodium-calcium exchanger 347 Asparagine (Asn) 203F, 204T 65F, 67F
Activator(s), metabolic systems 238–239 Aspartic acid (Asp) 203F, 204T Bayes’ rule 65
Activator module 417 ATH1 gene 310, 325 Bayes’ theorem 65
Acute lymphoblastic leukemia (ALL) 373–374 Atomic force microscopy (AFM) 174F Behavior(s), of discrete vs continuous models 29
Adaptation 273–274, 273F, 274F, 274T Atrial fibrillation 356 Behavioral conditioning 440
Adaptive evolution 407 Atrioventricular (AV) node 336 Bernard, Claude 10
Adaptive immune system 428–429, 429F Atrioventricular (AV) valves 333F, 334 Bertalanffy, Ludwig von 10
Adenine 171, 172F, 173F Atrium(a) 333–334, 333F Berzelius, Jöns Jacob 201
Ad hoc models 101–102, 101F, 102F Attractors 122–128 β-sheets 221, 222F
Adjacency matrices 55, 55F chaotic 126–128, 127F, 128F Bidirectional graphs 53, 55F
Adrenaline 336 defined 122 Bi-fan motifs 402, 402F
Affinity 139 limit cycles 123–125, 123F, 125F, 126F Bifurcations 119
AFM (atomic force microscopy) 174F Lorenz 126–127, 127F Bimolecular metabolic reactions 234–235
Agent-based modeling (ABM) 110, 437–438, 438F strange 126–128, 127F, 128F Binomial model 30, 32B
gene regulation 191 Automaticity 349 Biochemical reactions, metabolic systems 232–240
Aging Autonomic nervous system (ANS) 336 background 232–234, 233B
heart failure 355–356 Auto-regulation 402, 402F mathematical formulation of elementary reactions
inflammation 429 AV (atrioventricular) node 336 234–235
lens 217 AV (atrioventricular) valves 333F, 334 rate laws 235–240, 235F, 237F, 238F
Alanine (Ala) 203F, 204T Axioms 439–440 Biochemical systems theory (BST) 103
Albumin, serum 212–213 Generalized Mass Action (GMA) systems 103–104,
Algae 400, 418 104F
Algorithms for parameter estimation 135 metabolic pathways 239, 242
nonlinear systems 142–143 B S-system 104–105, 114–115
trehalose cycle 311
comprehensive grid search 142, 143–145, 144F
evolutionary 142–143, 146–148, 147F Bacillus anthracis 267 BioCyc 74, 242, 244B
genetic 142–143, 146–148, 147F Bacillus subtilis 410B Bioengineering 11
gradient 142, 143F Bacteria Biofilms 278, 433, 433F
other stochastic 148–149 actin 217 Biofuel 400, 411
ALL (acute lymphoblastic leukemia) 373–374 barrels 222F Bioinformatics 12, 74
Allostasis 373 biofilms 433, 433F DNA sequences 178
Allosteric inhibition 211, 239 carrier proteins 346 Markov models 89
Allosteric regulation 239 cyano- 208, 434, 434F Biological components, variations 9–10
α-helices 221, 221F cytokines 215 Biological process, gene products 176
Alzheimer’s disease 14 expressed sequence tag 193 Biological Revolution 3
Amino acids 202–204, 203T, 204T feedforward loop 403–404 Biological systems 1–18
codons 174–175 flagellar motor 213–214, 213F analyzing 3–4, 4F
Amorpha-4,11-diene 414, 415, 415F gene regulation 406 capability to manipulate 2–3
Amorphadiene 414 genetic engineering 409, 418 complicated 1, 2F
AND gate 402 genome confusing simple 8–10, 8F, 9F
Anfinsen, Christian 204 organization 188, 401 conversion into computational models 16
Animal breeding 407 size 174 design 399–423
ANN (artificial neural network) 374, 426 glycoproteins 218 case studies 411–418
ANS (autonomic nervous system) 336 inflammation 431 design principles 399, 404–406, 404F, 406F
Antagonistic pressures, heat stress response 308 keratin 217 drug development 414–415, 415F
Ant colony optimization (ACO) 148 lipo-oligosaccharides 209F elementary mode analysis 411–414, 413F
Antibodies 215, 215F metapopulations 432 future 419
monoclonal 382F–385F, 383–385, 384T operons 185, 188 gene circuits 415–418, 416F–418F
Antibody binding 382F–385F, 383–385, 384T plasmids 170, 178 goal-oriented manipulations 407–408
Antigen, immune cascade 404–406 population growth 284, 284F, 285, 286, 292 metabolic engineering 407–408, 411–414, 413F
Antimalarial drug 414–415 predation 288 natural 400–407
Anti-parallel strand 171, 172F quorum sensing 278 network motifs 399, 402–404, 402F, 403F
Antiporter 346–347 riboswitches 410B operating principles 406–407, 406F
Aplysia californica, neural networks 427, 427F stress signal sensing 403 optimization 400
Apolipoproteins 217 Toll-like receptors 215 overview 399–400
Apoptosis, differentiation 306 transcription factors 188–189 structural patterns 400–402, 401F
Applications, emerging 426–434 transposons 170 synthetic 399, 408–410, 410B, 411–418
Approximation(s) 29, 29F trehalose cycle 310 emergent properties 11, 12
canonical 101 tryptophan 170 engineering of 3
INDEX 459
examples 1, 2F Cardiac cells see Cardiomyocytes Clustering coefficient 55–56, 56F

