Genetic regulatory networks are dynamic systems which describe the interactions among gene products (mRNAs and proteins). The internal states of a genetic regulatory network consist of the concentrations of mRNA and proteins involved in it, which are very helpful in understanding its dynamic behaviors. However, because of some limitations such as experiment techniques, not all internal states of genetic regulatory network can be effectively measured. Therefore it becomes an important issue to estimate the unmeasured states via the available measurements. In this study, we design a state observer to estimate the states of genetic regulatory networks with time delays from available measurements. Furthermore, based on linear matrix inequality (LMI) approach, a criterion is established to guarantee that the dynamic of estimation error is globally asymptotically stable. A gene repressillatory network is employed to illustrate the effectiveness of our design approach.
1. Introduction
Recently nonlinear differential equations have been proposed to model genetic regulatory networks. Based on this model, stability of genetic regulatory networks has been intensively studied, which is believed useful in designing and controlling genetic regulatory networks. In [1], sufficient and necessary local delay-independent stability conditions are given for several types of simplified genetic regulatory networks with a single time delay. In [2, 3], we present some sufficient and necessary conditions of local delay-independent stability conditions for general genetic regulatory networks with a single time delay and multiple time delays. Some sufficient conditions for global stability of genetic regulatory networks have been derived based on LMI approaches [4–6] and M-matrix theorem [7, 8].
On the other hand, to understand the dynamic behavior of genetic regulatory networks, measurements of all internal states are very useful. The internal states of a genetic regulatory network consist of the concentrations of mRNA and proteins involved in it. However, because of some limitations such as experiment techniques, not all internal states of genetic regulatory network can be effectively measured. As a result, the internal states of genetic regulatory networks cannot be completely available. Therefore, the state estimation problem can play an important role in understanding the dynamic behaviors of genetic regulatory networks. The state estimation problem addressed is to estimate the states based on available output measurements such that the dynamic of estimation error is globally asymptotically stable. Actually, the state estimation methods have been very important in understanding, designing, and controlling dynamic systems such as engineering control system [9], neural networks [10, 11], and complex systems [12].
In this study, we will study the state estimation of genetic regulatory networks with time delays modeled by nonlinear differential equations. Section 2 briefly describes delayed genetic regulatory networks with SUM regulatory logic. In Section 3 we design a full-order state observer to estimate the states of delayed genetic regulatory networks. Some properties of this observer are discussed. In Section 4, based on LMI approach we establish a sufficient condition under which the dynamic of estimation error for designed state observer is asymptotically and delay-independently stable. In Section 5, a gene repressillatory network is employed to illustrate the effectiveness of our approach described in Section 4. Section 6 gives our conclusion of this study and points out some directions of future work.
2. Delayed Genetic Regulatory Networks
A delayed genetic regulatory network consisting of n mRNAs and n proteins can be described by the following equations:
(1)m˙i(t)=-kmimi(t)+ci(p(t-τp))p˙i(t)=-kpipi(t)+rimi(t-τm)fori=1,2,…,n,
where mi(t), pi(t)∈R+n represent the concentrations of mRNA i and protein i, respectively. kmi and kpi are positive real numbers that represent the degradation rates of mRNA i and protein i, respectively. ri is a positive constant representing the rate of translating mRNA i to protein i. ci(p(t,τp)) is a nonlinear function of p1(t-τp),…,pn(t-τp) representing the regulation function of gene i. Both τm and τp are positive constants indicating time delays of mRNAs and proteins, respectively.
The bottom equation in model (1) describes the translational process. The term rimi(t) reflects the fact that one kind of proteins is translated only from one kind of mRNA molecules. The top equation in model (1) describes the transcriptional process. One gene or mRNA is generally activated or repressed by multiple proteins in the transcriptional process indicated in the definition of ci(p(t)). In this paper, we take ci(p(t))=∑j=1ncij(pj(t)), which is called the “SUM” logic [13]. That is, each transcription factor acts additively to regulate gene i. The SUM logic is applicable if one gene can be regulated by several proteins independently by binding with different promoters or by a family of similar proteins independently binding to one promoter. In many natural gene networks, this SUM logic does exist [13]. The regulation function cij(pj(t)) is a function of the Hill form [14] as follows:
(2)cij(pj(t))=aij11+(pj(t)/bj)hj
if transcription factor j is a repressor of gene i, or
(3)cij(pj(t))=aij(pj(t)/bj)hj1+(pj(t)/bj)hj
if transcription factor j is an activator of gene i, where aij and bj are nonnegative constants and hj is the Hill coefficient representing the degree of cooperativity. In this study, assume that hj≥1. Note that
(4)11+(pj(t)/bj)hj=1-(pj(t)/bj)hj1+(pj(t)/bj)hj.
