Figures
Abstract
This paper addresses the problem of providing mathematical conditions that allow one to ensure that biological networks, such as transcriptional systems, can be globally entrained to external periodic inputs. Despite appearing obvious at first, this is by no means a generic property of nonlinear dynamical systems. Through the use of contraction theory, a powerful tool from dynamical systems theory, it is shown that certain systems driven by external periodic signals have the property that all their solutions converge to a fixed limit cycle. General results are proved, and the properties are verified in the specific cases of models of transcriptional systems as well as constructs of interest in synthetic biology. A self-contained exposition of all needed results is given in the paper.
Author Summary
The activities of living organisms are governed by complex sets of biochemical reactions. Often, entrainment to certain external signals helps control the timing and sequencing of reactions. An important open problem is to understand the onset of entrainment and under what conditions it can be ensured in the presence of uncertainties, noise, and environmental variations. In this paper, we focus mainly on transcriptional systems, modeled by Ordinary Differential Equations. These are basic building blocks for more complex biochemical systems. However, the results that we obtain are of more generality. To illustrate this generality, and to emphasize the use of our techniques in synthetic biology, we discuss the entrainment of a Repressilator circuit and the synchronization of a network of Repressilators. We answer the following two questions: 1) What are the dynamical mechanisms that ensure the entrainment to periodic inputs in transcriptional modules? 2) Starting from natural systems, what properties can be used to design novel synthetic biological circuits that can be entrained? For some biological systems which are always “in contact” with a continuously changing environment, entrainment may be a “desired” property. Thus, answering the above two questions is of fundamental importance. While entrainment may appear obvious at first thought, it is not a generic property of nonlinear dynamical systems. The main result of our paper shows that, even if the transcriptional modules are modeled by nonlinear ODEs, they can be entrained by any (positive) periodic signal. Surprisingly, such a property is preserved if the system parameters are varied: entrainment is obtained independently of the particular biochemical conditions. We prove that combinations of the above transcriptional module also show the same property. Finally, we show how the developed tools can be applied to design synthetic biochemical systems guaranteed to exhibit entrainment.
Citation: Russo G, di Bernardo M, Sontag ED (2010) Global Entrainment of Transcriptional Systems to Periodic Inputs. PLoS Comput Biol 6(4): e1000739. https://doi.org/10.1371/journal.pcbi.1000739
Editor: Herbert M. Sauro, University of Washington, United States of America
Received: July 6, 2009; Accepted: March 12, 2010; Published: April 15, 2010
Copyright: © 2010 Russo et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: GR and MdB acknowledge support from the European Union Project ‘Engineering Complexity in Biological Systems (COBIOS)’, VI Framework program (Grant no. 043379). EDS acknowledges support from grants NSF-DMS-0614371, NIH-1R01GM086881, and AFOSR-FA9550-08. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Periodic, clock-like rhythms pervade nature and regulate the function of all living organisms. For instance, circadian rhythms are regulated by an endogenous biological clock entrained by the light signals from the environment that then acts as a pacemaker [1]. Moreover, such an entrainment can be obtained even if daily variations are present, like e.g. temperature and light variations. Another important example of entrainment in biological systems is at the molecular level, where the synchronization of several cellular processes is regulated by the cell cycle [2].
An important question in mathematical and computational biology is that of finding conditions ensuring that entrainment occurs. The objective is to identify classes of biological systems that can be entrained by an exogenous signal. To solve this problem, modelers often resort to simulations in order to show the existence of periodic solutions in the system of interest. Simulations, however, can never prove that solutions will exist for all parameter values, and they are subject to numerical errors. Moreover, robustness of entrained solutions needs to be checked in the presence of noise and uncertainties, which cannot be avoided experimentally.
From a mathematical viewpoint, the problem of formally showing that entrainment takes place is known to be extremely difficult. Indeed, if a stable linear time-invariant model is used to represent the system of interest, then entrainment is usually expected, when the system is driven by an external periodic input, with the system response being a filtered, shifted version of the external driving signal. However, in general, as is often the case in biology, models are nonlinear. The response of nonlinear systems to periodic inputs is the subject of much current systems biology experimentation; for example, in [3], the case of a cell signaling system driven by a periodic square-wave input is considered. From measurements of a periodic output, the authors fit a transfer function to the system, implicitly modeling the system as linear even though (as stated in the Supplemental Materials to [3]) there are saturation effects so the true system is nonlinear. For nonlinear systems, driving the system by an external periodic signal does not guarantee the system response to also be a periodic solution, as nonlinear systems can exhibit harmonic generation or suppression and complex behavior such as chaos or quasi-periodic solutions [4]. This may happen even if the system is well-behaved with respect to constant inputs; for example, there are systems which converge to a fixed steady state no matter what is the input excitation, so long as this input signal is constant, yet respond chaotically to the simplest oscillatory input; we outline such an example in the Materials and Methods Section, see also [5]. Thus, a most interesting open problem is that of finding conditions for the entrainment to external inputs of biological systems modeled by sets of nonlinear differential equations.
One approach to analyzing the convergence behavior of nonlinear dynamical systems is to use Lyapunov functions. However, in biological applications, the appropriate Lyapunov functions are not always easy to find and, moreover, convergence is not guaranteed in general in the presence of noise and/or uncertainties. Also, such an approach can be hard to apply to the case of non-autonomous systems (that is, dynamical systems directly dependent on time), as is the case when dealing with periodically forced systems.
The above limitations can be overcome if the convergence problem is interpreted as a property of all trajectories, asking that all solutions converge towards one another (contraction). This is the viewpoint of contraction theory, [6], [7], and more generally incremental stability methods [8]. Global results are possible, and these are robust to noise, in the sense that, if a system satisfies a contraction property then trajectories remain bounded in the phase space [9]. Contraction theory has a long history. Contractions in metric functional spaces can be traced back to the work of Banach and Caccioppoli [10] and, in the field of dynamical systems, to [11] and even to [12] (see also [13], [8], and e.g. [14] for a more exhaustive list of related references). Contraction theory has been successfully applied to both nonlinear control and observer problems, [7], [15] and, more recently, to synchronization and consensus problems in complex networks [16], [17],[18]. In [19] it was proposed that contraction can be particularly useful when dealing with the analysis and characterization of biological networks. In particular, it was found that using non Euclidean norms, as also suggested in [6] (Sec. 3.7ii), can be particularly effective in this context [19], [20].