goal of deeper understanding of 3 Cardiac cycle 356F Coding regions 177
history of dealing with 1–2 blood flow 333F, 335 Codon(s) 174–175, 202–204, 204T
mathematics 83–134 ECG 339F Codon bias 175
nonlinear processes in 5, 5F, 8 metabolic responses to oxygen supply 342 Codon law 440
size vs complexity of 4–5, 4F movement of ions 343, 347 Codon usage bias 204
static structure and dynamics of 16 shape of heart 336 Coefficients 94
Biological target, drug development 379 Cardiac output 334 Cohen, Joel 60
Biological theory 439–441 Cardiomyocytes 334, 335–336 Collagen 216, 217F
Biological traits 176 action potentials 337, 348–350, 349F, 350F Collins, James 417–418
Biological variability, and disease 369–370 models 350–354, 353F Communication 13–15, 14F
modeling 370–372, 371F, 371T, 372F aging 356 Community proteomics 433
Biologics 380–381, 382F automaticity 349 Compartment models 94
Bioluminescence 278 contraction 332 pharmacokinetics 385–387, 385F–387F
Biomarkers 373–374, 374F electrochemistry 345–354 Competitive inhibition 239
disease 218, 373–376, 374F biophysical description 347–348 Complementary DNA (cDNA) 192
normal range 377, 377F models of action potentials 350–354, 353F Complexities, Lotka-Volterra model 103
Biomass 411, 413–414 repeated heartbeats 354, 354F Comprehensive grid search 143–145, 144F
Bionanotechnology 11, 13F resting potentials 348–350, 349F, 350F Comprehensive screening experiments 121
Biotechnology 11, 13F excitable 336, 337, 338–339, 338F Computational analysis 241–244, 243B, 244B
Bipartite graphs 54 funny currents 349 Computational models 16
Bistability heart failure 355–356 Computational power 435
gene circuits 416 modeling based on molecular events 356–361, Computational systems biology 375, 375F
signaling systems 261–263, 262F, 262T, 263F 356F, 357F, 359F, 360F Computer-aided approximation 12
Black box models, oscillations 339–343, 340F–342F oxidative stress 355–356 Computer-aided design (CAD) 410
Blood sugar level 373 resting potentials and action potentials 337, Computer-based reasoning systems 433
Blue Brain Project 426, 426F, 427F 348–350, 349F, 350F Computer languages 427, 435
Blue sky catastrophe 30, 31B sarcolemma 356–357, 357F Computer simulations 19, 38, 38F
Bolus experiments 120, 120F sinoatrial node 337 personalized medicine 376–377
Boolean networks 259–261, 259F, 260F β-Carotene 415 population growth 286
Bootstrap method 152–153 Carrier proteins 346–347 Computing environment 12–13, 13F
Bottom-up strategy 16–17 Carriers 211–214, 212F, 213F Concatamers 193
parameter estimation 136 Carrying capacity 284, 288, 288F Concentration gradient 345
Bradycardia 355 CaRU (calcium release unit) 361 Concept-map modeling 437, 438–439, 439F
Brady-tachy syndrome 355 Case studies 16, 17 Conceptual model 15
Brain, neurons 362 Catabolite activator protein (CAP) 186, 186F, 187 Conditional probabilities 64–65, 65F
Brain tumor 189, 193, 224F Causality analysis 62 Conductance 336, 351–353
Branch-and-bound methods 144–145 cDNA (complementary DNA) 192 Conformation 187, 204
Branched pathway 406, 406F CE (capillary electrophoresis) 247 Congestive heart failure 349F, 355, 356
Breeding 407 Cell cycle 60 Conjugation 409
BRENDA 240, 241B genomic regulation 274 Connections 16
pathway screening 392–393 heat stress response 306 Connectivity relationships 76–77
Brenner, Sydney 17 oscillator 276 trehalose cycle 308, 308F
BST see Biochemical systems theory (BST) regulatory network 260 Connexins 211, 212F
Bundle of His 336, 337–338 subpopulation analysis 292 Consistency 38–40, 39F
Burkholderia pseudomallei 175F Cell membrane Constrained optimization 73
depolarization 337 Constraints 70, 73
gated channels 353–354, 355 Continuous linear systems 93–100
heart function and failure 356 linear differential equations 94
C ion movements 345, 346
resting potential 348
linearized models 95–100, 95F, 96B, 97B, 97F,
99F, 100F
CAD (computer-aided design) 410 T-tubules 358 Continuous models 29, 84
Caenorhabditis elegans, neural networks 427, 427F Cell-to-cell communication 211 Continuous nonlinear systems 100–110
Calcium adenosine triphosphatase (Ca2+ ATPase), Cellular automata 437, 438 ad hoc models 101–102, 101F, 102F
plasma membrane 347 Cellular component, gene products 176 canonical models 101, 102–110, 103F–105F,
Calcium (Ca2+) channels Cellular stress response 305 107F, 108B, 109B
L-type 349, 356F, 357–361, 357F Central Dogma 169–171, 170F, 440 more complicated dynamical systems
T-tubules 356 Central limit theorem 358 descriptions 110
T-type 349 Central nervous system 11 Continuum approximation 336
Calcium dynamics, dyads 356–361, 356F, 357F, Chalfie, Martin 222 Control coefficients 74–75
359F, 360F Channel proteins 346 Controlled comparison 404–406, 406F
Calcium fluxes Chaotic attractors 126–128, 127F, 128F Conway, John 260
dyads 356F, 358 Chaotic dynamics 298 Cook, James 6F
slowing down 355 Chaotic process 93 Cooperativity 211
Somogyi-Stucki model 344, 344F Chaperones 205, 207F Coral growth 434, 435F
Calcium-induced calcium release (CICR) 343, 357 Chaperonins 205 Co-regulating genes 188
Calcium release unit (CaRU) 361 Charge gradient 345 Corn 415
cAMP see Cyclic adenosine monophosphate (cAMP) CheY protein 266, 266F Coronaries 335
Cancer 431–432, 431F Children, growth curves 47, 48F Correlative models 27, 27F
Canonical models 101, 102–110 Chitin 217 Corynebacterium glutamicum 189, 251, 408
concept-map modeling 438 Chloroplasts 170, 188 Crick, Francis 169, 171
generalized mass action (GMA) systems Christou, Paul 415 Critical point analyses 121–122, 122F
103–104, 104F Chromatin 178 Cross-bridges 338, 338F
limit-cycle oscillations 102 Chromatography 193, 219, 247 Crosstalk, MAPK cascade 273
lin-log model 107, 115 Chromosomes 178 Crowding effect 285
Lotka-Volterra 101, 102–103, 103F, 114 CICR (calcium-induced calcium release) 343, 357 Cryo-electron microscopy (cryo-EM) 206, 207F
power-law approximation 107, 108B, 109B Ciechanover, Aaron 205 Curved surface, linearization of 95, 95F
S-systems 104–105, 114–115 Cilia 213 Customized mark-up languages (XML) 12
toggle switch model for SOS pathway of E. coli Circadian clocks 275–278, 276F, 277F, 278T Cyanobacteria 208, 434, 434F
105–107, 105F, 107F cis-regulatory motifs 185 Cyclic adenosine monophosphate (cAMP) 187
trehalose cycle 310–311 Citric acid 399, 408 funny currents 349
CAP (catabolite activator protein) 186, 186F, 187 Clinical phase, drug development 380 heat stress response 307
Capacitance 351 Clinical trials 380 signaling systems 258
Capillary electrophoresis (CE) 247 Cliques 55, 60 Cysteine (Cys) 203F, 204T
Caprioli, Richard 222 Closed system model 29 Cystine 202
Carcinogenesis 431–432, 431F Clustering, heat stress response 317 Cytochromes 213
460 INDEX
Cytokine(s) 215, 428–429, 429F base pairs 171, 172F, 173F overview 83–85
Cytokine storm 428–429, 429F chemical and physical features 171–172, pathway screening 390–393, 391F
Cytoscape 60, 61F 172F–174F standard analyses 110–122
Cytosine 171, 172F, 173F codons 174–175, 202–204, 204T analysis of systems dynamics 119–122, 120F, 122F
double-stranded 170, 171, 174, 188, 193 parameter sensitivity 118–119
epigenetics 170, 178 stability analysis 115–117, 116F, 117F
eukaryotic packing 178, 179F, 180F steady-state analysis 110–115, 112F, 113F
D exons and introns 177, 177F
genetic variation 170
Dynamics
of biological systems 16
DAG (directed acyclic graph) 67–68 hybridization 171 model development 24–25
Damped oscillations 103 “junk” 176–177 Dystrophin 232
Data availability 25–26 key properties 171–178
Data mining 250, 381 noncoding 174, 175–178, 177F
Data-modeling pipeline 437–439, 438F, 439F nucleotides 171, 172F
Data warehouses 433
DD (duplication-divergence) model 56
origin-binding domains (OBDs) 173F
random mutations 30
E
Degree distribution 57, 58F replication 171–172, 173F EC (excitation-contraction) coupling 338
Degrees associated with nodes 54–55, 55F sequencing 174–175 EC (Enzyme Commission) number 241F
Demand theory 406, 440 size and organization 174–175, 175F Ecoinformatics 433–434
Denaturation 306 supercoiled 174, 174F Edelstein-Keshet, Leah 86
Deoxyribonucleic acid see DNA (deoxyribonucleic synthetic biology 408–410 Edges 53, 55
acid) transcription 169, 172, 176, 180 Effector cells 404–406, 404F, 406F
Deoxyribose 172F hierarchical control 189, 189F Efficacy 379
Dependencies among network components translation 176 Eigenvalues
62–63, 63F DNA adducts 178 linear differential equations 94
Dependent variables 33–34, 33f DNA chips 193–194, 193F stability analysis 116–117
Depolarization 337, 348, 350, 350F heat stress response 315 trehalose cycle 311
Design principles 399, 404–406, 404F, 406F DNA methyltransferases (DNMTs) 179F Elasticities 74, 76
Detection, metabolite analysis 247–249, 248F Dose 379 Elasticity coefficients 76
Deterministic models 29, 30, 31B Drosophila 194, 195F Elastins 216
population dynamics 291–292 Drought, plant response 2F Electrical charge 350–351
recursive 85–88, 85T, 86F Drug development 378–393 Electrical conductance 336, 351–353
Diabetes 29, 373, 381, 428 antimalarial 414–415 Electrical conductor 351
Diagram, model development 34–35, 34F, 35F biological systems designs 414–415, 415F Electrical gradient 345
Diastole 334 biological target 379 Electrical potential 351
Dicer 184 biologics 380–381, 382F Electrical resistance 351
Difference equations 91 clinical phase 380 Electric current 351
Difference gel electrophoresis 219 compartment models 385–387, 385F–387F Electrocardiogram (ECG, EKG) 339, 339F
Differential equations lead identification 381 Electrochemical gradients 347–348
linear 94 lead optimization 379 Electrochemical processes, heart 332
models of gene regulation 190 models Electrochemistry, cardiomyocytes 345–354
ordinary 91, 110 compartment 385–387, 385F–387F biophysical description 347–348
heart 358 dynamic 390–393, 391F models of action potentials 350–354, 353F
metabolic systems 234 pharmacokinetic 385–390, 386F, 387F, 388B–389B repeated heartbeats 354, 354F
models of gene regulation 190 receptor dynamics 382–385, 382F–385F, 384T resting potentials and action potentials 348–350,
population growth 284, 285–286 new chemical entities (NCEs) 378–379 349F, 350F
signaling systems modeled with 258, 261–278 pipeline 378–380, 379F, 381, 391 Electromotive force (EMF) 351
parameter estimation for systems of 153–159, 153F, post-approval phase 380 Electrophysiology, heart 332
154T, 156F–159F, 156T–158T preclinical phase 378–379, 379F Elementary mode analysis (EMA) 411–414, 413F
partial 110 process 378–380, 379F Ellington, Andrew 418
heart 334, 358 quantitative structure-activity relationships Embryonic cell cycle oscillator 276
models of gene regulation 191 (QSARs) 379 Emergent properties 11
signaling systems modeled with 261–278 role of systems biology 380–393 EMF (electromotive force) 351
adaptation 273–274, 273F, 274F, 274T computational target and lead identification 381 Entropy 63, 63F
bistability and hysteresis 261–265, 262F–264F, emerging 393 Environment
262T, 264T, 265B–266B pathway screening with dynamic models interactions of organisms with 432–434, 433F–435F
mitogen-activated protein kinase cascades 390–393, 391F model development 29
270–273, 270F–273F, 271T pharmacokinetic modeling 385–390, 386F, 387F, Environmental pressures 16
other 274–278, 276F, 277F, 278T 388B–389B EnvZ two-component sensing system 268, 406
stochastic 258, 261–278 receptor dynamics 382–385, 382F–385F, 384T ENZYME 240
two-component 266–269, 266F–270F, 268T toxicity testing 379 Enzyme(s) 209–211, 209F, 210F
stochastic 258, 261–278 Drug sensitivity 373–374 activity profile 52
Differentiation 84 Duplication-divergence (DD) model 56 catalyzing 72F, 73, 102, 102F, 139, 139F, 235, 235F
Differentiation apoptosis 306 Dyads, calcium dynamics 356–361, 356F, 357F, control of pathway systems 245–246
Diffusion process 234 359F, 360F dicer 184
Digitalis 347 Dynamical features, exploration and validation 38, drug development 379, 392, 414, 415F
Digitalis purpurea 347 40–42, 41B, 42F, 43F flux analysis 249
Directed acyclic graph (DAG) 67–68 Dynamic(al) models 83–134 functional loop 8–9, 8F
Directed evolution 209F, 407 continuous linear 93–100 gene circuits 418
Directed graphs 53, 53F, 55–56, 55F, 56F, 62 linear differential equations 94 gene expression 191–193, 315–318
Discrete linear systems models 85–91 linearized models 95–100, 95F, 96B, 97B, 97F, generalized mass action systems 107
recursive deterministic 85–88, 85T, 86F 99F, 100F gene regulation 185
recursive stochastic 84F, 88–91 continuous nonlinear 100–110 heat stress response 306–309, 307F, 312–318,
Discrete models 29, 84 ad hoc models 101–102, 101F, 102F 313F, 314F
Discrete nonlinear systems 91–93, 92F canonical models 101, 102–110, 103F–105F, 107F, hexokinase 231, 232F
Discrete process 41 108B, 109B trehalose cycle 308–309
Disease(s) more complicated dynamical systems immune system 214
biological variability and 369–370 descriptions 110 incompletely characterized organism 250
modeling 370–372, 371F, 371T, 372F defined 27–29, 28B independent variables 33, 121
complex 369, 428 discrete linear 85–91 kinases 214
delineation 377, 377F recursive deterministic 85–88, 85T, 86F kinetics 102, 102F, 160, 240–241
see also Biomarkers; Infectious disease modeling recursive stochastic 84F, 88–91 lac operon 186, 186F, 187
Disulfide bond 202 discrete nonlinear 91–93, 92F metabolic control analysis 75–78, 75F
DNA (deoxyribonucleic acid) other attractors 122–128 metabolic engineering 407–408
adducts 178 chaotic 126–128, 127F, 128F metabolic network analysis 251
barcoding 175, 175F limit cycles 123–125, 123F, 125F, 126F metabolic systems 232, 233, 233B, 243B
INDEX 461
reconstruction 74, 74F False-positive outcomes 377, 378 human 6, 176, 177
Michaelis-Menten rate law 236–238, 237F, 238F Farnesyl pyrophosphate (FPP) 414 information transfer 170
mitogen-activated protein kinase cascades FBA (flux balance analysis) 73, 411 noncoding DNA 174, 175–178, 177F
270–273 FDA (Food and Drug Administration) 378, 379F, open reading frame 177
multiscale analysis 318–326, 324F, 324T, 362 380, 382F phenotype 176
personalized medicine 373, 375 Feature selection, structure identification 160 transcription network 402
perturbations 371–372, 392 Feedback Gene annotation 171
phospholipase A 222F metabolic pathways 245, 245F, 246F Gene circuits 415–418, 416F–418F
population growth 283 nested 245, 246F Gene expression
power-law approximations of multivariate Feedback inhibition 32–33, 33F control of metabolic pathways 245–246
functions 109B metabolic pathways 245, 245F, 246F localization 194, 195F
protein activity and dynamics 224 Feedback loops measuring 191–194, 192F, 193F
protein folding 205 design of biological systems 403, 403F structural patterns 401
proteomics 220 positive 428 trehalose cycle 308, 309F, 315–318, 316F,
rate laws 236–239 Feedforward loops, design of biological systems 316T, 317F
ribozymes 183, 183F 403–404, 404F Gene Expression Atlas 194
sensitivity analysis 118 Fermentation 25, 400 Gene Expression Omnibus 194
signal transduction 258, 266 Fibroblasts 216 Gene interaction network 190
small RNAs 184 Fibronectins 213 Gene Ontology (GO) Consortium 176
substrate conversion 36–37 Filaments, heart 334 Generalized mass action (GMA) systems
synthetic biology 408, 410, 412–413 Finite element approaches, heart 335 103–104, 104F
transcription factors 189 First-degree heart block 355 structure identification 160
tryptophan pathway 170 Fish populations 434 trehalose cycle 311
Enzyme-catalyzed reaction 72F, 73, 102, 102F, 139, Fitness 142 see also Biochemical Systems Theory
139F, 235, 235F Fitting 137 Gene regulation 185–191
Enzyme Commission (EC) number 241F FitzHugh, Richard 332 lac operon 186–187, 186F, 187T
Epidemics 297 FitzHugh-Nagumo equations 354 models 190–191, 190F, 191F
Epidermal growth factor receptor 214, 214F Flagellar motor 213–214, 213F modes 187–188
Epigenetics 170, 178, 179F, 370 Flow chart 22, 22F transcription factors 169, 188–190, 189F
Epinephrine 336 Flow of information 32–33, 33F Gene regulatory networks 69
ERG9 gene 415 Flow of material 32, 32B as signaling systems 274
ERGO 242 Fluid mechanics, heart 335 Gene systems 169–200
ERK (extracellular-signal-regulated kinase) 270 Flux(es) 52, 70 Central Dogma 169–171, 170F
Escherichia coli metabolic systems 240, 242–244 DNA
analyzing metabolism in incompletely trehalose cycle 311 chemical and physical features 171–172,
characterized organism 250, 251 Flux analysis, metabolic 249 172F–174F
barcode 175F Flux balance analysis (FBA) 73, 411 epigenetics 170, 178
barrel structure 222F Flux network 70–73, 70F–72F eukaryotic packing 178, 179F, 180F
computation of volume 28B Food(a), transgenic 415 genes and noncoding 175–178, 177F
CheY signal transduction 266 Food and Drug Administration (FDA) 378, 379F, key properties 171–178
design of biological systems 400, 402, 406 380, 382F size and organization 174–175, 175F
elementary mode analysis 411–414 Force-directed layout 60, 61F gene expression
gene regulation 185, 190 F6P (fructose 6-phosphate) 72, 242, 309 localization 194, 195F
lac operon 186 FPP (farnesyl pyrophosphate) 414 measuring 191–194, 192F, 193F
metabolic engineering 407, 411, 412 Fractals 93, 122, 432 gene regulation 185–191
metabolic flux distribution 72, 72F Franklin, Rosalind 171 lac operon 186–187, 186F, 187T
modeling variability and disease 371 Frank-Starling law 334 models 190–191, 190F, 191F
network motifs 403 Free radicals 355 modes 187–188
resources for computational pathway analysis 242 Fructose 1,6-bisphosphate 8, 318 transcription factors 169, 188–190, 189F
single mutants 11 Fructose 6-phosphate (F6P) 72, 242, 309 outlook 196
SOS pathway 105–107, 105F, 107F Fugu rubripes 174 RNA
supercoiled DNA 174F Functional loop 8–9, 8F, 9F chemical and physical features 172F,
synthetic biology 408, 410 Funny currents 349 178–179, 181F
toggle switch model 105, 105F Future developments 16, 17 key properties 178–185
transcription factors 189 Fuzzy laws 440 messenger 179
two-component signaling system 266, 268 ribosomal 182–183, 183F
EST (expressed sequence tag) 193 small 183–184, 184F
Ethanol 399, 400, 411, 412, 414 transfer 170, 182, 182F
Euler, Leonard 54, 54F
Evolutionary algorithms, parameter estimation
G viruses 184–185, 185F
Genetic algorithms (GAs) 142–143, 146–148, 147F
142–143, 146–148, 147F GA(s) (genetic algorithms) 142–143, 146–148, 147F Genetic code 174, 440
Evolutionary programming 148 Gain analysis 118 Genetic engineering 409
Excitable cardiomyocytes 336, 337, 338–339, 338F trehalose cycle 311–312 Genetics, Mendelian 440
Excitation-contraction (EC) coupling 338 Galactose 186 Genetic variation 69, 370
Exons 177, 177F Galactose transacetylase 186, 186F Genome(s)
ExPASy database 251 Galactose utilization gene network 60, 61F bacterial 188, 401
Experimental systems biology 374–375, 375F β-Galactosidase 186, 186F human 6, 171, 184
Explanation 16 β-Galactoside permease 186, 186F integrative analysis 303–330
Explanatory models 27, 27F Galileo 62 size and organization 174–175, 175F
Exploration, mathematical model 24, 38, 40–42, 41B, Game of Life (Conway) 260 structural patterns 401, 401F
42F, 43F Gap junctions 211, 212F, 337 transposons 170–171
Exponential growth 283, 286 Gas chromatography with mass spectrometry Genome comparison 74, 74F
Exposome 372 (GC-MS) 247 Genome engineering, multiplex automated 407
Expressed sequence tag (EST) 193 Gates 402 Genomic networks 69
External validation 40 GC-MS (gas chromatography with mass Genomic regulation, cell cycle 274
Extracellular matrix 208 spectrometry) 247 Genomics
Extracellular-signal-regulated kinase (ERK) 270 Gel electrophoresis meta- 172, 433
Extraction, metabolite analysis 247, 251–252 difference 219 pharmaco- 373–374
Extrapolations 19 sodium dodecyl sulfate polyacrylamide 191 GFP (green fluorescent protein) 194,
two-dimensional (2D) 219, 219F 222, 223F
Gene(s) Gibbs sampler 68
biological variability 369–370 Gilman, Alfred 258
F modeling 370–372, 371F, 371T, 372F
Central Dogma 169–171, 170F
Gln (glutamic acid) 203F, 204T
Global optimum 144
Factor VIII 381 coding 176 Glu (glutamine) 203F, 204T
False-negative outcomes 377–378 defined 176 Glucokinase, trehalose cycle 308–309
462 INDEX
Glucose 231, 232F

