Journal of Theoretical Biology 340 (2014) 199208

Contents lists available at ScienceDirect

Journal of Theoretical Biology

journal homepage: www.elsevier.com/locate/yjtbi

The regulation of gene expression in eukaryotes: Bistability

and oscillations in repressilator models
Rui Dilo a,b,n
a
Nonlinear Dynamics Group, Instituto Superior Tcnico Av. Rovisco Pais, 1049-001 Lisbon, Portugal
b
Institut des Hautes tudes Scientiques, 35, route de Chartres, 91440 Bures-sur-Yvette, France

H I G H L I G H T S

We use delays with the mass action law to model protein synthesis in eukaryotes.

We have derived the properties of m-repressilator models, with and without delays.

If m is even, the m-repressilator model shows bistability.

If m is odd, the m-repressilator model has limit cycle solutions.

The 1-repressilator is the simplest genetic mechanisms with stable oscillations.

art ic l e i nf o a b s t r a c t

Article history: To model the regulation of gene expression in eukaryotes by transcriptional activators and repressors, we
Received 26 March 2013 introduce delays in conjugation with the mass action law. Delays are associated with the time gap
Received in revised form between the mRNA transcription in the nucleoplasm and the protein synthesis in the cytoplasm. After
6 September 2013
re-parameterisation of the m-repressilator model with the Hill cooperative parameter n, for n 1, the m-
Accepted 10 September 2013
repressilator is deducible from the mass action law and, in the limit n-1, it is a Boolean type model.
Available online 19 September 2013
With this embedding and with delays, if m is odd and n 41, we show that there is always a choice of
Keywords: parameters for which the m-repressilator model has sustained oscillations (limit cycles), implying that
Regulation of gene expression in eukaryotes the 1-repressilator is the simplest genetic mechanism leading to sustained oscillations in eukaryotes. If m
Delays
is even and n 4 1, there is always a choice of parameters for which the m-repressilator model has
m-repressilator
bistability.
Limit cycles
Bistability & 2013 Elsevier Ltd. All rights reserved.

1. Introduction mRNA molecules have to migrate through the nucleoplasm, cross-

In the cell, protein synthesis begins with the transcription of
mRNA from its DNA template. Then, the mRNA is translated into
protein in the ribosomes. This is the central dogma of molecular
biology, as discussed in Crick (1967, 1970). This mechanism of
protein synthesis is common to both prokaryote and eukaryote
organisms.
As prokaryote organisms lack a nuclear membrane, both DNA
molecules and ribosomes are located in the cytoplasm. In eukar-
yotes, the DNA molecules are in the nucleoplasm and the ribo-
somes are in the cytoplasm. This different localisations of the DNA
implies that, in eukaryotes, for protein translation to occur, the
n
Correspondence address: Nonlinear Dynamics Group, Instituto Superior
Tcnico Av. Rovisco Pais, 1049-001 Lisbon, Portugal. Tel.: 351 218 417 617;
fax: 351 218 419 123.
E-mail address: rui@sd.ist.utl.pt
ing the nuclear membrane.
The transcription of mRNA is done by a catalytic enzymatic
process involving the RNA polymerase enzyme that is continu-
ously produced in the cell. In this process, the DNA is the catalyst.
The RNA polymerase binds to the promoter regions of the DNA and
translation begins. Some genes have control regions where tran-
scription activators and repressors can bind, enabling the activa-
tion or inhibition of the mRNA transcription, Alberts et al. (2008).
At the molecular spatial scale and during short time scales, it is
assumed that RNA polymerase molecules move randomly in the
cytoplasm (prokaryotes) or in the nucleoplasm (eukaryotes).
In models, it is implicitly assumed that the mRNA motion is
intrinsically stochastic, Larson et al. (2009).
A general mathematical framework to describe quantitatively
the transcription, translation and protein synthesis in prokaryotes,
based on the mass action law, has been proposed in Alves and
Dilo (2005). This class of models is built with genes and proteins.
The mass action law assumes that reaction kinetics result from the

