REVIEWS
BIOLOGICAL ROBUSTNESS
Hiroaki Kitano
Abstract | Robustness is a ubiquitously observed property of biological systems. It is considered
to be a fundamental feature of complex evolvable systems. It is attained by several underlying
principles that are universal to both biological organisms and sophisticated engineering systems.
Robustness facilitates evolvability and robust traits are often selected by evolution. Such a
mutually beneficial process is made possible by specific architectural features observed in robust
systems. But there are trade-offs between robustness, fragility, performance and resource
demands, which explain system behaviour, including the patterns of failure. Insights into inherent
properties of robust systems will provide us with a better understanding of complex diseases and
a guiding principle for therapy design.
LYSIS
Part of a bacteriophage life
cycle in which its genome is
expressed to cause dissolution
of the bacterial host cell,
leading to manufacture of
more bacteriophage particles
and subsequent infection of
other cells.
LYSOGENY
Part of a bacteriophage life cycle,
during which its genetic material
is integrated into the genome of
its bacterial host, where it
remains in a latent state.
SEGMENTAL POLARITY
A pathway that regulates the
anteroposterior identity of
segments during insect
development.
Sony Computer Science
Laboratories, Inc., 3-14-13
Higashi-Gotanda,
Shinagawa, Tokyo 141-0022,
Japan, and The Systems
Biology Institute, Suite 6A,
M31, 6-31-15 Jingumae,
Shibuya, Tokyo 150-0001,
Japan.
e-mail:
[email protected]
doi:10.1038/nrg1471
826
The discovery of fundamental, systems-level principles
that underlie complex biological systems is a prime scientific goal in systems biology1,2. Robustness is a property that allows a system to maintain its functions despite
external and internal perturbations. It is one of the fundamental and ubiquitously observed systems-level phenomena that cannot be understood by looking at the
individual components. A system must be robust to
function in unpredictable environments using unreliable
components. Understanding the origin and principles of
robustness in biological systems will help us to put various biological phenomena into perspective; it will also
catalyse the formation of principles at the systems level.
In this article, I argue that robustness is a fundamental feature of evolvable complex systems. Complex biological systems must be robust against environmental
and genetic perturbations to be evolvable. Evolution
often selects traits that might enhance robustness of the
organism. Robustness is, therefore, ubiquitous in living
organisms that have evolved. However, systems that are
robust face fragility and performance setback as an
inherent trade-off. Identification of the basic architecture
for a robust system and the associated trade-offs is essential for understanding their faults and countermeasures
— diseases and therapies, respectively.
Robustness as an organizational principle
Robustness enables the system to maintain its functionalities against external and internal perturbations. This
| NOVEMBER 2004 | VOLUME 5
property has been widely observed across many species,
from the level of gene transcription to the level of systemic homeostasis. For example, fate decision of λ phage
— the result of which is the activation of either LYSIS or
LYSOGENY pathways — was once considered the result of
fine tuning of the binding affinity of promoters to corresponding regulatory factors. However, it has been shown
that it is the structure of the network, which involves both
positive and negative feedback, that is responsible for
making sustainable commitment, not the specific binding affinity — the fate-decision behaviour was shown to
be robust against point mutations in the promoter
region3. In addition, cooperative binding of repressors
that forms implicit local positive feedback also contributes to the stability of the switch4–6. Many examples of
robust properties can be observed in different biological
systems.
Escherichia coli is capable of chemotaxis over a
wide range of chemo-attractant concentrations
owing to integral intracellular feedback that ensures
perfect adaptation and that is independent of ligand
concentration7–9.
A biochemical network that is involved in the establishment of SEGMENTAL POLARITY in Drosophila melanogaster
has been shown to be robust against changes in initial
values and rate constants of molecular interactions,
enabling stable pattern formation10,11. Similar observations have also been made for MORPHOGEN-pattern
formations10,12,13.
www.nature.com/reviews/genetics
©2004 Nature Publishing Group
REVIEWS
Robust adaptation
(return to a periodic attractor)
Transition to a
new attractor
Stochastic process
influences the
trajectory
Robust adaptation
(return to a point attractor)
Unstable
Figure 1 | Robust reactions of the system: to stay or to change. The state of a system
can be shown as a point in the state space. In this case, the state space is simplified into two
dimensions. Perturbations forcefully move the point representing the system’s state. The state of
the system might return to its original attractor by adapting to perturbations, often using a negative
feedback loop. Bacterial chemotaxis is an example. There are basins of attractions in the state
space within which the state of the system moves back to that attractor. If the boundary is
exceeded, the system might move into an unstable region or move to other attractors. Positive
feedback can either move the system’s state away from the current attractor, or push the system
towards a new state. The cell cycle involves a combination of positive and negative feedbacks that
facilitate transition between two attractors (G1 and S/G2/M) creating a bistable system. Often,
stochastic processes affect transition between attractors, as seen in λ-phage fate decision, but
maintenance of a new state has to be robust against minor perturbations.
MORPHOGEN
A diffusible signal that acts at a
distance to regulate pattern
formation in a dose-dependent
manner.
ATTRACTOR
A point or an orbit in the phase
space where different states of
the system asymptotically
converge.
PHASE SPACE
A multi-dimentional space that
represents the dynamics of a
system. For a system with
N-variables, a phase space is a
2N dimensional space
composed of N-variables and
their time derivatives.
Diseases such as cancer and diabetes are manifestations of co-opted robustness, in which mechanisms that
normally protect our bodies are effectively taken-over to
sustain and promote the epidemic states14–16. As more
studies are done, it is becoming important to provide an
integrated perspective on the robustness of biological
systems.
The robustness of a system can manifest itself in one
of two ways: the system returns to its current ATTRACTOR
or moves to a new attractor that maintains the system’s
functions (FIG. 1). A return to the current attractor is
often called ‘robust adaptation’. The attractor can be
either static (a point attractor; a fixed point in the PHASE
SPACE that the trajectory of the system state approaches
asymptotically) or oscillatory (a periodic attractor; a
cyclic orbit in the phase space that the trajectory of the
system state approaches asymptotically).
A transition to a new attractor has to be made
robustly in response to stimuli so that the system
behaves consistently against perturbations. As seen in λphage fate decision, the stochastic process often influences the trajectory of transition and the attractor on
which the system eventually converges.
NATURE REVIEWS | GENETICS
Robustness is often misunderstood to mean staying unchanged regardless of stimuli or mutations, so
that the structure and components of the system, and
therefore the mode of operation, is unaffected. In fact,
robustness is the maintenance of specific functionalities of the system against perturbations, and it often
requires the system to change its mode of operation in
a flexible way. In other words, robustness allows
changes in the structure and components of the system owing to perturbations, but specific functions are
maintained.
In the following sections, I outline the mechanisms
that ensure the robustness of a system: system control,
alternative (or fail-safe) mechanisms, modularity and
decoupling.
System control. System control consists of negative
and positive feedback to attain a robust dynamic
response observed in a wide range of regulatory networks, including the cell cycle, the circadian clock
and chemotaxis7,17,18. Negative feedback is the principal mode of control that enables robust response
(or robust adaptation) to perturbations. Bacterial
chemotaxis is one of the most studied examples
of robust adaptation that uses negative feedback —
INTEGRAL FEEDBACK in particular — to attain the perfect
adaptation that allows chemotaxis to occur in response
to a wide range of stimuli7–9. Integral feedback, a particular control strategy, is essential to maintain robust
adaptation in both E. coli and Bacillus subtilis, despite
the fact that the network topologies are not the
same19.
Positive feedback contributes to robustness by
amplifying the stimuli, often producing bistability,
so that the activation level of a downstream pathway
can be clearly distinguished from non-stimulated
states, and so that these states can be maintained.
In D. melanogaster segment-polarity formation —
repetitive stripes of differential gene expression — is
observed along the antero-posterior axis of the developing embryo. The first stripe has to express wingless
(wg), the second stripe has to express engrained (en),
but the third stripe expresses neither. von Dassow and
colleagues10 created a computational model of this
system, initially without positive autoregulatory feedback on wg and en, but the model failed to reproduce
experimentally observed patterns. However, with two
positive feedbacks on wg and en activations, robust
pattern formation was reproduced10. Recently, Ingolia11
analysed this model and showed that the bistability
caused by positive feedback loops is responsible for
robust pattern formation11.
Positive feedback is also used in signal transduction
and the cell cycle to form switch-like behaviour of the
system by amplification of stimuli, and for fate decision (as seen in the λ phage), so that it initiates a transition and a new state of the system is made that is
more robust against noise and fluctuations of stimuli3,20–27. Many biological subsystems use the combination of these system controls to perform their functions
and to maintain robustness27,28.
VOLUME 5 | NOVEMBER 2004 | 8 2 7
©2004 Nature Publishing Group
REVIEWS
INTEGRAL FEEDBACK
A method of feedback control in
which control is proportional to
the integral of the systems’
output.
DIAUXIC SHIFT
The process of switching from
anaerobic to aerobic respiration.
CANALIZATION
The buffering or stabilization of
developmental pathways against
mutational or environmental
perturbations, by several genetic
factors.
828
An alternative, or fail-safe, mechanism (redundancy
and diversity). Robustness can be enhanced if there are
multiple means to achieve a specific function, because
failure of one of them can be rescued by others. Here,
I call this mechanism ‘alternative’, or ‘fail-safe’. This
concept encompasses redundancy, overlapping function and diversity, as the differing degrees of similarity between the various alternative means that are
available.
