betaFeedback

WorksComputer Science

Efficient Analysis of Chemical Reaction Networks Dynamics based on Input-Output Monotonicity

Abstract

Motivation: A Chemical Reaction Network (CRN) is a set of chemical reactions, which can be very complex and difficult to analyze. Indeed, dynamical properties of CRNs can be described by a set of non-linear differential equations that rarely can be solved in closed-form, but that can instead be used to reason on the system dynamics. In this context, one of the possible approaches is to perform numerical simulations, which may require a high computational effort. In particular, in order to investigate some dynamical properties, such as robustness or global sensitivity, many simulations have to be performed by varying the initial concentration of chemical species. Results: In order to reduce the computational effort required when many simulations are needed to assess a property, we exploit a new notion of monotonicity of the output of the system (the concentration of a target chemical species at the steady-state) with respect to the input (the initial concentration of another chemical species). To assess such monotonicity behavior, we propose a new graphical approach that allows us to state sufficient conditions for ensuring that the monotonicity property holds. Our sufficient conditions allow us to efficiently verify the monotonicity property by exploring a graph constructed on the basis of the reactions involved in the network. Once established, our monotonicity property allows us to drastically reduce the number of simulations required to assess some dynamical properties of the CRN.

Introduction

From the discovery of DNA structure, in 1953, there has been a growing interest in understanding the morphological and functional organization of living cells [16]. Cells are very complex to analyze because they consist of many components that interact with each other, through multiple sequences of chemical reactions, Chemical Reaction Networks (CRNs), which regulate the overall behavior. Besides, several fluctuations can alter the cell functionalities, such as internal errors propagation and variation in the concentrations of chemical species.

Bioinformatics and systems biology emerge as powerful tools to investigate CRN dynamics merging computational methods and real data. Through simulations, for instance, it is possible to mimic the internal dynamics of a natural system and, therefore, to predict its behaviour. Moreover, model-based analysis techniques can be used to interpret some less intuitive features of the system.

In this context, computer scientists developed many formalisms to study systems of interacting components, which can be applied to model and describe CRNs and, in general, biological systems. Among these formalisms, those that have been applied in systems biology [7] include Petri nets [6, 22, 18], Hybrid systems [1, 13] (Li et al. 2017), process calculi such as the π\pi-calculus [25] as well as many ad-hoc biologically inspired calculi [9] and rule-based systems [5].

Corresponding author: Lucia Nasti

[email protected]

The development of models that can help in predicting the system behaviour requires precise and detailed information about the set of initial conditions of the biological system under study. In many cases, obtaining such precise informations is unfortunately very challenging (or even impossible) because of the noisy nature of biological data [10]. Moreover, some parameters may be affected by fluctuations that alter ordinary system behavior. Finally, in many cases such precise informations simply cannot be measured.

In order to determine which model parameters are more critical in case of perturbations and approximate measurements, it is possible to apply sensitivity analysis methods, which give a measure of the behavioral change of the system under perturbation of one of its parameters. As underlined in [15, 31], there exist two fundamental approaches to sensitivity analysis, local and global. While the local sensitivity analysis investigates the effects of small perturbations, the global one studies the effects of large perturbations and determines also the most (or the least) influencing parameters.

Sensitivity analysis methods (in particular, the global ones) typically require performing many simulations, by varying the system parameters one by one. These methods are, in general, quite expensive, because of the number of parameters and the large range of values to be tested.

For these reasons, alternative dynamical properties of CRNs, such as monotonicity [3, 12] and steady-state reachability [11], have been studied. Establishing such properties, indeed, provides information on the CRN dynamics without the need of performing several numerical simulations [23].

Monotonicity, in particular, is a property stating that a given measurable aspect of the system dynamics increases (or decreases) with the increase of a given system parameter. Many more specific definitions of monotonicity exist [3, 12] and in some cases they can be tested simply by inspecting the structure of the system models. It is worth noting that the validation of many different biological properties, such as robustness [16, 26, 29], persistence [4], and adaptation [28], greatly benefits from the assessment of monotonicity properties of the network, since this typically allows reducing the number of cases to be analyzed through numerical simulations.