One of the objectives of this paper is to give a self-contained exposition, with all proofs included, of results in contraction theory as applied to entrainment of periodic signals, and, moreover, to show their applicability to problems of biological interest. We believe that contraction analysis should be recognized as an important component of the “toolkit” of systems biology, and this paper should be useful to other researchers contemplating the use of these tools.
For concreteness, we focus mainly on transcriptional systems, as well as related biochemical systems, which are basic building blocks for more complex biochemical systems. However, the results that we obtain are of more generality. To illustrate this generality, and to emphasize the use of our techniques in synthetic biology design, we discuss as well the entrainment of a Repressilator circuit in a parameter regime in which endogenous oscillations do not occur, as well as the synchronization of a network of Repressilators. A surprising fact is that, for these applications, and contrary to many engineering applications, norms other than Euclidean, and associated matrix measures, must be considered.
Mathematical tools
We consider in this paper systems of ordinary differential equations, generally time-dependent:(1)defined for
and
, where
is a subset of
. It will be assumed that
is differentiable on
, and that
, as well as the Jacobian of
with respect to
, denoted as
, are both continuous in
. In applications of the theory, it is often the case that
will be a closed set, for example given by non-negativity constraints on variables as well as linear equalities representing mass-conservation laws. For a non-open set
, differentiability in
means that the vector field
can be extended as a differentiable function to some open set which includes
, and the continuity hypotheses with respect to
hold on this open set.
We denote by the value of the solution
at time
of the differential equation (1) with initial value
. It is implicit in the notation that
(“forward invariance” of the state set
). This solution is in principle defined only on some interval
, but we will assume that
is defined for all
. Conditions which guarantee such a “forward-completeness” property are often satisfied in biological applications, for example whenever the set
is closed and bounded, or whenever the vector field
is bounded. (See Appendix C in [21] for more discussion, as well as [22] for a characterization of the forward completeness property.) Under the stated assumptions, the function
is jointly differentiable in all its arguments (this is a standard fact on well-posedness of differential equations, see for example Appendix C in [21]).
We recall (see for instance [23]) that, given a vector norm on Euclidean space (), with its induced matrix norm
, the associated matrix measure
is defined as the directional derivative of the matrix norm, that is,
For example, if
is the standard Euclidean 2-norm, then
is the maximum eigenvalue of the symmetric part of
. As we shall see, however, different norms will be useful for our applications. Matrix measures are also known as “logarithmic norms”, a concept independently introduced by Germund Dahlquist and Sergei Lozinskii in 1959, [24],[25]. The limit is known to exist, and the convergence is monotonic, see [24], [26].
We will say that system (1) is infinitesimally contracting on a convex set if there exists some norm in
, with associated matrix measure
such that, for some constant
,
(2)
Let us discuss informally (rigorous proofs are given later) the motivation for this concept. Since by assumption is continuously differentiable, the following exact differential relation can be obtained from (1):
(3)where, as before,
denotes the Jacobian of the vector field
, as a function of
and
, and where
denotes a small change in states and “
” means
, evaluated along a trajectory. (In mechanics, as in [27],
is called “virtual displacement”, and formally it may be thought of as a linear tangent differential form, differentiable with respect to time.) Consider now two neighboring trajectories of (1), evolving in
, and the virtual displacements between them. Note that (3) can be thought of as a linear time-varying dynamical system of the form:
once that
is thought of as a fixed function of time. Hence, an upper bound for the magnitude of its solutions can be obtained by means of the Coppel inequality [28], yielding:
(4)where
is the matrix measure of the system Jacobian induced by the norm being considered on the states and
. Using (4) and (2), we have that
Thus, trajectories starting from infinitesimally close initial conditions converge exponentially towards each other. In what follows we will refer to
as contraction (or convergence) rate.
The key theoretical result about contracting systems links infinitesimal and global contractivity, and is stated below. This result can be traced, under different technical assumptions, to e.g. [6], [13], [12], [11].
Theorem 1. Suppose that is a convex subset of
and that
is infinitesimally contracting with contraction rate
. Then, for every two solutions
and
of (1), it holds that:
(5)
In other words, infinitesimal contractivity implies global contractivity. In the Materials and Methods section, we provide a self-contained proof of Theorem 1. In fact, the result is shown there in a generalized form, in which convexity is replaced by a weaker constraint on the geometry of the space.
In actual applications, often one is given a system which depends implicitly on the time, , by means of a continuous function
, i.e. systems dynamics are represented by
. In this case,
(where
is some subset of
), represents an external input. It is important to observe that the contractivity property does not require any prior information about this external input. In fact, since
does not depend on the system state variables, when checking the property, it may be viewed as a constant parameter,
. Thus, if contractivity of
holds uniformly
, then it will also hold for
.
Given a number , we will say that system (1) is
-periodic if it holds that
Notice that the system
is
-periodic, if the external input,
, is itself a periodic function of period
.
The following is the basic theoretical result about periodic orbits that will be used in the paper. A proof may be found in [6], Sec. 3.7.vi.
Theorem 2. Suppose that:
is a closed convex subset of
;
is infinitesimally contracting with contraction rate
;
is
-periodic.
Then, there is a unique periodic solution of (1) of period
and, for every solution
, it holds that
as
.
In the Materials and Methods section of this paper, we provide a self-contained proof of Theorem 2, in a generalized form which does not require convexity.
A simple example
As a first example to illustrate the application of the concepts introduced so far, we choose a simple bimolecular reaction, in which a molecule of and one of
can reversibly combine to produce a molecule of
.
This system can be modeled by the following set of differential equations:(6)where we are using
to denote the concentration of
and so forth. The system evolves in the positive orthant of
. Solutions satisfy (stoichiometry) constraints:
(7)for some constants
and
.
We will assume that one or both of the “kinetic constants” are time-varying, with period
. Such a situation arises when the
's depend on concentrations of additional enzymes, which are available in large amounts compared to the concentrations of
, but whose concentrations are periodically varying. The only assumption will be that
and
for all
.
Because of the conservation laws (7), we may restrict our study to the equation for . Once that all solutions of this equation are shown to globally converge to a periodic orbit, the same will follow for
and
. We have that:
(8)Because
and
, this system is studied on the subset of
defined by
. The equation can be rewritten as:
(9)Differentiation with respect to
of the right-hand side in the above system yields this (
) Jacobian:
(10)Since we know that
and
, it follows that
for
. Using any norm (this example is in dimension one) we have that
. So (6) is contracting and, by means of Theorem 2, solutions will globally converge to a unique solution of period
(notice that such a solution depends on system parameters).