Redundancy generally refers to a situation in which
several identical, or similar, components (or modules)
can replace each other when another component fails.
Diversity, or heterogeneity, represents the other extreme,
whereby a specific function can be attained by different
means available in a population of heterogeneous components. Some of these phenomena are well known as
phenotypic plasticity29–31.
In some tissues, cells are surrounded by similar
neighbours, so that damaged cells are quickly replaced
by other cells. However, having multiple identical
components as alternatives is rare. An alternative, or
fail-safe, mechanism is usually attained by having
multiple heterogeneous components and modules
with overlapping functions. For example, Clb5 and
Clb6 of budding yeast are two relatively homologous
genes encoding B-type cyclins. They share 49.7%
identical residues and both gene products are involved
in the entry into the cell cycle32. Deletion of Clb6 has
little or no effect on the progression of the cell cycle,
and deletion of Clb5 caused a prolonged S-phase.
Deletion of both genes impedes timely initiation of
DNA replication. Recent findings strongly demonstrate that gene duplication, particularly wholegenome duplication followed by extensive gene loss
and specialization, is one of the crucial mechanisms of
evolutionary innovation33,34, providing support for the
long-standing hypothesis proposed by Susumu Ohno35.
If the function of duplicated gene pairs overlaps to
some extent, the duplication acts as an evolutionary
capacitor36,37. There is at least one indication from computational studies that under certain conditions these
pairs could be evolutionary stable38.
There are also numerous examples of alternative
mechanisms at the network level. Text-book examples
include glycolysis and oxidative phosphorylation39.
Although both processes produce ATP, oxidative
phosphorylation requires a constant supply of oxygen, whereas glycolysis can be either aerobic or anaerobic (although the latter is less efficient). DIAUXIC SHIFT
in yeast causes a drastic change in metabolic pathways, depending on whether glucose or ethanol is
available for energy metabolism40. The flux balance of
metabolic pathways is readjusted according to the
available resources in the environment, but either
pathway can, ultimately, produce essential materials
for survival and growth41,42. These are examples of
phenotypic plasticity that are often considered to be
the opposite of robustness. However, I argue that it is
more consistent to view phenotypic plasticity as a part
of robustness, because this plasticity enables organisms
to robustly adapt to a changing environment.
| NOVEMBER 2004 | VOLUME 5
It is interesting to note that although redundancy
through duplicated genes is frequently observed, there is
no reported case of duplicated circuits, despite the fact
that there are many circuits with similar topologies.
Investigation of networks that share similar topology, as
far as the degree of homologous genes involved in each
network is concerned, revealed that although genes
might be duplicated, the network as a whole is not
duplicated43. This indicates that similar network topologies represented in different contexts within and
between species are the result of convergent evolution,
rather than duplications43,44. This is consistent with the
fact that circuit-level alternative mechanisms are
attained by the differing implementation of overlapping
functions.
It is important to understand that alternative mechanisms are coupled with system control to ensure robustness. First, the existence of alternative mechanisms at
the system-component level allows regulatory feedback
to remain intact despite mutations. Second, switching
between alternative mechanisms has to be orchestrated
by specific controls so that the system behaviour is
properly maintained.
Modularity. Modularity is an effective mechanism for
containing perturbations and damage locally to minimize the effects on the whole system. Modules are
widely observed in various organisms, functioning as
possible biological design principles30,45,46 and as essential elements in engineering and industry47,48. Despite
intuitive consensus, the concept is still ambiguous and
therefore can be difficult to define46.
A cell is an obvious example of a module that constitutes multi-cellular systems; it interacts with the environment and other cells. Modules are often hierarchically
organized; a cell itself is composed of organelles, and, at
the same time, it is also a part of larger modules such as
tissues and organs.
Aside from physical modules such as a cell, there are
functional, spatial and temporal modules that can be
recognized as subsystems of metabolic networks, signal
transduction and developmental regulatory networks. A
bacterial flagellum and its control module, for example,
represents both a physical and logical module that is
robust and versatile49. Logical modules, as seen in segmental polarity networks10,11 and elsewhere50, are often
less obvious than physically partitioned and engineering
modules.
Decoupling. Decoupling isolates low-level variation from
high-level functionalities. For example, Hsp90 not only
fixes proteins that are mis-folded as a result of environmental stresses, but also decouples genetic variations
from the phenotype using the same mechanism, therefore providing a genetic buffer against mutations51–53.
These genetic buffers decouple the genotype from the
phenotype, and they provide robustness to cope with
mutation while maintaining a degree of genetic diversity.
These buffers have been shown to underpin the robustness of developmental processes (as in the case of
Waddington’s CANALIZATION54). Importantly, the mutations
www.nature.com/reviews/genetics
©2004 Nature Publishing Group
REVIEWS
Perturbations
Instructions
Flight control
surfaces and
propulsion
system
AFCS
Flight path
Feedback
Flight
control
computer
number 1
Flight
control
computer
number 2
Flight
control
computer
number 3
Figure 2 | Explaining robustness — the aeroplane example. The concept of robustness
is best described using the example of modern aeroplanes. Many commercial passenger
aeroplanes have an automatic flight control system (AFCS) that maintains a flight path
(direction, altitude and velocity of flight) against perturbations in atmospheric conditions. This
can be accomplished by a feedback control in which deviations from the defined flight path
are automatically corrected. AFCS is the crucial component that allows the robust
maintenance of the flight path by controlling the aeroplane’s flight-control surfaces (rudder,
elevator, flaps, aileron, etc) and the propulsion system (engines). AFCS is generally
composed of three modules with the same functions, thereby creating redundancy, although
each is designed differently (heterogeneity) to avoid a common mode failure. Three
computers are made that are modular, so that failure in one module does not affect the
functions of other parts of the system. This type of mechanism is implemented using digital
technologies that decouple low-level voltages from digital signal (ON/OFF of pulses), thereby
preventing noise from influencing its functions. Although this is a simplified explanation of the
actual system, the concept applies to details of the basic system as much as it does to the
more complex systems. Although there are differences between man-made systems and
biological systems , the similarities are overwhelming. Fundamentally, robustness is the basic
organizational principle of evolving dynamic systems, be it through evolution, competition, a
market niche or society.
NEUTRAL THEORY OF
EVOLUTION
A theory proposed by Motoo
Kimura which states that most
variations at the molecular level
are neutral to selection.
WEAK LINKAGE
A property of a process that
refers to the coupling of
processes; in this case, a process
depends minimally on other
components or processes;
example include neural relays or
signal transduction pathways, in
which individual components
often have a switch-like capacity
to exist in active or inactive
states.
that are masked by genetic buffering are selectively neutral (one of the premises of Kimura’s NEUTRAL THEORY OF
55,56
EVOLUTION
), providing a source of material for the evolution of the system during extreme perturbations.
Feedback controls sometimes compensate for
changes in rate constants of interactions within the network and changes in the initial state of the network, as is
the case for bacterial chemotaxis7–9 and D. melanogaster
segmentation10, or they might even mitigate the impact
of loss-of-function mutation. A computational study of
the cell cycle demonstrated that removing some genes
does not necessarily block the cell cycle, it might only
make it more fragile against perturbations17.
Bistability created through positive feedback sometimes results in the decoupling of fluctuations at a molecular level; for example, the decoupling of a number of
molecules that are involved in reactions from the committed state of the system. Therefore, dynamic networks
NATURE REVIEWS | GENETICS
often decouple genetic and environmental perturbations57. There is even a hypothesis, albeit computational,
that claims that this buffering is an intrinsic property of
complex networks58–60.
Another example of decoupling might take place
between information encoding and conversion of stimulus dosage into pulses of protein activations. When the
DNA is damaged, p53–MDM2-feedback loops generate
oscillatory behaviour; this behaviour was recently found
to be a potential converter of graded stimuli (degree of
DNA damage) to digital pulses (with a peak of p53 activation), so that it is only the number of pulses that matter after the conversion61,62, not the subtle changes in
concentration levels.
The mechanisms that underlie robustness can be
understood using an example of a sophisticated engineering object, such as an aeroplane (FIG. 2). There are
similar mechanisms in various other sophisticated engineering systems: they can be built on less than perfect
components, but have to cope with unpredictable environmental pressures, thereby indicating the universal
nature of robustness. It is interesting to consider
whether there is a fundamental system architecture for
successful robust systems, and what limitations and
risks are associated with these systems.
The origin of robustness
It is now increasingly recognized that robustness is
ubiquitous. So, what are the principles and mechanisms that lead to the emergence of robustness in biological systems? My theory is that robustness is an
inherent property of evolving, complex dynamic systems — various mechanisms incurring robustness of
organisms actually facilitate evolution, and evolution
favours robust traits. Therefore, requirements for
robustness and evolvability are similar. This implies
that there are architectural requirements for complex
systems to be evolvable, which essentially requires the
system to be robust against environmental and genetic
perturbations.