Consider, for example, the robustness property. It is observed in many biological systems and it expresses the ability of the system to preserve its functions despite the presence of perturbations (Kitano 2007). Without any assumption on monotonicity, verifying robustness would require, in general, to consider all possible perturbations, usually expressed as different initial states of the system. In particular, regarding the CRNs, it would be necessary to test the system behaviour by examining all the possible combinations of initial concentrations of chemical species and, in practice, this would require a huge (in principle, infinite) number of simulations [24].

The same reasoning applies also to the case in which the initial concentrations of a biological network are simply unknown. Without any assumption on monotonicity, understanding the qualitative system behaviour would require, in principle, to consider all possible initial concentrations.

In this paper we propose a sufficient condition for CRNs that, if satisfied, ensures that a form of monotonicity, called Input-Output monotonicity, holds. Given a species considered as the input of a CRN, and another species considered as the output, we say that input and output are in a monotonicity relation if the concentration of the output species always increases (or decreases) in response to an increase in the initial concentration of the input. If this monotonicity relationship holds, than it is possible to consider an interval of initial concentrations for the input species and obtain the corresponding concentrations interval for the output species by performing only two simulations, one for each extreme value of the input initial concentration [12].

The sufficient condition we propose is based on a condition on the structure of the CRN that can be efficiently evaluated, without the need of performing any simulation. Following the lines of [3], our condition is based on a graph representation of the CRN enriched with information about cooperation and competition among reactions and it is expressed as a set of constraints on the graph structure.

The paper is organized as follows. In the rest of this section we recall some background definitions of CRNs and related concepts. In § 2 we define Input-Output monotonicity and state Theorem 5 expressing our sufficient condition. In § 3, we apply our methodology to study the case of the ERK signaling pathway. In § 5 we draw our conclusions and discuss future work. Finally, in § 4, we give the proof of Theorem 5.

Chemical Reaction Networks

A Chemical Reaction Network (CRN) is a set of reactions that can be formally defined as follows.

Definition 1 (Chemical reaction network). Given an indexed set of chemical species S=(S1,S2,…,Ss)\mathcal{S}=(S_1,S_2,\ldots,S_s), a Chemical Reaction Network is an indexed set of reactions R=(R1,R2,…,Rr)\mathcal{R}=(\mathcal{R}_1,\mathcal{R}_2,\ldots,\mathcal{R}_r) involving such species. Each chemical reaction is denoted as

Ri:∑j=1sαijSj⟶∑j=1sβijSjorRi:∑j=1sαijSj⇌∑j=1sβijSj,\mathcal{R}_i:\sum_{j=1}^{s}\alpha_{ij}S_j\longrightarrow\sum_{j=1}^{s}\beta_{ij}S_j \qquad\text{or}\qquad \mathcal{R}_i:\sum_{j=1}^{s}\alpha_{ij}S_j\rightleftharpoons\sum_{j=1}^{s}\beta_{ij}S_j,

where αij\alpha_{ij} and βij\beta_{ij} are non-negative integers called the stoichiometric coefficients.

The arrow is used to indicate the direction in which a chemical reaction takes place (from reactants to products). When we have a single arrow (→\rightarrow), the reaction is irreversible, namely that a reaction transforming products back into reactants cannot take place. When there is a double arrow (⇌\rightleftharpoons) it means that it is possible to have both the forward and the backward transformation, and then the reaction is reversible. In a CRN, a reversible reaction can be equivalently represented as a pair of irreversible reactions with opposite directions.

The stoichiometric coefficient matrix Γ∈Rs×r\Gamma\in\mathbb{R}^{s\times r} is defined as

Γji=βij−αij.\Gamma_{ji}=\beta_{ij}-\alpha_{ij}.

Note that we allow in our networks the presence of promoters: a promoter is a species SjS_j (such as an enzyme) that affects the rate of a reaction Ri\mathcal{R}_i but does not get produced nor consumed by it, i.e., Γji=0\Gamma_{ji}=0.

The rate of a reaction is expressed as a function of the concentrations of the species in the network. The vector of species concentrations at a given time is denoted by S=(S1,S2,…,Ss)′∈RsS=(S_1,S_2,\ldots,S_s)^{\prime}\in\mathbb{R}^{s}. The vector of reaction rates (which are functions of SS) is denoted by R(S)=(R1(S),R2(S),…,Rr(S))′∈RrR(S)=(R_1(S),R_2(S),\ldots,R_r(S))^{\prime}\in\mathbb{R}^{r}.

