Some new directions in control theory inspired by

systems biology
E.D. Sontag
Abstract: This paper, addressed primarily to engineers and mathematicians with an interest in
control theory, argues that entirely new theoretical problems arise naturally when addressing
questions in the eld of systems biology. Examples from the authors recent work are used to
illustrate this point.
1 Introduction
We are in the midst of revolutionary developments in the
biological sciences. Literally each day brings new dis-
coveries, and proposals for novel organising principles,
which hold the promise of altering the understanding of life
itself. The potential impact of these developments, ranging
from the extension of life spans to the cure of diseases, is
impossible to overstate.
Concomitant with these advances, leading biologists have
recognised that new systems-level knowledge is urgently
required. Position papers emphasising this need regularly
appear in inuential publications such as Nature and
Science. The loosely dened eld of systems biology has
arisen, having as its goal the unravelling and conceptual-
isation of the basic dynamic processes, feedback control
loops, and signal processing mechanisms underlying life. As
evidence of the seriousness of this effort, consider the fact
that entire new academic departments have been recently
created, such as the Department of Systems Biology at
Harvard University Medical School, and new educational
programs have been established, such as MITs Compu-
tational and Systems Biology Ph.D. program. Even the
Institute for Advanced Studies, home to Albert Einstein and
John von Neumann, now has a Centre for Systems Biology.
The eld offers control theorists and engineers an
abundance of opportunities and challenges. These opportu-
nities and challenges may be broadly classied into several
categories, the rst three of which I will not discuss further
in this article:
(1) The role of control and signal processing techniques in
the design of instrumentation for high-precision biological
measurements and manipulation. This rich subject is hugely
varied in scope, and well covered in bioengineering
journals, such as for instance the IEEE Transactions on
Biomedical Engineering, on Medical Imaging, on Infor-
mation Technology in Biomedicine, and on Systems, Man
and Cybernetics.
(2) The use of existing techniques from identication, gain
quantication, sensitivity analysis, optimal control, and
other well-developed areas of control theory in the analysis
and solution of problems of interest to biologists. These
applications are also well-addressed in the literature: the
reader is referred to control journals such as IEE Proceed-
ings on Control Theory & Applications. Automatica, IEEE
Transactions on Automatic Control, or IEEE Transactions
on Control Technology, as well as journals focused on
mathematical biology or systems biology, such as Systems
Biology, IEEE Transactions on Computational Biology and
Bioinformatics, Mathematical Biosciences and Engineer-
ing, Mathematical Biosciences, Bulletin of Mathematical
Biology, Journal of Mathematical Biology, Mathematical
Medicine and Biology, and Biosystems. In addition, the
related elds of metabolic control theory [1, 2] and
biochemical systems theory [3] provide complementary
insights to control issues in biology.
(3) The abstraction, from biological research, of new ideas
for control and sensor engineering. Evolution has resulted in
systems that are highly fault-tolerant, non-linear, feedback-
rich, and truly hybrid (in the sense that the digital
information encoded in DNA controls chemical concen-
trations in cells). Recent advances in genomic research are
continually adding detailed knowledge of such systems
architecture and operation, and one may reasonably argue
that they will constitute a rich source of inspiration for
innovative solutions to problems of control and communi-
cation engineering, as well as sensor and actuator design and
integration.
(4) The formulation of entirely new theoretical control and
systems theory problems, motivated by systems biology
research. I will focus this article on this last point.
2 Cells as I=O systems
One may view cell life as a collection of wireless networks
of interactions among proteins, RNA, DNA, and smaller
molecules involved in signalling and energy transfer. These
networks process environmental signals, induce appropriate
cellular responses, and sequence internal events such as
gene expression, thus allowing cells and entire organisms to
perform their basic functions.
Research in molecular biology, genomics, and proteo-
mics has provided, and will continue to produce, a wealth of
data describing the elementary components of such net-
works, as well the mapping of intra- and inter-cellular
signalling networks. The genome encodes, through a
particular ordering of the four possible (A,T,C,G) bases in
q IEE, 2004
Systems Biology online no. 20045006
doi: 10.1049/sb:20045006
The author is with the Department of Mathematics and BioMaPS Institute
for Quantitative Biology, Hill Center, 110 Frelinghuysen Rd., Rutgers,
The State University of New Jersey, Piscataway, NJ 08854-8019, USA,
email: [email protected]
Paper rst received 14th April and in revised form 4th May 2004
Syst. Biol., Vol. 1, No. 1, June 2004 9
its DNA sequence, a parts list for the proteins that are
potentially present in every cell of a given organism.
Genomics research has as its objective the complete
decoding of this information, both the parts common for a
species as a whole as well as the cataloging of differences
among individual members. The shape of proteins is what
largely determines their function, and thus the elucidation of
their three-dimensional structure is a goal of proteomics
research. Proteins, which interact with each other through
lego-like tting of parts in lock and key fashion, are the
primary components of living things. Among other roles,
they form receptors that endow the cell with sensing
capabilities, actuators that make muscles move (myosin,
actin), detectors for the immune response, enzymes that
catalyse chemical reactions, and switches that turn genes on
or off. They also provide structural support, and help in the
transport of smaller molecules, as well as in directing the
breakdown and reassembly of other cellular elements such
as lipids and sugars. (An intermediate link between genetic
information and the proteins that DNA encodes for is RNA.
Until recently, RNA was not believed to be a direct player
in cell control mechanisms, but research into microRNA
conducted within the past two years is forcing a complete
rethinking of their role.) Massive amounts of data are being
generated by genomics and proteomics projects, facilitated
by sophisticated genetic engineering tools (gene knock-outs
and insertions, PCR), and measurement technologies (green
uorescent protein, microarrays, FRET), and there is a
widely recognised need to organise and interpret these data.
The control and systems-theory paradigm of input=output
systems, built out of simpler components that are inter-
connected according to certain rules, is a most natural one in
this context. Cells receive external information through
inputs that may be physical (UV or other radiation,
mechanical, or temperature) as well as chemical (drugs,
growth factors, hormones, nutrients), and their measurable
outputs include chemical signals to other cells, the move-
ment of agella or pseudopods, the activation of transcrip-
tion factors, and so forth. Each cell can be thought of, in
turn, as composed of a large number of subsystems,
involved in processes such as cell growth and maintenance,
division, and death. Indeed, an important theme in the
current molecular biology literature [4, 5] is the attempt to
understand cell behaviour in terms of cascades and feedback
interconnection of elementary modules.