Figure 1 shows the behavior of the dynamical system (9), using two different values of . Notice that the asymptotic behavior of the system depends on the particular choice of the biochemical parameters being used. Furthermore, it is worth noticing here that the higher the value of
, the faster will be the convergence to the attractor.
Time (minutes) on the -axis. The Figure shows the behavior of (9) for
(blue),
(green),
(red). Notice that an increase of
, causes an increase of the contraction rate, hence trajectories converge faster to the system unique periodic attractor. The other system parameters are set to:
,
Results
Mathematical model and problem statement
We study a general externally-driven transcriptional module. We assume that the rate of production of a transcription factor is proportional to the value of a time dependent input function
, and
is subject to degradation and/or dilution at a linear rate. (Later, we generalize the model to also allow nonlinear degradation as well.) The signal
might be an external input, or it might represent the concentration of an enzyme or of a second messenger that activates
. In turn,
drives a downstream transcriptional module by binding to a promoter (or substrate), denoted by
with concentration
. The binding reaction of
with
is reversible and given by:
where
is the complex protein-promoter, and the binding and dissociation rates are
and
respectively. As the promoter is not subject to decay, its total concentration,
, is conserved, so that the following conservation relation holds:
(11)We wish to study the behavior of solutions of the system that couples
and
, and specifically to show that, when the input
is periodic with period
, this coupled system has the property that all solutions converge to some globally attracting limit cycle whose period is also
.
Such transcriptional modules are ubiquitous in biology, natural as well as synthetic, and their behavior was recently studied in [29] in the context of “retroactivity” (impedance or load) effects. If we think of as the concentration of a protein
that is a transcription factor for
, and we ignore fast mRNA dynamics, such a system can be schematically represented as in Figure 2, which is adapted from [29]. Notice that
here does not need to be the concentration of a transcriptional activator of
for our results to hold. The results will be valid for any mathematical model for the concentrations,
, of
and
, of
(the concentration of
is conserved) of the form:
(12)
As explained in [29], the transcriptional component takes as input the concentration of protein and gives as output the concentration of protein
. The downstream transcriptional module takes as input the concentration of protein
.
An objective in this paper is, thus, to show that, when is a periodic input, all solutions of system (12) converge to a (unique) limit cycle (Figure 3). The key tool in this analysis is to show that uniform contractivity holds. Since in this example the input appears additively, uniform contractivity is simply the requirement that the unforced system (
) is contractive. Thus, the main step will be to establish the following technical result, see the Material and Methods:
Time in minutes on the -axis. The state of the system (green),
, is entrained to both
and to a repeating
sequence. System parameters are set to:
,
= 1,
.
Proposition 1. The systemwhere
(13)for all
, and
,
,
, and
are arbitrary positive constants, is contracting.
Appealing to Theorem 2, we then have the following immediate Corollary:
Proposition 2. For any given nonnegative periodic input of period
, all solutions of system (12) converge exponentially to a periodic solution of period
.
In the following sections, we introduce a matrix measure that will help establish contractivity, and we prove Proposition 1. We will also discuss several extensions of this result, allowing the consideration of multiple driven subsystems as well as more general nonlinear systems with a similar structure. (A graphical algorithm to prove contraction of generic networks of nonlinear systems can also be found in [18] where this transcriptional module is also studied.)
Proof of Proposition 1
We will use Theorem 2. The Jacobian matrix to be studied is:(14)As matrix measure, we will use the measure
induced by the vector norm
, where
is a suitable nonsingular matrix. More specifically, we will pick
diagonal:
(15)where
and
are two positive numbers to be appropriately chosen depending on the parameters defining the system.
It follows from general facts about matrix measures that(16)where
is the measure associated to the
norm and is explicitly given by the following formula:
(17)Observe that, if the entries of
are negative, then asking that
amounts to a column diagonal dominance condition. (The above formula is for real matrices. If complex matrices would be considered, then the term
should be replaced by its real part
.)
Thus, the first step in computing is to calculate
:
(18)Using (17), we obtain:
(19)Note that we are not interested in calculating the exact value for the above measure, but just in ensuring that it is negative. To guarantee that
, the following two conditions must hold:
(20)
(21)Thus, the problem becomes that of checking if there exists an appropriate range of values for
,
that satisfy (20) and (21) simultaneously.
The left hand side of (21) can be written as:(22)which is negative if and only if
. In particular, in this case we have:
The idea is now to ensure negativity of (20) by using appropriate values for
and
which fulfill the above constraint. Recall that the term
because of the choice of the state space (this quantity represents a concentration). Thus, the left hand side of (20) becomes
(23)The next step is to choose appropriately
and
(without violating the constraint
). Imposing
,
, (23) becomes
(24)Then, we have to choose an appropriate value for
in order to make the above quantity uniformly negative. In particular, (24) is uniformly negative if and only if
(25)We can now choose
with
. In this case, (24) becomes
Thus, choosing
and
, with
, we have
. Furthermore, the contraction rate
, is given by:
Notice that
depends on both system parameters and on the elements
,
, i.e. it depends on the particular metric chosen to prove contraction. This completes the proof of the Proposition.
Generalizations
In this Section, we discuss various generalizations that use the same proof technique.
Assuming 
activation by enzyme kinetics.
The previous model assumed that was created in proportion to the amount of external signal
. While this may be a natural assumption if
is a transcription factor that controls the expression of
, a different model applies if, instead, the “active” form
is obtained from an “inactive” form
, for example through a phosphorylation reaction which is catalyzed by a kinase whose abundance is represented by
. Suppose that
can also be constitutively deactivated. Thus, the complete system of reactions consists of
together with
where the forward reaction depends on
. Since the concentrations of
must remain constant, let us say at a value
, we eliminate
and have:
(26)
We will prove that if is periodic and positive, i.e.
, then a globally attracting limit cycle exists. Namely, it will be shown, after having performed a linear coordinate transformation, that there exists a negative matrix measure for the system of interest.
Consider, indeed, the following change of the state variables:(27)The system dynamics then becomes:
(28)As matrix measure, we will now use the measure
induced by the vector norm
. (Notice that this time, the matrix
is the identity matrix).
Given a real matrix , the matrix measure
is explicitly given by the following formula (see e.g. [23]):
(29)(Observe that this is a row-dominance condition, in contrast to the dual column-dominance condition used for
.)