Following Angeli et al. [2], we assume that all entries of the Jacobian of R(S)R(S) (which we denote by DR∈Rr×sDR\in\mathbb{R}^{r\times s}) have a well-defined constant sign that does not depend on SS (although they may become zero for certain values of SS), and that

DRijΓji≤0for all i,j.(1)DR_{ij}\Gamma_{ji}\leq0\quad\text{for all }i,j. \tag*{(1)}

The assumption (1) is a natural one because in most cases the rate of a reaction increases with the quantity of reactants. This assumption covers, in particular, the most common case, described by the well-established law of mass action, in which the rate of a reaction is proportional to the concentration of each of its reactants.

Following a deterministic approach, the evolution of the concentrations in time is usually described by a system of differential equations:

dSdt=Γ⋅R(S),S∈R≥0s.\frac{dS}{dt}=\Gamma\cdot R(S),\qquad S\in\mathbb{R}_{\geq0}^{s}.

Example 1 (Enzyme kinetics). We consider the simple enzymatic reaction network

E+S⇌k−1k1ES→k2E+P.E+S \xrightleftharpoons[k_{-1}]{k_1} ES \xrightarrow{k_2} E+P.

In this CRN, the set of chemical species is S=(E,S,ES,P)\mathcal{S}=(E,S,ES,P). The enzyme EE, binding the substrate SS, forms a complex ESES, which releases the product PP and the free enzyme EE. According to mass-action kinetics, the rate of each reaction is directly proportional to the concentrations of its reactants, via a coefficient marked next to each arrow (k1,k−1k_1,k_{-1} and k2k_2, respectively). In our framework, this system is represented with a reversible reaction R1\mathcal{R}_1 and an irreversible one R2\mathcal{R}_2, with rate vector

R(S)=[k1[E][S]−k−1[ES]k2[ES]].R(S)=\begin{bmatrix} k_1[E][S]-k_{-1}[ES]\\ k_2[ES] \end{bmatrix}.

Here, we have used the symbol [E][E] to denote the concentration of EE, and so on.

The differential equations describing the behavior of the CRN are

d[E]dt=−k1[E][S]+k−1[ES]+k2[ES],d[S]dt=−k1[E][S]+k−1[ES],d[ES]dt=+k1[E][S]−k−1[ES]−k2[ES],d[P]dt=+k2[ES].\begin{aligned} \frac{d[E]}{dt}&=-k_1[E][S]+k_{-1}[ES]+k_2[ES],\\ \frac{d[S]}{dt}&=-k_1[E][S]+k_{-1}[ES],\\ \frac{d[ES]}{dt}&=+k_1[E][S]-k_{-1}[ES]-k_2[ES],\\ \frac{d[P]}{dt}&=+k_2[ES]. \end{aligned}

As we can notice, a CRN is usually described by a set of non-linear differential equations, which makes difficult the analysis of the dynamics of the system. Indeed, different initial concentrations of the species involved in the reactions can affect the internal and external behavior of the network. Besides, the exact values of parameters are often unknown. Then, in order to investigate the behaviour of the system, we need to study it by performing many simulations under all possible combinations of chemical species concentrations.

To reduce the computational effort, one of the possible approaches is to study the qualitative behaviour of CRNs, without making assumptions on the structure of the dynamical equations involved. In this context, establishing some kind of monotonicity property can help to answer questions concerning the network asymptotic dynamics, such as which are the functionalities of specific chemical pathways or how parameter variations influence the network. Indeed, monotonicity describes the capacity of a system to respond in a natural way to perturbation on its components.

According to Angeli et al. [3, 2] a system is monotonic if its forward flow preserves some order defined on the state space. The system dynamics is expressed in terms of reaction coordinates, for which all the reaction processes are taken into account as a unique flux. As a consequence, with this approach, while it is possible to study how internal or external perturbations influence the fluxes generated by the reactions, it is not possible to address (and therefore understand) how different concentrations of the chemical species involved in the network influence the overall behaviour. The main advantage of Angeli’s approach is that the authors provide efficiently verifiable sufficient conditions to establish if a system is globally monotonic, based on a graphical representations of reactions. In particular, they investigate the dynamics of the system using a particular kind of graph, the so called RR-graph. Their sufficient condition for global monotonicity is the following: the network is globally monotonic if each closed path of the RR-graph contains an even number of negative edges (positive loop property). In the next section, we shall see a formal definition of this graph and comment on this condition.