0022-5193/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jtbi.2013.09.010
200 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208
random motion of the molecules present in the media, Gillespie
(1977, 1992), Gardiner (1997), van Kampen (2007), and all the
model parameters have a biochemical meaning. However, as in
eukaryotes ribosomes are separated from the DNA template by the
nuclear wall, the mobility of mRNA is constrained. The transit from
inside to outside the nucleus is done through pores in the nuclear
membrane, and no active transport process is known. The nuclear
wall has the effect of delaying both the entry of proteins into the
nucleus and the transport of mRNA chains to ribosomes. Mathe-
matical modelling of single cell gene expression with stochastic
techniques shows the existence of delays between transcription
and translation in eukaryotes, Larson et al. (2009). Experiments
inducing protein synthesis in Drosophila have quantied that the
temporal delay between the mRNA transcription and the protein
synthesis can be in the range 812 min, O'Brien and Lis (1993).
To model the regulation of gene expression in eukaryotes by
transcriptional activators and repressors, we propose here to
introduce delays in conjugation with the mass action law. Several
authors have introduced delays in systems biology models,
Allwright (1977), Mackey and Glass (1977), Murray (1989), Fall
et al. (2002), Chen and Aihara (2002), Chen et al. (2004), Zhu et al.
(2007), Smith (2011), among others. These models have been used
to describe the intrinsic delays between the production of a
substance and its action on a specic target. The stability analysis
of specic low dimensional linear models evolving delay effects
between mRNA transcription and protein production has been
analysed by Monk (2003) and Mincheva and Roussel (1977).
In these examples, delays were important to explain the oscilla-
tions or even chaos in the observed phenomena. On the other
hand, in some of these models, the functional response of the
systems was described by Hill type functional forms, which have
neither a direct correspondence nor a theoretical justication
within the microscopic dynamical mechanisms associated with
the mass action law.
It is known that some biochemical systems described by the
mass action law have a stable steady state in phase space, but
when modied to a Hill type functional description they show
oscillatory behaviour. The best well-known example is the
3-repressilator model, Elowitz and Leibler (2000). In its original
formulation, the m-repressilator model is described by a system of
2m-dimensional systems of rst order equations. However, based on
biological arguments and the use of a steady state approximation, the
model equations can be reduced to a m-dimensional system of rst
order equations, reducing considerably the dynamical complexity of
the problem. For the 2m-dimensional m-repressilator systems, the
analysis of the dynamics is based essentially on numerical experi-
ments. For example, Mller (2006) reported the possibility of the
existence of limit cycle behaviour for the 2m dimensional repressi-
lator model, and Strelkowa and Barahona (2011) observed oscillatory
and monotonous transients to a steady state for the same model. For
these equations, the dimensionality of the phase space makes
difcult the general analysis of the model properties.
The 3-dimensional version of the (3-)repressilator model has a
steady state for the corresponding mass action description, but,
1 repressilator 2 repressilator
P1 P2
Fig. 1. Graphs describing the different types of repressilator interactions. Pi represent proteins and the end symbol a indicates that the interactions are inhibitory.
in the framework of the Hill type response functions, it shows
oscillatory behaviour with limit cycles in phase space, Buse et al.
(2010); Kuznetsov and Afraimovich (2012). Due to the intrinsic
difculty in calibrating models of protein synthesis with experi-
mental or observational data, it is not known if the oscillations
emerging from the Hill specic parameterisation have a biological
meaning or if they are a mathematical artefact. In the enzyme
kinetics literature, the Hill approach is generally criticised as non-
realistic, (Bizwanger, 2008, p. 3243). However, there is an effort
to justify by thermodynamic arguments the cooperative effects
associated with the Hill parameterisation, Marzen et al. (2013).
One of the goals here is to have a biologically meaningful and
physically consistent framework for the modelling of the regula-
tion of gene expression in eukaryotes. To be consistent with the
mechanism of molecular interactions and dynamics, we follow a
strategy where only bi-molecular kinetic mechanism based on the
mass action law is considered. Then, delays are introduced when-
ever we consider both the action of transcription activators and
repressors on genes and the corresponding protein synthesis.
Within this approach, we analyse mechanisms that can eventually
lead to biological oscillations (damped or persistent) and are
described with a family of parameterised models containing the
mass action law, the Hill type response functions and the Boolean
formalism. The repressilator model is one of the models that best
adapts to these conditions.
The transition from a differential equation approach to a
Boolean approach has been introduced by Glass and Kauffman
(1973) in order to simplify the biological interpretation of net-
works of interactions. In this approach, sigmoidal type functions
are transformed into step functions with a compatibility analysis
of the parameters of the model. This approach has been further
developed by Davidich and Bornholdt (2008) for a model of the
cell cycle. In their paper, the Boolean formalism is obtained in the
continuous limit n-1 of the Hill cooperative parameter n.
This paper is organised as follows. In Section 2, using the
mass action law, we derive the 1-repressilator model. In this
model, a protein is produced and, as its concentration increases,
it represses its own production, Fig. 1. Re-parameterising this
model with the Hill extra parameter n, for n 1, the model is
deductible from the mass action law. For n 4 1, we have a Hill type
empirical description, and, in the limit n-1, we obtain a Boolean
type model. Then, for these different cases with and without
delays, the asymptotic solutions of the family of parameterised
models have been analysed. The introduction of delays can be
used to distinguish between eukaryotic and prokaryotic regula-
tion. In Section 3, we analyse the general m-repressilator model,
Fig. 1, and we narrow this analysis to the 2- and 3-repressilator
models, with and without delays. In Section 4, we analyse the
general solutions of the Boolean m-repressilator models for
m Z 2. In this case, we show that all the asymptotic solutions are
bistable or oscillatory stable of limit cycle type, depending if m is
even or odd, respectively. Finally, in Section 5, we summarise
and discuss the main conclusions of the paper. To simplify the ow
of this exposition, all the formal proofs were moved to Appendix A.
3 repressilator m repressilator
P2 P2
P1 P1 P1
P3 Pm
 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208 201

2. The 1-repressilator model 1

n 6 n
The 1-repressilator model describes the production of a protein 0.75
that represses itself by binding to some operator site in its n 1
template DNA, Fig. 1. Following the mathematical formalism