Then system (1) can be rewritten as follows:
(5)m˙(t)=-Kmm(t)+Gg(p(t-τp))+Lp˙(t)=-Kpp(t)+Rm(t-τm),
where m(t)=(m1(t),…,mn(t)) and p(t)=(p1(t),…,pn(t)); Km=diag(km1,…,kmn), Kp=diag(kp1,…,kpn), and R=diag(r1,…,rn); G=(Gij) is an n×n stoichiometric matrix representing regulatory relationships of the network, which is defined as follows: Gij=0 if transcription factor j does not directly regulate gene i, Gij=aij if transcription factor j directly activates gene i, and Gij=-aij if transcription factor j directly represses gene i; L=(l1,…,ln) where li is a constant and is defined as li=∑j∈Repaij, where Rep is the set of repressors of gene i. g=(g1,…,gn) where gj(u)=(u/bj)hj/[1+(u/bj)hj] is a monotonically increasing function. Obviously these functions with hj≥1 have the continuous derivatives for u≥0. From calculus, we have
(6)θj=maxu≥0gj′(u)=(hj-1)(hj-1)/hj(hj+1)(hj+1)/hj4bjhj>0.
3. State Observer
In practice, the information about the network states is often incomplete from the experimental measurements. For example, the concentrations of proteins might be immeasurable because of the limitation of measurement techniques. Our purpose of this study is to develop an efficient estimation system (called a state observer) in order to estimate the network states from the available measurements. In this paper, assume that measurements are the linear combinations of mRNA and protein concentrations and thus the output can be expressed as follows:
(7)z(t)=C[m(t)p(t)],
where z(t) is an m-dimensional vector representing the measurements and C is an m×2n observation matrix. Unless the rank of matrix C in (7) is 2n, the states of system (5) cannot be exactly estimated from the static observation equation (7) only. In practice, the rank of matrix C in (7) is less than 2n. To approximately estimate the states of a dynamic system, a dynamic system similar to the original one is designed to estimate the states. In this paper, the full-order state estimator of network (5) is designed as follows:
(8)[m^˙(t)p^˙(t)]=-[Km00Kp][m^(t)p^(t)]+[Gg(p^(t-τp))+LRm^(t-τm)]+D(z(t)-C[m^(t)p^(t)]),
where m^(t) and p^(t) are the estimation of states and D is 2n×m estimate gain matrix to be determined.
Let the estimation error be
(9)x(t)=m(t)-m^(t)y(t)=p(t)-p^(t).
Then from (5), (8), and (9), the error system can be described as follows:
(10)[x˙(t)y˙(t)]=-(K+DC)[x(t)y(t)]+[Gf(t-τp)Rx(t-τm)],
where K=diag(Km,Kp) and f(t)=g(p(t))-g(p^(t)).
Now designing the state estimator for network (5) is reduced to find the estimate gain matrix D such that the error system (10) is globally asymptotically stable. From (6), we have
(11)0≤fj(t)yj(t)≤θjforj=1,2,…,n.
Furthermore, for any nonnegative diagonal Λ=diag(λ1,…,λn)≥0, from (11) it follows that
(12)-fT(t-τp)2Λf(t-τp)+yT(t-τp)2ΛΘf(t-τp)≥0,
where Θ=diag(θ1,…,θn).