bacteria 153
H Hill rate law (HRL) 238, 238F
pharmacokinetics 387
13C-labeled 250, 325, 325F trehalose cycle 311
Half-life 94, 379
diabetes 29 HDACs (histone deacetylases) 179F His, Wilhelm 336
enzymes 209 HDLs (high-density lipoproteins) 217 Histidine (His) 203F, 204T
heat stress response 306, 308, 308F Health Histidine kinase 267, 268
accounting for dynamics 314–315, 315F delineation 377, 377F Histone(s) 178, 179F
analysis 312–314 predictive 372–378 Histone deacetylases (HDACs) 179F
gene expression 317 Heart 331–367 Hits, small-molecule 379
metabolic pathway model 311–313, 312F aging effects 355–356 Hodgkin, Alan Lloyd 332, 350, 353
multiscale analysis 318–323, 319F–321F anatomy 331, 332F, 333–334, 333F Hodgkin-Huxley model 350, 353, 353F,
trehalose puzzle 325 electrochemistry in cardiomyocytes 345–354 354, 354F
lac operon 186–187, 186F biophysical description 347–348 Homeostasis 373
metabolic control analysis 242 models of action potentials 350–354, 353F Homogeneous population 289
metabolic engineering 414 repeated heartbeats 354, 354F Hormones 214, 217
metabolic flux distribution 72, 72F, 73 resting potentials and action potentials 348–350, HPLC (high-performance liquid chromatography)
metabolic networks 70, 231, 232F, 233 349F, 350F 247
metabolomic data generation 247 failing 355–361 HRL see Hill rate law (HRL)
simple system 8, 10 modeling based on molecular events 356–361, HSPs (heat-shock proteins) 205, 306–307
small metabolite 169–170 356F, 357F, 359F, 360F Hubs 56, 57F, 60
Glucose 6-phosphate (G6P) hierarchy of scales and modeling approaches Human genome sequencing 6
Escherichia coli 72, 72F 332–339 Human Protein Atlas 201
glycolysis 8 basics of anatomy 333–334, 333F Hunter, Peter 332
trehalose cycle 308, 309–310, 311, 314, 316, at cell level 337–339, 338F, 339F Huxley, Andrew 332, 338, 350, 353
317, 318, 320F at organ level 334–335, 335F HXTs (hexose transporters) 308
Glutamic acid (Gln) 203F, 204T at tissue level 335–337 Hybridization 171
Glutamine (Glu) 203F, 204T outlook for physiological multiscale modeling Hyperuricemia 391–393
Glycans 217 361–362 Hypotheses 15
Glycine (Gly) 203F, 204T overview 331–332, 332F Hypoxanthine-guanine phosphoribosyltransferase
Glycogen 308 pacemaker 342, 342F, 354–355 (HGPRT) 392
Glycogen synthase 323 simple models of oscillations 339–345, 339F Hysteresis 263–266, 264F, 264T, 265B–266B
Glycolysis black box 339–343, 340F–342F
computational pathway analysis 242 meaningful 343–345, 344F, 345F
enzymes 209 Heartbeat(s)
in incompletely characterized organism 250–251
simple systems 8
repeated 354, 354F I
simulation of perturbed 341–342, 341F
static metabolic network 70 Heart block 355 Identifiability 160
time-series data 252 Heart failure 355–361 Ile (isoleucine) 203F, 204T
trehalose cycle 308, 308F, 309–310, 311, 316, 317, 318 modeling based on molecular events 356–361, Immune cascade 404–406, 404F, 406F
Glycoproteins 217–218 356F, 357F, 359F, 360F Immune system(s)
GMA systems see Generalized mass action (GMA) Heat-shock proteins (HSPs) 205, 306–307 innate and adaptive 428–429, 429F
systems Heat stress proteins 215, 215F, 216F
Goal-oriented manipulations 407–408 analysis 312–314, 313F, 313T, 314F Immunoglobulins 215, 215F
Goals, model development 24 glucose depletion 314–315, 315F Immunoprecipitation 225
GO (Gene Ontology) Consortium 176 Heat stress response 303–330 Independent variables 33–34, 33f
Golden Rice 415 brief review 306–310, 307F–309F trehalose cycle 312, 313, 313F
Goldman-Hodgkin-Katz equation 348, 360 gene expression 308, 309F, 315–318, 316F, 316T, 317F Inducers 185–186
Gompertz growth law 284–286 modeling analysis 310–318 Inducible promoters 409
Gout 391 analysis of heat stress 312–314, 313F, 313T, 314F Infection 428–429, 429F
G6P see Glucose 6-phosphate (G6P) design and diagnosis of metabolic pathway Infectious disease modeling 23
G-protein(s) 257–258 model 310–312, 312F asymptomatics 43–45, 44F, 45F
G-protein-coupled receptor (GPCR) 257–258, gene expression 315–318, 316F, 316T, 317F canonical 102, 103
258F, 349 glucose dynamics 314–315, 315F consistency and robustness 39F, 40
Gradient methods, parameter estimation 142, multiscale analysis 318–326 diagram 35F
145–146, 147F model design 320–324, 323T, 324F, 324T model selection 26
Granger causality 62 trehalose puzzle 215F–327F, 324–326 model structure 27
Graphs in vivo NMR profiles 318–320, 319F–321F parameter estimation 37
adjacency matrices 55, 55F origin of models 304–305 Inflammation 428–432, 429F, 430F
bidirectional 53, 55F Heme 212F Information
bipartite 54 Hemoglobin 212, 212F coarse vs fine-grained 437
clique 55, 60 Henri-Michaelis-Menten rate law see Michaelis- mutual 62–63, 63F
clustering coefficient 55–56, 56F Menten rate law Information theory 62
defined 53 Hepatocytes 362 Inhibition
degree distribution 57, 58F Heraclitus of Ephesus 52 allosteric 211, 239
degrees 55, 55F Hershko, Avram 205 competitive, uncompetitive, and noncompetitive
directed acyclic 67–68 Heterogeneity, medicine 369–370 239
directed vs. undirected 53, 53F, 55–56, 55F, 56F, 62 modeling 370–372, 371F, 371T, 372F Inhibitors, metabolic systems 238–240
edges 53, 55 Heterogeneous information merged into conceptual Initial value 94
hubs 56, 57F, 60 model 15 Innate immune system 428–429, 429F
interaction 53–61 Hexokinase 231, 232F Inputs 8–9, 8F, 9F
nodes 53, 55, 55F, 56B trehalose cycle 308–309 model development 24
properties 54–58, 55F–58F, 56B Hexose 321F, 411, 412, 414 persistent changes 121
random 57–58 Hexose transporters (HXTs) 308 Insulin
signaling networks as 258, 259–261 HGPRT (hypoxanthine-guanine phosphoribosyl- crystal structure 208F
theory 54, 54F transferase) 392 design of biological systems 399, 408
vertex 53 High-density lipoproteins (HDLs) 217 diabetes modeling 428
Green fluorescent protein (GFP) 194, 222, 223F High-performance liquid chromatography drug development 381
Grid search, comprehensive 143–145, 144F (HPLC) 247 metabolic network analysis 251
GroEL/ES protein complex 205 High-throughput methods, of molecular biology model design 29, 31
Growth curves, children 47, 48F 11, 12F Integrative analysis, genome, protein, and metabolite
Growth dynamics 284 Hill-climbing methods 142, 145–146, 147F data 303–330
Growth laws 284–286 Hill function Interacting populations 292–299
Growth rate 284 heart 344 general modeling strategy 292, 292F
Guanine 171, 172F, 173F metabolic systems 240 more complex models 297–299, 298F, 298T
Guare, John 56 parameter estimation 36, 37, 37F, 160 phase-plane analysis 292–297, 293F–297F, 293T
Gut microbiome 432–433 signaling systems 262F, 265B Interaction(s) 16
INDEX 463
Interaction graphs 53–61 Levinthal’s Paradox 204–205 metabolic 74, 74F, 250
edges 53, 55 Lewin, Kurt 440 stoichiometric connectivity 242
nodes 53, 55, 55F, 56B Life cycle 16 MAPK (mitogen-activated protein kinase) cascades
properties 54–58, 55F–58F, 56B Ligand binding 384 270–273, 270F–273F, 271T
theory 54, 54F Limit-cycle oscillations 102 Markov chain models 84F, 88–91
vertex 53 Limit cycles 123–125, 123F, 125F, 126F Markov chain Monte Carlo (MCMC) algorithms 68
Interaction networks, Bayesian reconstruction 63–69, Limiting distribution 90 Markov models 84F, 88–91
65F, 67F Linear chains 8, 8F Mark-up languages 12
Introns 177, 177F Linear differential equations 94 Mass action kinetics 235
Inverse parameter estimation 439 Linearization 95 heart function 359–360, 360F
Inversion recombination elements 410 Linearized models 95–100 Mass-action models 103, 268
Ion(s) 332, 345 approximation 95–98, 96B, 97B, 97F see also Generalized Mass Action systems
Ion channels curved surface 95, 95F Mass action systems, generalized 103–104, 104F
heart 346, 348, 350, 353, 361 Jacobian 98–100, 99F, 100F structure identification 160
protein systems 221, 221F operating point 95 trehalose cycle 311
Ion transport 211, 337, 346–347 perturbation 95 Mass spectrometry (MS) 220, 220F, 223F
Irreversible step, trehalose cycle 308 Linear models 84 metabolite analysis 247–248, 248F
Isobaric tags for relative and absolute quantitation Linear regression 27, 27F Mathematica 435
(iTRAQ) technique 219 misuse 137–138, 137F Mathematical model(s) 19–50
Isoleucine (Ile) 203F, 204T parameter estimation 136–141 analysis 37–42
Isoprenoid pathway 407, 414 several variables 138–141, 138F–141F, 139T–141T approximations 29, 29F
Isozymes single variable 136–138, 137F binomial 30, 32B
control of metabolic pathways 246 Linear relationship, analysis of heat stress 312 blue sky catastrophe 30, 31B
trehalose cycle 308, 317 Linear systems 84 computer simulations 19, 38, 38F
Iterations 46 continuous 93–100 consistency and robustness 38–40, 39F
iTRAQ (isobaric tags for relative and absolute linear differential equations 94 correlative vs explanatory 27, 27F
quantitation) technique 219 linearized models 95–100, 95F, 96B, 97B, 97F, data availability 25–26
99F, 100F defined 19
discrete 85–91, 85T, 86F deterministic 29, 30, 31B
recursive deterministic models 85–88, 85T, 86F diagnostics 37–40, 39F
J recursive stochastic models 84F, 88–91
parameter estimation 136–141
diagram 34–35, 34F, 35F
discrete vs continuous 29
Jackknife method, parameter estimation 152–153 several variables 138–141, 138F–141F, equations 35–36
Jacob, François 186 139T–141T exploration and validation of dynamical features
Jacobian of system 98–100, 99F, 100F single variable 136–138, 137F 38, 40–42, 41B, 42F, 43F
Julia sets 93 Lin-log model 107, 115 extensions and refinements 43–45, 44F, 45F
“Junk DNA” 176–177 Lipid Maps 242 external validation 40
c-Jun N-terminal kinase (JNK) 270 Lipo-oligosaccharides 209F flow chart of process 22, 22F
Lipoproteins 217 flow of information 32–33, 33F
Liver flow of material 32, 32F
compartment model 385–386, 386F, 387F goals, inputs, and initial exploration 24–26, 24F
K drug development 379, 385–387, 388B, 390
enzyme activity profile 52
issues to address 22–23, 23T
large-scale model assessments 45–46, 47F
Kai system 277, 277F lipoproteins 217 open vs closed 29
Kauffman, Stuart 260 multiscale modeling 362 overview of 19–24, 20F, 22F, 23T
Keasling, Jay 414 Virtual Liver Network 333, 356, 362 overfitting 40, 160
KEGG database Local minimum 146 parameters 31–32, 34, 36–37, 37F
metabolic network reconstruction 74 Local motifs 221 Poisson 30, 33B
metabolic systems 241F, 242, 243B, 250, 251 Logistic function, graph 101, 101F probabilistic or stochastic 29–30, 32B, 33B
trehalose cycle 308, 311 Logistic growth function 283, 284 processes and interactions 31
Kekulé von Stradonitz, Friedrich August 304 Logistic map 91–93, 92F questions of design 46–47
Keratin 216–217 Lorenz, Edward 126 questions of scale 24–25, 24F
Kinases 214 Lorenz attractor 126–127, 127F selection and design 26–37, 26F
Kinetic models 160, 251, 277 Lotka, Alfred 296 sensitivity 39–40, 39F
Kinetic order 103 Lotka-Volterra (LV) model 101, 102–103, 103F, 114 simplicity vs complexity 47–49, 48F
trehalose cycle 311 interacting populations 296–298, 298F, 298T spatial aspects 29
KInfer 242 Low-density lipoproteins (LDLs) 217 static vs dynamic 27–29, 28B
Kirchhoff’s laws 351 Low-diameter networks 56B steady state 38–39
Kitano, Hiroaki 440 L-type calcium (Ca2+) channels (LCCs) 349, 356F, structure 27–30
Kyoto Encyclopedia of Genes and Genomes see 357–361, 357F system components 30–35
KEGG database Lung(s) use and applications 43–49
and heart 333, 334, 362 variables 30–31, 33–34, 33F
infection and inflammation 428, 431 Mathematics of biological systems 83–134
pharmacokinetic modeling 385, 389, 390 continuous linear 93–100
L pneumonia 285
Lung cancer 372
linear differential equations 94
linearized models 95–100, 95F, 96B, 97B, 97F,
lacI regulator gene 186, 186F Lung Physiome 333 99F, 100F
lac operon 186–187, 186F, 187T Lwoff, André Michel 186 continuous nonlinear 100–110
Lactococcus lactis 252 Lymphocytes, suppressor 404, 405 ad hoc models 101–102, 101F, 102F
Lactose 185–187 Lysine (Lys) 203F, 204T, 251 canonical models 101, 102–110, 103F–105F, 107F,
Laplace transform 94 108B, 109B
Large-scale model assessments 45–46, 47F more complicated dynamical systems
Latin hypercube technique 121, 144 descriptions 110
Latin square technique 121
Laws, biological 29, 46, 440
M discrete linear 85–91
recursive deterministic models 85–88, 85T, 86F
LCCs (L-type Ca2+ channels) 349, 356F, 357–361, 357F Machine learning 12, 148 recursive stochastic models 84F, 88–91
LDLs (low-density lipoproteins) 217 Malaria 414–415 discrete nonlinear 91–93, 92F
Lead(s) 379 MALDI (matrix-assisted laser desorption/ionization) other attractors 122–128
Lead identification, computational 381 222, 223F chaotic 126–128, 127F, 128F
Lead optimization 379 MALDI-TOF (matrix-assisted laser desorption/ limit cycles 123–125, 123F, 125F, 126F
Least-squares method, parameter estimation 135 ionization-time of flight) 220 overview 83–85
Left ventricular hypertrophy 356 Malthus, Thomas Robert 283 standard analyses 110–122
Lens 217, 218F Mandelbrot sets 93 analysis of systems dynamics 119–122, 120F, 122F
Leslie model, population dynamics 289–292, 291B Map(s) parameter sensitivity 118–119
Leucine (Leu) 203F, 204T concept 437, 438–439, 439F stability analysis 115–117, 116F, 117F
Leukemia, acute lymphoblastic 373–374 logistic 91–93, 92F steady-state analysis 110–115, 112F, 113F
464 INDEX
Mating process, genetic algorithms 142–143, 147 Metadata 433 Multidimensional protein identification technology
MATLAB 435 Metagenomics 172, 433 (MudPIT) 219
Matrices 84 Metapopulations 286, 432–433, 433F Multidisciplinary team 15
Matrix-assisted laser desorption/ionization (MALDI) Methicillin-resistant Staphylococcus aureus Multiple linear regression, parameter estimation
222, 223F (MRSA) 170 138–141, 138F–141F, 139T–141T
Matrix-assisted laser desorption/ionization-time of Methionine (Met) 203F, 204T Multiple organ failure 431