Differentiation of (28) yields the Jacobian matrix:Thus, it immediately follow from (29) that
is negative if and only if:
(30)
(31)The first inequality is clearly satisfied since by hypotheses both system parameters and the periodic input
are positive. In particular, we have:
By using (27) (recall that
), the right hand side of the second inequality can be written as:
Since all system parameters are positive and
, the above quantity is negative and upper bounded by
.
Thus, we have that , where:
The contraction property for the system is then proved. By means of Theorem 2, we can then conclude that the system can be entrained by any periodic input.
Simulation results are presented in Figure 4, where the presence of a stable limit cycle having the same period as is shown.
Time in minutes on the -axis. The system state (green),
, is entrained to the periodic input (blue):
. The zoom on
min highlights that trajectories starting from different initial conditions converge towards the attracting limit cycle. System parameters are set to:
,
,
,
,
.
Multiple driven systems.
We may also treat the case in which the species regulates multiple downstream transcriptional modules which act independently from each other, as shown in Figure 5. The biochemical parameters defining the different downstream modules may be different from each other, representing a situation in which the transcription factor
regulates different species. After proving a general result on oscillations, and assuming that parameters satisfy the retroactivity estimates discussed in [29], one may in this fashion design a single input-multi output module in which e.g. the outputs are periodic functions with different mean values, settling times, and so forth.
A schematic diagram of the transcriptional modules given in (12).
We denote by the various promoters, and use
to denote the concentrations of the respective promoters complexed with
. The resulting mathematical model becomes:
(32)
We consider the corresponding system with no input first, assuming that the states satisfy and
for all
.
Our generalization can be stated as follows:
Proposition 3. System (32) with no input (i.e. ) is contracting. Hence, if
is a non-zero periodic input, its solutions exponentially converge towards a periodic orbit of the same period as
.
Proof. We only outline the proof, since it is similar to the proof of Proposition 2. We employ the following matrix measure:(33)where
(34)and the scalars
have to be chosen appropriately (
).
Hence, the inequalities to be satisfied are:
(37)and
(38)
Clearly, the set of inequalities above admits a solution. Indeed, the left hand side of (38) can be recast aswhich is negative definite if and only if
for all
. Specifically, in this case we have
Also, from (37), as
for all
, we have that (37) can be rewritten as:
Since
, we can impose
(with
) and the above inequality becomes
Clearly, such inequality is satisfied if we choose
sufficiently small; namely:
Following a similar derivation to that of the previous Section, we can choose
with
. In this case, we have:
Thus,
, where
The second part of the Proposition is then proved by applying Theorem 2.
In Figure 6 the behavior of two-driven downstream transcriptional modules is shown. Notice that both the downstream modules are entrained by the periodic input , but their steady state behavior is different.
Time in minutes on the -axis. Outputs
(top) and
(bottom) of two transcriptional modules driven by the external periodic input
. The parameters are set to:
,
,
,
for module
and
,
,
for module
.
Notice that, by the same arguments used above, it can be proven that(39)is contracting.
Transcriptional cascades.
A cascade of (infinitesimally) contracting systems is also (infinitesimally) contracting [6], [30] (see Materials and Methods for an alternative proof). This implies that any transcriptional cascade, will also give rise to a contracting system, and, in particular, will entrain to periodic inputs. By a transcriptional cascade we mean a system as shown in Figure 7. In this figure, we interpret the intermediate variables as transcription factors, making the simplifying assumption that TF concentration is proportional to active promoter for the corresponding gene. (More complex models, incorporating transcription, translation, and post-translational modifications could themselves, in turn, be modeled as cascades of contracting systems.)
Each box contains the transcriptional module described by (12).
More abstract systems.
We can extend our results even further, to a larger class of nonlinear systems, as long as the same general structure is present. This can be useful for example to design new synthetic transcription modules or to analyze the entrainment properties of general biological systems. We start with a discussion of a two dimensional system of the form:(40)In molecular biology,
would typically represent a nonlinear degradation, for instance in Michaelis-Menten form, while the function
represents the interaction between
and
. The aim of this Section is to find conditions on the degradation and interaction terms that allow one to show contractivity of the unforced (no input
) system, and hence existence of globally attracting limit cycles.
We assume that the state space is compact (closed and bounded) as well as convex. Since the input appears additively, we must prove contractivity of the unforced system.
Theorem 3. System (40), without inputs , evolving on a convex compact subset of phase space is contracting, provided that the following conditions are all satisfied, for each
:
;
;
does not change sign;
.
Notice that the last condition is automatically satisfied if , because
.
As before, we prove contraction by constructing an appropriate negative measure for the Jacobian of the vector field. In this case, the Jacobian matrix is:(41)Once again, as matrix measure we will use:
(42)with
(43)and
appropriately chosen.
Using (42) we have(44)Following the same steps as the proof of Proposition 1, we have to show that:
(45)
(46)
Clearly, if for every
and
, the first inequality is satisfied, with
To prove the theorem we need to show that there exists
and
satisfying (46). For such inequality, since
does not change sign in
by hypothesis, we have two possibilities:
,
;
,
.
In the first case, the right hand side of (46) becomes(47)Choosing
, with
, we have:
Specifically, if we now pick
where
and
, we have that the above quantity is uniformly negative definite, i.e.
In the second case, the right hand side of (46) becomes
(48)Again, by choosing
, with
, we have the following upper bound for the expression in (48):
(49)Thus, it follows that
provided that the above quantity is uniformly negative definite. Since, by hypotheses,
(50)then
. The proof of the Theorem is now complete.
From a biological viewpoint, the hardest hypothesis to satisfy in Theorem 3 might be that on the derivatives of . However, it is possible to relax the hypothesis on
if the rate of change of
with respect to
, i.e.
, is sufficiently larger than
. In particular, the following result can be proved.
Theorem 4. System (40), without inputs , evolving on a convex compact set, is contractive provided that:
,
;
,
;
.
Proof. The proof is similar to that of Theorem 3. In particular, we can repeat the same derivation to obtain again inequality (46). Thence, as no hypothesis is made on the sign of , choosing
we have
(51)Thus, it follows that, if
, then
such that
, implying contractivity. The above condition is satisfied by hypotheses, hence the theorem is proved.