Results

INPUT-OUTPUT MONOTONICITY DEFINITION

Global monotonicity, proposed in [3], is a very strong property, since it is based on an unique ordering on the whole reaction network. Unfortunately, such a strong property does not hold on most realistic chemical reaction networks.

For this reason, our goal is to assess monotonicity properties between species concentrations that would allow us to infer the behaviour of the network under different initial conditions and to define sufficient conditions for chemical reactions networks that are easy to test and guarantee that such monotonicity properties hold.

The monotonicity property we are interested in can be summarised as follows. Given two species, that we call input and output species of the network, we say that there is a monotonicity relation if the concentration of output species at any time either increases or decreases due to an increase in the initial concentration of the input species. If established, our monotonicity property would allow us to substantially reduce the number of simulations required to study the system. Indeed, if two species are in a monotonicity relation and we are interested in studying the dynamics of the output when the input varies, we can avoid to simulate the chemical reaction network for all possible values of the initial concentrations of the input species. To formally define the previous intuition, we give a new definition of monotonicity, namely the Input-Output monotonicity, and then, following the approach in [2], we give sufficient conditions that guarantee the monotonicity relation of the input and output species.

The following two definitions describe the concept of monotonicity we are interested in: it describes whether the output species reacts in a monotonic way to the increase of the input concentration.

We consider two initial states S0S^0, S‾0\overline{S}^0 such that S‾I0>SI0\overline{S}_I^0 > S_I^0 for one particular species II (the input species), and Sk0=S‾k0S_k^0 = \overline{S}_k^0 for all other species k≠Ik \ne I. With Si(t)S_i(t) we indicate the solution of the ODEs for the species SiS_i with initial value S(0)=S0S(0) = S^0, and with S‾i(t)\overline{S}_i(t) the solution with initial value S‾(0)=S‾0\overline{S}(0) = \overline{S}^0.

Definition 2 (Positive Input-Output Monotonicity). Given a set of reactions R\mathcal{R}, species SO≠SIS_O \ne S_I is positively monotonic with respect to SIS_I in R\mathcal{R} if, for any two initial states S0,S‾0S^0,\overline{S}^0 as above, S‾O(t)≥SO(t)\overline{S}_O(t) \ge S_O(t), for every time t∈R≥0t \in\mathbb{R}_{\ge0}.

Definition 3 (Negative Input-Output Monotonicity). Given a set of reactions R\mathcal{R}, species SO≠SIS_O \ne S_I is negatively monotonic with respect to SIS_I in R\mathcal{R} if, for any two initial states S0,S‾0S^0,\overline{S}^0 as above, S‾O(t)≤SO(t)\overline{S}_O(t) \le S_O(t), for every time t∈R≥0t \in\mathbb{R}_{\ge0}.

DIRECTED SR-GRAPH AND R-GRAPH

We define two graphs associated to a reaction network. The directed SR-graph models the interplay between species and reactions.

Definition 4 (Directed SR-graph). Given a finite set of reactions R\mathcal{R} over a set of species S\mathcal{S}, the directed SR-graph is the directed graph ⟨S∪R,E⟩\langle\mathcal{S} \cup\mathcal{R}, E\rangle, where E⊂(S∪R)×(S∪R)E \subset(\mathcal{S} \cup\mathcal{R}) \times(\mathcal{S} \cup\mathcal{R}) is defined as follows.

  • If Γjk≠0\Gamma_{jk} \ne0, i.e., Rk\mathcal{R}_k affects the concentration of Sj\mathcal{S}_j, then we include the edge (Rk,Sj)(\mathcal{R}_k, \mathcal{S}_j) in EE.

  • If DRij≠0D R_{ij} \ne0, i.e., Sj\mathcal{S}_j affects the speed of Ri\mathcal{R}_i, then we include the edge (Sj,Ri)(\mathcal{S}_j, \mathcal{R}_i) in EE.

No other edges exist in EE (in particular, there no edges in S×S\mathcal{S} \times\mathcal{S} nor in R×R\mathcal{R} \times\mathcal{R}.

In particular, for each pair (Ri,Sj)(\mathcal{R}_i,\mathcal{S}_j) so that Sj\mathcal{S}_j is involved in Ri\mathcal{R}_i, one of these three cases applies.