Evolvability requires flexibility in generating diverse
phenotypes by means of producing non-lethal mutations45,63,64. Kirschner and Gerhart define evolvability,
or evolutionary adaptability, as a capacity to generate
heritable and selectable phenotypic variations that
consists of features that “…reduce the potential lethality of mutations and the number of mutations needed
to produce phenotypically novel traits”63. They argue
that flexible versatile proteins, WEAK LINKAGE, EXPLORATORY
SYSTEMS and compartmentalization are central features
that foster evolvability63. They also argue that the
emerging architecture is composed of highly conserved core processes that are co-selected with various
other processes, some of which bring about phenotypically novel traits, which is consistent with the BOW-TIE
architecture.
These features can be translated into architectural
requirements of the system that are consistent with
robustness. First, mechanisms that preserve the components and interactions against mutation must be
capable of generating genetic variation. Second, there
VOLUME 5 | NOVEMBER 2004 | 8 2 9
©2004 Nature Publishing Group
REVIEWS
must be modules that robustly maintain their functions against external perturbations and mutations.
Third, there must be highly conserved core processes,
which are also modular, that have fundamental functions, such as metabolism, the cell cycle and transcriptional machinery, where various modules can be
interfaced to create diverse phenotypes. The overall
architecture that meets these requirements is probably
a modularized nested bow-tie, or hour-glass, structure
where various input and output modules are connected to a conserved core (FIG. 3); incidentally, the
same architecture underlies the World Wide Web65.
This bow-tie structure occurs in various aspects of biological systems, from global structure to specific mechanisms such as transcription and translation
processes66. The architectural features of a modularized bow-tie structure are optimal for enhancing the
robustness of the various aspects of a system.
Buffering
The first step towards robustness is to protect components and interactions from perturbations. As well as
correcting misfolded proteins, chaperones also impose a
particular conformation on some proteins that are genetically varied; masked genetic variation is exposed when
these mechanisms are impeded53. Networks also contribute to this type of buffering57. This is particularly the
case when a network is robust against external perturbations, because regulatory feedbacks can also provide
robustness against perturbations of various internal
parameters. The genetic buffering mechanism — also
referred to as the evolutionary capacitor — attained
either by chaperons or networks, is one of the fundamental mechanisms that provides robustness and evolvability.
It has been increasingly noted that robustness against
mutation might evolve as a side effect of robustness
against environmental perturbations and EMERGENT
67
PROPERTIES of complex networks .
Robust modules
In principle, a modular system allows the generation of
diverse phenotypes by the rearrangement of its intermodule connections and relatively independent evolution of each module through mutations. If the system
is well integrated without modular structure, change
in any part of the system might have a significant
impact on other parts of the system, and slight changes
in stimuli or noise might result in an unexpected and
undesirable outcome. The system will be intractable,
and it is difficult for these systems to generate new
phenotypes without lethal effects. Notably, modularization mitigates this problem by allowing each module to function autonomously. However, simply having
modules is not enough to ensure evolvability. It must
also be robust against various perturbations. This feature is essential because modules need to be able to
cope with changes in stimuli from adjacent modules
that might evolve independently or that function in a
Output process
(responses and
products)
Input process
(signalling and
nutrients)
Output process
(responses and
products)
Highly conserved
core processes
Output process
(responses and
products)
Input process
(signalling and
nutrients)
Diverse reactions
Diverse stimuli and
environmental conditions
System control
Input process
(signalling and
nutrients)
EXPLORATORY SYSTEMS
Systems that are based on
epigenetic variations and
selection; such as angiogenesis
and nerve outgrowth.
BOW-TIE
A structure that has various
inputs (fan-in) and outputs
(fan-out) that are connected by a
knot, resembling a bow-tie.
EMERGENT PROPERTY
A feature that is characteristic of
system-level dynamics that
cannot be attributed to any of its
components. The existence of an
emergent property indicates that
the whole is more than just the
sum of the parts.
830
Input process
(signalling and
nutrients)
Conserved versatile
interfaces (weak linkage)
Output process
(responses and
products)
Figure 3 | The architectural framework of robust evolvable systems. The bow-tie (or hour-glass) structure has many inputs
and outputs that are connected through a conserved core and versatile weak linkage with the extensive system control governing
the system’s dynamics. Core processes and versatile interfaces overlap or merge in some cases. This bow-tie structure appears
at various levels in the system, such as metabolism, signal transduction, transcription and translation66. In signal transduction,
diverse stimuli are initially received by receptors, different isoforms of G-proteins are activated, but converge mainly to second
messengers that have a limited variety and cause weak linkage. Then, modulations in second messengers influence core
processes to trigger expression of different genes and, ultimately, different reactions. However, this process is not a simple flow of
information, as extensive local and global feedback regulations are imposed at every step. During metabolism, diverse nutrients
are processed into precursors that core metabolic pathways then covert into basic cellular ‘currencies’ such as ATP and NADH, as
well as activating biosynthesic pathways to produce amino acids, nucleotides, sugar, and so on. Transcription and translation also
represent structures where common machineries are used to decode a wide range of genetic information and produce diverse
proteins, but versatile mechanisms themselves make up the conserved core. Various processes are interfaced with core
processes through versatile interfaces, so that novel processes can be added and removed easily without seriously affecting other
parts of the system.
| NOVEMBER 2004 | VOLUME 5
www.nature.com/reviews/genetics
©2004 Nature Publishing Group
REVIEWS
different context. This robustness is attained through
system control and alternative mechanisms that are
inherent to each module.
The contribution of system control to module
robustness is particularly important in development,
whereby new morphologies can be explored during
evolution. A similar type of robustness to that of the
segment-polarity network in D. melanogaster was also
identified through the establishment of the BMP (bone
morphogenetic protein) morphogen gradient68. The
organizing centre is a good example of a robust buffer
against variation in development69,70. There is a subset of
developmental processes that is tolerant against variations in initial values and that robustly forms patterns
that serve as a basis for further elaborations in developmental patterning and that are amenable to reuse in a
different context71,72. Hox genes exemplify the power of
modularity where changes in the Hox cluster alone
affects basic body plan — a network of genes downstream of the Hox gene direct development of a given
part of the body autonomously from other parts of the
body73–78.
Signal-transduction pathways have an important
role in generating phenotypic diversity during evolution. Hedgehog (Hh), wingless-related (Wnt), TGF-β,
receptor tyrosine kinase (RTK), Notch, JAK-STAT
(signal tranducers and activators of transcription)
and nuclear hormone pathways are signal-transduction pathways that are widely used in various aspects
of development. Co-option of existing signal-transduction pathways to new processes is considered one
of the crucial features in evolution. For example, the
Hh-signalling pathway that is used in wing-pattern
formation is co-opted in butterfly-eyespot pattern formation71. Several signal-transduction cascades combine negative and positive feedback loops and are
robust against perturbations so that normal cellular
physiology and developmental processes can be maintained25,26,28,79,80. This intrinsic robustness of the pathway enables co-option, so that new morphologies can
be generated.
GIANT STRONG COMPONENT
SUB-NETWORK
A sub-network in which there
are a large number of
components that have extensive
internal connections.
The origin of modularity. The origin of modularity is
still controversial. It is certainly an evolved property, not
necessarily a selectable trait by itself, and it enhances
flexible generation of various phenotypes during development81. At the same time, modular structures and
modular regulatory networks within a single cell
(including bacteria) that demonstrate enormous diversity and evolution indicate that flexibility of development is not the only reason for modularity. There are
two possible reasons for why modularity might have
emerged.
First, emergence of modularity of gene regulation
might be required to handle diverse and complex stimuli and responses49. It is essential that signalling networks and reaction-related networks are modularized to
some extent, to cope with various external perturbations without losing specificity to stimuli versus
responses, and to prevent the effects of environmental
perturbations from spreading system-wide.
NATURE REVIEWS | GENETICS
Second, if a modularized phenotype has selective
advantage owing to robustness against environmental
perturbations, modular developmental processes
could be co-selected because they are better at generating modular phenotypes. This hypothesis assumes that
the modular phenotype has a selective advantage and the
modular developmental processes are better at generating the modular phenotype; both assumptions have still
to be tested. If these two reasons hold, modularity might
have originated to ensure robustness against environmental perturbations, but congruent with flexibility of
development.
The architectural framework
There are two architectural features that facilitate evolvability and robustness63,64: highly conserved core
processes that are interfaced with diverse inputs (signalling and nutrients) and outputs (reactions and
products); and versatile mechanisms that underlie
essential processes of the system, so that any new
processes that properly interface with these mechanisms can use them — this is also known as ‘weak linkage’63. These features represent the bow-tie architecture
at different levels, ranging from global topology to specific processes. Below, I argue that the bow-tie structure
that is actually observed in biological systems facilitates
robustness and evolvability.
Bow-tie structure in biological systems. Genome-wide
analysis has revealed intriguing characteristics of the
biological networks that support bow-tie structure.
It has been proposed that metabolic networks and
protein-interaction networks form a scale-free network
(a property of scale-free networks that predicts that proteins prefer to form links with other proteins that
already have the highest number of links)82,83. Scale-free
networks tolerate random removal of nodes, but not
systematic removal of nodes with high connectivity84.
Ma and Zeng85 considered the directionality of reactions
in the analysis of metabolic networks of 65 fully
sequenced genomes. They found that the these networks do not exhibit scale-free structure, but rather a
‘bow-tie’ structure, in which a large, highly connected
core cluster is interfaced with less connected in- and
out-clusters. The core of the network is a GIANT STRONG
COMPONENT (GSC) SUB-NETWORK, the components of which
are tightly connected85. Further comparisons of the
metabolic networks between Streptococcus pneumoniae
and Pyrococcus furiosus showed conservation of essential
metabolic pathways, such as the TCA (tricarboxylic
acid) cycle, pentosephosphate pathway and glycolysis
pathway, within a GSC sub-network; the conserved core
pathway is robust against perturbations85.