Applications to synthetic biology
We introduced above a methodology for checking if a given transcriptional module can be entrained to some periodic input. The aim of this section is to show that our methodology can serve as an effective tool for designing synthetic biological circuits that are entrained to some desired external input.
In particular, we will consider the synthetic biological oscillator known as the Repressilator [31], for which an additional coupling module has been recently proposed in [32]. A numerical investigation of the synchronization of a network of non-identical Repressilators was independently reported in [33].
We will show that our results can be used to isolate a set of biochemical parameters for which one can guarantee the entrainment to any external periodic signal of this synthetic biological circuit. In what follows, we will use the equations presented in [32] to model the Repressilator and the additional coupling model.
Entrainment using an intra-cellular auto-inducer.
The Repressilator is a synthetic biological circuit that consists of three genes that inhibit each other in a cyclic way [31]. As shown in Figure 8, gene lacI (associated to the state variable in the model) expresses protein LacI (
), which inhibits the transcription of gene tetR (
). This translates into protein TetR (
), which inhibits transcription of gene cI (
). Finally, the protein CI (
), translated from cI, inhibits expression of lacI, completing the cycle.
A schematic representation of the three-genes Repressilator circuit.
In Figure 9 a modular addition to the three-genes circuit is presented. The module was first presented in [32] and makes the Repressilator circuit sensitive to the concentration of the auto-inducer (labeled as in the model) which is a small molecule that can pass through the cell membrane. Specifically, the module makes use of two proteins: (i) LuxI, which synthesizes the auto-inducer; (ii) LuxR, with which the auto-inducer synthesized by LuxI forms a complex that activates the transcription of various genes.
This module is used for forcing the original circuit with some external signal (represented by an extra-cellular molecule in the bottom panel).
We model the above circuit with the simplified set of differential equations proposed in [32]. Specifically, the dynamics of the mRNAs are(53)
Notice that the above equations are dimensionless. This is done by: (i) measuring time in units of mRNA lifetime (which is assumed equal for the three genes), and (ii) expressing the protein levels in units of their Michaelis constant. The parameter represents the dimensionless transcription rate in the absence of self-repression, while
denotes the maximum contribution of the auto-inducer to the expression of lacI.
The dynamics of the proteins are described by(54)The parameters
,
,
represent the ratios between the mRNAs and the respective proteins' lifetimes and
,
,
represent the protein decay rate.
The last differential equation of the model from [32] keeps track of the evolution of the intra-cellular auto-inducer. It is assumed that the proteins TetR and LuxI have equal lifetimes. This in turn implies that the dynamics of such proteins are identical, and hence one uses the same variable to describe both protein concentrations. Thus, the dynamics of the auto-inducer are given by:where
is the rate of degradation of
.
We now model the forcing on the intracellular auto-inducer concentration by adding an external input to the above dynamical equation. The equation for
becomes:
(55)where
can be thought as a diffusion rate.
We will now use the analytical methodology developed in the previous sections, to properly tune the biochemical parameters of the Repressilator circuit, whose mathematical model consists of the set of differential equations (53), (54), (55), so that it shows entrainment to the periodic input . That is, the measured output (e.g.
), oscillates asymptotically with a period equal to that of
. Of course, the periodic orbit of the output will depend on the particular choice of the parameters.
In what follows, we assume that all the system parameters can be varied except for the self-degradations that we assume to be fixed as, in practice, they are difficult to modify.
In this case, the Jacobian matrix to be studied is(56)
The matrix measure that we will use to prove contraction iswhere
is a
diagonal matrix having on the main diagonal the positive arbitrary scalars
. Computation of
yields
(57)Thus, from the definition of
given in (29), we have that there exists some
such that
,
if and only if there exists a set of scalars
,
, such that
(58a)
(58b)
(58c)
(58d)
(58e)
(58f)
(58g)
It is easy to check that the nonlinear terms in the above equations satisfy the following inequalities:and
for all
. Hence, the system of inequalities (58a)–(58g) are satisfied, if the following set is fulfilled:
(59a)
(59b)
(59c)
(59d)
(59e)
(59f)
(59g)
The system can then be proved to be contracting for a given set of biochemical parameters, if there exists a set of scalars ,
satisfying the above inequalities. For example, if the repressilator parameters are chosen so that
(60)then it is trivial to prove that, for any constant value
, the set of scalars
, for
, satisfies (59a)–(59g). Indeed, in Figure 10 we provide a set of biochemical parameters for which the circuit is contracting and shows entrainment to the periodic input
. (These parameters, except for the maximal transcription rate
, are in the same ranges as those used in [31], [32]. These parameters are also close to those used in [33] and [34]. The reason for picking an
much smaller than in [32], is that we need to slow down transcription so as to eliminate intrinsic oscillations; otherwise the entrainment effect cannot be shown. This lowering of
by two orders of magnitude is also found in other works, for example in [35], where the same model is studied, with
somewhat larger but of the same order of magnitude as here.)
Time (minutes) on the -axis. Behavior of
when the input
is applied. Notice that when no forcing is present
converges to a non oscillatory regime behavior. System parameters are tuned in order to satisfy (72). Specifically:
,
,
,
,
,
,
.
Note that using the set of inequalities (59a)–(59g) as a guideline, it is possible to find other parameter regions where the system is still contracting but exhibit some other desired properties. For instance, to tune (e.g. increase) the amplitude of the output oscillations shown in Figure 10, a possible approach can be that of increasing the biochemical parameter so as to make stronger the effect of the auto-inducer on the dynamics of the gene
(variable
in the model).
Again we can prove that the set of inequalities (59a)–(59g) is satisfied for arbitrarily large, if we set
, for
and choose
such that
and
Now, due to biochemical constraints the parameter
is considerably smaller than
and
(in our simulations the ratio is of about two orders of magnitude). Therefore, whatever the value of
, it suffices to set
and
, with
being a positive arbitrary constant, to get a solution to (59a)–(59g) and hence guarantee the system to be contracting.
Figure 11 shows the behavior of the system output with the modified parameters confirming that with this choice of parameters the oscillation amplitude is indeed larger as expected.
Time (minutes) on the -axis. Behavior of
when: (i) the input
is applied; (ii) no forcing is present. System parameters are the same as that used in Figure 10, except
.
Observe the nonlinear character of the oscillation depicted in Figure 11, which is reflected in the lack of symmetry in the behavior at minima and maxima of cI. Our theory predicts the existence (and uniqueness) of such a nonlinear oscillations. None of the usual techniques, based on linear analysis, can explain such behavior.