fx
0.5
developed in Alves and Dilo (2005) and the discussion therein,
in prokaryote organisms, the kinetic diagrams describing the
0.25
1-repressilator protein production model are
k1
GG P 0.
0 1n
1
k2
P G GP x
k2
Fig. 2. Graphs of the family of regulatory functions f(x) dened in (6), describing
k3
P 1 the self-repressive behaviour of a regulatory protein. For n 1, the regulatory
function f(x) is derived from the mass action law. For n 4 1, the regulatory function
where P, G and GP represent, respectively, the protein, its gene has the Hill or inverted sigmoidal form and, for n 1, f(x) is a Boolean or step
function.
template and the gene with the protein bound to the DNA
repressor site. The rst diagram in (1) describes the protein free
production, and the last diagram represents protein degradation.
The second diagram represents the inactivation of the gene by the Eq. (6) describes a mechanism of protein production with self-
protein. The constants ki are the rates at which production, repression in a prokaryote organism, and, for n 1, it is derivable
inhibition and degradation occur. from the mass action law. If n 4 1, the physical mechanisms
From (1) and the mass action law, Alves and Dilo (2005), Dilo associated with the mass action law and leading to (6) are no
and Muraro (2010), the time evolution of the protein concentra- longer true. In Fig. 2, we show, for several values of n, the graph of
tions is described by the system of ordinary differential equations the regulatory function f(x). For n 4 1, the function f(x) has the
8 inverted sigmoidal Hill form and, for n-1, f(x) converges to a
_
< G k  2 GP k2 G:P
> Boolean or step function. Due to the particular form of the function
_ P  k  2 GP k2 G:P
G 2 f(x), for large values of n, it is sometimes argued that f(x) describes
>
:_
P k  2 GP  k2 G:P k1 G  k3 P a threshold mechanism. This threshold occurs for xth 1= 1=n ,
and f xth 1=2.
with the conservation law By a simple analysis, it follows that Eq. (6) has a unique stable
Gt GP t G0 3 xed point or steady state. In fact, the xed point equation can be
written in the form 1 x xn 1 hx. If 4 0, and by the
and G0 is the initial concentration of the gene G. As usual, the monotonicity of h(x) for x Z 0, it follows that the xed point is
small dots represent derivatives in order to time t. Applying the unique. For any value of n Z 1, and as -1, the protein concen-
steady state simplication G _ 0, we obtain k  2 GP  k2 G  P 0.
pto a steady state. For n 1, the xed point is
tration converges
Solving this equation together with the conservation law (3) in xn  1 1 4 =2 and, in the limit n-1, xn -1.
order to GP and of G, we obtain Let us assume now that the protein x is produced in an
8 eukaryote organism. In this case, there exists a delay between
>
> k  2 G0
>G
< the protein synthesis and the effective transcriptional inhibition.
k  2 k2 P
To model this effect, Eq. (6) is transformed into the new equation
>
> k2 G0P
>
: GP k k P:
2 2
dx 1
 x 7
Introducing these last relations into (2), we obtain the equation for d 1 xn  s
the 1-repressilator model
which is an ordinary differential equation with delays. In this case,
k1 G0
P_  k3 P; 4 the delay is represented by s, being s the mean (renormalised)
k2
1 P time that the protein takes to reach the interior of the nucleus of
k2 the cell of the eukaryote organism. To be more precise, in this
Eq. (4) describes the production of the self-inhibiting or self- cases, we must represent the delayed contribution for the kinetics
k2
repressing protein P. by the diagram P Gr GP , explicitly indicating that the delay
To make Eq. (4) dimensionless, we introduce the new variables, r s=k3 is associated with the reaction ratio k2.
x Pk3 =k1 G0, k3 t and the new constant, k1 k2 G0 From the mathematical point of view, Eq. (7) denes an innite
=k  2 k3 4 0. Substitution of these variables into (4) leads to the dimensional dynamical system. In fact, the solutions of Eq. (7)
1-repressilator dimensionless model equation depend on the choice of continuous real functions t :
 s; 0-R Z 0 , where xt t, for t A s; 0. The innite dimen-
dx 1
x; 5 sionality of the dynamical system comes from the innite dimen-
d 1 x
sionality of the set of admissible -functions. In the following, we
In order to include the Hill type functional form in the will consider solutions of delay equations with initial conditions
1-repressilator model, we introduce the Hill parameter n into t  0, for t A  s; 0.
(5), obtaining, For sufcient small values of the delay s, the asymptotic solutions
of Eqs. (6) and (7) coincide, Smith (2011), and the steady state
dx 1
 x f x  x; 6 solution xn is stable. For larger delays, to determine the stability of
d 1 x n
the steady state xn, we have to linearise Eq. (7) around xn.
Some authors call n the cooperative parameter, Kuznetsov and Let xn be the unique xed point of the delay equation (7). The
Afraimovich (2012). In this way, the 1-repressilator model is point with coordinate xn  xn ; n is the unique positive real root
embedded in a larger class of models. of the equation 1=1 xn  x 0. The delay equation obtained by
202 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208
linearising (7) around the steady state xn is
dx
 nxn n 1 x  s  xn x  xn : 8 and of the parameter
d
With y x  xn and the new parameter nxn n 1 n1  xn
; n 4 0, Eq. (8) becomes
dy
 y  s y: 9
d such that
The stability of the steady state xn of Eq. (7) is determined by
the stability of the zero steady state of Eq. (9). Using standard
techniques of stability analysis for delay equations, Hale (1977)
and Smith (2011), we now construct the characteristic equation
associated with (9). Assuming that the solution of Eq. (9) has the
form y e and after substitution of this solution into Eq. (9),
we obtain the characteristic equation
p 1 e  s ; 10
Lemma 1. Let p be the characteristic equation of the linear delay
equation (9), with 4 0 and s 40, and let 1 be the unique solutions
of the equation,  tan s , with A  =2s; =s. Let 0
e  s  1 =s and 1  1= cos 1 s 40. Then, the roots of the character-
istic equation (10) are:
(a) If o 0 , p has two real negative roots.
(b) If 0 , p has one real negative root.
(c) If 0 o o 1 , p has two complex conjugate roots with
Real o0.
(d) If 1 , p has two complex conjugate roots with Real 0.
(e) If 4 1 , p has two complex conjugate roots with Real 4 0.
The above lemma establishes the conditions of applicability of
the Hopf bifurcation theorem to the dimensionless non-linear
delay equation (7).
Theorem 2. Consider the delay equation (7), with 40, n Z 1, s 4 0,
and the positive parameter n1  xn ; n, where xn ; n is the
coordinate of the unique positive xed point of (7). Let 0 4 0 and
1 4 0 be the constants as dened in Lemma1, and let xt t 0,
for t A  s; 0, be the initial data. Then, we have
(a) If r 0 , then the xed point x is asymptotically stable.
n
(b) If 0 o o 1 , then the xed point xn is (oscillatory) asymptoti-
cally stable.
(c) If 1 , then the xed point xn has a supercritical Hopf
bifurcation and xn is asymptotically stable.
(d) If 4 1 , then for initial conditions away from the unstable xed
point, the delay equation (7) has an asymptotically stable
periodic solution.
As the conditions on of Theorem 2 encapsulate the para-
metric information about the coordinates of the xed point
xn ; n, we must be more specic.
Corollary 3. The delay equation (7) with the choice n 1 and initial
data xt t 0, for t A  s; 0, has no asymptotically stable
periodic solutions but, for sufciently large values of s, it can have
damped oscillations, asymptotically converging to the xed point
xn ; 1. If n 4 1, for sufciently small and sufciently large delays s,
Eq. (7) with initial data xt t 0, for t A  s; 0, has an
asymptotically stable periodic solution (limit cycle), and the xed
point xn ; n is unstable.
Corollary 4. In the limit n-1, the solutions of the delay equation (7),
with initial data, xt t 0, for t A s; 0, converge monotonically
to the steady state xn 1.
From Theorem 2, we can construct the bifurcation diagram of
the solutions of the delay equation (7) as a function of the delay s
n1  xn ; n, where xn ; n is the
coordinate of the unique xed point of Eq. (7), Fig. 3. From
Corollary 3, it follows that, for n 1, the delay equation (7) has
no asymptotically stable oscillatory solutions. However, choosing a
Hill parameter n 4 1, it is always possible to choose values of n
n1  xn ; n 4 1 s, where 1 s is the parameter
that determines a supercritical Hopf bifurcation. In this case, the
delay equation (7) has asymptotically stable oscillatory solutions
(limit cycle solutions), Fig. 3.
In Table 1, we summarise the type of asymptotic solutions of
the model equation (7), with and without delays. Both the mass
action law and the Boolean based models do not have asympto-
tically oscillatory stable solutions, with or without delays. How-
ever, for the same model, but with the Hill parameter n 4 1 and
delay s 4 0, for a suitable choice of the parameters and n, it is
always possible to have stable oscillations in the protein concen-
tration. In the limit n-1, Corollary 4 applies.
In Fig. 4, we show an oscillatory solution of Eq. (7), for the
parameter values 1:0, n 10.0 and s 20.0. In this case, all the
solutions of Eq. (7) oscillate for delays s 4 2:0. In the same gure,
we also show a damped oscillatory solution obtained with n 1.
Assuming that the Hill functional form describes genetic
transcriptional regulation, we conclude that the 1-repressilator
model is the prototype of the simplest genetic mechanism being
able to produce oscillations in eukaryote organisms. On the other
hand, if this self-repressing mechanism of protein synthesis does
not show persistent oscillations, two situations can occur. Either
we are in a particular region of parameters for which natural
delays are not enough to induce oscillations, or the Hill functional
form does not accurately describe the mechanism of protein
synthesis. These facts stress the importance of the calibration
4
0 1
3 limit cycles
2
Hopf supercritical
1
damped oscillations
stable f.p.
0.0 0.5 1.5 2.0 2.5 3.0
s
Fig. 3. Bifurcation diagram for the delay equation (7), for solutions reached from
the initial condition xt t 0, for t A  s; 0. The functions 0 s and 1 s are
dened in Lemma 1. In the region marked damped oscillations, the asymptotic
solutions of the delay equation converge to the stable xed point (f.p.) xn. Limit
cycle solutions are asymptotically stable oscillatory solutions of the delay equation.
By Corollary 3, if n 1, there are no oscillations. If n 41, it is always possible to nd
a value of s such that Eq. (7) has a limit cycle solution.
Table 1
Asymptotic solutions of the 1-repressilator model equation (7) as a function of the
delay s and for different values of the Hill parameter n.
Hill parameter s0 s40
n 1 (mass action) Steady state Steady state damped
no oscillations oscillations
n 41 (Hill form) Steady state Steady state damped
no oscillations and stable oscillations
Hopf bifurcation
n 1 (Boolean form) Steady state Steady state
no oscillations no oscillations
 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208 203