As a simple illustration, consider the diagram shown in
Fig. 1, extracted from the paper on cancer research [6],
which describes the wiring diagram of the growth signalling
circuitry of the mammalian cell. Of course, such a diagram
leaves out a lot of information, some known but omitted for
simplicity, and some unknown: much of the system has not
been identied yet, and the numerical values of most
parameters as well as the functional forms of interactions
are only very approximately known. However, data is being
collected at an amazing rate and better and better models are
being constantly obtained. Many of the natural systems-
theoretic questions that one would normally pose for such a
system are precisely those that leading biologists are asking,
if sometimes in a different language: What is special about
the information-processing capabilities, or input=output
behaviours, of such networks, and how does one character-
ise these behaviours? How do the different signal transduc-
tion pathways interact? How does one nd the forms of
reactions, and values of parameters (identication, reverse
Fig. 1 Signalling circuitry of the mammalian cell, from [6]. The illustration shows signalling pathways for growth, differentiation, and
apoptosis commands, which instruct the cell to die. Highlighted in red are some of the genes known to be functionally altered in cancer cells.
The analogy with electronic integrated circuits is striking. Reproduced from [6] with permission from Elsevier
Syst. Biol., Vol. 1, No. 1, June 2004 10
engineering)? Once these forms of reactions are known,
how does one estimate time-varying internal states,
such as the concentrations of proteins and other chemical
substances, from input=output experiments (observer
problem)? What subsystems appear repeatedly? Where lie
the main sensitivities affecting robustness of the system?
What is the reason that there are cascades and feedback
loops? More generally, what can one say, if anything, about
stability, oscillations, and other dynamical properties of
such complex systems? In addition to analysis questions
there are, of course, also synthesis ones, dealing with the
control of cellular systems through drugs or genetic
modications.
Many publications address the above types of problems
for cell signalling systems, but the eld is still in its infancy,
and a major and long-term research effort will continue
toward their solution. For the rest of this article, however,
I would like to focus not on what existing control theory and
tools can do for systems biology but rather on a sort of
converse, namely: do new questions in control theory arise
from looking at problems in systems biology? (Not
surprisingly, my answer will be a resounding yes!.) This
is of interest for several reasons. First of all, control theory
draws one of its main strengths from its origins in concrete
applications. Abstraction is necessary and useful, but should
be grounded on concepts whose motivation lies outside of
pure theory. Second, it is by being open-minded and
receptive to the possibility of modifying our paradigmatic
problems that the control theory (and more generally the
mathematical) community may have an even greater impact
upon this new eld.
One of the main messages that I wish to convey it is that,
often, problems in systems biology sound like standard
problems in control theory, but that on closer inspection,
they differ in fundamental ways, and that these differences
are challenging and worth exploring. As I hope to convince
the reader through the brief discussion of selected examples,
new and interesting theoretical questions do arise. These
examples are unabashedly picked from my own research,
and, obviously, they constitute but a small sample of the
new problems that suggest themselves. For example, no
mention is made of multi-scale aspects in time and space,
of the need for effective hybrid techniques to handle the
interface of discrete and continuous components, or of the
study of stochastic aspects, which are particularly important
in systems involving small numbers of molecules.
3 A sample of questions
In the rest of this paper, I will briey describe a few new
research directions in control theory that were motivated by
the study of particular systems biology problems. The rst
one deals with the analysis of the dynamical behaviour of
interconnections of monotone systems with well-dened
characteristics, and is discussed in more detail than the rest.
3.1 A mix of qualitative and quantitative
modelling
The analysis of signalling networks constitutes one of the
central questions in systems biology. There is a pressing
need for powerful mathematical tools to help understand,
quantify, and conceptualise their information processing
and dynamic properties.
The most widespread approach to the modelling of such
complex systems involves the use of biological knowledge
in the design of large-scale simulation models. Extensive
simulations, often combined with bifurcation analysis and
dimensionality-reducing techniques such as time-scale
decompositions, are used in order to explore the space of
parameters and initial conditions. A shortcoming of this
approach is that it is virtually impossible to experimentally
validate the form of the non-linearities used in reaction
terms. Moreover, even when such forms are known, the
accurate estimation of coefcients (parameters) in vivo is
extremely hard if not impossible even in principle,
since enzyme and other chemical concentrations may
greatly vary from cell to cell. Coupled with the need for a
huge number of simulations in order to explore state spaces
of large dimension, and the fact that numerical algorithms
cannot be guaranteed to produce accurate results, due to
problems of local vs global convergence, numerical error,
and so forth, this argues for the necessity of more effective
theoretical tools. In addition, theory has the virtue of
providing fundamental understanding, by focusing on the
essential principles underlying given behaviours.
Of course, it would be na ve to search for a general theory
that would encompass all interesting phenomena exhibited
by cell networks. The range of potential non-linear
dynamical behaviours is too wide for that. But there is
hope, substantiated by the results to be described next, that
large classes of signalling systems may be protably studied
by rst decomposing them into several subsystems, each of
which is endowed with certain qualitative mathematical
properties which in conjunction with a relatively small
amount of quantitative data allow the behaviour of the
entire, reconstituted system, to be easily deduced from the
behaviour of its parts. This paradigm of decomposition and
reconnection has always been one of the basic principles in
systems theory and control engineering. Recently, a certain
class of subsystems that appear to be very suitable for the
analysis of enzymatic cascades and feedback loops was
identied in [79], in work done with David Angeli (see
also [10]). This novel approach emerged originally from our
study of possible multi-stability and oscillations in feedback
loops involving MAPK cascades.
Mitogen-activated protein kinase (MAPK) cascades
represent a biological module, or subcircuit, which is
ubiquitous in eukaryotic cell signal transduction processes
[1113] and is a critical component of pathways involved in
cell proliferation, differentiation, movement, and death.
They are activated by diverse stimuli (cytokines, growth
factors, neurotransmitters, hormones, cellular stress, cell
adherence), and their outputs drive different cellular
responses, including DNA transcription. The MAPK
cascade motif appears in different forms in distinct
organisms from yeast to humans, and even in any given
cell, and different MAPK cascades involve different
chemicals, but their basic architecture is conserved.
A MAPK pathway is a three-component module, consisting
of three kinases that establish a sequential activation
pathway comprising a MAPK kinase kinase (MKKK), a
MAPK kinase (MKK), and the nal MAPK itself [13].