Entrainment using an extra-cellular auto-inducer.
We now consider the case in which the extracellular auto-inducer can change due to an external signal as well as diffusion from intracellular auto-inducer, as represented in Figure 9. A new variable must be introduced, to keep track of the extracellular auto-inducer concentration. The only difference in the new model with respect to the previous one is that the differential equation for becomes:
(61)Notice that the parameter
measures the diffusion rate of the auto-inducer across the cell membrane, i.e.
, with
representing the membrane permeability,
its surface area and
the cell volume. In the above equation,
denotes the concentration of the extra-cellular auto-inducer, whose dynamics are given by:
(62)where
, with
denoting the total extracellular volume, while
stands for the decay rate.
In analogy with the previous section, we will ensure entrainment of the dynamical system consisting of (53), (54), (61), (62), by tuning the biochemical parameters of this new circuit. Again, the guidelines for engineering the parameters will be provided by the tools developed in the previous sections.
Following the schematic of the previous section, we will prove that there exists and a
constant diagonal matrix
, such that
, where
is the system Jacobian.
If we denote with ,
the diagonal elements of
, we obtain the following block-structure for the matrix
:
(63)where
is given in (57) and:
(64)
Thus, we have that if and only if there exist some
,
such that inequalities (58a)–(58f) are all satisfied and additionally:
(65a)
(65b)
Again, we can find sets of biochemical parameters in order to satisfy the above inequalities and hence ensure global entrainment of the circuit to some external input. For example, if we set(66)then, as in the previous section, it is trivial to show that setting all
to the same identical value satisfies the set of inequality required to prove contraction and hence guarantees entrainment. Notice that the last constraint in (66) is automatically satisfied by the physical (i.e. positivity) constraints on the system parameters.
In Figure 12, the behavior of the circuit is shown with the parameters chosen so as to satisfy the constraints given in (66).
Time (minutes) on the -axis. Behavior of
when the input
is applied. Notice that when no forcing is present, the steady state behavior is non-oscillatory. System parameters are:
,
,
,
,
,
.
Entraining a population of Repressilators.
Consider, now, a population of Repressilator circuits, which are coupled by means of an auto-inducer molecule. We can think of such a network as having an all-to-all topology, with the coupling given by the concentration of the extracellular auto-inducer,
. The aim of this section is to show that the methodology proposed here can also be used as an effective tool to guarantee the synchronization of an entire population of biochemical oscillators onto some entraining external periodic input.
We denote with the subscript the state variables of the
-th circuit in the network, which is modelled using the equations reported in [32] as:
(67)
Figure 13 shows a simulation of a population of Repressilators modeled as in (67), with biochemical parameters tuned as in the previous Section: all the circuits composing the network evolve asymptotically towards the same synchronous evolution, which has period equal to that of the input signal . The interested reader is referred to the Materials and Methods for the proof.
Behavior of a population of Repressilator modeled as in (80). Time (minutes) on -axs. Notice that all the circuits synchronize with a steady-state evolution having the same period as
. System parameters are chosen as in Figure 11, with
.
Materials and Methods
All simulations are performed in MATLAB (Simulink), Version 7.4, with variable step ODE solver ODE23t. Simulink models are available upon request. The proofs of the results are as follows.
-reachable sets
We will make use of the following definition:
Definition 1. Let be any positive real number. A subset
is
-reachable if, for any two points
and
in
there is some continuously differentiable curve
such that:
,
and
,
.
For convex sets , we may pick
, so
and we can take
. Thus, convex sets are
-reachable, and it is easy to show that the converse holds as well.
Notice that a set is
-reachable for some
if and only if the length of the geodesic (smooth) path (parametrized by arc length), connecting any two points
and
in
, is bounded by some multiple
of the Euclidean norm,
. Indeed, re-parametrizing to a path
defined on
, we have:
Since in finite dimensional spaces all the norms are equivalent, then it is possible to obtain a suitable
for Definition 1.
Remark 1. The notion of -reachable set is weaker than that of convex set. Nonetheless, in Theorem 5, we will prove that trajectories of a smooth system, evolving on a
-reachable set, converge towards each other, even if
is not convex. This additional generality allows one to establish contracting behavior for systems evolving on phase spaces exhibiting “obstacles”, as are frequently encountered in path-planning problems, for example. A mathematical example of a set with obstacles follows.
Example 1. Consider the two dimensional set, , defined by the following constraints:
Clearly,
is a non-convex subset of
. We claim that
is
-reachable, for any positive real number
. Indeed, given any two points
and
in
, there are two possibilities: either the segment connecting
and
is in
, or it intersects the unit circle. In the first case, we can simply pick the segment as a curve (
). In the second case, one can consider a straight segment that is modified by taking the shortest perimeter route around the circle; the length of the perimeter path is at most
times the length of the omitted segment. (In order to obtain a differentiable, instead of merely a piecewise-differentiable, path, an arbitrarily small increase in
is needed.)
Proof of Theorem 1
We now prove the main result on contracting systems, i.e. Theorem 1, under the hypotheses that the set , i.e. the set on which the system evolves, is
-reachable.
Theorem 5. Suppose that is a
-reachable subset of
and that
is infinitesimally contracting with contraction rate
. Then, for every two solutions
and
it holds that:
(68)
Proof. Given any two points and
in
, pick a smooth curve
, such that
and
. Let
, that is, the solution of system (1) rooted in
,
. Since
and
are continuously differentiable, also
is continuously differentiable in both arguments. We define
It follows that
Now,
so, we have:
(69)where
. Using Coppel's inequality [28], yields
(70)
,
, and
. Notice the Fundamental Theorem of Calculus, we can write
Hence, we obtain
Now, using (70), the above inequality becomes:
The Theorem is then proved.
Proof of Theorem 1. The proof follows trivially from Theorem 5, after having noticed that in the convex case, we may assume .
Proof of Theorem 2
In this Section we assume that the vector field is
-periodic and prove Theorem 2.
Before starting with the proof of Theorem 2 we make the following:
Remark 2. Periodicity implies that the initial time is only relevant modulo . More precisely:
(71)Indeed, let
,
, and consider the function
, for
. So,
where the last equality follows by
-periodicity of
. Since
, it follows by uniqueness of solutions that
, which is (71). As a corollary, we also have that
(72)where the first equality follows from the semigroup property of solutions (see e.g. [21]), and the second one from (71) applied to
instead of
.