1.4 Lemma 5. Consider the delay equation (14) with i 4 0 and i 4 0.

Then, we have
1.2
1.0 (a) If m is odd, Eq. (14) has a unique xed point with positive
x 0.8 coordinates.
0.6 (b) If m is even, Eq. (14) has at least one xed point with positive
0.4 coordinates.
0.2
0.0 By Lemma 5, the systems of Eqs. (14) have at least one xed
0 50 100 150 200 point with positive coordinates. We denote the generic xed point
t by xn xn1 ; ; xnm . Linearizing the systems of Eqs. (14) around the
Fig. 4. Solution of the delay equation (7), for the parameter values 1:0, n 10.0, xed point xn, we obtain
s 20.0 (thick line) and initial condition xt t 0, for t A  s; 0. For this choice
y_ Byt  r Ayt 15
of and n, by Corollary 3 and Theorem 2, the solution of equation (7) oscillates for
delays s 42:0. For the case n 1 and in agreement with Corollary 3, the delay
where yt xt  xn ,
equation (7) shows damped oscillations (thin line).
0 1
0  1 0 0
B 0 0  2 0 C
B C
and validation of the mathematical models for the mechanisms B C
BBB C;
associated with the central dogma of molecular biology. C
B 0  C
@ 0 0 m1 A
 m 0 0 0
3. The m-repressilator model 0 1
 1 0 0
B C
We consider m proteins Pi , i 1; ; m, continuously produced B 0  2 0 C
ABB
C 16
by its gene template Gi but repressed by protein Pi 1 . This @ CA
sequential process can be viewed as a circular chain of repressor 0 0  m
proteins where Pm is repressed by P1 , Fig. 1. The kinetic diagrams
for this process are and i n i xni 1 n  1 =1 i xni 1 n 2 n i 2i xni 1 n  1 xni 2 . The
coordinates of the xed point xn are also functions of the constants
k3i  2
Gi Gi Pi i , i and n.
k3i  1 The characteristic equation of the linear delay equation (15) is
Pi 1 Gi r GPi
k  3i 1
i 1; ; m 11 easily calculated and we obtain (see Smith, 2011 for details)
Pi 1 Gi GPi
k3i p detI  A e  r B
Pi
m m
i   1m e  mr i ; 17
and the arrow subscript r indicates that this mechanism is delayed. i1 i1
As we have discussed in the introduction, the delay is due to the
re-entry of the protein into the nucleus. The stability of the xed points of the delay equations (14) and
Following the same approach as in the previous section, after (15) is determined through the analysis of the roots of the
substitution of conservation laws and making the steady state characteristic equation (17). As, by Lemma 5, there are differences
simplication, the equations for the time evolution of the proteins between m-repressilator models, eventually affecting the stability
are of asymptotic solutions, we now analyse separately the cases m 2
and m 3.
k3i  2 Gi 0
P_ i  k3i Pi ; i 1; ; m 12
k
1  3i 1 Pi 1
k3i  1 3.1. The 2-repressilator model
where Pm 1 P1 . Eqs. (12) describe the time evolution of proteins
involved in the m-repressilator model, Fig. 1, and have been We rst consider the case of the 2-repressilator model, Fig. 1,
derived from the mass action law. without delays, r 0. For the case m 2, the characteristic equation
With the new variables, xi Pi =k3i  2 Gi 0, and the new (17) associated with the positive xed points reduces to
constants i k3i  1 k3i  2 Gi 0=k  3i 1 and i k3i , with i 1; ; p 1 2  1 2 e  2r ; 18
m, the system of Eqs. (12) reduces to
With r 0, i 4 0 and i 40, a simple geometric analysis shows
1
x_ i  i xi ; i 1; ; m; 13 that the characteristic polynomial (18) has two real roots, both
1 i xi 1
negative or one negative and the other positive. In the rst case,
To include the Hill parameter n Z1 and the delays associated with the positive xed points of Eqs. (14) are of stable node type and,
the repressing proteins, we transform (13) into the new delay in the second case, the xed points are of saddle type and are
equation unstable.
1 Eqs. (14) have two other xed points with coordinates xnn1
x_ i  i xi t; i 1; ; m 14 x1=1 ; 0 and xnn2 0; 1=2 . These xed points can be saddle
1 i xni 1 t  r
points or stable nodes.
where xm 1 x1 and r Z 0, and the delay term has been intro- If m 2, r 0 and n 1 in Eqs. (14), due to the convexity and
duced only in the terms involving the rate constant k3i  1 , as monotonicity of the vector eld components, Eqs. (14) have a
indicated in (11). The systems of Eqs. (14) are the reparameterized unique xed point with positive coordinates of stable node type,
m-repressilator model with delays. Fig. 5(a).
204 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208

10 10

8 8 B
Ws

6 6
A
x2

x2
4 4 Wu

2 2
Wu
Ws
0 0
0 2 4 6 8 10 0 2 4 6 8 10
x1 x1
Fig. 5. Solutions in phase space of Eqs. (14) and (19), for m2, r 0, and kinetic parameters, 1 1=8, 2 1=5, 1 0:3 and 2 0:1. Bullets represent xed points. In (a),
n 1, we show four solutions converging to the positive stable steady state. In (b), n 1, we show the stable (Ws) and the unstable (Wu) manifolds of the point xs 1; 1. In
this case, the set Ws, dened in (20), splits the phase space in two invariant regions denoted by A and B. In both cases (a) and (b), 1 o 1 and 2 o 1.