Once matrix D is determined, the estimations m^(t) and p^(t) are numerically calculated from (8). That the same technique can be applied for solving (1) directly is the same as solving system (8) with D=0, which results in an estimation error system (10) with D=0. If the system (1) is unstable and the values of m^(0) and p^(0) are different from their true counterparts, then the estimation errors will be exponentially increased. Even if the values of m^(0) and p^(0) are the exact same as their true counterparts, the round-off errors can also cause the estimation errors to be exponentially increased. Therefore, in practice it is important to design matrix D to make sure the estimation error system is stable. Theorems 1 and 2 in next section will guarantee that, for any values of m^(0) and p^(0), the estimation errors will be asymptotically converged to zero.
4. Main Results and Proofs
In this section we will first derive the conditions under which the error system (10) is globally asymptotically stable for a given estimate gain matrix.
Theorem 1.
For a given estimate gain matrix D, the error system (10) has a unique equilibrium state x=0 and y=0 and is globally asymptotically stable if there exist 2n×2n positive definite matrices P and n×n positive definite matrices Q and S and positive diagonal matrix Λ=diag(λ1,…,λn)>0, such that the following LMI holds:
(13)Ω=[Ω11Ω120Ω12TΩ22Ω230Ω23T-S]<0,
where Ω11=-(K+DC)TP-P(K+DC)+diag(Q,S), Ω22=-diag(2Λ,Q), Ω12=Pdiag(G,R), and Ω23=[ΛΘ,0]T.
Proof.
Consider the following Lyapunov-Krasovskii functional:
(14)V(x(t),y(t))=V1(x(t),y(t))+V2(x(t),y(t)),
where
(15)V1(x(t),y(t))=[xT(t)yT(t)]P[x(t)y(t)],V2(x(t),y(t))=∫t-τmtxT(u)Qx(u)du+∫t-τptyT(u)Sy(u)du.
Differentiating Vi(x(t),y(t)) defined above along the trajectories of system (10), we have
(16)V˙1(x(t),y(t))=2[xT(t)yT(t)]P[x˙(t)y˙(t)]=-2[xT(t)yT(t)]P(K+DC)[x(t)y(t)]+2[xT(t)yT(t)]×Pdiag(G,R)[f(t-τp)x(t-τm)],V˙2(x(t),y(t))=xT(t)Qx(t)-xT(t-τm)Qx(t-τm)+yT(t)Sy(t)-yT(t-τp)Sy(t-τp).
Taking inequality (12) into consideration, we have
(17)V˙(x(t),y(t))≤V˙1(x(t),y(t))+V˙2(x(t),y(t))-fT(t-τp)2Λf(t-τp)+yT(t-τp)2ΛΘf(t-τp)=[xT(t)yT(t)](-2P(K+DC)+diag(Q,S))[x(t)y(t)]+2[xT(t)yT(t)]Pdiag(G,R)[f(t-τp)x(t-τm)]-[fT(t-τp)xT(t-τm)]diag(2Λ,Q)[f(t-τp)x(t-τm)]+2yT(t-τp)[ΛΘ0][f(t-τp)x(t-τm)]-yT(t-τp)Sy(t-τp)=ξT(t)Ωξ(t)<0,
where ξ(t)=[(xT(t),yT(t)),(fT(t-τp),xT(t-τp)),yT(t-τp)]T.
From Lyapunov-Krasovskii theory [15], the error system (10) is globally asymptotically stable. From (10), x=0 and y=0 are an equilibrium state. To prove the uniqueness of the equilibrium state of the error system (10), here we use proof-by-contradiction technique. Note that Lyapunov-Krasovskii functional (18) associated with the error system (10) is independent of the equilibrium state. Therefore if the error system (10) has another equilibrium state, it is also globally asymptotically stable, which is not possible.
In Theorem 1, for a given estimate gain matrix D, the stability condition of the error dynamic system (10) is established in terms of linear matrix inequality (LMI) which can be solved by standard MATLAB function. If matrix D is unknown, matrix inequality (13) becomes nonlinear in matrices, P, D, Q, S, and Λ, which is not easy to be solved. However, let PD=-T; then matrix inequality (13) becomes linear in matrices, P, T, Q, S, and Λ. Therefore, we have the following theorem.
Theorem 2.