flight (MALDI-TOF) 220 Method of mathematically controlled comparison Multiplex automated genome engineering 407
Matrix equation, population dynamics 289–290 (MMCC) 404–406, 406F Multiscale analysis, heat stress response 318–326
Matrix proteins 208 Methotrexate (MTX) 388B–389B model design 320–324, 323T, 324F, 324T
MBDs (methyl-CG-binding domains) 179F Methyl-CG-binding domains (MBDs) 179F trehalose puzzle 215F–327F, 324–326
MCA (metabolic control analysis) 74–78, 75F, 77F, 107 Methyl groups 178 in vivo NMR profiles 318–320, 319F–321F
metabolic pathways 242 Metropolis-Hastings method 68 Multiscale modeling 25, 361–362, 436–437
MCMC (Markov chain Monte Carlo) algorithms 68 Mevalonate pathway 414, 415D Multistage cancer model 432
Medicine 369–378 MI (myocardial infarction) 339 Multivariate functions, power-law approximations
biological variability and disease 369–370 Michaelis-Menten function 109B
modeling 370–372, 371F, 371T, 372F ad hoc 102 Multivariate systems 31
personalized (precision) 372–378 MAPK cascade 270F, 271 Muscle cells 211, 217, 335, 338
data needs and biomarkers 373–374, 374F parameter estimation 36, 139, 140 Mutations
mathematical models 374–378, 377F PBPK models 390 analysis of system dynamics 121
traditional vs. 374, 375F power-law approximations 109B biologic variability 370
predictive health 372–378 Michaelis-Menten process cyanobacteria 434
Membrane-bound receptors, drug development biological variability and disease 370–372, 371F, drug development 415
382, 382F 371T, 372F evolutionary programming 148
Membrane potential 346–349 mathematical model 235–236 genetic algorithms 146, 147, 147F, 148
Mendel, Gregor 176, 440 pharmacokinetics 387 heat stress response 312
Mendelian genetics 440 power-law approximations 109B immune system 215
Mesarović, Mihajlo 10 Michaelis-Menten rate law (MMRL) metabolic engineering 407
Messenger proteins 214, 214F metabolic systems 236–238, 237F, 238F metabolic systems 231
Messenger RNA (mRNA) 179 parameter estimation probability 32B, 33B
Metabolic analysis linear systems 139–140, 139F–141F, 140T, 141T random 30
flux 249 nonlinear systems 143–144, 144F rate 33B
incompletely characterized organism 250–251 trehalose cycle 311 sensitivity analysis 118
network 251 Microarrays 193–194, 193F sickle cell disease 104, 372
resources 241–244, 243B, 244B drug sensitivity 374 signaling systems 277
Metabolic burden 408 heat stress response 315 simulated annealing 149
Metabolic control analysis (MCA) 74–78, 75F, 77F, 107 Microbial biofilms 433, 433F small-world networks 59–60
lin-log model 107,115 Microbial metapopulations 432–433 transposons 170–171
metabolic pathways 242 Microbial systems 399, 418, 419 vectors 142–143
Metabolic engineering 407–408, 411–414, 413F Microbiome 432–433 Mutual information 62–63, 63F
elementary mode analysis 411–414, 413F Micro-RNA (miRNA) 177, 183, 184, 409 Mycoplasma pneumoniae 250
Metabolic flux analysis 249 Middle-out analyses 17 Myocardial cells see Cardiomyocytes
Metabolic map 74, 74F, 250 Mill, John Stuart 62 Myocardial infarction (MI) 339
Metabolic network(s) 52, 69–78 miRNA (micro-RNA) 177, 183, 184, 409 Myocardium 335
Bayesian methods 69 Mitochondria 1 Myocytes, heart see Cardiomyocytes
metabolic control analysis 74–78, 75F, 77F Mitochondrial oxidative phosphorylation 77 Myofibrils 338
reconstruction 73–74, 74F Mitogen-activated protein kinase (MAPK) cascades Myofilaments 338, 338F
stoichiometric 70–73, 70F–72F 270–273, 270F–273F, 271T Myosin 338, 338F
variants of stoichiometric analysis 73 MMCC (method of mathematically controlled
Metabolic network analysis 251 comparison) 404–406, 406F
Metabolic pathway(s), structural patterns 401 MMRL see Michaelis-Menten rate law (MMRL)
Metabolic pathway model, design and diagnosis
310–312, 312F
Model(s)
dynamic(al) see Dynamic(al) models
N
Metabolic profile 240 mathematical see Mathematical model(s) Na+-Ca2+ (sodium-calcium) exchanger 347
Metabolic reactions origin of 304–305 Nagumo, Jin-Ichi 332
mathematical formulation 234–235 static network see Static network models Na+-K+ ATPase (sodium-potassium adenosine
rate laws 235–240, 235F, 237F, 238F Model extensions 43–45, 44F, 45F triphosphatase) 347
Metabolic syndrome 428 Modeling Na+-K+ (sodium-potassium) pump 346–347
Metabolic systems 231–255 defined 19 Nanomedicine 11, 13F
biochemical reactions 232–240 needs 435–439 NCEs (new chemical entities) 378–379
background 232–234, 233B Modeling approach 16 Negatively inducible operons 187
mathematical formulation of elementary Modeling tools 16–17 Negatively repressible operons 187
reactions 234–235 Modules 20 Nernst equation 347–348
rate laws 235–240, 235F, 237F, 238F Modulons 188 Nested feedback 245, 246F
data generation methods 246–249 Molecular biology Network(s) 51
detection methods 247–249, 248F DNA sequences 178 Bayesian reconstruction 63–69, 65F, 67F
flux analysis 249 high-throughput methods of 11, 12F flux 70–73, 70F–72F
sampling, extraction, and separation methods Molecular chaperones 205, 207F gene regulatory 69
247, 251–252 Molecular dynamics 222 low-diameter 56B
from data to systems models 250–252 Molecular function, gene products 176 metabolic 52, 69–78
extraction of dynamic models from experimental Molecular inventories 16, 17 Bayesian methods 69
data 251–252 Molecular systems 53, 376 metabolic control analysis 74–78, 75F, 77F
incompletely characterized organism 250–251 Monoclonal antibody 382F–385F, 383–385, 384T reconstruction 73–74, 74F
metabolic network analysis 251 Monod, Jacques 186 stoichiometric 70–73, 70F–72F
metabolites 231–232, 232F Monte Carlo simulation 45–46, 47F variants of stoichiometric analysis 73
overview 231–232 personalized medicine 376 motifs 53, 402–404, 402F
pathways and pathway systems 240–246 Mori, Hirotada 11 neural 426–427, 426F, 427F
biochemistry and metabolomics 240–241, 241B Motifs, network 53 protein-protein interaction (PPI) 53, 60, 69
control 244–246, 245F, 246F design of biological systems 399, 402–404, random 56B
resources for computational analysis 241–244, 402F, 403F regular 56B
243B, 244B mRNA (messenger RNA) 179 regulated 54
Metabolism, defined 232 MRSA (methicillin-resistant Staphylococcus aureus) scale-free 57–58, 58F
Metabolites 231–232, 232F 170 signaling, Bayesian analysis 66–69, 67F
integrative analysis 303–330 MS (mass spectrometry) 220, 220F, 223F small-world 58–60, 59B, 61F
Metabolomics 232, 240–241 metabolite analysis 247–248, 248F biological variability and disease 372
data generation methods 246–249, 248F, 249F MTX (methotrexate) 388B–389B static 51–82
MetaCyc database 74 Mucins 217 Bayesian reconstruction 63–69
INDEX 465
dependencies among components 62–63, 63F Operons 185 heart 334, 358
interaction graphs 53–60 lac 186–187, 186F, 187T models of gene regulation 191
metabolic 52, 69–78 regulation 187–188 Particle swarm optimization (PSO) 148
small-world 58–60, 59B, 61F structural patterns 401, 401F Path lengths 59
stoichiometric 70–73, 70F–72F über- 188, 401, 401F Pathogen(s)
strategies of analysis 52–53 Optimization 54 antibodies 215
stoichiometric 70–73, 70F–72F constrained 73 immune system 428, 429, 430, 431
types of generic 51–52, 51F design of biological systems 400 lipo-oligosaccharides 209F
undirected 62 lead 379 quorum sensing 278, 418
Network interference 261 metabolic engineering 407 Toll-like receptors 216F
Network motifs, design of biological systems 399, parameter estimation 135 Pathogenicity 418
402–404, 402F, 403F Orbit 123 Pathway screening, dynamic models 390–393, 391F
Neural networks 426–427, 426F, 427F Ordinary differential equations (ODEs) 91, 110 Pathway Studio 242
Neurobiology 426–427, 426F, 427F heart 358 Pathway systems
Neuron(s) metabolic systems 234 gene expression 317
action potential 11, 350 models of gene regulation 190 generic 33F
multiscale modeling 362, 436 population growth 284, 285–286 metabolic 232, 240–246
neural networks 426–427, 427F signaling systems modeled with 258, 261–278 analysis 52, 53
signaling systems 261 ORF (open reading frame) 177 design 407, 408, 411–412, 413F
Neuronal signal transduction 426 Organism interactions with environment 432–434, genome size 174
Neurotransmitters 426 433F–435F reconstruction 74
New chemical entities (NCEs) 378–379 OR gate 402 screening with dynamic models 390–393
NMR spectroscopy see Nuclear magnetic resonance Origin-binding domains (OBDs) 173F simple 8
(NMR) spectroscopy Oscillation patterns Pavlov, Ivan Petrovich 440
Noble, Denis 332 analysis of system dynamics 119, 121 PBD (platform-based design) 410
Nodes 53 chaotic 128 PBPK (physiologically based pharmacokinetic)
degrees associated with 54–55, 55F circadian clocks 275–278, 276F, 277F models 385, 387–390, 387F, 388B–389B
number of paths between 55, 56B damped 9, 9F, 103, 155 PCR (polymerase chain reaction), reverse
Noise, parameter estimation 149, 150F, 156–157, 156F, gene circuits 417, 418F transcriptase 192–193, 192F
157F, 157T heart 332, 336, 338, 339, 339F, 345, 345F, 347 PDEs see Partial differential equations (PDEs)
Noisy channels, reliability of communication through limit-cycle 102, 123–125, 123F, 125F, 126F, 438 PDIs (protein disulfide isomerases) 205
62–63 parameter sensitivity 118 Pentose(s) 411, 412
Noncoding DNA 174, 176–178, 177F periodicity 93 Pentose phosphate pathway (PPP) 309, 316F
Noncoding RNA 409 predator-prey 296, 296F Peptide(s) 204
Noncompetitive inhibition 239 sine-cosine 93 Peptide bonds 202, 202F
Nonlinear models 84 stable, S-system 340, 340F Period doubling 93
Nonlinear partial differential equation, heart 334 unpredictable deterministic model 31B Permeases 346–347
Nonlinear processes 5, 5F, 8 Oscillations models 339–345, 339F Persistent changes in structure or input 121
Nonlinear regression 27, 27F black box 339–343, 340F–342F Personalized medicine 372–378
parameter estimation 142, 145–146, 145F meaningful 343–345, 344F, 345F data needs and biomarkers 373–374, 374F
Nonlinear systems 84 repeated heartbeats 354 mathematical models 374–378, 377F
continuous 100–110 Overfitting, structure identification 160 traditional vs. 374, 375F
ad hoc models 101–102, 101F, 102F Oxidative stress 355 Perturbation(s)
canonical models 101, 102–110, 103F–105F, 107F, analyses of biological systems models 110
108B, 109B analysis of systems dynamics 121
more complicated dynamical systems Bayesian reconstruction of interaction networks 63
descriptions 110
discrete 91–93, 92F
P bistability and hysteresis 262
data-modeling pipeline 439
parameter estimation 141–153 p53 protein 27 design of biological systems 405, 416, 417
comprehensive grid search 142, 143–145, 144F Pacemaker, effectiveness of simple 342, 342F drug development 391, 392
genetic algorithms 142–143, 146–148, 147F Pacemaker activity 336 heart 339, 345, 360
nonlinear regression 142, 145–146, 145F Pacemaker cells 337 heat stress response 312
other stochastic algorithms 148–149 Pacemaker system abnormalities 355 inflammation 429
search algorithms 142–143, 143F Parameter(s) limit cycles 123, 124, 125F
typical challenges 149–153, 150F–152F, defined 31–32 linearized models 95, 98
150T, 152T model development 34 metabolic control analysis 74, 75, 76
Noradrenaline 336 population 284 modeling variability and disease 371–372
Northern blot 192, 192F sensitivity analysis 118–119 other biological networks 69
NTH1/2 gene 310, 325 Parameter estimation 135–167 parameter sensitivity 118, 119
Nuclear magnetic resonance (NMR) spectroscopy bottom-up strategy 136 phase-plane analysis 296, 297
206, 206F identifiability 160 population dynamics under external 288–289,
heat stress response 318–320, 319F–321F branch-and-bound method 144–145 288F, 289F
metabolite analysis 247, 248–249, 248F, 249F evolutionary algorithms 142–143, 146–148, 147F recovery from 38
Nucleo-bases, DNA and RNA 171, 172F, 173F linear systems 136–141 signaling networks 69, 277, 277F
Nucleosides 171, 172F several variables 138–141, 138F–141F, 139T–141T stability analysis 115, 116, 116F, 117
Nucleosome 178, 179F, 180F single variable 136–138, 137F static network models 51
Nucleotides 171, 172F in model selection and design 36–37, 37F steady-state analysis 111
Nullcline, interacting populations 294–295, noise 149, 150F, 156–159, 156F–159F, 157T, 158T Peskin, Charles 332
294F, 295F nonlinear systems 141–153 Petri net analysis 54
Numerical mathematics 12 comprehensive grid search 142, 143–145, 144F PFK1/2 genes 310, 317
genetic algorithms 142–143, 146–148, 147F Pharmacogenomics 373–374
nonlinear regression 145–146, 145F Pharmacokinetic modeling 385–390, 386F, 387F,
other stochastic algorithms 148–149 388B–389B
O search algorithms 142–143, 143F
typical challenges 149–153, 150F–152F,
Pharmacokinetics 388B–389B
Pharmacokinetic tests 379
OBDs (origin-binding domains) 173F 150T, 152T Phase-plane analysis 292–297, 293F–297F, 293T
Objective function, parameter estimation 135 overview 135–136 Phenotype 176
Oceans 433–434, 434F, 435F simulated annealing 148 Phenylalanine (Phe) 203F, 204T
ODEs see Ordinary differential equations (ODEs) sloppiness 151 Phosphate-deoxyribose backbone 172, 173F
Ohm’s law 351 SSE (sum of squared errors) 135, 137 Phosphofructokinase 310, 317
Oligonucleotides 193–194, 381 structure identification 160 Phospholipase A 222F
Open reading frame (ORF) 177 systems of differential equations 153–159, 153F, Phosphoribosylpyrophosphate synthase (PRPPS) 392
Open system model 29 154T, 156F–159F, 156T–158T Photosynthesis 24, 25, 73, 208, 434, 436
Operating point (OP) 95 time-series-based 136, 153 Phylogeny 171
Operating principles 406–407, 406F uncertainty 151–152 Physiologically based pharmacokinetic (PBPK)
see also Design principles Partial differential equations (PDEs) 110 models 385, 387–390, 387F, 388B–389B
466 INDEX
Physiological modeling of heart 331–367 pharmacokinetics 387 P wave 350, 350F