In the extreme case of n-1, Eqs. (14), for m 2 and r 0,
reduce to
(
x_ 1 gx2  1 x1
19
x_ 2 gx1  2 x2
where gx 1, for x r 1, and gx 0, otherwise.
If 1 Z 1 or 2 Z 1, the vector eld (19) has one xed point. If
1 Z 1 and 2 Z 1, the coordinates of the xed point are
1=1 ; 1=2 . If 1 Z 1 and 2 o 1, the coordinates of the xed
point are 0; 1=2 . If 1 o 1 and 2 Z 1, the coordinates of the xed
point are 1=1 ; 0. In any of these cases, the xed point is of stable
node type.
If 1 o 1 and 2 o 1, the vector eld (19) has two xed points
with coordinates, xnn1 1=1 ; 0 and xnn2 0; 1=2 . These xed
points are of stable node type. Therefore, for 1 o 1 and 2 o 1, the
dynamics of the system of Eqs. (19) is bistable.
In the bistable case, due to the integrability of Eq. (19), the
boundary of the basin of attraction of the two xed points can be
easily found. For 1 o 1 and 2 o 1, we can dene the set
 
W s x0 ; y0 A R2Z 0 : x0 e  1 t 1; y0 e  2 t 1; for some t Z 0
    
1 1 1 1
[ x0 ; y0 A R2Z 0 :   x0 e  1 t 1;   y0 e  2 t 1;
1 1 2 2
 and these xed points do not change their stability.
for some t Z 0 ;
Geometrically, the set Ws behaves as the stable manifold of the
point xs 1; 1. The point in phase space with coordinates
xs 1; 1 is not a xed point, and, as is easily derived from (20),
is attained in nite time.
Analogously, we can dene the set of points that converge
backwards in time to xs, dening a kind of unstable manifold Wu
of xs. In Fig. 5(b) we show xs 1; 1, Ws and Wu. The two branches
of Wu end up at xnn1 and xnn2 . The set Ws separates the positive
quadrant of phase space into two invariant regions, marked A and
B in Fig. 5(b). Initial conditions x1 t; x2 t 1 t; 2 t in A or B,
for t A r; 0, lead to solutions of Eqs. (19) with asymptotic limits
xnn1 or xnn2 , respectively.
Thus, if 1 o 1 and 2 o 1, for some Hill parameter value n 4 1,
there exists a bifurcation from a system with one stable steady
state to a system with two stable steady statesbistable system.
To analyse the transition to the bistable system as the Hill
parameter n is varied, we analyse the coordinates of the positive
stable xed points of Eqs. (14). By a simple manipulation, the xed
points of Eqs. (14) with m 2 are solutions of the equations
8
>
> 1 1
>
> x1  n Fx1
>
>
<
1 1
1
1 1
2 1 2 x1
n
21
>
>
>
> 1 1
>
: x2 2 1 2 xn
>
1
(see the proof of Lemma 5 in Appendix A). From the rst equation
in (21), it follows that, for any n Z1, F0 4 0 and F1=1 o 1=1 .
Therefore, for any n Z1, the equation x Fx has at least one
solution in the interval 0; 1=1 . If 1 o1 and 2 o1, in the limit
n-1 and in the same interval, the same equation has three
solutions, depending on the graph of F(x). For the general case,
these solutions have not a closed form. However, we can even-
tually assume that the change in the number of positive xed
point changes the topology of the orbits in phase space and also
affects the stability of the xed points xnn1 x1=1 ; 0 and
xnn2 0; 1=2 . In this case, for n 4 1, expanding Eqs. (14) around
the xed points xnn1 x1=1 ; 0 and xnn2 0; 1=2 , the Jacobian
matrices of the system are non-degenerate and, for nite n 4 1,
20 To analyse the change in the number of solution of the rst
equation in (21), we proceed numerically. In Fig. 6, we show the
solutions of the rst equation in (21) as a function of the Hill
parameter n and for two different choices of the kinetic para-
meters i and i . The solutions of the rst equation in (21) give the
rst coordinates of the xed points of Eqs. (14), for m 2. From the
numerical analysis shown in Fig. 6, it follows that, for this choice of
parameters and when the Hill parameter n is varied, the system of
Eqs. (14), with m 2 and r 0, has a saddle-node bifurcation. The
value of the parameter n where the bifurcation occurs depends on
all the kinetic parameters.
Increasing n for values above the saddle-node bifurcation values,
Eq. (14), for m2 and r0, has one xed point of saddle node type
and two xed points of stable node type. The xed point of saddle
type has a stable and an unstable manifold similar to the ones
constructed previously in the n-1 case. The stable manifold
partitions the rst quadrant of phase space into two invariant
regions and initial conditions on these separated regions lead to
solutions with different asymptotic limits. In this case, the topology
 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208 205

10 10
8 8
6 6
x1

x1
4 4
2 2
0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
n n
Fig. 6. Bifurcation diagram of the positive xed points of Eqs. (14), for m2, r 0, 1 o 1 and 2 o 1. We represent the rst coordinate of the xed points of Eqs. (14) as a
function of the Hill parameter n. The dotted lines represent unstable xed points, and full lines represent stable xed points. In (a), the kinetic parameters are, 1 1=8,
2 1=5, 1 0:3 and 2 0:1, and the saddle-node bifurcation occurs for n C 2:04. In (b), the kinetic parameters are, 1 1=8, 2 1=5, 1 1:8 and 2 0:1, and the saddle-
node bifurcation occurs for n C 2:24. For values of n above the saddle-node bifurcation, Eq. (14), for m 2 and r 0, has one xed point of saddle node type and two xed
points of stable node type.