If there exist 2n×2n positive definite matrices P and n×n positive definite matrices Q and S, positive diagonal matrix Λ=diag(λ1,…,λn)>0, and an 2n×m matrix T such that the following LMI
(18)Ω=[Ω11Ω120Ω12TΩ22Ω230Ω23T-S]<0
holds, where Ω11=-KP-PK+CTTT+TC+diag(Q,S), and sub-matrices Ω22, Ω12, and Ω23 are the same as in Theorem 1, then with the estimator gain matrix
(19)D=-P-1T
the error system (10) has a unique equilibrium state x=0 and y=0 and is globally asymptotically stable.
Proof of Theorem 2 is straightforward from Theorem 1 and thus is omitted here.
5. An Illustration Example
In this section, we employ the gene repressilatory network to show the effectiveness and correctness of our theoretical results. The gene repressilatory network consists of three genes and three proteins (lacl, tetR, and cl), each repressing the transcription of its downstream partner [16] as shown in Figure 1. This network without time delays has been studied theoretically and experimentally in [16]. The delay-independent local and global stability of this gene repressilatory network with time delays has widely been studied in [1–8].
Structure of gene repressilatory network.
The mathematical model of this gene repressilatory network with time delay is described by the following equation:
(20)m˙i(t)=-kmmi(t)+a1+pi-1h(t-τp),p˙i(t)=-kppi(t)+rmi(t-τm),
where km, a, kp, and r are positive constants and subscript 0=3.
In this study we consider gene repressillatory network (20) with the values of parameters set as follows: h=2, km=1.2, a=2.5, kp=1, and r=0.8. For system (20) with these parameter specifications, we have
(21)Km=[1.20001.20001.2],G=[0-2.5000-2.5-2.500],Kp=[100010001],R=[0.80000.80000.8].
And θj=33/8 for j=1,2,3.
Case A.
Assume that the concentration of all proteins is unable to be measured. The observation matrix C is
(22)C=[100000010000001000].
By using MATLAB LMI toolbox, we solve LMIs (18) with the above data for P, T, Q, S, and Λ and obtain
(23)P=[1.87500000001.87500000001.87500000008.38070000008.38070000008.3807],T=-[10.124800010.124800010.1248000000000],Q=[9.28100009.28100009.2810],S=[7.94970007.94970007.9497],Λ=[6.99630006.99630006.9963].
Therefore, we have
(24)D=-P-1T=[5.40005.40005.4000000000].
Figure 2 depicts the estimation errors of protein concentrations of delayed genetic regulatory network (20) with specified parameters in the caption of Figure 2. From Figure 2, it can be seen that in six minutes the estimated protein concentrations are exactly the same as the true protein concentrations although they are not measured. The time that needs to exactly estimate the true states depends on the initial errors between the true states and estimated states (which are random guesses in practice). In Figure 2, the initial errors of protein estimations range from 0.1 to 0.6. If the initial errors are zero, the estimated state would be the exact true states from beginning on.
Estimation errors of protein concentrations of system (20) with specified parameters and τp=τm=0.5 minutes while mRNA concentrations are available.
Case B.
Assume that the concentration of all proteins is able to be measured while we would like to estimate the gene expressions. The observation matrix C becomes
(25)C=[000100000010000001].
By using MATLAB LMI toolbox, we solve LMIs (18) with above data for P, T, Q, S, and Λ and obtain
(26)P=[19.374700000019.374700000019.374700000021.987900000021.987900000021.9879],T=-[00000000053.145100053.145100053.1451],Q=[17.659200017.659200017.6592],S=[75.901700075.901700075.9017],Λ=[69.218900069.218900069.2189].
Therefore, we have
(27)D=-P-1T=[0000000002.41700002.41700002.4170].
Figure 3 depicts the estimation errors of mRNA concentrations of delayed genetic regulatory network (20) with specified parameters in the caption of Figure 3 and knowing protein concentrations. From Figure 3, it can be seen that in about ten minutes the estimated mRNA concentrations can pretty well approximate the true mRNA concentrations although they are not measured. In Figure 3, the initial errors of protein estimations range from 0.2 to 0.5.