electrochemistry in cardiomyocytes 345–354 population growth 286 Pyrimidines, DNA and RNA 171, 172F, 173F
biophysical description 347–348 trehalose cycle 311 Pyrococcus furiosus 175F
models of action potentials 350–354, 353F Power-law model, biological variability and disease Pyrrolysine 203F
repeated heartbeats 354, 354F 370–372, 371F, 371T, 372F Pyruvate 8, 210, 210F
resting potentials and action potentials 348–350, PPI (protein-protein interaction) networks 53, 60, 69,
349F, 350F 224–225, 224F, 225F
failing heart 355–361 PPP (pentose phosphate pathway) 309, 316F
modeling based on molecular events 356–361,
356F, 357F, 359F, 360F
Precision, adaptation 273–274, 273F
Precision medicine see Personalized medicine
Q
hierarchy of scales and modeling approaches Preclinical phase, drug development 378–379, 379F QRS complex 350, 350F
332–339 Predation, population dynamics 288 QS (quorum sensing) 278, 418
basics of heart anatomy 333–334, 333F Predator-prey systems 296, 296F, 297–298, 298t QSARs (quantitative structure-activity relationships)
at cell level 337–339, 338F, 339F Prediction(s) 16, 19 379
at organ level 334–335, 335F Predictive health 372–378 QSSA (quasi-steady-state assumption) 236
at tissue level 335–337 Preferential attachment 56 QT interval 350, 350F
outlook for physiological multiscale modeling PR interval 350, 350F QTL (quantitative trait loci) 176
361–362 Pro (proline) 203F, 204T Qualitative analysis, interacting populations 292–297,
overview 331–332, 332F Probabilistic laws 440 293F–297F, 293T
simple models of oscillations 339–345, 339F Probabilistic models 29–30, 32B, 33B Qualitative behavior, validation of model 40
black box 339–343, 340F–342F Probabilities, conditional 64–65, 65F Quantitative structure-activity relationships (QSARs)
meaningful 343–345, 344F, 345F Probability distribution, binomial 32B 379
Physiome Project 333, 356, 362 Prokaryotes Quantitative trait loci (QTL) 176
Piwi-interacting RNAs (piRNAs) 184 codon usage bias 204 Quasi-steady-state assumption (QSSA) 236
Plant breeding 407 flagella 213 Quaternary structure 221
Plasma membrane calcium adenosine genes 188 Quenching of metabolic activity 247
triphosphatase (PMCA) 347 MAPK cascades 270 Quorum sensing (QS) 278, 418
Plasmids 170, 178 synthetic biology 408
Plasmodium falciparum 414, 415 transcription factors 190
Platform-based design (PBD) 410 Prokaryotic cells 432
PMCA (plasma membrane calcium adenosine
triphosphatase) 347
Proline (Pro) 203F, 204T
Promoters 179F
R
Pneumonia 285 inducible 409 Radical oxygen species 429
Poisson model 30, 33B Proteases 106, 208, 245, 381, 382F, 410 Radioactive decay 234
Polychaos dubium 174 Proteasome 205 Random events 292
Polymerase chain reaction (PCR), reverse heat stress response 307 Random graphs 57–58
transcriptase 192–193, 192F Protein(s) 201–230 Random mutations 30
Polysaccharide, trehalose cycle 308 chemical and physical features 202–206 Randomness 29–30
Population(s) amino acids 202–204, 203T, 204T Random networks 56B
homogeneous 289 conformation (folding) 204–205 Random variables 29
interacting 292–299 experimental determination and visualization Rate constants 94, 311
general modeling strategy 292, 292F 206–208, 206F–208F Rate laws 235–240, 235F, 237F, 238F
more complex models 297–299, 298F, 298T peptide bonds 202, 202F Reactome 242
phase-plane analysis 292–297, 293F–297F, 293T post-transcriptional modifications 205 Receptor(s) 214
meta- 286 primary structure 221 Receptor dynamics, drug development 382–385,
sub- 286 quaternary structure 221 382F–385F, 384T
analysis 289–292, 291B secondary structure 221 Receptor-gated channels 346
Population dynamics 34 shapes 208, 208F Receptor tyrosine kinases (RTKs) 214
ad hoc models 101 size 204 Recursion 290
under external perturbations 288–289, 288F, 289F tertiary structure 202, 204–205, 221, 221F Recursive deterministic models 85–88, 85T, 86F
metapopulations 286 turnover 205 Recursive stochastic models 84F, 88–91
complex models 297–299, 298F, 298T integrative analysis 303–330 Red blood cells
nullcline 294, 295, 294F, 295F overview 201 drug development 414
predator-prey systems 296, 298 research challenges 218–225 heart 334
Population growth 283–286, 284F activity and dynamics 224–225, 224F, 225f trends over time 86, 86F, 87, 88, 252
exponential 283 localization 222, 223F, 224F Redox 317
logistic 283 proteomics 218–220, 219F, 220F Reductionism 5–8, 6F, 7F, 11, 14–15, 14F
complex phenomena 286, 287B structure and function prediction 220–222, Refractory period 336
stochastic events 284 221F, 222F Regular networks 56B
traditional models 284–286 roles and functions 208–218 Regulated networks 54
trends 286, 287B enzymes 209–211, 209F, 210F Regulatory signals 8–9, 8F, 9F
Population parameters 284 immune system 215, 215F, 216F Regulons 188
Population systems 283–302 signaling and messenger 214, 214F structural patterns 401, 401F
analysis of subpopulations 289–292, 291B structure 216–218, 217F, 218F Repeller 122
interacting populations 292–299 transporters and carriers 211–214, 212F, 213F Replication 171–172, 173F
general modeling strategy 292, 292F “Protein cages” 13F Repolarization 348, 349, 350, 350F
more complex models 297–299, 298F, 298T Protein disulfide isomerases (PDIs) 205 lengthening of action potential 355
phase-plane analysis 292–297, 293F–297F, 293T Protein folding 13, 204–206, 208, 221 Repressilator system 276, 417, 417F
population dynamics under external perturbations heat stress response 306, 310 Repressor(s) 185–186
288–289, 288F, 289F Protein-protein interaction (PPI) networks 53, 60, 69, Repressor module 417
population growth 283–286, 284F 224–225, 224F, 225F Residual error 113
exponential 283 Protein scaffolds 178, 180F Resistance 351
logistic 283 Proteome 218 Resting potentials 337, 348–350
more complex phenomena 286, 287B Proteomics 218–220, 219F, 220F Reverse engineering 63
traditional models 284–286 community 433 Reverse transcriptase polymerase chain reaction
Positively inducible operons 187 metabolic pathways 245 (RT-PCR) 192–193, 192F
Positively repressible operons 187 PRPPS (phosphoribosylpyrophosphate synthase) 392 Rheumatoid arthritis 381, 429
Post-approval phase, drug development 380 Pseudomonas aeruginosa 278 Ribonucleic acid see RNA (ribonucleic acid)
Post-transcriptional modifications (PTMs) 205 PSO (particle swarm optimization) 148 Ribose 172F
Potassium channels 221F, 307, 349, 353 PTMs (post-transcriptional modifications) 205 Ribosomal RNA (rRNA) 182–183, 183F
Potassium leak channel 346 Pulmonary artery 333F, 334 Ribosomes 1, 182–183, 205
Potrykus, Ingo 415 Pulmonary vein 333F, 334 Riboswitches 410, 410F
Power function 85 Pumps 346–347 Ribozymes 183, 183F
Power-law approximation 107, 108B, 109B Purine(s), DNA and RNA 171, 172F, 173F Richards function 284, 285
Power-law distribution 57, 58F Purine metabolism 391–393, 391F Rinzel, John 332
Power-law functions 103 Purkinje fibers 336, 337–338 RISC (RNA-induced silencing complex) 184
INDEX 467
RNA (ribonucleic acid) Shock states 431 Sodium-potassium adenosine triphosphatase