Another indication that there is such a division
into conserved core processes and those that provide
novel functions comes from comparative studies that
use functional annotation of 150 fully sequenced
genomes86,87. The Gene Ontology’s biological process
hierarchy88 was used to assign functional categories to
each gene, and the proportion of genes in each functional category, for all 150 species, was counted. The
VOLUME 5 | NOVEMBER 2004 | 8 3 1
©2004 Nature Publishing Group
REVIEWS
results were presented in terms of a scaling exponent —
1.0 is when the total number of genes in a category is
doubled, and less than 1.0 is when the number of genes
in a category that have been increased is less than the
increase in the total number of genes. A cross-species
study carried out by van Nimwegen on 65 bacterial
genomes indicates that basic biological processes have
a scaling exponent of less than 1.0 (these include the
cell cycle, DNA repair and replication, and protein
biosynthesis), but processes that might generate evolutionary novelty tend to have a scaling exponent over
1.0 (these processes include transcriptional regulation, signal transduction, ion transport, two-component systems and cell communication)86. This tendency
can also be seen in an extended study of 15 archaeal,
116 bacterial and 10 eukaryotic genomes by van
Nimwegen87. These results show that there are highly
conserved core processes, such as biosynthesis and DNA
replication, and processes that have been added with
increase in genome size, which might be responsible for
the generation of various cell-types and morphological
features, such as signal transduction, transcriptional
regulation and intercellular communications. This
implies that new pathways might be constantly added to
the wings of the bow-tie structure as the genome size
increases.
‘Weak linkage’64 enables the addition of new processes to the existing core process using common versatile mechanisms that operate on diverse inputs and
outputs, such as ion channels, G-proteins and transcription machinery. If transcription machinery is different
for each gene, addition of new genes and new transcriptional regulations requires the invention of customized
transcription machinery that makes evolution almost
impossible. GTP-binding proteins and the downstream
cyclic AMP and calcium dynamics, as well as ion channels also represent systems with common underlying
mechanisms that allow new repertoires to be added64,89.
Inputs from various receptor channels, such as RTK and
GPCR (G-protein-coupled receptor), converge mainly on
second messengers, such as cAMP and calcium ions,
which mediates various cellular responses, such as cell
movement, cell growth, metabolism and so on89. Cdc42, a
member of the Rho-family of GTPases, is another common regulator on which various RTK and GPCR pathways converge and mediate different cellular responses90.
Most signal-transduction cascades converge to modulate
limited numbers of second messengers, but signalling
pathways are often diverse and include cross-talk91,92.
Bow-tie is robust. Whether bow-tie structure provides
robustness against external perturbations depends on
the robustness of the conserved core and the global regulations imposed. Ma and Zeng argue that the GSC, a
conserved core of the metabolic network, is robust
against mutations because there are multiple routes
between any pair of nodes within the GSC85.
Various stimuli activate signal-transduction pathways that converge to modulate second messengers,
which in turn activate various cellular responses. Here,
second messengers are the conserved core of the bow-tie
832
| NOVEMBER 2004 | VOLUME 5
architecture and have to be maintained robustly. For
example, cAMP is produced from adenylyl cyclases by
ATP. Various adenylyl cyclase isoforms are involved in
multiple pathways, possibly creating alternative pathways.
ATP is supplied by a robust metabolic core mechanism
that is also a bow-tie structure. Here, the bow-tie structure at the metabolic level supports the bow-tie structure
at the level of signal transduction.
In addition to the robustness of the conserved core,
the bow-tie architecture might provide an advantage in
generating coordinated response to various stimuli. So,
a bow-tie architecture improves robustness against
external perturbation by having many inputs connecting to the robust core where numerous reactions are
mediated. Direct association between stimuli and reactions, without the use of the robust core, requires
extensive individual controls to achieve a coordinated
response, and disruption of any such regulation could
seriously undermine system behaviour. Unless each
stimulus reaction can be regarded as independent,
making coordination unnecessary, control through the
common robust core might provide better system
robustness.
However, do these mechanisms allow for the accommodation of phenotypic diversity? In fact, versatile
mechanisms are essential to accommodate various possible input stimuli and reactions in a consistent manner.
For example, the addition of new signal transductions
only needs to be interfaced with existing machinery
without inventing an entire cascade. Also, recent findings of differential tissue and cell-type specific expression of GPCRs in the human and mouse93 provide
explanations on how various cells might use the conserved-core and weak-linage mechanisms. Expressionpattern analyses using real time PCR on 100 randomly
selected endoGPCRs (GPCRs for ligands of endogenous
origins) revealed that each endoGPCR is expressed in
numbers of different tissues, with each tissue having a
unique combination of endoGPCRs93. This indicates
the possibility that a relatively small set of second messengers are used in different contexts to allow differentiation into various cell types and a range of cellular
responses. The same argument applies to a limited set of
versatile genetic networks, also referred to as a toolkit72,
because they can be used in different contexts.
Robustness trade-offs
I have argued that robustness is a fundamental feature that enables complex systems to evolve, and that
evolution enhances robustness of organisms. One
possibility following on from this, is to increase the
complexity of organisms through successive addition
of regulatory systems, such as diverse regulation, signal-transduction pathways, RNA regulation94–96 and
histone modifications97, to enhance robustness against
specific environmental perturbations and to allow
exploration into unoccupied niches98. However, the
introduction of various control feedback loops generates trade-offs by causing instability when unexpected
perturbations are encountered, leading to catastrophic
failure. Carlson and Doyle tried to generalize these
www.nature.com/reviews/genetics
©2004 Nature Publishing Group
REVIEWS
HIGHLY OPTIMIZED TOLERANCE
THEORY
A theory about the dynamic
properties of systems that are
designed, or evolved, to be
optimal (either towards a global
optimum or sub-optimum). The
theory predicts whether systems
that are robust against certain
perturbations are fragile against
unexpected perturbations.
SELF-ORGANIZED CRITICALITY
A phenomenon whereby certain
systems reach a crucial state
through their intrinsic
dynamics, independently of the
value of any control parameters.
issues in their HIGHLY OPTIMIZED TOLERANCE (HOT) THEORY by
arguing that systems that have evolved to be robust
against general perturbations are extremely fragile
against certain types of rare perturbations99,100. In
dynamic systems, these evolutionary optimizations are
achieved by successively adding feedback controls to
the system. It is important to recognize that although
the HOT theory describes behaviour of complex systems designed or evolved towards optimality against
perturbations, other self-organization models such as a
scale-free network83,101 by preferential attachment and
102
SELF-ORGANIZED CRITICALITY
describe systems with stochastic additions of complexity without design or evolution involved. The nature of robustness and fragility
that is predicted by the HOT model is generally consistent with observed properties of designed complex systems and is notably different from other models99,100.
Csete and Doyle further argued that trade-offs
between robustness and fragility indicate that robustness is a conserved quantity, a concept that also applies
to biological systems103.
Trade-offs between robustness and fragility can be
intuitively understood by using the aeroplane example
again. The Wright brothers’ aeroplane is not robust
against atmospheric perturbations, unlike modern
commercial aeroplanes. However, modern aeroplanes
are extremely fragile against unusual perturbations,
such as total power failure — because their flight-control
system is totally dependent on electricity. The Wright
brothers’ aeroplane, on the other hand, is not affected by
this type of failure as it does not use electrical controls in
the first place. Although they might seem simple, it is
important to appreciate these trade-offs because diseases
are often manifestations of fragility.
Fragility is not the only cost of improved robustness.
In electronic-circuit design, the use of negative-feedback
control achieves an improved fidelity or amplification
for a certain range of inputs by reducing overall gain of
the amplifier. So, the use of negative-feedback control
achieves robustness within a certain range of inputs at
the cost of some aspects of performance and with the
creation of fragility elsewhere. The trade-offs between
robustness, fragility and performance can be observed
in biological systems at different levels. Bacteria, for
example, should be able to swim faster without negative
feedback, but this would sacrifice their precision in following a chemical gradient: the use of negative feedback improves the bacteria’s ability to follow a chemical
gradient, at the cost of reduced swim speed.
Alternative mechanisms and modularity also
enhance robustness, but at the cost of increased resource
demands. For example, the probability that a function
with a single component will fail is p. The probability of
the function with two components that fail to back up
each other will be reduced to (1 – p)2. This reduction is
accompanied by doubling of the resource requirement.
This type of trade-off is effectively mitigated in biological systems because components are not identical
copies, but tend to have overlapping functions. Having
identical copies as alternatives is only efficient when failure rate or turn-over rate, is expected to be high.
NATURE REVIEWS | GENETICS
Modularity also causes trade-off between robustness,
flexibility and resource demands. Merging modules to
share common circuits and components reduces
resource needs. However, preventing the spread of perturbation and flexibility of rearrangement in a robust
way is seriously compromised.