Table 2 There exists a constant 0 , with 0 r 0 o 1 , such that, if

Asymptotically stable solutions of the 2-repressilator equation (14) as a function of
the delay r and for different values of the Hill parameter n. For sufciently large
values of n, the saddle-node bifurcation always occurs but it depends on the values
of the parameters in Eqs. (14).
Hill parameter r 0 r 40
n 1 (mass action) Steady state Steady state
n 4 1 (Hill form) Steady state bistable Steady state bistable
saddle-node bifurcation saddle-node bifurcation
n 1 (Boolean form) Steady state bistable Steady state bistable
of the phase space orbits is similar to the ones found in the
n-1 case.
All the analysis we have done so far is for the case without delays
in Eqs. (14). With delays, r 4 0, the number of xed points of Eqs. (14)
is the same. Also, from a straightforward geometric analysis and as
1 2 4 0, the exponential term in the characteristic polynomial (18)
does not affect the stability results previously obtained. Therefore, the
introduction of delays in the 2-repressilator model does not lead to
signicative qualitative differences between the asymptotic solutions
of the differential equation (14), for m2, provided x1 t; x2 t
1 t; 2 t 0; 0, for t A  s; 0.
In Table 2, we summarise the type of asymptotic solutions of
the model equation (14), for m 2. In all the cases, all the solutions
converge asymptotically in time to a steady state solution. As we
have seen in the bifurcation analysis, if the Hill parameter n is
large enough and if, 1 o 1 and 2 o1, it is possible to have
bistable behaviour and the choice of the asymptotic state depends
on the initial conditions.
3.2. The 3-repressilator model
By Lemma 5, the system of Eqs. (14), for m 3, has only one
xed point, and, by (17), the characteristic equation associated to
the xed point is
p 1 2 3 1 2 3 e  3r ; 22
As, 1 2 3 4 0 and with r 0, a simple geometric analysis shows
that the characteristic equation (22) has three roots. These three
roots can be real and negative or, one real and negative and the
two others complex conjugate.
Lemma 6. Consider the system of Eqs. (14), with m 3, r 0, i 4 0
and i 4 0, for i 1; 2; 3. Let 1 1 2 1 3 2 3 . For ssilator model with 1 o 1 and 2 o 1 has bistability, with or without
1 2 3 1 , the system of equations has a supercritical Hopf
bifurcation. If 1 2 3 4 1 , then the system of equations has an
asymptotically stable periodic solution in phase space (limit cycle).
0 o 1 2 3 o 1 , the solutions of the system equation are oscilla-
tory asymptotically stable. If, 0 o 1 2 3 r 0 , then the solutions of
the system equation are monotonically asymptotically stable.
We have tested numerically the results of Lemma 6. For the choice
of parameters, 1 0:1, 2 0:2, 3 0:3, 1 0:3, 2 0:1 and
3 0:5, we have obtained 0 0:00039 and 1 0:06. In this case,
the 3-repressilator model has damped oscillations for 1:04 n 2:8
and has a Hopf bifurcation for the Hill parameter n C 2:8. For the same
kinetic parameters, if n 4 2:8, the asymptotic solutions of the
3-repressilator model show stable oscillations (limit cycles).
The introduction of delays in the 3-repressilator model does
not change these results signicantly, as systematic numerical
exploration of the roots of the characteristic equation (22) shows.
For n 1 and large r ( 4 2), numerical tests show that it is possible
to have damped oscillations, but for much larger values of r an
Hopf bifurcation is never reached.
For n-1, the 3-repressilator model has the Boolean form
8
> x_ gx2  1 x1
< 1
x_ 2 gx3  2 x2 23
>
: x_ gx  x
3 1 3 3
where gx 1, for x r1, and gx 0, otherwise. If 1 o 1, 2 o1
and 3 o 1, the system of Eqs. (23) has no xed points. Systematic
numerical analysis of the geometry of the orbits in phase space
suggests that the asymptotic solution of the system of Eqs. (23)
converges always to a function periodic in time, a limit cycle type
solution. In Fig. 7, we show such an asymptotically periodic
solution of the system of Eqs. (23). In the next section, these cases
will be analysed in more detail.
For n-1, but 1 Z 1, 2 Z 1 and 3 Z 1, there exists one xed
point of stable node type, with coordinates 1=1 ; 1=2 ; 1=3 .
As in the case m 2, delays do not introduce qualitative changes
in the dynamics of the 3-repressilator model.
In Table 3, we summarise the type of asymptotic solutions of
the 3-repressilator model.
4. Boolean m-repressilator model for m Z 2
In the previous sections, we have studied Boolean m-repressilator
models for m r 3. We have shown that the 1-repressilator model
with positive delays can have limit cycle solutions. The 2-repre-
delays, and has no limit cycle solutions. The 3-repressilator model,
with i o 1, for i 1; 2; 3, and sufciently large Hill parameter n, has
asymptotically stable periodic solutions. This shows that for m2
206 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208

x2 4 Eqs. (24) converge to an asymptotically stable periodic function of

time or limit cycle. However, this system has no xed points in
2
phase space.
0 For the particular case where i Z 1, for every i 1; ; m,
case (c), the coordinates of the xed point are, xi 1=i , for
i 1; ; m.
4 In this Boolean case and for the initial condition x1 t; x2 t; ;
xm t 0; ; 0, for t A  r; 0, delays do not introduce topological
x3 modications on the solutions of the piecewise linear system of
Eqs. (24).
2

0 5. Conclusions

2 We have shown that the m-repressilator model is one of the

x1 simplest genetic mechanisms showing bistability (m 2) and
4 stable oscillations (m 1 and m 3). In general, we have shown
that for m Z2, there exists always a choice of the Hill parameter n
Fig. 7. Limit cycle solution of the 3-repressilator model equations (23), for the
parameter values, 1 0:1, 2 0:2, 3 0:3. such that

(a) If m is even, the asymptotic solutions of the m-repressilator

Table 3 model are bistable, and do not show damped or persistent
Asymptotically stable solutions of the 3-repressilator model as a function of the
oscillations.
delay r and for different values of the Hill parameter n.
(b) If m is odd, the asymptotic solutions of the m-repressilator
Hill parameter r0 r40 model are periodic of limit cycle type.

n 1 (mass action) Steady state Steady state damped

In case (a), upon variation of the Hill parameter n, bistability is
oscillations
n 4 1 (Hill form) Steady state damped Steady state damped
obtained through a saddle-node bifurcation. In case (b), limit cycle
and stable oscillations and stable oscillations solutions are obtained through a Hopf Bifurcation. The parameter
Hopf bifurcation Hopf bifurcation values for the appearance of bistability also depend on the
n 1 (Boolean form) Steady state Steady state degradations rates i and strengths of repression effects i . We
stable oscillations stable oscillations
note that, in the m-repressilator models, the effect of cooperativity
introduced through the Hill parameter n and delays can have the
same dynamical effect. This is true for m Z 2, as can be seen from
and m3, the asymptotic solutions of the system of Eqs. (14) in the
the qualitative results in Tables 2 and 3.
limiting case n-1 have the same phase space topology as the
The introduction of delays can be used to distinguish between
solutions of the same family of equations for nite but large values of
the eukaryote from the prokaryote protein regulation. The
the Hill parameter n.
1-repressilator model with delay is the simplest mechanism that
We analyse now the general properties of the asymptotic solu-
can originate oscillations in eukaryote organisms or in organisms
tions of the m-repressilator model in limit n-1 and for m Z 2.
where there exists a sufciently large delay between mRNA
In this case, by (14), the Boolean m-repressilator model has the form
transcription and protein synthesis. As the 1-repressilator model
x_ i gxi 1 t  r  i xi t; i 1; ; m 24 involves the synthesis of only one protein, due to its simplicity,
it may be important in synthetic biology applications. The impor-
where xm 1 x1 , gx 1, for x r 1, and gx 0, otherwise. tance of the 2-repressilator model relies on the fact that bistability
is a switch mechanism tuned by initial conditions.
Theorem 7. Consider the system of Eqs. (24) with r 0 and i 4 0, Some of the results obtained in this paper depend on the
for i 1; ; m. Then, we have parameterisation obtained through the Hill cooperative parameter
n. If n 1, neither oscillations nor bistability exists. On the contrary,
(a) If m Z 2, m is even and i o 1, for every i 1; ; m, then the for n 4 1 and with delays, these phenomena do appear. Synthetic
system of Eqs. (24) has two xed points with coordinates, biology models have been used to validate qualitatively the Hill
1=1 ; 0; 1=3 ; ; 0 embedding of the 3-repressilator model, Elowitz and Leibler
(2000). For other models involving activation and repression of
0; 1=2 ; 0; ; 1=m : genes, it is not straightforward to discover how to do similar
These xed point are of stable node type. embeddings. Confrontation of the literature on enzyme kinetics
(b) If m Z 3, m is odd and i o 1, for every i 1; ; m, then the with the systems biology approach leads to the conclusion that
system of Eqs. (24) has no xed points. experiments leading to the quantitative calibration of models of
(c) If m Z 1 and i Z 1, for every i 1; ; m, then the system of protein synthesis are indeed needed. Synthetic biology gives a
Eqs. (24) has a unique xed point with positive coordinates, good experimental framework to calibrate and validate experi-
mentally m-repressilator models.
1=1 ; 1=2 ; ; 1=m
and this xed point is of stable node type.
Acknowledgments
For the case (a) in Theorem 7, m is even, and the solutions of
the piecewise linear system of Eqs. (24) show bistability. For the I would like to thanks the comments and suggestions of one of
case (b), the numerical solutions of the piecewise linear system of the anonymous reviewers of this paper. I would like to thank IHS
 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208 207