(A kinase is an enzyme that catalyses the transfer of
phosphate groups from a high-energy phosphate-containing
molecule, such as ATP or ADP, to a substrate.) The diagram
in Fig. 2 illustrates one standard biological model for this
process. The output of the system is represented by the
activated form of MAPK, denoted as MAPK
++
in the
diagram. The precise meaning of this activation need not
concern us here, but we remark that MAPK must be
phosphorylated on both a threonine and tyrosine residue
for its activation, a dual phosphorylation catalysed by
activated-MKK. The concentration MAPK
++
(t) of this
activated form is controlled by the amount MKK
++
(t) of
Syst. Biol., Vol. 1, No. 1, June 2004 11
activated MKK present at any given time. (The intermediate
MAPK
+
(t) corresponds to a partially-activated form of
MAPK, where only one phosphorylation has taken place.)
A reverse reaction of dephosphorylation takes place as well,
not controlled by activated MKK, but controlled by a
phosphatase (an enzyme that removes phosphate groups),
whose concentration is assumed to be constant in the
time-scale being studied. Similarly, the concentrations of
partially and completely activated MKKs are controlled by
the concentration of activated MKKK. The input u to the
complete subsystem represents the concentration of a
different kinase or of a small GTP-binding protein, that
phosphorylates, and hence activates, MKKK.
Differential equation models of MAPK signalling
modules take as states the concentrations of MAPK(t),
MAPK
+
(t); and other variables; thus, corresponding to the
diagram in Fig. 2 there would be a set of eight differential
equations. (Variations in the literature include those in
which the rst subsystem is simplied to be one-
dimensional, or conversely, in which one includes
additional states corresponding to elementary intermediate
chemical reactions.) For example (see the references cited
earlier), the formula for dMAPK=dt would include a term
f (MAPK
+
(t)) to represent the effect of dephosphoryla-
tion of MAPK
+
; where f is some non-decreasing function
with f (0) = 0; and a term g(MKK
++
(t); MAPK(t)) to
represent the effect of phosphorylation by MKK
++
; where g
is non-decreasing function on each argument and g(0; ) =
g(; 0) = 0: There are, in addition, conservation laws. For
instance, MAPK(t) MAPK
+
(t) MAPK
++
(t) must be
constant, since these variables represent modied forms of
the same enzyme, the amount of which remains constant
unless one takes into account the slower time-scale
processes of transcription and translation of the protein
and its degradation.
The initial motivation for the work reported in [79] had
to do with the possible onset of oscillations, due to a Hopf
bifurcation, when inhibitory (negative) feedback is
introduced from the output to the input of a MAPK module,
cf [14, 15]. Such oscillations are not typically observed in
cells in their natural states, even though several inhibitory
feedbacks are known to exist, so it was natural to ask what
conditions on the feedback strength would guarantee that no
oscillations result. Small-gain theorems provide a routine
control theory approach to such problems, so a rst reaction
when confronted with this problem might be to attempt to
apply one such result (linearly for local analysis, or H

gain, or an ISS-small gain theorem, for example). However,

a technical problem arises: in this application and
typically in biological applications the location of
equilibria are changed by feedback. While in engineering
we are used to feedback laws of the form u = k(x) where k
preserves the equilibrium of interest (so, for instance, for the
open-loop system _ xx = x u; the equilibrium x = 0 is
preserved under u = kx), in biological applications the
equilibrium location may depend on the numerical values of
feedback gains. For example, consider the system _ xx =
x u; but suppose that the inhibitory feedback has the
form u =
a
bx
: Then the open-loop equilibrium at x = 0 will
change under feedback; for instance it becomes x = 1 for
_ xx = x
4
3x
: This shift of steady-state locations is
especially hard to handle with classical methods when we
wish to study global, as opposed to simply local, asymptotic
stability. Even more challenging, one should remember that
the exact forms of the non-linearities in the original system,
and certainly parameter values, are very hard to estimate; in
general, only rough forms are known. For instance, in the
diagram shown in Fig. 1, the only information given about
interactions is whether they are activating (arrows ) or
inhibitory ( symbols). Thus, a problem that resembles a
standard one in control theory turns out, on further
inspection, to be rather different in technical details.
These considerations led us to the introduction of a
different type of gains Cauchy gains for systems [16]
and more generally to the theory of monotone I=O systems
with characteristics in [7]. Using these concepts, we could
provide new forms of small-gain theorems that are very
well suited for MAPK cascades, and as it turned out, a large
number of other applications in biology (see for instance
[1719]).
Moreover, once we developed the basic results for
monotone systems with characteristics, we realised that
Fig. 2 Schematic of a MAPK signaling module. The output of the system is represented by the concentration of the activated form of
MAPK, denoted as MAPK
++
: Vertical arrows indicate that this concentration is in turn driven by the time-varying concentration of MKK
++
;
which in turn depends on the previous level of the cascade
Syst. Biol., Vol. 1, No. 1, June 2004 12
a wealth of other interesting questions, ranging from the
detection of bistability and hysteresis in systems biology
models to the study of delay-induced oscillations, could be
handled in the same fashion (see for instance [8, 9, 20,
21]). The key idea is to decompose a network into
subsystems each of which has the following two proper-
ties: input=output (I=O) monotonicity and existence of I=O
characteristics.
I=O monotonicity is dened mathematically through an
order-preservation property for ows (see below for more
details). The concept represents a generalisation of the notion
of monotone autonomous system studied by Hirsch, Smale,
Smith, and others [2225]. It can often be veried by simply
studying the sign structure of an incidence graph associated
to the Jacobian of the dynamics; in other words, by simply
inspecting which connections in a diagram such as the one
shown in Fig. 1 are activating () and which are inhibitory
(): Interconnections of monotone systems remain mono-
tone, and monotone systems to be precise, a slightly
stronger strict version, in which an appropriate margin is
required have very appealing global convergence proper-
ties. Almost all bounded trajectories converge to equilibria,
and therefore no limit cycles, chaotic behavior, and so forth,
can occur. In particular, all one-dimensional systems are
automatically monotone. An analogy to passivity was
suggested, to us by Rodolphe Sepulchre: just as statements
about interconnections of integrators often generalise to
statements about the same interconnections of passive
systems of arbitrary dimensions, statements about inter-
connections of scalar systems often generalise to statements
about monotone systems in arbitrary dimensions. One might
even speculate that I=O monotone systems may end up
playing a role in biological modelling not unlike that of
passive systems in mechanical and electrical systems.
Existence of characteristics means that there is a
monostable steady-state response to any constant external
input. This is, of course, a very strong constraint, but we
should remember that this constraint is only imposed on the
elementary components of our interconnections indeed,
the whole point of the theory is to predict what more
interesting dynamical behaviour emerges when several such
components are interconnected. In any event, we have
devoted considerable efforts to the characterisation of sub-
systems as appear in systems biology problems for which
this property is always satised, including receptor-ligand
models [26], certain general types of chemical reactions
[27], and very general processive modications such as
phosphorylation as with the special case of MAPK modules,
but not necessarily restricted to two steps, see [9, 28].