Although there are several trade-offs in robust complex systems, trade-offs involving system controls are
the most important ones. System controls define
dynamic properties of the biological systems that illustrate how complex biological systems behave when perturbed and coordinate how alternative components and
modules are re-routed to ensure robustness of the
whole system. Owing to the intrinsic trade-offs that
have been discussed, it is not possible to simply increase
general robustness of the system without a sacrifice in
performance and increased resource demands.
A system-level view of disease and therapy
Properties of robust evolvable systems have direct consequences on our understanding of diseases and therapy design. First, robust systems, whether biological or
engineered, are most vulnerable when the system’s
fragility is exposed. For example, Diabetes mellitus can
be thought of as an exposed fragility of the system that
has acquired robustness against near-starvation, a high
energy-demand lifestyle and high risk of infection, but
it is unusually perturbed by over-nutrition and a low
energy-demand lifestyle16. Second, the system is relatively tolerant of the removal of some components or
cells, because of available alternative mechanisms and
the robustness of the bow-tie architecture. However,
the system is vulnerable when components behave
inappropriately but are not being removed. Third, the
epidemic state might exhibit robustness against natural and therapeutic countermeasures if intrinsic
mechanisms for robustness in our bodies are co-opted.
The worst-case scenario is when components of the
system behave inappropriately without being removed,
and malfunctioning components and their behaviour
are robustly maintained against countermeasures: the
mechanisms that support robustness of the host are
also used to protect the failed components. This is
important because the intrinsic dynamics of diseases
and appropriate countermeasures are different depending on which scenario the system failure follows. For
example, cholera toxin interacts with the Gsα subunit to
trigger the symptoms of the disease105,106. It can be easily
removed by antibiotics, because the intrinsic robustness
of the host has not been hijacked by the pathogen and is
not involved in the infection.
By contrast, cancer and HIV infection represent a
worse scenario — they are maintained and even promoted through the intrinsic robustness mechanisms of
host system. Counter measures for these diseases could
include: actively perturbing specific interactions or
components to maintain or reduce robustness, finding a
point of fragility (an ‘Achilles’ heel’) that is inherent in
robust systems and re-establishing control of the epidemic state by introducing a counter-acting decoy (a
‘Trojan horse’) or a new regulatory feedback.
VOLUME 5 | NOVEMBER 2004 | 8 3 3
©2004 Nature Publishing Group
REVIEWS
CONTROL THEORY
The theory about the design of
optimal control methods for
engineered objects. It is one of
the most successful fields in
which mathematical principles
are directly applied to practical
products, such as aeroplanes,
hard disks, automobiles, robotics
and chemical plants, and enables
them to function properly.
Usually, the theory is concerned
with how feedback control can
be used in various cases to attain
optimal design behaviour.
834
Cancer is a highly robust disease in which the
tumour proliferates and metastasizes, in some cases
despite much therapeutic effort. Although anti-cancer
drugs might temporarily reduce tumour mass, it
relapses in many cases and cure is still rare. The difficulty of treating tumours is attributed to acquired
robustness, partly owing to the co-option of intrinsic
mechanisms for robustness14,15. The high level of genetic
heterogeneity in the tumour forms a fail-safe through
genetic diversity and multiple feedback loops in both the
cellular and the tumour-host environment and accounts
for the robustness of cancer. The genetic heterogeneity
within a tumour cluster, caused by chromosome instability, generates a high level of genetic heterogeneity in survival and proliferation. This genetic heterogeneity allows
some malignant cells to tolerate therapy and re-form a
tumour.
Mechanisms that maintain normal functions of
the body also function to enhance tumour robustness
against therapy. For example, drug resistance is
caused by upregulation of MDR1 and other genes,
the products of which pump out toxic chemicals
from the cell; so a function that protects us under
normal conditions, is exploited by the tumour to protect the malignant cells. Low oxygen supply (hypoxia)
during the tumour-cluster development is countered
by metabolic shift from oxygen-dependent TCA cycle
to glycolysis, as well as activating a feedback loop by
upregulating HIF1, which upregulates VEGF (vascular endothelial growth factor)to promote angiogenesis, and MMP (matrix metalloproteases), uPAR
(urokinase-type plasminogen activator) and CRCX4
(a chemokine receptor) to promote tumour-cell
metastasis107; a series of responses that protects the
body against the effects of hypoxia.
From the robustness perspective, possible clinical
strategies that could be used against cancer are the control of the robustness of the tumour and finding out the
point of fragility that is inherent in the robust system. To
control robustness, therapy should be directed to induce
tumour dormancy by selectively inducing cell-cycle
arrest, rather than aiming at tumour eradication,
because one source of robustness is genetic heterogeneity created through somatic recombination. Apart from
specific cases where tumour cells are relatively homogeneous, tumour-mass reduction might not be an appropriate therapeutic goal, because of the high risk of
relapse if the reduced tumour gained a greater level of
heterogeneity, including those cells that are highly resistant to therapeutic efforts. Controlling tumour robustness would be a genuine measure of therapeutic efficacy,
so that the risk of relapse could be well controlled. The
other approach is to find the fragility of tumours.
Tumours acquire robustness against a range of therapies, which must be accompanied by the development
of extreme fragility.
The question remains how to find fragility that is
therapeutically effective and practical. It is important to
investigate what the system has been optimized for and
to identify sources of robustness. As fragility is a byproduct of robustness, the fragile point of the system
| NOVEMBER 2004 | VOLUME 5
must be associated with a mechanism that gives rise to
enhanced robustness. For example, the robustness of a
tumour is sustained by chromosome instability, intracellular feedback loops and host-tumour interactions.
Possible countermeasures include control of the cell
cycle through the combined use of several drugs, possibly using RNAi, selectively destabilizing or stabilizing
unstable chromosomes, selectively delivering engineered genes108 to re-establish control of host-tumour
interactions or introducing artificial genetic circuits109 to
conditionally express tumour-suppressor genes, should
be considered. Importantly, these countermeasures have
to be carefully designed to specifically explore fragility
or to control robustness.
HIV predominantly infects CD4-positive T cells and
replicates when the cell activates its anti-virus responses110–113. Infected cells are not removed because
infection is hardly detectable. This is a typical case of
what happens when malfunctioning components (the
infected cell, in this case) are not removed from the system. In addition, HIV creates diverse genetic alternation.
HIV’s strategy is to hijack the robust immune-response
mechanism, which makes AIDS difficult to cure. Robust
persistence of the epidemic state through generation of
genetic diversity and various feedback loops is similar to
the strategy found in cancer. Possible therapeutic
approaches already discussed might apply to the effective treatment of AIDS. Interestingly, one strategy proposed to combat HIV involves forcing HIV into a latent
state, instead of trying to remove it, by introducing a
decoy — a conditionally replicating HIV-1 vector
(crHIV-1)114–116. crHIV-1 contains cis, but not trans, elements that are necessary for virus packaging, and carries
antiviral genes that inhibit wild-type HIV-1 functions116.
If successful, crHIV-1 might re-establish control of the
system and put HIV-1 virus in a state of prolonged
latency.
Towards a theory of biological robustness
Given the importance of robustness for the understanding of the principles of life and its medical implications,
it is an intriguing challenge to formulate a mathematically solid, and possibly unified theory of biological
robustness that might serve as a basic organizational
principle of biological systems (early attempts date back
to the middle of the last century117,118). Such a unified
theory could be a bridge between the fundamental principles of life, medical practice, engineering, physics and
chemistry. This is a difficult challenge in which a number of issues have to be solved, particularly to establish
mathematically well-founded theories. However, the
impact would be enormous.
First, the solid quantitative index of robustness has to
be established. This index has to be able to equate with
experimentally measurable quantities, so that hypotheses
on the degree of robustness can be tested experimentally.
There are several concepts on the stability of the system
in CONTROL THEORY119, and some attempt has been made to
apply these ideas to biological robustness120. The practical application of these concepts to biological systems
has been limited, mainly owing to the enormous degree
www.nature.com/reviews/genetics
©2004 Nature Publishing Group
REVIEWS
SHANNON’S CHANNEL-CODING
THEOREM
A theorem by Claude Shannon
which indicates that for a given
channel there exists a code that
will permit the error-free
transmission across the channel
at a rate R, provided R≤C, where
C is the channel capacity. This
means that the probability of
error will not equal zero when
R>C, that is, transmission is
larger than channel capacity.
LE CHATELIER–BRAUN’S
PRINCIPLE
A thermodynamics principle
which states that if a dynamic
equilibrium is disturbed by
changing the conditions, then
the system tends to adjust to a
new equilibrium counteracting
the change.
BELOUSOV-ZHABOTINSKI
REACTIONS
This is a chemical reaction that is
widely used to demonstrate
transition from the nearequilibrium state to the farfrom-equilibrium state. When a
low level of heat is applied, it is
dissipated without affecting the
qualitative characteristics of the
medium, but when additional
heat is applied, the system
undergoes a drastic change, and
a circulating flow of chemicals
emerges.
of dimensionality and non-linearity that is intrinsic to
biological systems. Measuring perturbations for all
these dimensions is impractical. A practical, theoretically solid index needs to be developed to provide
guidelines for selecting smaller numbers of parameters to be perturbed so that the system’s response can
be measured. Haken pointed out that only a few parameters are important to describe system behaviour
near the bifurcation point121, but this approach is too
limited to analyse various types of behaviour of biological systems. A number of mathematical methods
are being developed that might be of benefit for the
mathematical analysis of the robustness of biological
systems122,123. On the experimental front, comprehensive mutant creation and dosage-controlled perturbations are already feasible for some organisms. As there
are often practical aims to investigate robustness of
the system — for example, to induce cell-cycle arrest
— experiments can be carefully designed to perturb
relevant genes and parameters, and published data
can be systematically collected to obtain practical, sufficient indeces of robustness for specific aspects of the
system124.