for hospitality and support, and Fundao para a Cincia e a Proof of Corollary 3. By Theorem 2 and Lemma 1, the delay
Tecnologia for support. equation (7) has a stable oscillatory solution if it is possible to nd
a value for the parameter 1 such that, s1 A  =2; .
As 1  1= cos s1 , this implies that 1 A 1; 1. As, for n 1,
p
Appendix A 1  xn ; 1 1 1=2  1 1=4 2 o 1, for every 4 0, no
solutions with 4 1 4 1 exist. As 4 0, it is always possible to
Proof of Lemma 1. Let g  1  e  s . As g o 0 for every nd sufciently large values of s such that, e  s  1 =s o o 1 .
A R, the characteristic equation (10) has one real negative root n Therefore, by Lemma 1(c), damped oscillatory solutions exist. This
only if, o 0 and g 1. These conditions imply that
n n proves the rst part of the corollary.
n log s=s, with s o 1 and s 4 0. With the characteristic To prove the second part, we rst note that 1  xn ; n r 1,
equation condition g , we obtain
n n since xn is a solution of 1=1 xn x. So, as ; n
8 n1 xn ; n rn, in the limit -0, ; n n. On the other hand,
> 1 with  tan s1 1 and u s1 , we have  tan u 1 =s, and in
< e1s
>
s the limit s-1, the solution of  tan s1 1 converges to
> n 1
> s1 s1 , and, in the same limit, 1 -1. Therefore, we have proved
: log s 
s s that, for any n 4 1, there exists always a sufciently small and a
and (b) is proved. If oe  1  s =s, the characteristic equation has sufciently large delay s, such that the delay equation (7) has an
two negative real roots and (a) is also proved. asymptotically stable periodic solution (limit cycle).
For 4 e  1  s =s, the characteristic equation has two imaginary
Proof of Corollary 4. In the limit n-1, the delay equation (7)
root. This is straightforwardly proved by expanding the character-
reduces to x_ gx  x, where gx 1 for x r 1, and gx 0,
istic equation in series around . In this case, if  e  1  s =s is
n

otherwise. For the initial condition, xt t 0, with t A

positive and sufciently small, these complex roots have negative
 s; 0, the solution is xt 1  e  t , for t Z0, converging mono-
real parts. Let us show now that, if increases, these two
tonically to the steady state xn 1.
characteristic roots cross the imaginary axis of the complex plane.
For that, we assume that there exists a solution of the character-
Proof of Lemma 5. For m 1, the lemma is straightforward. Let us
istic equation of the form, i, with 4 0. So, the real and
prove it in general for any nite m Z1. The coordinates of the xed
imaginary parts of the characteristic equation obey to
points of Eq. (14), if they exist, are solutions of the equations
(
1 cos s 0 1 1
sin s ; xi f i 1 xi 1 ; i 1; ; m
i 1 i xni 1
Solving the above equations for and , for every s 4 0, there where xm 1 x1 , i 4 0 and i 40, for i 1; ; m. The functions
exists a countable number of solutions in and . With 4 0 and f i 1 x : R -0; 1=i  are strictly decreasing, positive and smooth
4 0, the solutions in are the unique roots of the equations, and f i 1 0 1=i . Then, the rst coordinate of a xed point of
 tan sq m q, where q A   =2s q =s; q =s with q 1; Eq. (14) is a root of the equation
3; (odd), and q 1= cos q s. For q 1, (c) and (d) are proved.
x1 f 2 Jf 3 Jf m Jf m 1 x1 Fx1
If q 41, the other roots cannot be reached by a continuous pass in
the complex plane and so 1 is the unique value of the parameter where Fx1 : R -0; 1=1  and F0 4 0.
for which the eigenvalues of the characteristic polynomial cross If m is odd, Fx1 is strictly decreasing, positive and smooth.
the imaginary axis. This proves (e). As FF0 F0 F0F, where, F o 0 and A 0; F0, then,
we have FF0 F0 o 0 and FF0 o F0. So, by the continuity
Proof of Theorem 2. The proofs of (a) and (b) follow from Lemma 1
of Fx1 , there exists a unique xn1 A F0; FF0, such that
and the analysis of the roots of the characteristic polynomial of the
Fxn1 xn1 . This solution is the rst positive coordinate of the
linearised delay equation (9). (c) and (d) are a result of the Hopf
unique xed point of Eq. (14). The other positive coordinates are
bifurcation theorem for delay equations, Hale (1977), Smith (2011).
the unique solutions of, xi f i 1 xi 1 , with i 1; ; m, where
To show that the Hopf bifurcations is indeed supercritical, we
xm 1 x1 .
calculate now the derivatives at bifurcation, 1 . Let i
If m is even, Fx1 is strictly increasing, positive and smooth. But
be a root of the characteristic equation p. The real part of is a root
as Fx1 has values in the interval 0; 1=1 , we have F1=1 r 1=1 .
of the function h 1 e  s cos s, and  . As we have
As F0 4 0, then, in the interval 0; 1=1 , there exists at least one
seen in the proof of Lemma 1, h 0 has the solution 0. So
solution of the equation Fx1 x1 . The other coordinates are
developing h in Taylor series around 0 and keeping only the
determined as previously.
rst order terms, h 1 cos s 1  s cos s, the local
behaviour of near 0 is Proof of Lemma 6. Under the conditions of the lemma, the
1 cos s system of equations has a Hopf bifurcation if the characteristic
 polynomial (22) has one negative real root  1 and two pure
1 s cos s
imaginary roots, 2;3 7 i. In this case, p pn
3

and 1 1 . Comparing the characteristic polynomial (22)