chemical and physical features 172F, 178–179, 181F Short-interfering RNA (siRNA) 183, 184, 184F (Na+-K+ ATPase) 347
double-stranded 183–184, 192 Sickle cell anemia 204, 372 Sodium-potassium (Na+-K+) pump 346–347
key properties 178–185 Sick sinus syndrome 355 Somogyi-Stucki model 344–345, 344F, 345F
messenger 179 Signal(s) 257 Source antigen 404–406
micro- 177, 183, 184, 409 Signaling cascade Southern, Edwin 191
noncoding 409 defined 66 Spatial scale 10
piwi-interacting 184 MAPK 270, 273 Specialized ontologies 433
ribosomal 182–183, 183F metabolic pathway analysis 242 Sphingolipids 306
short-interfering 183, 184, 184F model design 32, 32F SR (sarcoplasmic reticulum) 343
silencing 183, 184, 184F, 409 network motifs 402, 403 SR Ca2+ ATPase (SERCA) pumps 357
single-stranded 179, 181F, 182, 183F, 184 Toll-like receptors 216F Srienc, Friedrich 411
small 183–184, 184F Signaling molecules SSE (sum of squared errors) 135
small activating 184 cAMP 187 S-systems 104–105, 114–115
small-interfering 183, 184, 184F, 409 DNA splicing 177 oscillations 340, 340F
small temporal 177, 183, 184 drug development 379 structure identification 160
transfer 170, 182, 182F gene regulation 190 Stability analysis 115–117, 116F, 117F
viruses 184–185, 185F G-proteins 257, 258 trehalose cycle 311
RNA-induced silencing complex (RISC) 184 immune response 215 Staphylococcus aureus 278
RNA interference (RNAi) pathways 183–184 quorum sensing 278, 418 methicillin-resistant 170
RNAi (RNA interference) pathways 183–184 simple system 8 START codon 204T
RNA polymerase 180 Signaling networks see Signaling systems State, mathematical model 38
RNA-Seq 194 Signaling pathways, structural patterns 401 Static condition 52
RNA viruses 184–185, 185F Signaling proteins 214, 214F Static graph model 261
Robotics 11 Signaling systems 257–281 Static models 27–29, 28B
Robustness Bayesian analysis 66–69, 67F Static network models 51–82
biological networks 401–402 G-protein-coupled receptor (GPCR) 257–258, 258F causality analysis 62
heat stress response 307, 307F modeled with differential equations 261–278 dependencies among network components
mathematical model 38–40, 39F, 118 adaptation 273–274, 273F, 274F, 274T 62–63, 63F
Rodbell, Martin 258 bistability and hysteresis 261–265, 262F–264F, graphs
Rose, Irwin 205 262T, 264T, 265B–266B adjacency matrices 55, 55F
rRNA (ribosomal RNA) 182–183, 183F mitogen-activated protein kinase cascades bipartite 54
RTKs (receptor tyrosine kinases) 214 270–273, 270F–273F, 271T clique 55, 60
RT-PCR (reverse transcriptase polymerase chain other 274–278, 276F, 277F, 278T clustering coefficient 55–56, 56F
reaction) 192–193, 192F two-component 266–269, 266F–270F, 268T defined 53
RuBisCo 208 overview 257–258, 258F degree distribution 57, 58F
Ryanodine receptor (RyR) 357–361, 357F second messengers 258 degrees 55, 55F
static models 259–261 edges 53, 55
Boolean networks 259–261, 259F, 260F hubs 56, 57F, 60
network interference 261 interaction 53–61
S Signal transduction
cellular stress response 305
nodes 53, 55, 55F, 56B
properties 54–58, 55F–58F, 56B
SA (simulated annealing) 148–149 defined 214, 257 theory 54, 54F
Saccharomyces cerevisiae 57F, 185 enzymes 258 vertex 53
SAGE (serial analysis of gene expression) 193 extracellular matrix 208 metabolic 52, 69–78
Sampling, metabolite analysis 247 G-proteins 258 fluxes 70
SA (sinoatrial) node 336 heart 338, 346, 347, 362 metabolic control analysis 74–78, 75F, 77F
SAR(s) (structure-activity relationships), heat stress response 306 reconstruction 73–74, 74F
quantitative 379 importance 66 stoichiometric networks 70–73, 70F–72F
Sarcolemma 356–357, 357F lipids 242 variants of stoichiometric analysis 73
Sarcoplasmic reticulum (SR) 343 MAPK cascades 272, 273 mutual information 62–63, 63F
Sarcoplasmic reticulum (SR) calcium (Ca2+) metabolic systems 231 networks 51
adenosine triphosphatase (ATPase) (SERCA) network motifs 402, 403 Bayesian reconstruction 63–69, 65F, 67F
pumps 357 networks see Signaling systems generic 51–52, 51F
saRNA (small activating RNA) 184 neuronal 426 regulated 54
Savageau, Michael 11, 404, 440 neurotransmitters 211 scale-free 57–58, 58F
SBML (systems biology mark-up language) 12, 435 personalized medicine 375 small-world 58–60, 59B, 61F
Scaffold proteins 214 protein states 32F stoichiometric 70–73, 70F–72F
Scale(s) 10 receptor tyrosine kinases 214 optimization 54
model development 24–25, 24F Signal transduction pathways 262 overview 51
Scale-free networks 57–58, 58F Signal transduction protein 266 Petri net 54
Scatter plot 136 Silencing RNA (siRNA) 183, 184, 184F, 409 strategies of analysis 52–53
Screening experiments, comprehensive 121 Simulated annealing (SA) 148–149 Static structure of biological systems 16
SDS-PAGE (sodium dodecyl sulfate polyacrylamide Simulations 19, 38, 38F, 305 Statistical association analysis 374
gel electrophoresis) 191 Sine-cosine oscillations 93, 123–124, 123F Statistical association models 436
Search algorithms 113–114 Single nucleotide polymorphism (SNP) 369–370 Statistical mechanics 234
Secondary structure, RNA 179, 181F Sinoatrial (SA) node 336 Steady state
Second messengers 258 SIR model 35–36, 41B defined 110–111
Selenocysteine 203F linearization 99–100, 100F mathematical model 38–39
Semantic web tools 433 population dynamics 289 population growth 284
Semilunar outflow valves 333F, 334 siRNA (small-interfering RNA) 183, 184, 184F, 409 trehalose cycle 311
Sensitivity Size scale 10 Steady-state analysis 110–115, 112F, 113F
adaptation 273–274, 273F Sliding filament model 338, 338F Steepest-descent methods 142, 145–146, 147F
mathematical model 39–40, 39F Sloppiness, parameter estimation 151 Stimulons 188
parameters 118–119 Small activating RNA (saRNA) 184 Stimulus, validation of model 40
trehalose cycle 311–312 Small-interfering RNA (siRNA) 183, 184, 184F, 409 Stochastic differential equations 258, 261–278
Separation Small-molecule hits 379 Stochastic events 284
metabolite analysis 247 Small RNAs 183–184, 184F Stochasticity 91, 110, 261
of timescales 358–359 Small temporal RNAs (stRNAs) 177, 183, 184 Stochastic models 29–30, 32B, 33B
SERCA (SR Ca2+ ATPase) pumps 357 Small-world networks 58–60, 59B, 61F recursive 84F, 88–91
Serial analysis of gene expression (SAGE) 193 biological variability and disease 372 Stochastic processes 358
Serine (Ser) 203F, 204T Smoothing, parameter estimation 154 Stochastic simulations 45
Serum albumin 212–213 SNP (single nucleotide polymorphism) 369–370 Stoichiometric analysis 241–242, 243B, 244B
Seven Bridges of Königsberg 54, 54F Sodium-calcium (Na+-Ca2+) exchanger 347 Stoichiometric connectivity map 242
Shannon, Claude 62 Sodium dodecyl sulfate polyacrylamide gel Stoichiometric networks 70–73, 70F–72F
Shimomura, Osamu 222 electrophoresis (SDS-PAGE) 191 Stoichiometry 70
468 INDEX
STOP codon 204T