Second, a theory that embraces the various aspects of
robustness has not yet been formulated. Control theory
is often used to explain robustness that involves feedback regulation, but this only covers one aspect of
robustness. Furthermore, the control theory assumes
that there is a certain set point, determined by the
designer, that the system’s state will approach, even
when perturbed. Of course, there is no such designer in
biological systems, and set point is implicit in the equilibrium state of the system, which often changes dynamically. A new theory is required to reflect these features
of biological systems.
Recent efforts to integrate control theory and
SHANNON’S CHANNEL-CODING THEOREM might provide an
interesting framework for feedbacks125,126. This theorem tries to formalize cases where information capacity
in the feedback loop is limited, so that the feedback signal is potentially impeded by fidelity, noise and delay.
This better reflects the reality of biological systems, in
which signals are transmitted with noise, delay and
compromised fidelity.
Third, attempts to relate the dynamics of life to thermodynamics have been of limited success. The behaviour of physical and chemical systems near equilibrium
have been investigated in the past. LE CHATELIER–BRAUN’S
PRINCIPLE provides the basis for response against perturbations for systems in equilibrium127. This principle
indicates the emergence of a compensatory feedback
to cope with perturbations. Efforts have been made to
extend thermodynamics from equilibrium to explain
biological phenomena128–130, but apply only to relatively
simple chemical reactions under a medium, such as in
BELOUSOV-ZHABOTINSKI (BZ) REACTIONS. Signal transduction,
for example, brings about dramatic changes in the state
of the cell depending on the intensity and types of stimuli, resembling transition from the near-equilibrium
state to a state that is far from equilibrium. However, this
is not the result of simple chemical reactions under an
NATURE REVIEWS | GENETICS
unstructured medium, as seen in BZ reactions. Cells
are highly structured and have explicit interactions
and physical structures that are controlled by gene
expression, cytoskeleton and other regulatory systems
that are optimized to be robust and evolvable. The
greatest challenge will be to formulate theories that
account for thermodynamics in heterogeneous and
structured systems. Some efforts are being made to
find a theoretical framework by assuming network
structure as a basis of mathematical description 131.
However, the results are still too abstract to be practically applied to biological systems. Progress in this field
will help to connect biology, chemistry, physics and
mathematics in a coherent manner.
Conclusion
Robustness is a fundamental property of biological
systems. It facilitates evolvability, and evolution selects
robust traits. I have argued that there are specific architectural features required to make organisms robust
and that they might be universal to any robust and
evolvable complex system. System controls, modularity, alternative mechanisms and decoupling serve as
basic mechanisms to provide robustness to the system,
but these mechanisms need to be organized into
coherent architecture to be effective at the level of the
organism. Enhancement of robustness against perturbations can be made through the combination of these
mechanisms, but system control is the prime mechanism for coping with environmental perturbations
that require proper dynamics. Therefore, evolution of
organisms can be viewed, at least in one aspect, as evolution of control systems. Modularity, alternative
mechanisms and decoupling, in part, support the
robust maintenance of control loops, but are also controlled by control loops either explicitly or implicitly.
There is an intriguing possibility that genetic buffering
and modularity originated from robustness against
environmental perturbations, and subsequently
evolved to have wider applicability. It is important to
realize that systems that are evolved to be robust
against certain perturbations are extremely fragile to
unexpected perturbations. This robust yet fragile
trade-off is fundamental to complex dynamic systems.
It is important to understand architectural features of
robust and evolvable systems and the intrinsic nature
of robustness and fragility because they dictate a mode
of system failures and effective countermeasures,
which, as previously discussed, has direct implications
for understanding disease and devising effective therapies. Failures and viable countermeasures of these systems are often counter-intuitive, which implies that a
theoretically motivated robustness perspective might
provide new therapeutic approaches. The emerging
field of systems biology has been trying to identify system-level properties, but simple use of large data sets
and computation would not effectively provide insights
into biological systems. The perspective on biological
robustness would provide effective guiding principles
for understanding many biological phenomena, and
for therapy design.
VOLUME 5 | NOVEMBER 2004 | 8 3 5
©2004 Nature Publishing Group
REVIEWS
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
836
Kitano, H. Systems biology: a brief overview. Science 295,
1662–1664 (2002).
Kitano, H. Computational systems biology. Nature 420,
206–210 (2002).
Little, J. W., Shepley, D. P. & Wert, D. W. Robustness of a
gene regulatory circuit. EMBO J. 18, 4299–4307 (1999).
Ptashne, M. A Genetic Switch: Gene Control and Phage λ
(Blackwell Scientific, Oxford, 1987).
Zhu, X. M., Yin, L., Hood, L. & Ao, P. Calculating biological
behaviors of epigenetic states in the phage λ life cycle.
Funct. Integr. Genomics 4, 188–195 (2004).
Santillan, M. & Mackey, M. C. Why the lysogenic state of
phage λ is so stable: a mathematical modeling approach.
Biophys. J. 86, 75–84 (2004).
Alon, U., Surette, M. G., Barkai, N. & Leibler, S. Robustness
in bacterial chemotaxis. Nature 397, 168–171 (1999).
A seminal research paper on the robust adaptation
observed in bacterial chemotaxis.
Barkai, N. & Leibler, S. Robustness in simple biochemical
networks. Nature 387, 913–917 (1997).
Yi, T. M., Huang, Y., Simon, M. I. & Doyle, J. Robust perfect
adaptation in bacterial chemotaxis through integral
feedback control. Proc. Natl Acad. Sci. USA 97, 4649–4653
(2000).
von Dassow, G., Meir, E., Munro, E. M. & Odell, G. M. The
segment polarity network is a robust developmental
module. Nature 406, 188–192 (2000).
A study that suggests that the modular robust
network contributes to pattern formation during
embryogenesis in D. melanogaster.
Ingolia, N. T. Topology and robustness in the Drosophila
segment polarity network. PLoS Biol. 2, e123 (2004).
Barkai, N. & Shilo, B. Modeling pattern formation: counting
to two in the Drosophila egg. Curr. Biol. 12, R493 (2002).
Houchmandzadeh, B., Wieschaus, E. & Leibler, S.
Establishment of developmental precision and proportions
in the early Drosophila embryo. Nature 415, 798–802
(2002).
Kitano, H. Cancer robustness: tumour tactics. Nature 426,
125 (2003).
Kitano, H. Cancer as a robust system: implications for
anticancer therapy. Nature Rev. Cancer 4, 227–235 (2004).
A theoretical proposal to approach cancer treatment
from the aspect of robustness.
Kitano, H. et al. Metabolic syndrome and robustness
trade–offs. Diabetes 53, (Suppl. 3), 1–10 (2004).
Morohashi, M. et al. Robustness as a measure of plausibility
in models of biochemical networks. J. Theor. Biol. 216,
19–30 (2002).
Borisuk, M. T. & Tyson, J. J. Bifurcation analysis of a model
of mitotic control in frog eggs. J. Theor. Biol. 195, 69–85
(1998).
Rao, C. V., Kirby, J. R. & Arkin, A. P. Design and diversity in
bacterial chemotaxis: a comparative study in Escherichia
coli and Bacillus subtilis. PLoS Biol. 2, e49 (2004).
McAdams, H. H. & Arkin, A. It’s a noisy business! Genetic
regulation at the nanomolar scale. Trends Genet. 15, 65–69
(1999).
Rao, C. V., Wolf, D. M. & Arkin, A. P. Control, exploitation
and tolerance of intracellular noise. Nature 420, 231–237
(2002).
McAdams, H. H. & Shapiro, L. Circuit simulation of genetic
networks. Science 269, 650–656 (1995).
Arkin, A., Ross, J. & McAdams, H. H. Stochastic kinetic
analysis of developmental pathway bifurcation in phage λinfected Escherichia coli cells. Genetics 149, 1633–1648
(1998).
McAdams, H. H. & Arkin, A. Stochastic mechanisms in gene
expression. Proc. Natl Acad. Sci. USA 94, 814–819 (1997).
Bagowski, C. P., Besser, J., Frey, C. R. & Ferrell, J. E. Jr. The
JNK cascade as a biochemical switch in mammalian cells:
ultrasensitive and all-or-none responses. Curr. Biol. 13,
315–320 (2003).
Ferrell, J. E. Jr. Self-perpetuating states in signal
transduction: positive feedback, double-negative feedback
and bistability. Curr. Opin. Cell. Biol. 14, 140–148 (2002).
Tyson, J. J., Chen, K. & Novak, B. Network dynamics and
cell physiology. Nature Rev. Mol. Cell Biol. 2, 908–916
(2001).
Freeman, M. Feedback control of intercellular signalling in
development. Nature 408, 313–319 (2000).
A comprehensive review of how feedback control
contributes to robustness and pattern formation
during development.
Agrawal, A. A. Phenotypic plasticity in the interactions and
evolution of species. Science 294, 321–326 (2001).