When coupled, these two properties, monotonicity and
existence of characteristics, allow the characterisation of
limit sets of arbitrary trajectories, under all possible external
forcing signals. When several systems with these two
properties are interconnected, possibly in feedback, and
possibly destroying monotonicity of the overall system
when negative feedback is used, the resulting system is not
monotone anymore this characterisation allows the
reduction of global stability questions to simple graphical
tests involving external signals. These graphical tests do not
require knowledge of the dynamics, but merely of the steady-
state response to constant inputs. Steady-state responses
called sometimes dose response curves, or activity assays for
receptor binding are routinely provided by biologists; they
constitute the quantitative information that our analysis
requires. Monotonicity, on the other hand, is a qualitative
property which is characterised by the topology of the
network (and the signs of interactions).
3.1.1 Monotone systems: A continuous-time
nite-dimensional system, in the standard sense of control
theory:
_ xx = f (x; u); y = h(x)
(one also studies delay-differential systems, reaction-
diffusion PDEs, and more abstract ows in metric spaces)
is monotone if there are non-trivial orders in the state,
input, and output spaces, such that

1
_
2
& u
1
_ u
2
= x(t;
1
; u
1
) _ x(t;
2
; u
2
)
\t _ 0
with respect to the state and input orders, and the output
map h preserves the order as well. As usual, x(t; ; u)
means the solution at time t if the initial state is at t = 0
and the input is u(); and u
1
_ u
2
for controls means that
u
1
(t) _ u
2
(t) for all t (see [7, 8] for more technical details).
Orders are typically dened by positivity cones, and one
may check monotonicity in innitesimal terms, not
requiring solution of differential equations (see the
references). A very special but most important case is
that of monotonicity with respect to cones that happen to
be orthants in Euclidean space. Suppose that a system is
sign-denite, meaning that we can draw unambiguous
sign-graphs for the Jacobians of f and h. More precisely,
(@f
i
=@x
j
)(x; u) has a constant sign e
ij
0; ; for all
(x, u) and all i ,= j (we may ignore self-loops), and, for all
i, j and (x, u), (@f
i
=@u
j
)(x; u) has a constant sign a
ij

0; ; and (@h
i
=@x
j
)(x) has a constant sign b
ij

0; ; : A system is monotone with respect to some
orthant if and only if its incidence graph does not contain
any negative cycles. Thus, for instance, the following
system (evolving on non-negative states and with non-
negative inputs) is monotone:
_ xx
1
= x
1

1
1 u
_ xx
2
= x
1
x
2
x
3
u
_ xx
3
= x
2
x
1
x
3
y = x
3
x
1
since there are no negative-parity cycles (including cycles
when directions of arrows are ignored) in its incidence
graph shown in Fig. 3. It is easy to show, using this
graphical characterization, that many systems of interest in
systems biology are indeed monotone (see [79, 28]).
3.1.2 Steady-state responses: The only quan-
titative component needed for the theory is the steady-state
response to step inputs. For our elementary subsystems, we
Fig. 3 Incidence graph for system _ xx
1
=x
1

1
1u
; _ xx
2
=x
1
x
2

x
3
u; _ xx
3
=x
2
x
1
x
3
with output y = x
3
x
1
: For example, the
negative arrow from u to x
1
represents the fact that
d_ xx
1
du
< 0
Syst. Biol., Vol. 1, No. 1, June 2004 13
require not just monotonicity but also that, for each constant
input u(t) = v there be a globally asymptotically stable state
x
v
of _ xx = f (x; v): As we discussed earlier, this is in principle
a very strong constraint, but we have shown that many
modules in systems biology have this property. For further
analysis, all that is required is the knowledge of the graph of
the steady-state output y = h(x
v
) as a function of the
constant input value being applied. We call this map the
I=O characteristic of the system.
3.1.3 Applications: In [7], we provided a small-
gain theorem for global stability of negative feedback loops
involving monotone systems with well-dened steady-state
responses. Leaving out some technical details, we may
summarise the results as follows. Suppose that _ xx = f (x; u);
y = h(x) is monotone and has characteristic y = k(u); and
we wish to study the effect of a feedback : y u that is
monotone decreasing. We assume for simplicity here that
the inputs and outputs are scalar. The result was that the
closed-loop has a globally attractive equilibrium provided
that the scalar discrete time iteration
u
i1
= F(u
i
) (F := k: u u)
has a globally attractive equilibrium uu (see Fig. 4). As with
classical small-gain theorems, phase is irrelevant, and
arbitrary delays are allowed in the feedback loop with no
change in the result. This was applied to the analysis of the
examples given in the literature of oscillations in MAPK
cascades.
Another early application of this new approach was to the
analysis of bistability (and more generally multi-stability).
A bistable system is one that admits two discrete, alternative
stable steady states. Early biological examples of bistable
systems included the lambda phage lysis-lysogeny switch
and the hysteretic lac repressor system [29, 30]. More recent
examples have included several MAPK cascades in animal
cells [31, 32], as well as cell cycle regulatory circuits in
Xenopus and S. cerevisiae [33, 34]. Bistable systems are
thought to be involved in the generation of switch-like
biochemical responses [31, 35, 36], the establishment of
cell cycle oscillations and mutually-exclusive cell cycle
phases [34, 35], the production of self-sustaining biochemi-
cal memories of transient stimuli [37, 38], and the rapid
lateral propagation of receptor tyrosine kinase activation
[39]. Our papers [810] showed how often and in
particular, in the case of MAPK signaling cascades
bistability and hysteretic behaviour can be analysed in terms
of decompositions into modules each of which is I=O
monotone and admits an I=O characteristic. The method led
us to establish properties of MAPK cascades at a level of
generality largely independent of functional forms and
parameter values, leading to conclusions regarding the
global stability of behaviour.