Strange attractor 126–128, 127F, 128F
Top-down analyses 16–17
Toxicity 379
V
Stress response elements (STREs) 307 T6P (trehalose 6-phosphate) 310 Validation 38, 40–42, 41B, 42F, 43F
Stress response system 6–8, 7F, 305 TPS1/2 genes 312 Valine (Val) 203F, 204T
see also Heat stress response T6P synthase (TPS1) 310, 313–317 van der Pol, Balthasar 124
stRNAs (small temporal RNAs) 177, 183, 184 Training set, parameter estimation 152 van der Pol oscillator 124, 125F
Structural patterns 400–402, 401F Traits, biological 176 Variability 30
Structure, persistent changes 121 Trajectories 123 and disease 369–372, 371F, 371T, 372F
Structure-activity relationships (SARs), Transcription 169, 172, 176, 180 Variables 30–31
quantitative 379 hierarchical control 189, 189F dependent 33–34, 33F
Structure identification 160 Transcription factors (TFs) 169, 188–190, 189F independent 33–34, 33F
Structure proteins 216–218, 217F, 218F heat stress response 307 trehalose cycle 312, 313, 313F
ST segment 350, 350F Transcriptome 193, 194 random 29
Subpopulations 286 Transfer RNA (tRNA) 170, 182, 182F Variations in biological components 9–10
analysis 289–292, 291B Transforming growth factor β (TGF-β) 430 Vectors 84, 142–143, 409
Sudden cardiac death 361 Transgenic foods 415 Vena cava 333F, 334
Sugar, DNA and RNA 171, 172F Transient cell cycle arrest 306 Ventricles 333–334, 333F
Sum of squared errors (SSE) 135 Transient decrease, translation initiation 306 Ventricular fibrillation 335F
Supercoiled DNA 174, 174F Translation 176 Ventricular tachycardia 335F, 355
Superposition 84 Translation initiation, transient decrease 306 Verhulst, Pierre-François 283
Suppressor lymphocytes 404, 405 Transmembrane potassium channel 221F Vertex 53
Symporter 346 Transmembrane proteins 211 Virchow, Rudolf 440
Synechococcus 434, 434F Transporters 211–214, 212F, 213F Virtual Liver Network 333, 362
Synergism 11, 15 Transposable elements 170–171 Virtual Physiological Rat Project 356, 362
Synthetic biology 399, 408–410, 410B Transposons 170–171 Virulence 267, 278, 418
case studies 411–418 Trehalose 306, 308, 308F Viruses 1
System components, model development 30–35 formation 310 Vitamin deficiency 415
System matrix 116 Trehalose 6-phosphate (T6P) 310 Voigt, Christopher 418
Systems analysis, collecting information in 5–6, 6F Trehalose cycle Voltage 351
Systems biology brief review 308–310, 308F, 309F Voltage-gated channels 346
application of 16–17 connectivity 308, 308F Volterra, Vito 296
communicating 13–15, 14F gene expression 308, 309F, 315–318, 316F, 316T,
drug development 378–393 317F
emerging applications 426–434 importance 310
emerging topics 425–443
history of 10–13, 12F, 13F
modeling analysis 310–318
analysis of heat stress 312–314, 313F, 313T, 314F
W
in mainstream 10–13, 12F, 13f design and diagnosis of metabolic pathway Watson, James 171
medicine 369–378 model 310–312, 312F Watson, Thomas John 425
modeling needs 435–439 gene expression 315–318, 316F, 316T, 317F Weaver, Warren 62
reductionism and 5–8, 6F, 7F, 11, 14–15, 14F glucose dynamics 314–315, 315F Western blot 191
theories 439–441 multiscale analysis 318–326 “Wild type” bacteria 16
Systems Biology for Energy and Environment 434 model design 320–324, 323T, 324F, 324T Winslow, Raimond 332
Systems biology mark-up language (SBML) 12, 435 trehalase puzzle 215F–327F, 324–326 Wobble 175, 202
Systems dynamics, analysis 119–122, 120F, 122F in vivo NMR profiles 318–320, 319F–321F
Systems ecology 60 regulation 308, 309F
Systole 334 response to glucose bolus 312, 312F
Szymanski, J. S. 275 temperature effect 313–314, 313F
Trial 29
X
tRNA (transfer RNA) 170, 182, 182F XML (customized mark-up languages) 12
T7 RNA polymerase (T7R) 262, 263 X-ray crystallography 206, 206F
T Troponin 338, 343
Tryptophan (Trp) 170, 203F, 204T
Tachyarrhythmia 355 Tsien, Roger Y. 222
Target identification, computational 381
Taylor, Brook 95, 96B
T tubules 356
Tumor cells 292
Y
Taylor approximation theorem 95, 96B Tumor development 431–432, 431F Yeast
canonical models 107, 108B, 109B epigenetics 178 alcohol fermentation 25, 400
linearized models 95–98, 96B, 97B, 97F proteins 208 artemisinic acid production 414–415, 415F
parameter sensitivity 118–120 Tumor growth 286, 292, 417 cell cycle 60, 260, 274
TCS (two-component signaling) systems 266–269, Tumor suppression protein 27 drug development 414–415, 415F
266F–270F, 268T Tumor suppressor genes 179F gene network 61F
Temperature effect, trehalose cycle 313–314, 313F Turns 221 gene regulation 185, 190
Template-and-anchor model 362 T wave 350, 350F heat stress response 303–330
Temporal scale 10 Two-color microarray 194 brief review 306–310, 307F–309F
Termite mounds 1, 2F Two-component signaling (TCS) systems 266–269, gene expression 308, 309F, 315–318, 316F, 316T,
Tertiary structure, protein 202, 204–205, 221, 221F 266F–270F, 268T 317F
Text mining 308, 381 Two-dimensional (2D) gel electrophoresis 219, 219F modeling analysis 310–318
TFs (transcription factors) 169, 188–190, 189F Tyrosine (Tyr) 203F, 204T multiscale analysis 318–326
heat stress response 307 origin of models 304–305
TGF-β (transforming growth factor β) 430 integrative analysis of genome, protein, and
T-helper cells 438F metabolite data 303–330
Theorems 439
Theory of biology 439–441
U intracellular metabolites 247
metabolic network analysis 251
Thermodynamic hypothesis 204 Über-operons 188, 401, 401F metabolic pathway 243B, 244B
Thomas oscillator 128, 128F Ubiquitin 205 signaling cascade 273
Threonine (Thr) 203F, 204T Ubiquitination 225, 245 sphingolipid 190
Thymine 171, 172F, 173F UCP2 gene 415 transcription regulatory network 56, 57F, 402
Timescales 10 UDPG (uridine diphosphate glucose) 310 trehalose cycle 310
separation of 358–359 Uncertainty, parameter estimation 151–152 two-hybrid assay 224
Time-series-based parameter estimation 136, 153 Uncompetitive inhibition 239 Young-Laplace equation 334
Tissue necrosis factor (TNF) 430 Undirected graphs 53, 53F, 55–56, 55F, 56F, 62
TNF (tissue necrosis factor) 430 Uniporter 346
Toggle switch model Uracil 171, 172F, 179
gene circuits 415–416, 416F, 417
SOS pathway in E. coli 105–107, 105F, 107F
Uric acid 391, 392
Uricase 392
Z
Toll-like receptors (TLRs) 215, 216F Uridine diphosphate glucose (UDPG) 310 ZWF1 gene 316