West-Eberhard, M. J. Developmental Plasticity and
Evolution (Oxford Univ. Press, Oxford, 2003).
Schlichting, C. & Pigliucci, M. Phenotypic Evolution: A
Reaction Norm Perspective (Sinauer Associates Inc.,
Sunderland, Massachusetts, 1998).
32. Schwob, E. & Nasmyth, K. CLB5 and CLB6, a new pair of B
cyclins involved in DNA replication in Saccharomyces
cerevisiae. Genes Dev. 7, 1160–1175 (1993).
33. Kellis, M., Birren, B. W. & Lander, E. S. Proof and
evolutionary analysis of ancient genome duplication in the
yeast Saccharomyces cerevisiae. Nature 428, 617–624
(2004).
34. Langkjaer, R. B., Cliften, P. F., Johnston, M. & Piskur, J.
Yeast genome duplication was followed by asynchronous
differentiation of duplicated genes. Nature 421, 848–852
(2003).
35. Ohno, S. Evolution by Gene Duplication (Springer, Berlin,
1970).
36. Gu, X. Evolution of duplicate genes versus genetic
robustness against null mutations. Trends Genet. 19,
354–356 (2003).
37. Gu, Z. et al. Role of duplicate genes in genetic robustness
against null mutations. Nature 421, 63–66 (2003).
38. Nowak, M. A., Boerlijst, M. C., Cooke, J. & Smith, J. M.
Evolution of genetic redundancy. Nature 388, 167–171
(1997).
39. Berg, J., Tymoczko, J. & Stryer, L. Biochemistry 5th edn (W.
H. Freeman, 2002).
40. DeRisi, J. L., Iyer, V. R. & Brown, P. O. Exploring the
metabolic and genetic control of gene expression on a
genomic scale. Science 278, 680–686 (1997).
41. Edwards, J. S. & Palsson, B. O. Robustness analysis of the
Escherichia coli metabolic network. Biotechnol. Prog. 16,
927–939 (2000).
42. Edwards, J. S., Ibarra, R. U. & Palsson, B. O. In silico
predictions of Escherichia coli metabolic capabilities are
consistent with experimental data. Nature Biotechnol. 19,
125–130 (2001).
43. Conant, G. C. & Wagner, A. Convergent evolution of gene
circuits. Nature Genet. 34, 264–266 (2003).
44. Teichmann, S. A. & Babu, M. M. Gene regulatory network
growth by duplication. Nature Genet. 36, 492–496
(2004).
45. Hartwell, L. H., Hopfield, J. J., Leibler, S. & Murray, A. W.
From molecular to modular cell biology. Nature 402, 47–52
(1999).
An influential article that stresses the importance of a
system-based approach in biology, with particular
emphasis on modularity.
46. Schlosser, G. & Wagner, G. (eds.) Modularity in
Development and Evolution (Univ. Chicago Press, Chicago,
2004).
An edited collection of papers on modularity in
biological systems that provides up-to-date
discussions of the issue.
47. Baldwin, C. & Clark, K. Design Rules, Vol. 1: The Power of
Modularity (MIT Press, Cambridge, Massachusetts, 2000).
48. Simon, H. The Sciences and the Artificial 3rd edn (MIT
Press, Cambridge, Massachusetts, 1996).
49. McAdams, H. H., Srinivasan, B. & Arkin, A. P. The evolution
of genetic regulatory systems in bacteria. Nature Rev.
Genet. 5, 169–178 (2004).
50. Spirin, V. & Mirny, L. A. Protein complexes and functional
modules in molecular networks. Proc. Natl Acad. Sci. USA
100, 12123–12128 (2003).
51. Rutherford, S. L. & Lindquist, S. Hsp90 as a capacitor for
morphological evolution. Nature 396, 336–342 (1998).
52. Queitsch, C., Sangster, T. A. & Lindquist, S. Hsp90 as a
capacitor of phenotypic variation. Nature 417, 618–624
(2002).
53. Rutherford, S. L. Between genotype and phenotype: protein
chaperones and evolvability. Nature Rev. Genet. 4, 263–274
(2003).
This article describes issues to do with genetic
buffering, particularly involving hsp90, and outlines
the implications of such phenomena.
54. Waddington, C. H. The Strategy of the Genes: a Discussion
of Some Aspects of Theoretical Biology (Macmillan, New
York, 1957).
55. Kimura, M. Preponderance of synonymous changes as
evidence for the neutral theory of molecular evolution.
Nature 267, 275–276 (1977).
56. Kimura, M. The neutral theory of molecular evolution. Sci.
Am. 241, 98–100, 102, 108 passim (1979).
57. Hartman, J. L. 4th, Garvik, B. & Hartwell, L. Principles for the
buffering of genetic variation. Science 291, 1001–1004
(2001).
58. Bergman, A. & Siegal, M. L. Evolutionary capacitance as a
general feature of complex gene networks. Nature 424,
549–552 (2003).
59. Kitami, T. & Nadeau, J. H. Biochemical networking
contributes more to genetic buffering in human and mouse
metabolic pathways than does gene duplication. Nature
Genet. 32, 191–194 (2002).
60. Siegal, M. L. & Bergman, A. Waddington’s canalization
revisited: developmental stability and evolution. Proc. Natl
Acad. Sci. USA 99, 10528–10532 (2002).
| NOVEMBER 2004 | VOLUME 5
61. Lev Bar-Or, R. et al. Generation of oscillations by the p53Mdm2 feedback loop: a theoretical and experimental study.
Proc. Natl Acad. Sci. USA 97, 11250–11255 (2000).
62. Lahav, G. et al. Dynamics of the p53-Mdm2 feedback loop
in individual cells. Nature Genet. 36, 147–150 (2004).
63. Kirschner, M. & Gerhart, J. Evolvability. Proc. Natl Acad. Sci.
USA 95, 8420–8427 (1998).
An influential article that discusses the various
molecular and system mechanisms that contribute to
the evolvability of organisms. Together with their book
(reference 64), basic issues and perspectives are
presented.
64. Gerhart, J. & Kirschner, M. Cells, Embryos, and Evolution:
Toward a Cellular and Developmental Understanding of
Phenotypic Variation and Evolutionary Adaptability
(Blackwell Science, Malden, Massachusetts, 1997).
65. Broder, A. et al. in The Nineth International World Wide Web
Conference 309–320 (Elsevier Science, Amsterdam, , 2000).
66. Csete, M. E. & Doyle, J. Bow ties, metabolism and disease.
Trends Biotechnol. 22, 446–450 (2004).
67. de Visser, J. et al. Evolution and detection of genetics
robustness. Evolution 57, 1959–1972 (2003).
A summary of workshop discussions on evolvability
and robustness, which are relevant to current
discussions.
68. Eldar, A. et al. Robustness of the BMP morphogen gradient
in Drosophila embryonic patterning. Nature 419, 304–308
(2002).
69. Nieuwkoop, P. D. Pattern formation in artificially activated
ectoderm (Rana pipiens and Ambystoma punctatum). Dev.
Biol. 7, 255–279 (1963).
70. Nieuwkoop, P. D. Inductive interactions in early amphibian
development and their general nature. J. Embryol. Exp.
Morphol. 89, S333–S347 (1985).
71. Pires-daSilva, A. & Sommer, R. J. The evolution of signalling
pathways in animal development. Nature Rev. Genet. 4,
39–49 (2003).
72. Carroll, S., Grenier, J. & Weatherbee, S. From DNA to
Diversity: Molecular Genetics and the Evolution of Animal
Design (Blackwell, Oxford, 2001).
This book describes the role of toolkit and co-option
in development.
73. Wagner, G. P., Amemiya, C. & Ruddle, F. Hox cluster
duplications and the opportunity for evolutionary novelties.
Proc. Natl Acad. Sci. USA 100, 14603–14606 (2003).
74. Gehring, W. J. Master Control Genes in Development and
Evolution: The Homeobox Story (Yale Univ. Press, New
Haven; London, 1998).
75. Lewis, E. B. A gene complex controlling segmentation in
Drosophila. Nature 276, 565–570 (1978).
76. Kaufman, T. C., Seeger, M. A. & Olsen, G. Molecular and
genetic organization of the antennapedia gene complex of
Drosophila melanogaster. Adv. Genet. 27, 309–362 (1990).
77. Struhl, G. A homoeotic mutation transforming leg to antenna
in Drosophila. Nature 292, 635–638 (1981).
78. Halder, G., Callaerts, P. & Gehring, W. J. Induction of ectopic
eyes by targeted expression of the eyeless gene in
Drosophila. Science 267, 1788–1792 (1995).
79. Taniguchi, T. & Takaoka, A. A weak signal for strong
responses: interferon-α/β revisited. Nature Rev. Mol. Cell
Biol. 2, 378–386 (2001).
80. Bhalla, U. S. & Iyengar, R. Robustness of the bistable
behavior of a biological signaling feedback loop. Chaos 11,
221–226 (2001).
81. Wagner, G. P. & Altenberg, L. Complex adaptations and the
evolution of evolvability. Evolution 50, 967–976 (1996).
82. Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N. &
Barabasi, A. L. The large-scale organization of metabolic
networks. Nature 407, 651–654 (2000).
83. Barabasi, A. L. & Oltvai, Z. N. Network biology:
understanding the cell’s functional organization. Nature Rev.