One way in which multi-stability arises is through
positive-feedback loops involving monotone systems with
well-dened steady-state responses. We assume once again
that the inputs and outputs are scalar (the multivariate case is
treated in [40]). The main result in [8] (again omitting some
technical details) was as follows. Suppose that _ xx = f (x; u);
y = h(x) is monotone and has characteristic y = k(u); and
we wish to study the effect of a feedback : y u which is
strictly increasing. We plot the characteristic k together with
the graph of the inverse of the mapping ; see Fig. 5 (where
we use (y) = y for simplicity, that is, we only illustrate
unity feedback). Next, we label the intersection points
between these two plots with a symbol S or U depending
on whether the derivative k
/
is less than that of the inverse
mapping
1
or not. The theorem guarantees then that there
are as many steady states in the complete closed loop system
_ xx = f (x; (y)) as there are intersections (this part is quite
trivial to show), and that among these states, those
corresponding to the intersections marked S are global
attractors, in the sense that all trajectories, except for those
originating in a set of measure zero in the state space,
converge to them. (A technical condition, not hard to check,
of strong monotonicity is needed as well, see [8].) Note that
this conclusion is only based on a qualitative piece of data,
the monotonicity of the open-loop system which one
may typically check using information from incidence
graphs and one quantitative piece of data, the plot of
the characteristic which is often readily available from
experiments. In [10], we illustrated this technique
through the analysis of bistability of a positive-feedback
loop involving progesterone-induced Xenopus oocyte
maturation. These results represent a broad generalisation
of classical work on feedbacks on cascades of
Fig. 4 Small-gain theorem for monotone system with I=O
characteristic k, under feedback : The spiderweb diagram,
shows convergence of the discrete one-dimensional iteration to
an input=output value uu; this is enough to guarantee that the
closed-loop system, which may have arbitrary dimensions, is
globally asymptotically stable
Fig. 5 Multi-stability theorem for monotone system with I=O
characteristic k, under feedback (here taken to be the identity).
The intersection points between the plots of k and
1
are in a one-
to-one correspondence with internal steady states; those marked
S correspond to those states (here, two) which attract almost all
trajectories
Syst. Biol., Vol. 1, No. 1, June 2004 14
one-dimensional systems such as described in [41] and
references there.
Ordinary differential equation models such as those
considered above implicitly assume that reactions proceed
in a well-mixed environment. While this is a reasonable
assumption when diffusion is fast compared to the time
scales of reactions (see for instance [42], Ex. 11.5), it is
other times important to explicitly incorporate spatial
inhomogeneity. This leads to reaction-diffusion partial
differential equations: instead of a dynamics _ xx = f (x)
(omitting now for simplicity the inputs u), one must
consider equations such as
@x
@t
= DDx f (x); where now x
= x(t; q) depends on both time t and space variables q, Dx is
the Laplacian of the vector x (that is, its ith coordinate
is
P
n
j=1
@
2
x
i
=@q
2
j
) with respect to the space variables, and
D is a diagonal matrix of positive diffusion constants.
It is a surprising and non-trivial fact that many stability
conclusions derived for the ordinary differential equation
_ xx = f (x) remain valid, with little or no change, for the PDE
_ xx = DDx f (x); subject to no-ux (Neumann) boundary
conditions, provided that the system _ xx = f (x) is (strongly)
monotone. For example, stable steady states x
0
of _ xx = f (x)
remain stable as homogeneous states (x(q) = x
0
) of the
PDE, see [25], Remark 7.6.1. Thus, diffusive instability,
which is the basis of Turings mechanism for pattern
formation, cannot arise in monotone systems. Unstable
states of _ xx = f (x) are also unstable when viewed as
homogeneous steady states of the associated diffusion
equation, since they are unstable already with respect to
homogeneous perturbations. On the other hand, in principle
there could be inhomogeneous (not constant on space
variables q) steady states of the diffusion system. However,
if the shape of the space in which the variables q evolve is
convex (for example, if a cell is modelled by a circle,
sphere, or ellipsoid), then there are no stable inhomo-
geneous states, i.e. the stable points of the ODE and PDE
systems coincide, see e.g. [4346]. These facts, applied to
the closed-loop system, provide an extension to the
reaction-diffusion case of the discussed results regarding
multi-stability and global behaviour of trajectories under
positive feedback.
3.2 Some other problems
To close, let me quickly mention a few other new theoretical
questions in control theory that arose from looking at
systems biology problems.
3.2.1 Disturbance rejection with signal
detection: One strength of control theory is that it
predicts that certain structures must be present in systems, in
order for regulation objectives to be met. Such insight may
help guide biological research, as it suggests a search for the
corresponding structures. Conversely, the absence of a
critical subsystem might be an indication that the biological
entity being studied does not regulate its behaviour in some
hypothesised sense.
The authors of [47] brought this point of view to a
wide biological audience. The setup was that of E. coli
chemotaxis: these bacteria move in response to the temporal
sensing of a chemical concentration gradient; for constant
concentrations, there is no net movement. (Actually the
story is more complicated, since the bacteria really perform
a kind of stochastic line search, with random tumbles to
reorient themselves when no gradient has been detected, but
this simplied version sufces for the point that we wish to
make.) It is experimentally known that the response, as
measured by the activity of motors driving agella, has a
zero DC gain, and a zero at the origin suggests a feedback
structure containing an integrator. One may view this
behaviour as disturbance rejection of step signals, and thus
motivated, the paper [47] proposed that the internal model
principle (IMP) due to Francis and Wonham predicts the
existence of an integrator in the E. coli chemotatic
subsystem; a possible mechanism see also [48] was
described. Recall that the classical IMP tells us that if a
controller regulates a system against external disturbances
in some family such as steps and if this regulation is
structurally stable in a precise mathematical sense, then the
controller must necessarily contain a subsystem such as
an integrator that can itself generate all such disturb-
ances, and which is driven by a suitable error signal.
But there is a potential drawback when attempting to use
this theorem in biological applications. The structural
stability criterion may be impossible to check: it would
imply, for instance, regulation even in the presence of direct
connections from inputs, such as cell receptors, to outputs,
such as agellar motors. Furthermore, the distinction
between controller and plant in a cell is not at all
obvious.
In addition, there is another characteristic that dis-
tinguishes the more classical engineering disturbance
rejection setup from the problem of interest in cell biology:
signal detection is of great importance. The cell should
detect the signal before regulating in fact, a large
overshoot is often desirable, in order to start up a
downstream subsystem. (On the other hand, adaptation
after the signal is processed is also often important: the
reaction might be metabolically too expensive to be kept
on for very long). The analogy would be the design of an
automobile active suspension system in which a require-
ment is that passengers hit the car roof as hard as possible,
when a bump in the road is encountered, before the control
system takes over and stabilises the oscillation.
Thus, once again, we are faced with a problem that
resembles a standard one in control theory, yet which
mathematically differs in a subtle fashion. Motivated by
this, in [49], we provided an internal model theorem that did
not require the assumption of structural stability, nor an a
priori requirement for the system to be partitioned into
separate plant and controller components. In lieu of
structural stable regulation, a signal detection criterion,
expressed through zero-dynamics and relative degree for
non-linear systems, was imposed in order to force the
existence of an internal model. Much work remains to be
done on this subject; the paper [49] barely scratched the
surface.