Define nowwhere
. The following Lemma will be useful in what follows.
Lemma 1. for all
and
.
Proof. We will prove the Lemma by recursion. In particular, the statement is true by definition when . Inductively, assuming it true for
, we have:
as wanted.
Theorem 6. Suppose that:
is a closed
-reachable subset of
;
is infinitesimally contracting with contraction rate
;
is
-periodic;
.
Then, there is an unique periodic solution of (1) having period
. Furthermore, every solution
, such that
, converges to
, i.e.
as
.
Proof. Observe that is a contraction with factor
:
for all
, as a consequence of Theorem 5. The set
is a closed subset of
and hence complete as a metric space with respect to the distance induced by the norm being considered. Thus, by the contraction mapping theorem, there is a (unique) fixed point
of
. Let
. Since
,
is a periodic orbit of period
. Moreover, again by Theorem 5, we have that
. Uniqueness is clear, since two different periodic orbits would be disjoint compact subsets, and hence at positive distance from each other, contradicting convergence. This completes the proof.
Proof of Theorem 2. It will suffice to note that the assumption in Theorem 6 is automatically satisfied when the set
is convex (i.e.
) and the system is infinitesimally contracting.
Notice that, even in the non-convex case, the assumption can be ignored, if we are willing to assert only the existence (and global convergence to) a unique periodic orbit, with some period
for some integer
. Indeed, the vector field is also
-periodic for any integer
. Picking
large enough so that
, we have the conclusion that such an orbit exists, applying Theorem 6.
Cascades
In order to show that cascades of contracting systems remain contracting, it is enough to show this, inductively, for a cascade of two systems.
Consider a system of the following form:where
and
for all
(
and
are two
-reachable sets). We write the Jacobian of
with respect to
as
, the Jacobian of
with respect to
as
, and the Jacobian of
with respect to
as
,
We assume the following:
- The system
is infinitesimally contracting with respect to some norm (generally indicated as
), with some contraction rate
, that is,
for all
and all
, where
is the matrix measure associated to
.
- The system
is infinitesimally contracting with respect to some norm (which is, in general different from
, and is denoted by
), with contraction rate
, when
is viewed a a parameter in the second system, that is,
for all
,
and all
, where
is the matrix measure associated to
.
- The mixed Jacobian
is bounded:
, for all
,
and all
, for some real number
, where “
” is the operator norm induced by
and
on linear operators
. (All norms in Euclidean space being equivalent, this can be verified in any norm.)
We claim that, under these assumptions, the complete system is infinitesimally contracting. More precisely, pick any two positive numbers and
such that
and let
We will show that
, where
is the full Jacobian:
(73)with respect to the matrix measure
induced by the following norm in
:
Since
for all
and
, we have that, for all
and
:
where from now on we drop subscripts for norms. Pick now any
and a unit vector
(which depends on
) such that
. Such a vector
exists by the definition of induced matrix norm, and we note that
, by the definition of the norm in the product space. Therefore:
where the last inequality is a consequence of the fact that
for any nonnegative numbers with
(convex combination of the
's). Now taking limits as
, we conclude that
as desired.
Entraining a population of Repressilators: proof
The general principle that we apply to prove entrainment of a population of Repressilators is as follows.
Assume that the cascade system(74)with
being an exogenous input, satisfies the contractivity assumptions of the above Section. Then, consider the interconnection of
identical systems which interact through the variable
as follows:
(75)Suppose that
is a solution of (75) defined for all
, for some input
. Then, we have the synchronization condition:
, as
.
Indeed, we only need to observe that every pair is a solution of (74) with the same input
Furthermore, if
is a
-periodic function, the
interconnected dynamical systems synchronize onto a
-periodic trajectory.
The above principle can be immediately applied to prove that synchronization onto a -periodic orbit is attained for the Repressilator circuits composing network (67) (see also [19]).
Specifically, let and
; we have that
is a solution of (67). We notice that any pair
is a solution of the following cascade system
(76)Thus, as shown above, contraction of (76) implies synchronization of (67). Differentiation of (76) yields the Jacobian matrix
(77)where
and
denote the partial derivatives of decreasing and increasing Hill functions with respect to the state variable of interest and
,
.
Note that the Jacobian matrix has the structure of a cascade, i.e.
with:
,
. Thus, to prove contraction of the virtual system (76) it suffices to prove that there exist two matrix measures,
and
such that:
;
;
where . Clearly, since
is a positive real parameter, the second condition above is satisfied (with
being any matrix measure). Now, notice that matrix
has the same form as the Jacobian matrix of the Repressilator circuit (56). Hence, if the parameters of the Repressilator are chosen so that they satisfy (66), then there exist a set of positive real parameters
,
, such that
(that is, the first condition above is also satisfied with
).
Thus, we can conclude that (76) is contracting. Furthermore, all the trajectories of the virtual system converge towards a -periodic solution (see Theorem 6). This in turn implies that all the trajectories of network (67) converge towards the same
-periodic solution. That is, all the nodes of (67) synchronize onto a periodic orbit of period
.
A counterexample to entrainment
In [5] there is given an example of a system with the following property: when the external signal is constant, all solutions converge to a steady state; however, when
, solutions become chaotic. (Obviously, this system is not contracting.) The equations are as follows:
where
and
. Figure 14 shows typical solutions of this system with a periodic and constant input respectively. The function “rand” was used in MATLAB to produce random values in the range
.
The following randomly-chosen input and initial conditions are used: ,
,
,
,
. Green: inputs are
(left panel) and
(randomly picked, right panel). Blue:
. Note chaotic-like behavior in response to periodic input, but steady state in response to constant input.
Discussion
We have presented a systematic methodology to derive conditions for various types of biochemical systems to be globally entrained to periodic inputs. For concreteness, we focused mainly on transcriptional systems, which constitute basic building blocks for more complex biochemical systems. However, the results that we obtained are of more generality. To illustrate this generality, and to emphasize the use of our techniques in synthetic biology design, we discussed as well the entrainment of a Repressilator circuit in a parameter regime in which endogenous oscillations to not occur, as well as the synchronization of a network of Repressilators. These latter examples serve to illustrate the power of the tools even when a large amount of feedback is present.