C1 if Sj\mathcal{S}_j is a promoter for Ri\mathcal{R}_i that does not get consumed, i.e., DRij≠0D R_{ij} \ne0 but Γji=0\Gamma_{ji} = 0, then the edge Sj→Ri\mathcal{S}_j \to\mathcal{R}_i exists only in this direction;

C2 if Sj\mathcal{S}_j is produced or consumed by Ri\mathcal{R}_i but does not affect its rate (for instance the product of an irreversible reaction), i.e., DRij=0D R_{ij} = 0 but Γji≠0\Gamma_{ji} \ne0, then the edge Ri→Sj\mathcal{R}_i \to\mathcal{S}_j exists only in this direction;

C3 in all other cases in which Sj\mathcal{S}_j is involved Ri\mathcal{R}_i, the edge Ri↔Sj\mathcal{R}_i \leftrightarrow\mathcal{S}_j is bidirectional.

Definition 5 (Consistent labeling). A consistent labeling of the set of reactions R\mathcal{R} is a map σ:R→{+1,−1}\sigma: \mathcal{R} \to\{+1,-1\} such that

σ(i)DRijΓjkσ(k)≥0,∀i≠k, ∀j.(2)\sigma(i)D R_{ij}\Gamma_{jk}\sigma(k) \ge0,\quad\forall i \ne k,\ \forall j. \tag*{(2)}

The second graph that we define is a signed but undirected graph that represents the constraints for the existence of a consistent labeling.

Definition 6 (R-graph). Given a finite set of reactions R\mathcal{R} over a set of species S\mathcal{S}, the R-graph of R\mathcal{R} is the signed graph ⟨R,E+,E−⟩\langle\mathcal{R}, E_+, E_-\rangle, where E+⊆(R×R)E_+ \subseteq(\mathcal{R} \times\mathcal{R}) and E−⊆(R×R)E_- \subseteq(\mathcal{R} \times\mathcal{R}) are defined as follows:

  • (Ri,Rk)∈E+(\mathcal{R}_i,\mathcal{R}_k) \in E_+ if i≠ki \ne k and there is a species Sj\mathcal{S}_j such that DRijΓjk>0D R_{ij}\Gamma_{jk} > 0.

  • (Ri,Rk)∈E−(\mathcal{R}_i,\mathcal{R}_k) \in E_- if i≠ki \ne k and there is a species Sj\mathcal{S}_j such that DRijΓjk<0D R_{ij}\Gamma_{jk} < 0.

Note that an edge in E+∪E−E_+ \cup E_- exists if and only if there is a path Ri←Sj←Rk\mathcal{R}_i \leftarrow\mathcal{S}_j \leftarrow\mathcal{R}_k in the directed SR-graph, so the R-graph is the graph obtained from the SR-graph by removing the vertices corresponding to species and connecting reactions directly.

In biological terms, edges of the R-graph describe the relationship between reactions behavior. If the reactions cooperate, helping each other, the reaction nodes are linked by a positive edge. This occurs when the chemical species produced by a reaction is the reactant of an another one. Otherwise, if the reactions compete, for example when they share the same reactants, the reaction nodes are linked by a negative edge.

Note that these definitions are very similar, but subtly different, from those in [3, 2]: indeed, in those papers the case of enzymes (C1) is excluded a priori, and the SR-graph is constructed in an undirected version based only on αij\alpha_{ij} and βij\beta_{ij}, neglecting the difference between cases C2 and C3.

Example 2. Consider the network

R1:A→C,R2:B→C,R3:C→D,(3)\mathcal{R}_1 : A \to C,\qquad\mathcal{R}_2 : B \to C,\qquad\mathcal{R}_3 : C \to D, \tag*{(3)}

where all reactions are irreversible. According to our definition, there is no edge between R1\mathcal{R}_1 and R2\mathcal{R}_2 in the R-graph. Under the definition in [3, 2], there is a negative edge between them, instead.

The following graph-theoretical characterization appears in [3].

Lemma 1. Given a finite set of reactions R\mathcal{R} over a set of species S\mathcal{S}, a consistent labeling of R\mathcal{R} exists if and only if every loop in the R-graph includes an even number of edges in E−E_- (the positive-loop property).

Indeed, to construct a consistent labeling, we can assign any value to σ(R1)\sigma(\mathcal{R}_1) and then traverse the graph assigning signs so each encountered vertex so that (2) holds; if the graph has the positive-loop property we will never encounter a contradiction.