3.2.2 Systems identication and reverse
engineering: One of the most active current research
areas in systems biology is that of reverse engineering of
gene and protein networks. From measurements of how
certain measured variables concentrations of proteins,
amount of RNA being transcribed, etc. change in
response to probing inputs, one wishes to determine the
internal structure of the system. At this abstract level of
description, this appears to be a perfect candidate for the use
of systems identication and realisation theory techniques.
And indeed, there is substantial work being carried out,
using control theoretic techniques to approach such reverse
engineering problems. In the spirit of this paper, however,
let me point out some twists which often distinguish this
type of problem from more standard formulations of
systems identication.
One of the main features of biological problems is that it
is often extremely expensive, if not impossible, to apply
Syst. Biol., Vol. 1, No. 1, June 2004 15
arbitary test signals to the system. More often than not,
biologists will only be able to consider step inputs, or
perhaps a small combination of such inputs and ramps.
Although sufcient for local identication of linearised
models, this is certainly not enough information to
characterise the behaviour of a non-linear system. Thus,
an interesting question has to do with the amount of
experimentation needed in order to identify a system if only
a nitely-parameterised family of inputs is available; the
paper [50] looked into this problem and established that
the best possible answer, assuming exact measurements,
is 2r 1 experiments, where r is the number of parameters.
Moreover, in a precise mathematical sense, a generic set of
such experiments sufces. This resembles, but is technically
different in a fundamental way, from similar results for
embeddings of chaotic systems or for generic observability.
An even more important set of questions has to do with the
fact that often only steady-state measurements are available.
In mathematical terms, there is an unknown system _ xx =
f (x; p); where the vector of state variables evolves in some
manifold M and p is a vector of parameters, and for each
particular value of p in some neighbourhood of a default
value p
0
; one may have measurements of a steady state (p)
satisfying f ((p); p) = 0: In cellular networks, for instance,
the state vector might represent the concentrations of certain
proteins, mRNA, or other substances at different time
instants, the parameters might represent the concentration
levels of certain enzymes that are maintained at a constant
value during a particular experiment, and the experiment
measures concentrations after the system has relaxed to a
steady state. As discussed earlier, in cell biology applications
it is often the case that the equations dening the system
that is, the functions f
i
describing the vector eld are
unknown, even in general form, but one wishes nonetheless
to determine the interaction graph of the system, that is to
say, to know which variables directly inuence which
variables, as well as the relative strengths of these
interactions. Suppose the steady-state measurements induce
a smooth map from parameters to states, at least locally, so
that perturbation experiments, corresponding to the possible
vector elds Y in parameter space induce, by push-forward,
vector elds X =
+
(Y) on the state space, which in turn
determine a distribution T := spanX _ TM: Under suit-
able genericity conditions, dimT = n 1; and with the
main technical assumption described in [51], one can obtain
that df
i
T
l
(df
i
; T) = 0) for appropriate coordinates of
the motion, which determines each df
i
projectively. That is,
the interaction graph is completely characterised, and
strengths are identied up to a constant common multiple.
See [51] for details, as well as [52] for further results and [53]
for new problems in combinatories and computational
complexity also suggested by this problem.
3.2.3 Adaptive control of bifurcation para-
meters: Some biological systems are believed to
operate at a critical point between stability and instability,
which brings up the issue of how bifurcation parameters
may be automatically tuned to this critical value.
The papers [5456] were motivated by two such instances
from the literature: neural integration in the nervous
system, and hair cell oscillations in the auditory system. In
both examples, the question arises as to how the required
ne-tuning may be achieved and maintained in a robust
and reliable way. As with the other questions illustrated
here, this led to a new type of problem in control theory,
related but in fact different from other work on bifurcation
control in the literature. We formulated this question in the
language of adaptive control, and presented solutions in
some simple instances.
3.2.4 Robust stability from structure: Designers
strive to engineer as much robustness as possible into control
systems. However, few systems perform acceptably under
truly large variations in parameters. In biology, in contrast,
there is often a very large variability in intracellular
concentrations of chemicals, due to, for instance,
unequal division among daughter cells during mitosis, gene
duplications, or mutations. If functions critical to the survival
of the organism are not affected, this means that evolution
must have selected for appropriately robust structures. Their
study is of great theoretical interest, and may also potentially
suggest ideas for novel robust designs in engineering
applications.
One such type of problem arose during a fascinating
series of talks [57] given by a mathematical immunologist,
Carla Wofsy. Among other tasks, the immune system is
charged with the destruction and elimination of invading
organisms and of the toxic products that they produce, as
well as the destruction of virus-infected or mutated cells.
One of the most challenging problems in the study of the
immune system is to understand how it manages to
distinguish among self and other while still being able to
react fast. One approach to this question was proposed by
McKeithan [58], who suggested that a chain of modi-
cations of the T-cell a type of immune cell receptor
complex, via tyrosine phosphorylation and other reactions,
may give rise to both increased sensitivity and selectivity of
response. This process, which he called kinetic proofreading
because of its analogy to an older model proposed by
Hopeld for DNA error correction, was modelled in [58] for
T-cell receptor (TCR) and peptide-major histocompatibility
complex (MHC) interactions by means of a certain set of
non-linear ordinary differential equations. The steady states
that result are interpreted as a signal of antigen recognition.
Thus, it was natural to ask if the steady states of this system
are unique, if they are asymptotically stable, and so forth.
No standard technique from control theory seemed
appropriate for handling this example, but it soon became
apparent that the beautiful theory of chemical network
stability studied by Feinberg, Horn, and Jackson in the
1970s [59, 60] provides a rich approach to establishing
uniqueness and asymptotic stability of equilibria for the
system in [58] as well as, it turns out, many other types of
problems in receptor-ligand dynamics [26]. The theorems
that result are valid essentially for all parameters appearing
in the system description, as long as the interconnection
structure of the system satises certain rules. In [27], we
extended the results in [59, 60] to handle global stability as
well as to provide estimates of robustness to unmodelled
dynamics, and also gave solutions to a stabilisation problem
that one may formulate for such systems; an observer theory
was subsequently developed in work with Madalena Chaves
[61, 62]. In this case, a systems biology problem did not
suggest an entirely new theoretical development, but instead
led to bringing into control theory, tools from an area of
mathematical chemistry that had not attracted so much
previous attention in the control theory community.