A First Course in Systems Biology (2018)

Uploaded by

Copyright:

Available Formats

A First Course in Systems Biology (2018)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A First Course in Systems Biology (2018)

Uploaded by

Copyright:

Available Formats

A F IR S T COUR SE IN

© 2018 by Garland Science, Taylor & Francis Group, LLC

This book contains information obtained from authentic

Library of Congress Cataloging-in-Publication Data

Printed in the United States of America

Visit our website at http://www.garlandscience.com

Instructor Resources Website

Accessible from www.garlandscience.com, the Instructor’s Resource Site

Guy Grant, University of Bedfordshire

chapter 1: biological systems 1 3.2 Small-World Networks 58

4.11 Analysis of Systems Dynamics 119 6.13 Transcription Factors 188

chapter 5: Parameter estimation 135

From Data to Systems Models 250 chapter 11: Integrative Analysis of

Figure 1.1 Biological systems abound at

Sph SphK Malate

GCR1 GPA1 AGB1

Figure 1.3 Analyzing a biological system

Figure 1.4 The size of a network or

(A) Figure 1.5 Biological phenomena are often

$100 investment $120 return $10,000 investment $12,000 return

1 tablespoon of fertilizer 50 blossoms 100 tablespoons of fertilizer dead roses!

REduCTIOnISM And SYSTEMS BIOLOGY

Figure 1.6 Collecting information is the

predictions regarding a gas in a container, even if trillions of molecules are involved.

H2O O2 ABIOTIC STRESS AND BIOTIC RESPONSES

photosynthesis WRKY MYC2 ERF

MVB DREB SA-responsive JA-responsive ET-responsive

nutrients, microbes CYCB

transcription factors hormones CYCD

kinases carbohydrates CDKA

EVEn SIMPLE SYSTEMS CAn COnFuSE uS

X = Input − aEX 0.5 ,

Figure 1.11 Simulation results

different responses. If the effect of Z on TF

short-day long-day Figure 1.12 Modern high-throughput

The third frontier is the co-evolution of mathematical, physical, and computa-

Figure 1.13 “Protein cages” are

turnip yellow papillomavirus human hepatitis murine

Figure 1.14 Advances in computer power,

COMMunICATInG SYSTEMS BIOLOGY

Figure 1.15 Information about isolated

THE TASK BEFORE uS

At the core of any computational analysis in systems biology is a mathematical

Modeling in systems biology is really a special way of looking at the world.

Figure 2.1 Modeling resembles a fine

GOALS, INPUTS, scope, goals, objective

TABLE 2.1: ISSUES TO PONDER DURING THE MODELING PROCESS

complicated and requires so many assumptions that it may no longer describe

GOALS, INPUTS, AND INITIAL EXPLORATION

2.1 Questions of Scale

level is biochemical or physiological, and the corresponding timescale is that of

2.2 Data Availability

provided us with rough, relative estimates of concentrations and metabolic fluxes

MODEL SELECTION AND DESIGN

Figure 2.4 Mathematical modeling can be

• We do not intend to account for the spatial distribution of cases of infection

2.3 Model Structure

BOX 2.1: STATIC AND DYNAMIC MODELS

Figure 2 To model the

Figure 3 Static model

X = Input − aEX 0.5 ,

The formulation in (2.7) is only an approximation of real-world, repeated I = −rQ I . (1)