In addition, we can find a necessary condition for this labeling to exist.

Lemma 2 (Rule of 2). For each species SjS_j in a reaction network, determine a number n(Sj)n(S_j) as follows. Consider all reactions in which SjS_j is involved, and determine which of the cases C1, C2, C3 above holds.

  • Take the number of reactions in case C3.

  • Add 1 if there are one or more reactions that fall in case C1.

  • Add 1 if there are one or more reactions that fall in case C2.

Call this number n(Sj)n(S_j). If, for some species SjS_j, n(Sj)>2n(S_j)>2, then R\mathcal{R} does not admit a consistent labeling.

Proof. Suppose by contradiction that a consistent labeling exists, but that n(Sj)>2n(S_j)>2 for a certain species SjS_j. Then, we can find reactions ii, kk, ℓ\ell so that Ri\mathcal{R}_i falls in case C1 or C3, Rk\mathcal{R}_k falls in case C2 or C3, and Rℓ\mathcal{R}_\ell falls in case C3. Thanks to the definition of the three cases and to (2),

σ(i)DRijΓjℓσ(ℓ)>0,\sigma(i)D R_{ij}\Gamma_{j\ell}\sigma(\ell)>0,
σ(i)DRijΓjkσ(k)>0,\sigma(i)D R_{ij}\Gamma_{jk}\sigma(k)>0,
σ(ℓ)DRℓjΓjℓσ(ℓ)>0.\sigma(\ell)D R_{\ell j}\Gamma_{j\ell}\sigma(\ell)>0.

Taking the product of these three inequalities, one obtains DRℓjΓjℓ>0D R_{\ell j}\Gamma_{j\ell}>0, which contradicts our assumption (1).

Example ((2), continued). In the network (3), the only species involved in more than one reaction is CC. We have n(C)=2n(C)=2, as it is involved in R3\mathcal{R}_3 under case C3 (1), and in R1,R2\mathcal{R}_1,\mathcal{R}_2 under case C2 (+1). Thus Lemma 2 does not prohibit the existence of a consistent labeling. Indeed, one can verify that σ(R1)=σ(R2)=σ(R3)=1\sigma(\mathcal{R}_1)=\sigma(\mathcal{R}_2)=\sigma(\mathcal{R}_3)=1 is a consistent labeling. The same network (3), does not admit a consistent labeling with the definitions in Angeli et al. [3, 2], instead, as Condition 2 in (Angeli et al. [3], Proposition 1) fails with their definition.

MONOTONICITY IN REACTION COORDINATES

Let x(t)=∫0tR(S(τ)) dτx(t)=\int_0^t R(S(\tau))\,\mathrm{d}\tau be the vector such that xix_i is the extent of the iith reaction. The vector x(t)x(t) solves, with initial condition x(0)=0x(0)=0, the system of differential equations

ddtx(t)=R(S0+Γx(t)),(4)\frac{d}{dt}x(t)=R(S^0+\Gamma x(t)), \tag*{(4)}

where S0S^0 is the vector of initial concentrations.

The following result appears in Angeli et al. [3], and continues to hold with our definition of R-graph (as proved in § 4).

Theorem 3. If the RR-graph has the positive-loop property, then the system (4) is orthant-cooperative (or orthant-monotone), i.e., for a certain diagonal matrix Σ\Sigma, given two initial conditions x0x^0 and x‾0\overline{x}^0 such that Σ(x0−x‾0)≥0\Sigma(x^0-\overline{x}^0)\geq0, then the solutions x(t),x‾(t)x(t),\overline{x}(t) of (4) with those initial conditions satisfy Σ(x(t)−x‾(t))≥0\Sigma(x(t)-\overline{x}(t))\geq0 for each time t≥0t\geq0.

An explicit choice for this matrix Σ\Sigma is obtained by taking Σii=σ(i)\Sigma_{ii}=\sigma(i), where σ\sigma is a consistent labeling of R\mathcal{R} (which exists thanks to the positive-loop property).

Note that the physical meaning of Theorem 3 is not apparent, since x0=0x^0=0 (i.e., empty history of reactions occurred before the initial time) is the only initial condition that has a direct relevance to the application to CRNs. Rather, this theorem is used in (Angeli et al. [3, 2]) to obtain statements about the asymptotic behavior of the reaction network.