Estimation errors of mRNA concentrations of system (20) with specified parameters and τp=τm=0.5 minutes while protein concentrations are available.
6. Conclusion and Future Work
In this paper, we have studied the state estimation of genetic regulatory networks with time delays. Based on LMI approach, a full-order state observer is designed to estimate the states from incomplete measurements so that the state estimation error is globally asymptotically stable. The theorems presented in this paper have been illustrated by the gene repressillatory network. The simulation results have verified that our designed observer can effectively estimate the unmeasured states. In this study, we assume that all parameters of genetic regulatory networks are available. In practice, some of parameters in networks may be unknown. One direction of our future work is to employ the extended Kalman filter [17] to estimate the known parameters and state of the systems simultaneously. Parameter uncertainties and noise perturbations exist in genetic regulatory networks [4, 6, 7, 16, 18] and measured outputs, which can affect the performance of state observer. The second direction of our future work is to design robust state observer for genetic regulatory networks with parameter uncertainties and noises. Typically measured outputs are sampled at a series of time points although state variables of genetic regulatory networks are continuous. The third direction of our future work is to design a state observer for genetic regulatory networks with discretized outputs.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.
Acknowledgments
This work was supported by Base Fund of Beijing Wuzi University and Fund for Beijing Excellent Team for Teaching Mathematics through LPT and by Natural Sciences and Engineering Research Council of Canada (NSERC) through FXW. The authors would like to thank the reviewers for their comments and suggestions.
ChenL.AiharaK.Stability of genetic regulatory networks with time delay20024956026082-s2.0-003657634010.1109/TCSI.2002.1001949WuF.-X.Delay-independent stability of genetic regulatory networks with time delays20091213192-s2.0-65249183854WuF.-X.Stability analysis of genetic regulatory networks with multiple time delaysProceedings of the 29th Annual International Conference of IEEE-EMBS, Engineering in Medicine and Biology Society (EMBC '07)August 2007fra138713902-s2.0-5764917643710.1109/IEMBS.2007.4352557LiC.ChenL.AiharaK.Stability of genetic networks with SUM regulatory logic: Lur'e system and LMI approach20065311245124582-s2.0-3454841178210.1109/TCSI.2006.883882LiC.ChenL.AiharaK.Synchronization of coupled nonidentical genetic oscillators20063137442-s2.0-3364551635310.1088/1478-3975/3/1/004RenF.CaoJ.Asymptotic and robust stability of genetic regulatory networks with time-varying delays2008714–68348422-s2.0-3864908380210.1016/j.neucom.2007.03.011WuF.-X.Global and robust stability analysis of genetic regulatory networks with time-varying delays and parameter uncertainties2011543913982-s2.0-7996098587310.1109/TBCAS.2011.2124459WuF.-X.Delay-independent stability of genetic regulatory networks20112211168516932-s2.0-8045517460910.1109/TNN.2011.2165556SinhaA.2007New York, NY, USACRC PressWangZ.HoD. W. C.LiuX.State estimation for delayed neural networks20051612792842-s2.0-1384425669510.1109/TNN.2004.841813HeY.WangQ.-G.WuM.LinC.Delay-dependent state estimation for delayed neural networks2006174107710812-s2.0-3374691385610.1109/TNN.2006.875969LiuY. Y.SlotineJ. J.BarabasiA. L.Observability of complex systems2013110724602465KalirS.ManganS.AlonU.A coherent feed-forward loop with a SUM input function prolongs flagella expression in Escherichia coli2005110.1038/msb41000102-s2.0-33746428728NielsenJ.VilladsenJ.LidenG.20032ndNew York, NY, USAKluwer Academic/Plenum PublishersKolmanovskiiV. B.MyshkisA. D.1999Dodrecht, The NetherlandsKluwer AcademicElowitzM. B.LeibierS.A synthetic oscillatory network of transcriptional regulators200040367673353382-s2.0-003468817310.1038/35002125SimonD.2006New Jersey, NJ, USAWiley-InterscienceTianL. P.ShiZ. K.LiuL. Z.WuF. X.M-Matrix based stability conditions for genetic regulatory networks with time-varying delays and noise perturbations201375214222