4 Conclusions
I have argued, through a selection of examples, that new
problems in control theory arise most naturally when
looking at systems biology problems. In my experience, it
is necessary to listen carefully to the questions that our
biology colleagues are asking. Many of these questions may
Syst. Biol., Vol. 1, No. 1, June 2004 16
sound like routine questions in control theory, but they often
turn out to lead to challenging and interesting, and
ultimately highly rewarding, new directions of theoretical
research.
5 Acknowledgments
I wish to thank Patrick de Leenheer, Sarah Koskie, and
Robert Dinerstein for useful comments on a draft of this
paper, as well as all my collaborators, especially David
Angeli.
6 References
1 Heinrich, R., and Schuster, S.: The regulation of cellular systems
(Chapman & Hall, New York, 1996)
2 Schuster, S.: Analysis of structural robustness of metabolic networks,
Syst. Biol., 2004, 1
3 Savageau, M.A.: Biochemical systems analysis. a study of function and
design in molecular biology (Addison-Wesley, Reading, MA, 1976)
4 Hartwell, L.H., Hopeld, J.J., Leibler, S., and Murray, A.W.: From
molecular to modular cell biology, Nature, 1999, 402, pp. C47C52
5 Lauffenburger, D.A.: Cell signaling pathways as control modules:
Complexity for simplicity, Proc. Natl. Acad. Sci. USA, 2000, 97,
pp. 50315033
6 Hanahan, D., and Weinberg, R.A.: The hallmarks of cancer, Cell,
2000, 100, pp. 5770
7 Angeli, D., and Sontag, E.D.: Monotone control systems, IEEE Trans.
Autom. Control, 2003, 48, pp. 16841698
8 Angeli, D., and Sontag, E.D.: Multistability in monotone I/O systems,
Syst. Control Lett., 2004, 51, pp. 185202
9 Angeli, D., and Sontag, E.D.: Interconnections of monotone systems
with steady-state characteristics, in de Queiroz, M., Malisoff, M., and
Wolenski, P. (Eds.): Optimal control, stabilization, and nonsmooth
analysis (Springer-Verlag, Heidelberg, 2004), pp. 135154
10 Angeli, D., Ferrell, J.E., Jr., and Sontag, E.D.: Detection of multi-
stability, bifurcations, and hysteresis in a large class of biological
positive-feedback systems, Proc. Natl. Acad. Sci. USA, 2004, 101,
pp. 18221827
11 Huang, C.-Y.F., and Ferrell, J.E.: Ultrasensitivity in the mitogen-
activated protein kinase cascade, Proc. Natl. Acad. Sci. USA, 1996, 93,
pp. 1007810083
12 Lauffenburger, D.A.: A computational study of feedback effects on
signal dynamics in a mitogen-activated protein kinase (MAPK)
pathway model, Biotechnol. Prog., 2001, 17, pp. 227239
13 Widmann, C., Spencer, Q., Jarpe, M.B., and Johnson, G.L.: Mitogen-
activated protein kinase: conservation of a three-kinase module from
yeast to human, Physiol. Rev., 1999, 79, pp. 143180
14 Kholodenko, B.N.: Negative feedback and ultrasensitivity can bring
about oscillations in the mitogen-activated protein kinase cascades,
Eur. J. Biochem, 2000, 267, pp. 15831588
15 Shvartsman, S.Y., Wiley, H.S., and Lauffenburger, D.A.: Autocrine
loop as a module for bidirectional and context-dependent cell
signaling, preprint, MIT Chemical Engineering Department, Decem-
ber 2000
16 Sontag, E.D.: Asymptotic amplitudes and Cauchy gains: A small-gain
principle and an application to inhibitory biological feedback, Syst.
Control Lett., 2002, 47, pp. 167179
17 De Leenheer, P., Angeli, D., and Sontag, E.D.: Small-gain theorems
for predator-prey systems. Proc. 1st Multidisciplinary Int. Symp. on
Positive Systems: Theory and Applications, pp. 191198
18 De Leenheer, P., Angeli, D., and Sontag, E.D.: A feedback perspective
for chemostat models with crowding effects, in Benvenuti, L., De
Santis, A., and Farina, L. (Eds.): First Multidisciplinary International
Symposium on Positive Systems: Theory and Applications (Springer-
Verlag, Heidelberg, 2003), pp. 167174
19 Angeli, D., De Leenheer, P., and Sontag, E.D.: A small-gain theorem
for almost global convergence of monotone systems, Syst. Control
Lett., in press
20 Angeli, D., and Sontag, E.D.: An analysis of a circadian model using
the small-gain approach to monotone systems, arXiv math.OC/
0312062, Dec. 2003
21 Enciso, G., and Sontag, E.D.: On the stability of a model of
testosterone dynamics, J. Math. Biol., to appear
22 Hirsch, M.: Differential equations and convergence almost everywhere
in strongly monotone ows, Contemp. Math., 1983, 17, pp. 267285
23 Hirsch, M.: Systems of differential equations that are competitive or
cooperative II: Convergence almost everywhere, SIAM J. Math. Anal.,
1985, 16, pp. 423439
24 Smale, S.: On the differential equations of species in competition,
J. Math. Biol., 1976, 3, pp. 57
25 Smith, H.L.: Monotone dynamical systems: An introduction to the
theory of competitive and cooperative systems, Math. Surveys
Monographs, 1995, 41
26 Chaves, M., Sontag, E.D., and Dinerstein, R.J.: Steady-states of
receptor-ligand dynamics: A theoretical framework, J. Theor. Biol.,
2004, 227, pp. 413428
27 Sontag, E.D.: Structure and stability of certain chemical networks and
applications to the kinetic proofreading model of T-cell receptor
signal transduction, IEEE Trans. Autom. Control, 2001, 46,
pp. 10281047
28 De Leenheer, P., Angeli, D., and Sontag, E.D.: A tutorial on monotone
systems- with an application to chemical reaction networks. Proc. 16th
Int. Symp. on Mathematical Theory of Networks and Systems, Leuven,
Belgium, 2004, to appear
29 Noviek, A., and Wiener, M.: Enzyme induction as an all-
or-none phenomenon, Proc. Natl. Acad. Sci. USA, 1957, 43,
pp. 553566
30 Ptashne, M.: A Genetic Switch: Phage l and Higher organisms
(Cell Press and Blackwell Scientic Publications, Cambridge, MA,
1992)
31 Ferrell, J.E., Jr., and Machleder, E.M.: The biochemical basis of an all-
or-none cell fate switch in Xenopus ooeytes, Science, 1998, 280,
pp. 895898
32 Bhalla, U.S., Ram, P.T., and Iyengar, R.: MAP kinase phosphatase as a
locus of exibility in a mitogen-activatcd protein kinase signaling
network, Science, 2002, 297, pp. 10181023
33 Cross, F.R., Archambault, V., Miller, M., and Klovstad, M.: Testing a
mathematical model of the yeast cell cycle, Mol. Biol. Cell, 2002, 13,
pp. 5270
34 Pomerening, J.R., Sontag, E.D., and Ferrell, J.E., Jr.: Building a cell
cycle oscillator: hysteresis and bistability in the activation of Cdc2,
Nat. Cell Biol., 2003, 5, pp. 346351
35 Sha, W., Moore, J., Chen, K., Lassaletta, A.D., Yi, C.S., Tyson, J.J.,
and Sible, J.C.: Hysteresis drives cell-cycle transitions in
Xenopus laevis egg extracts, Proc. Natl. Acad. Sci. USA, 2003, 100,
pp. 975980
36 Bagowski, C.P., Besser, J., Frey, C.R., and Ferrell, J.E., Jr.: The JNK
Cascade as a biochemical switch in mammalian cells: ultrasensitive and
all-or-none responses, Curr. Biol., 2003, 13, pp. 315320
37 Lisman, J.E.: A mechanism for memory storage insensitive to
molecular turnover: a bistable autophosphorylating kinase, Proc.