2 2 2

1 s cos s with pn , and solving for and 1 , we obtain, 1 2 3 1 , where

 ; 1 1 2 1 3 2 3 . For 1 2 3 Z 1 , the Hopf bifurca-
1  s cos s2
tion theorem applies. The second part of the lemma follows from
For 0, cos s 0 cos 1 s  1=1 , we obtain the fact that for 1 2 3 0, the characteristic polynomial (22) has
  three real negative roots.
1
4 0;
0 1 1 s
Proof of Theorem 7. To calculate the coordinates of the xed
By the Hopf bifurcation theorem, this last condition guarantees that points of the system of Eqs. (24), the right hand sides of Eqs. (24)
the Hopf bifurcation at 1 is supercritical, proving (c) and (d). are equated to zero. So, we have gxi 1  i xi 0, where gx 1,
208 R. Dilo / Journal of Theoretical Biology 340 (2014) 199208

for x r 1, and gx 0, otherwise. The solution of this equation is Allwright, D.J., 1977. A global stability criterion for single control loops. Journal of
8 Mathematical Biology 4, 363373.
>
< 1 if xi 1 r 1 Alves, F., Dilo, R., 2005. A simple framework to describe the regulation of gene
xi i 25 expression in prokaryotes. Comptes Rendus Biologies 328, 429444.
>
: 0 if x Bizwanger, H., 2008. Enzyme Kinetics. Wiley, Weinheim.
i 1 4 1; Buse, O., Prez, R., Kuznetsov, A., 2010. Dynamical properties of the repressilator
model. Physical Review E 81, 066206.
Chen, L., Aihara, K., 2002. A model of periodic oscillations for genetic regulatory
To prove (a), we assume now that i o 1, for every i 1; ; m. systems. IEEE Transactions on Circuits and Systems Part I: Fundamental Theory
and Applications 49 (10), 14291436.
In this case, we have two possible solutions Chen, L., Wang, R., Kobayashi, T., Aihara, K., 2004. Dynamics of gene regulatory
  networks with cell division cycle. Physical Review E 70, 011909.
1
yn1 ; Crick, F., 1967. The genetic code. Proceedings of the Royal Society of London, Series
1 B 167, 331347.
Crick, F., 1970. The central dogma of molecular biology. Nature 227, 561563.
yn1 0; ; Davidich, M., Bornholdt, S., 2008. The transition from differential equations to
Boolean networks: a case study in simplifying a regulatory network model.
Using the same rule (25) for i 2, we obtain Journal of Theoretical Biology 255, 269277.
  Dilo, R., Muraro, D., 2010. A software tool to model genetic regulatory networks.
1 Applications to the modeling of threshold phenomena and of spatial patterning
yn1 ; 0;
1 in Drosophila. PLoS One 5 (5), e10743.
Elowitz, M.B., Leibler, S., 2000. A synthetic oscillatory network of transcriptional
  regulators. Nature 403, 335338.
1 Fall, C., Marland, E., Wagner, J., Tyson, J., 2002. Computational Cell Biology.
yn2 0; ; ;
2 Springer-Verlag, New York.
Gardiner, C.W., 1997. Handbook of Stochastic Processes. Springer-Verlag, Berlin.
Otherwise we arrive at a contradiction. In fact, making the choice Gillespie, D.T., 1977. Exact stochastic simulation of coupled chemical reactions.
yn1 1=1 ; 1=2 ; , then x2 1=2 4 1, which contradicts the Journal of Physical Chemistry 81, 23402361.
Gillespie, D.T., 1992. A rigorous derivation of the chemical master equation. Physica
rst condition in (25). For yn2 the argument is similar. So, by A 188, 404425.
induction Glass, L., Kauffman, S.A., 1973. The logical analysis of continuous, nonlinear
  biochemical control networks. Journal of Theoretical Biology 39, 103129.
1 1 Hale, J., 1977. Theory of Functional Differential Equations. Springer-Verlag, New York.
yn1 x1 ; x2 0; x3 ;
1 3 van Kampen, N.G., 2007. Stochastic Processes in Physics and Chemistry. North-
Holland, Amsterdam.
  Kuznetsov, A., Afraimovich, V., 2012. Heteroclinic cycles in the repressilator model.
1
yn2 x1 0; x2 ; x3 0; ; Chaos, Solitons & Fractals 45, 660665.
2 Larson, D.R., Singer, R.H., Zenklusen, D., 2009. A single molecule view of gene
expression. Trends in Cell Biology 19 (11), 630637.
If m is even, and as xm 1 x1 and m 1 1 , both solutions are Mackey, M.C., Glass, L., 1977. Oscillation and chaos in physiological control systems.
compatible with the parameter signs in the correspondence table. Science 197, 287289.
Marzen, S., Garcia, H.G., Phillips, R., 2013. Statistical mechanics of MonodWyman
Therefore, the system of Eqs. (24) has two xed points. As the Changeux (MWC) models. Journal of Molecular Biology 425, 14331460.
Jacobian matrix of the Eqs. (24) calculated at the xed points is Mincheva, M., Roussel, M.R., 1977. Graph-theoretic methods for the analysis of
diagonal with negative entries ( i ), both xed points are of chemical and biochemical networks. II. Oscillations in networks with delays.
Journal of Mathematical Biology 55, 87104.
stable node type. This proves (a) for m Z2.
Monk, N.A.M., 2003. Oscillatory expression of Hes1, p. 53, and NF-kappaB driven by
If m is odd, the solutions for xm 1 x1 are not compatible with transcriptional time delays. Current Biology 13, 14091413.
the parameter signs in the correspondence table, and the system Mller, S., Hofbauer, J., Endler, L., Flamm, C., Widder, S., Schuster, P., 2006.
A generalised model of the repressilator. Journal of Mathematical Biology 53,
of Eqs. (24) has no xed points. This proves (b).
905937.
To prove (c), we use the same argument for i Z 1, for every Murray, J.D., 1989. Mathematical Biology. Springer-Verlag, Berlin.
i 1; ; m. The result follows from the fact that, i Z 1; i 1 O'Brien, T., Lis, J.T., 1993. Rapid changes in Drosophila transcription after an
Z1xi 1=i ; xi 1 1=i 1 , for every i 1; ; m. instantaneous heat shock. Molecular and Cellular Biology 13, 34563463.
Smith, H., 2011. An Introduction to Delay Differential Equations with Applications
to the Life Sciences. Springer-Verlag, New York.
References Strelkowa, N., Barahona, M., 2011. Transient dynamics around unstable periodic
orbits in the generalized repressilator model. Chaos 21, 023104.
Zhu, R., Ribeiro, A.S., Salahub, D., Kauffman, S.A., 2007. Studying genetic regulatory
Alberts, B., Johnson, A., Lewis, J., Raff, M., Roberts, K., Walter, P., 2008. Molecular networks at the molecular level: delayed reaction stochastic models. Journal of
Biology of the Cell, 5th ed. Taylor & Francis, Abingdon. Theoretical Biology 246, 725745.