Our key tool is the use of contraction theory, which we believe should be recognized as an important component of the “toolkit” of systems biology. In all cases conditions are derived by proving that the module of interest is contracting under appropriate generic assumptions on its parameters. A surprising fact is that, for these applications, and contrary to many engineering applications, norms other than Euclidean, and associated matrix measures, must be considered. Of course, more than one norm may be appropriate for a given problem: for example we can pick different 's in our weighted norms, and each such choice gives rise to a different estimate of convergence rates. This is entirely analogous to the use of Lyapunov functions in classical stability analysis: different Lyapunov functions provide different estimates.
Ultimately, and as with any other method for the analysis of nonlinear systems, such as the classical tool of Lyapunov functions, finding the “right” norm is more of an art than a science. A substantial amount of trial and error, intuition, and numerical experimentation may be needed in order to come up with an appropriate norm, and experience with a set of already-studied systems (such as the ones studied here) should prove invaluable in guiding the search.
Acknowledgments
The authors wish to thank Dr Diego di Bernardo, Telethon Institute of Genetics and Medicine (TIGEM), Naples (Italy) for suggesting entrainment of Repressilators as a possible suitable application of the tools presented in this paper and Prof J J Slotine, MIT, Boston (USA) for all the insightful discussions about contraction theory.
References
- 1. Gonze D, Bernard S, Walterman C, Kramer A, Herzerl H (2005) Spontaneous synchronization of coupled circadian oscillators. Biophysical Journal 89: 120–129.
- 2. Tyson JJ, Csikasz-Nagy A, Novak B (2002) The dynamics of cell cycle regulation. Bioessays 24: 1095–1109.
- 3. Mettetal JT, Muzzey D, Gomez-Uribe C, van Oudenaarden A (2008) The frequency dependence of osmo-adaptation in Saccharomyces Cerevisiae. Science 319: 482–484.
- 4.
Kuznetsov YA (2004) Elements of applied bifurcation theory. Springer-Verlag (New York).
- 5.
Sontag ED (2009) An observation regarding systems which converge to steady states for all constant inputs, yet become chaotic with periodic inputs. Technical report, Dept. of Mathematics, Rutgers University. http://arxiv.org/abs/0906.2166.
- 6. Lohmiller W, Slotine JJE (1998) On contraction analysis for non-linear systems. Automatica 34: 683–696.
- 7. Lohmiller W, Slotine JJE (2000) Nonlinear process control using contraction theory. AIChe Journal 46: 588–596.
- 8. Angeli D (2002) A Lyapunov approach to incremental stability properties. IEEE Transactions on Automatic Control 47: 410–321.
- 9.
Pham QC, Tabareau N, Slotine JJE (2009) A contraction theory approach to stochastic incremental stability. IEEE Transactions on Automatic Control (to appear).
- 10.
Granas A, Dugundji J (2003) Fixed Point Theory. Springer-Verlag (New York).
- 11. Hartman P (1961) On stability in the large for systems of ordinary differential equations. Canadian Journal of Mathematics 13: 480–492.
- 12. Lewis DC (1949) Metric properties of differential equations. American Journal of Mathematics 71: 294–312.
- 13. Pavlov A, Pogromvsky A, van de Wouv N, Nijmeijer H (2004) Convergent dynamics, a tribute to Boris Pavlovich Demidovich. Systems and Control Letters 52: 257–261.
- 14. Lohmiller W, Slotine JJ (2005) Contraction analysis of non-linear distributed systems. International Journal of Control 78: 678–688.
- 15.
Jouffroy J, Slotine JJE (2004) Methodological remarks on contraction theory. pp. 2537–2543.
- 16.
Slotine JJE, Wang W, Rifai KE (2004) Contraction analysis of synchronization of nonlinearly coupled oscillators.
- 17. Wang W, Slotine JJE (2005) On partial contraction analysis for coupled nonlinear oscillators. Biological Cybernetics 92: 38–53.
- 18.
di Bernardo M, Russo G, Slotine JJ (2009) An algorithm to prove contraction, consensus and network synchronization.
- 19. Russo G, di Bernardo M (2009) How to synchronize biological clocks. Journal of Computational Biology 16: 379–393.
- 20.
Russo G, di Bernardo M (2009) An algorithm for the construction of synthetic self synchronizing biological circuits. pp. 305–308.
- 21.
Sontag ED (1998) Mathematical Control Theory. Deterministic Finite-Dimensional Systems. Springer-Verlag (New York).
- 22. Angeli D, Sontag ED (1999) Forward completeness, unboundedness observability, and their Lyapunov characterizations. Systems and Control Letters 38: 209–217.
- 23.
Michel AN, Liu D, Hou L (2007) Stability of Dynamical Systems: Continuous, Discontinuous, and Discrete Systems. Springer-Verlag (New-York).
- 24.
Dahlquist G (1959) Stability and error bounds in the numerical integration of ordinary differential equations. Transanctions of the Royal Institute Technology (Stockholm).
- 25. Lozinskii SM (1959) Error estimate for numerical integration of ordinary differential equations. I. Izv Vtssh Uchebn Zaved Matematika 5: 222–222.
- 26. Strom T (1975) On logarithmic norms. SIAM Journal on Numerical Analysis 12: 741–753.
- 27.
Arnold VI (1978) Mathematical methods of classical mechanics. Spriger-Verlag (New York).
- 28.
Vidyasagar M (1993) Nonlinear systems analysis (2nd Ed.). Prentice-Hall (Englewood Cliffs, NJ).
- 29. Del Vecchio D, Ninfa AJ, Sontag ED (2008) Modular cell biology: Retroactivity and insulation. Nature Molecular Systems Biology 4: 161.
- 30. Slotine J (2003) Modular stability tools for distributed computation and control. International Journal of Adaptive Control and Signal Processing 17: 397–416.
- 31. Elowitz MB, Leibler S (2000) A synthetic oscillatory network of transcriptional regulators. Nature 403: 335–338.
- 32. Garcia-Ojalvo J, Elowitz MB, Strogatz SH (2004) Modeling a synthetic multicellular clock: Repressilators coupled by quorum sensing. Proceedings of the National Academy of Science 101: 10955–10960.
- 33. Zhou T, Zhang J, Yuan Z, Xu A (2007) External stimuli mediate collective rhythms: artificial control strategies. PLoS ONE 3: e231.
- 34. Wang R, Chen L, Aihara K (2006) Synchronizing a multicellular system by external input: an artificial control strategy. Bioinformatics 22: 1775–1781.
- 35. Li C, Chen L, Aihara K (2007) Stochastic synchronization of genetic oscillator networks. BMC Systems Biology 1: 6.