Natl. Acad. Sci. USA, 1985, 82, pp. 30553057
38 Xiong, W., and Ferrell, J.E., Jr.: A positive-feedback-based bistable
memory module that governs a cell fate decision, Nature, 2003, 426,
pp. 460465
39 Reynolds, A.R., Tischer, C., Verveer, P.J., Rocks, O., and Bastiaens,
P.I.: EGFR activation coupled to inhibition of tyrosine phosphatases
causes lateral signal propagation, Nat. Cell Biol., 2003, 5, pp. 447453
40 Enciso, G., and Sontag, E.D.: Monotone systems under positive
feedback: multistability and a reduction theorem, Syst. Control Lett.,
to appear
41 Tyson, J.J., and Othmer, H.G.: The dynamics of feedback control
circuits in biochemical pathways, in Rosen, R., and Snell, F.M. (Eds.):
Progress in theoretical biology (Academic Press, New York, 1978),
vol. 5, pp. 162
42 Jones, D.S., and Sleeman, B.D.: Differential equations and
mathematical biology (George Allen, London, 1983)
43 Kishimoto, K., and Weinberger, H.F.: The spatial homogeneity of
stable equilibria of some reaction-diffusion systems on convex
domains, J. Differ. Equ., 1985, 58, pp. 1521
44 Casten, R., and Holland, C.: Stability properties of solutions to systems
of reaction-diffussion equations, SIAM J. Appl. Math., 1977, 33,
pp. 353364
45 Casten, R., and Holland, C.: Instability results for reaction diffusion
equations with Neumann boundary conditions, J. Differ. Equ., 1978,
27, pp. 266273
46 Levin, S.A.: Non-uniform stable solutions to reaction-diffusion
equations: applications to ecological pattern formation, in Haken, H.
(Ed.): Pattern formation by dynamic systems and pattern recognition
(Springer-Verlag, Berlin, 1979), pp. 210222
47 Yi, T.-M., Huang, Y., Simon, M.I., and Doyle, J.: Robust perfect
adaptation in bacterial chemotaxis through integral feedback control,
Proc. Natl. Acad. Sci. USA, 2000, 97, pp. 46494653
48 Iglesias, P.A., and Levchenko, A.: A general framework for achieving
integral control in chemotactic biological signaling mechanisms.
Proc. IEEE Conf. on Decision and Control, Orlando, USA, 2001,
pp. 843848
49 Sontag, E.D.: Adaptation and regulation with signal detection implies
internal model, Syst. Control Lett., 2003, 50, pp. 119126
50 Sontag, E.D.: For differential equations with r parameters, 2r 1
experiments are enough for identication, J. Nonlinear Sci., 2002, 12,
pp. 553583
51 Kholodenko, B.N., Kiyatkin, A., Bruggeman, F., Sontag, E.D.,
Westerhoff, H., and Hock, J.: Untangling the wires: a novel strategy
to trace functional interactions in signaling and gene networks, Proc.
Natl. Acad. Sci. USA, 2002, 99, pp. 1284112846
52 Sontag, E.D., Kiyatkin, A., and Kholodenko, B.N.: Inferring dynamic
architecture of cellular networks using time series of gene expression,
protein and metabolite data, Bioinformatics, 2004, in press
53 Berman, P., DasGupta, B., and Sontag, E.D.: Randomized approxi-
mation algorithms for set multicover problems with applications to
reverse engineering of protein and gene networks. Proc. 7th Int.
Workshop on Approximation Algorithms for Combinational Optimiz-
ation Problems, Cambridge, MA, Aug. 2004, to appear in Lec. Notes
Comp. Sci.
54 Moreau, L., and Sontag, E.D.: Balancing at the border of instability,
Phy. Rev. E, 2003, 68, pp. 020901(14)
55 Moreau, L., Arcak, M., and Sontag, E.D.: Feedback tuning of
bifurcations, Syst. Control Lett., 2003, 50, pp. 229239
Syst. Biol., Vol. 1, No. 1, June 2004 17
56 Moreau, L., Arcak, M., and Sontag, E.D.: How feedback can tune a
bifurcation parameter towards its unknown critical bifurcation value.
Presented at IEEE Conf. on Decision and Control, Maui, USA, 2003
57 Wofsy, C.: Modeling receptor aggregation. Presented at Workshop on
Mathematical Cellular Biology, Pacic Institute for the Mathematical
Sciences, UBC, Vancouver, August 1999
58 McKeithan, T.W.: Kinetic proofreading in T-cell receptor signal
transduction, Proc. Natl. Acad. Sci. USA, 1995, 92, pp. 50425046
59 Feinberg, M.: Chemical reaction network structure and the stabiliy of
complex isothermal reactors - I. The deciency zero and deciency
one theorems, Review Article 25, Chem. Eng. Sci., 1987, 42,
pp. 22292268
60 Horn, F.J.M., and Jackson, R.: General mass action kinetics, Arch.
Ration. Mech. Anal., 1972, 49, pp. 81116
61 Chaves, M., and Sontag, E.D.: State-Estimators for chemical reaction
networks of Feinberg-Horn-Jackson zero deciency type, Eur.
J. Control, 2002, 8, pp. 343359
62 Chaves, M., and Sontag, E.D.: Observers for certain chemical reaction
networks. Proc. European Control Conf., Porto, Portugal, 2001,
pp. 37153720
Syst. Biol., Vol. 1, No. 1, June 2004 18