Genet. 5, 101–113 (2004).
A good summary of the scale-free network in biology
by the originator of the idea.
84. Albert, R., Jeong, H. & Barabasi, A. L. Error and attack
tolerance of complex networks. Nature 406, 378–382
(2000).
85. Ma, H. W. & Zeng, A. P. The connectivity structure, giant
strong component and centrality of metabolic networks.
Bioinformatics 19, 1423–1430 (2003).
86. van Nimwegen, E. Scaling laws in the functional content of
genomes. Trends Genet. 19, 479–484 (2003).
87. van Nimwegen, E. in Power Laws, Scale–free Networks,
and Genome Biology (ed. Koonin, E. V.) (Landes Bioscience,
in the press).
88. Ashburner, M. et al. Gene ontology: tool for the unification of
biology. The Gene Ontology Consortium. Nature Genet. 25,
25–29 (2000).
89. Jordan, J. D., Landau, E. M. & Iyengar, R. Signaling
networks: the origins of cellular multitasking. Cell 103,
193–200 (2000).
www.nature.com/reviews/genetics
©2004 Nature Publishing Group
REVIEWS
90. Jordan, J. D. & Iyengar, R. Modes of interactions between
signaling pathways. Biochem. Pharmacol. 55, 1347–1352
(1998).
91. Hermans, E. Biochemical and pharmacological control of
the multiplicity of coupling at G-protein-coupled receptors.
Pharmacol. Ther. 99, 25–44 (2003).
92. Werry, T. D., Wilkinson, G. F. & Willars, G. B. Mechanisms of
cross-talk between G-protein-coupled receptors resulting in
enhanced release of intracellular Ca2+. Biochem. J. 374,
281–296 (2003).
93. Vassilatis, D. K. et al. The G protein-coupled receptor
repertoires of human and mouse. Proc. Natl Acad. Sci. USA
100, 4903–4908 (2003).
94. Mattick, J. S. RNA regulation: a new genetics? Nature Rev.
Genet. 5, 316–323 (2004).
95. Mattick, J. S. Non-coding RNAs: the architects of eukaryotic
complexity. EMBO Rep. 2, 986–991 (2001).
96. Mattick, J. S. Challenging the dogma: the hidden layer of
non-protein-coding RNAs in complex organisms. Bioessays
25, 930–939 (2003).
97. Schreiber, S. L. & Bernstein, B. E. Signaling network model
of chromatin. Cell 111, 771–778 (2002).
98. Carroll, S. B. Chance and necessity: the evolution of
morphological complexity and diversity. Nature 409,
1102–1109 (2001).
99. Carlson, J. M. & Doyle, J. Highly optimized tolerance: a
mechanism for power laws in designed systems. Phys. Rev.
E 60, 1412–1427 (1999).
100. Carlson, J. M. & Doyle, J. Complexity and robustness. Proc.
Natl Acad. Sci. USA 99 (Suppl 1), 2538–2545 (2002).
An introductory article on the highly optimized
tolerance (HOT) theory.
101. Barabasi, A. L. & Albert, R. Emergence of scaling in random
networks. Science 286, 509–512 (1999).
102. Bak, P., Tang, C. & Wiesenfeld, K. Self-organized criticality.
Phys. Rev. A 38, 364–374 (1988).
103. Csete, M. E. & Doyle, J. C. Reverse engineering of biological
complexity. Science 295, 1664–1669 (2002).
An inspiring piece that discusses the complexity of
biological systems from engineering and controltheory perspective.
104. Lamport, L., Shostak, R. & Pease, M. The Byzantine
generals problem. ACM Trans. Program. Lang. Sys. 4,
382–401 (1982).
105. Cassel, D. & Pfeuffer, T. Mechanism of cholera toxin action:
covalent modification of the guanyl nucleotide-binding
protein of the adenylate cyclase system. Proc. Natl Acad.
Sci. USA 75, 2669–2673 (1978).
106. Moss, J. & Vaughan, M. Guanine nucleotide-binding
proteins (G proteins) in activation of adenylyl cyclase:
lessons learned from cholera and ‘travelers’ diarrhea’.
J. Lab. Clin. Med. 113, 258–268 (1989).
107. Harris, A. L. Hypoxia — a key regulatory factor in tumour
growth. Nature Rev. Cancer 2, 38–47 (2002).
108. Bingle, L., Brown, N. J. & Lewis, C. E. The role of tumourassociated macrophages in tumour progression:
implications for new anticancer therapies. J. Pathol. 196,
254–265 (2002).
109. Hasty, J., McMillen, D. & Collins, J. J. Engineered gene
circuits. Nature 420, 224–230 (2002).
110. McMichael, A. J. & Rowland-Jones, S. L. Cellular
immune responses to HIV. Nature 410, 980–987 (2001).
111. McCune, J. M. The dynamics of CD4+ T-cell depletion in
HIV disease. Nature 410, 974–979 (2001).
112. Richman, D. D. HIV chemotherapy. Nature 410,
995–1001 (2001).
113. Pomerantz, R. J. HIV: a tough viral nut to crack. Nature
418, 594–595 (2002).
114. Dropulic, B., Hermankova, M. & Pitha, P. M. A
conditionally replicating HIV-1 vector interferes with wildtype HIV-1 replication and spread. Proc. Natl Acad. Sci.
USA 93, 11103–11108 (1996).
115. Weinberger, L. S., Schaffer, D. V. & Arkin, A. P.
Theoretical design of a gene therapy to prevent AIDS but
not human immunodeficiency virus type 1 infection.
J. Virol. 77, 10028–10036 (2003).
116. Mautino, M. R. & Morgan, R. A. Gene therapy of HIV-1
infection using lentiviral vectors expressing anti-HIV-1
genes. AIDS Patient Care STDS 16, 11–26 (2002).
117. von Bertalanffy, L. General System Theory: Foundations,
Development, Applications (George Braziller Inc., New
York, 1976).
118. Wiener, N. Cybernetics: or Control and Communication
in the Animal and the Machine (MIT Press, Cambridge,
Massachusetts, 1948).
119. Doyle, J., Glover, K., Khargonekar, P. & Francis, B.
State-space solutions to standard H2 and H1 control
problems. IEEE Trans. Automat. Control 34, 831–847
(1989).
120. Ma, L. & Iglesias, P. A. Quantifying robustness of
biochemical network models. BMC Bioinformatics 3, 38
(2002).
121. Haken, H. Synergetics — An Introdution 2nd edn
(Springer, Berlin, 1978).
122. Prajna, S. & Papachristodoulou, A. in Proceedings of
American Control Conference 2779–2784 (IEEE, Denver,
Colorado, 2003).
123. Prajna, S., Papachristodoulou, A. & Parrilo, P. A. in
Proceedings of IEEE Conference on Decision and Control
741–746 (IEEE, Las Vegas, 2002).
124. Chen, K. C. et al. Integrative analysis of cell cycle
control in budding yeast. Mol. Biol. Cell 15, 3841–3862
(2004).
NATURE REVIEWS | GENETICS
Acknowledgements
I would like to thank members of the Sony Computer Science laboratories, Inc. and ERATO-SORST Kitano Symbiotic Systems
Project for their fruitful discussions, John Doyle and Marie Csete for
critical reading of the initial version of this article, a number of colleagues who discussed the article, and anonymous referees for
informative comments. This research is, in part, supported by
the ERATO-SORST programme (run by the Japan Science and
Technology Agency) of the Systems Biology Institute, the Center
of Excellence programme, the special coordination funds
(Ministry of Education, Culture, Sports, Science, and Technology)
to Keio University and the Air Force Office of Scientific Research
(AFOSR/AOARD).
Competing interests statement
The author declares no financial interests.
Online links
DATABASES
The following terms in this article are linked online to:
Entrez: http://www.ncbi.nih.gov/Entrez/
wg | en | Clb5 | Clb6
FURTHER INFORMATION
Gene Ontology: http://www.geneontology.org/
Kitano’s web page:
http://www.csl.sony.co.jp/person/kitano/kitano.html
Access to this interactive links box is free online.
VOLUME 5 | NOVEMBER 2004 | 8 3 7
©2004 Nature Publishing Group
View publication stats
125. Martins, N. C. & Dahleh, M. A. in Forty-Second
Annual Allerton Conference on Communication,
Control, and Computing (Univ. Illinois, UrbanaChampaign, 2004).
126. Martins, N. C., Dahleh, M. A. & Elia, N. in IEEE Conference
on Decision and Control (IEEE, Nassau, The Bahamas, in
the press).
127. Prigogine, I. & Defay, R. Chemical Thermodynamics (Everett,
Longmans Geeen, London, 1954).
128. Prigogine, I., Lefever, R., Goldbeter, A. & HerschkowitzKaufman, M. Symmetry breaking instabilities in biological
systems. Nature 223, 913–916 (1969).
129. Prigogine, I., Nicolis, G. & Babloyantz, A. Nonequilibrium
problems in biological phenomena. Ann. NY Acad. Sci. 231,
99–105 (1974).
130. Nicolis, G. & Prigogine, I. Self-Organization in
Non–Equilibrium Systems: From Dissipative Structures to
Order Through Fluctuations (J. Wiley & Sons, New York,
1977).
131. Ao, P. Potential in stochastic differential equations: novel
construction. J. Phys. A 37, L25–L30 (2004).