Handling reversible reactions

As in [3], in the case of reversible reactions, we construct the graph by considering only one of the two possible orientations. The choice of reaction orientation does not influence Angeli’s monotonicity result that we will exploit in the proof of our theorems. Moreover, even if this choice changes the way we process the R-graph, as we will discuss later, it does not influence the final result of our method.

Note that, as we already pointed out, the choice of the orientation of reversible reactions does not influence the existence of a consistent labeling.

Lemma 4. Given a set of reactions R\mathcal{R} and a reaction Ri∈RR_i \in\mathcal{R}, let RirevR_i^{\mathrm{rev}} be the reaction obtained from RiR_i by swapping reactants with products. There exists a consistent labeling for the R-graph of R\mathcal{R} if and only if there exists a consistent labeling for the R-graph of R′=R∖{Ri}∪{Rirev}\mathcal{R}' = \mathcal{R} \setminus\{R_i\} \cup\{R_i^{\mathrm{rev}}\}.

Proof. Let ⟨R,E+,E−⟩\langle\mathcal{R}, E_{+}, E_{-}\rangle be the R-graph of R\mathcal{R}. The R-graph of R′\mathcal{R}' can be obtained from the previous one as follows: ⟨R′,E+′,E−′⟩\langle\mathcal{R}', E_{+}', E_{-}'\rangle, where E+′=E+∖{(Ri,Rj)∈E+}∪{(Rirev,Rj)∣(Ri,Rj)∈E−}E_{+}' = E_{+} \setminus\{(R_i,R_j) \in E_{+}\} \cup\{(R_i^{\mathrm{rev}},R_j) \mid(R_i,R_j) \in E_{-}\} and E−′=E−∖{(Ri,Rj)∈E−}∪{(Rirev,Rj)∣(Ri,Rj)∈E+}E_{-}' = E_{-} \setminus\{(R_i,R_j) \in E_{-}\} \cup\{(R_i^{\mathrm{rev}},R_j) \mid(R_i,R_j) \in E_{+}\}. Hence, the R-graph of R′\mathcal{R}' is the same as the one of R\mathcal{R}, but in which edges that were connected to RiR_i are now connected to RirevR_i^{\mathrm{rev}} with opposite sign. Given a consistent labeling for the R-graph of R\mathcal{R} it is possible to derive a consistent labeling for R′\mathcal{R}' by simply assigning to RirevR_i^{\mathrm{rev}} an opposite label with respect to R\mathcal{R}, and vice-versa.

Sufficient condition for input-output monotonicity

Our main result is the following, proved in § 4.

Theorem 5. Let a set of chemical reactions R\mathcal{R} be given, with two distinct species SIS_I and SOS_O designated as the input and the output species.

Augment the network by adding two dummy reactions: an irreversible reaction with output SIS_I

RIN:∅→SI,R_{\mathrm{IN}} : \varnothing\to S_I,

and a reaction which has SOS_O as promoter

ROUT:SO→SO.R_{\mathrm{OUT}} : S_O \to S_O.

If R∪{RIN,ROUT}\mathcal{R} \cup\{R_{\mathrm{IN}},R_{\mathrm{OUT}}\} has a consistent labeling (i.e., if its R-graph has the positive-loop property), then SOS_O is monotonic with respect to SIS_I. In particular, it is positively monotonic if σ(IN)σ(OUT)=+1\sigma(\mathrm{IN})\sigma(\mathrm{OUT}) = +1, and negatively monotonic if σ(IN)σ(OUT)=−1\sigma(\mathrm{IN})\sigma(\mathrm{OUT}) = -1.

To show how Theorem 5 works, consider again our running Example 1, where we want to study the monotonicity between the species S and the species P, which we consider as the input and the output of the CRN.

We build the stoichiometric matrix of the network, considering only one direction for the reversible reaction R1R_1

Γ=R1R2E−1+1S−10ES+1−1P0+1\Gamma= \begin{array}{c|cc} & R_1 & R_2 \\ E & -1 & +1 \\ S & -1 & 0 \\ ES & +1 & -1 \\ P & 0 & +1 \end{array}

and the matrix DRDR

DR=ESESPR1+k1[S]+k1[E]−k1[ES]0R200+k20.DR = \begin{array}{c|cccc} & E & S & ES & P \\ R_1 & +k_1[S] & +k_1[E] & -k_1[ES] & 0 \\ R_2 & 0 & 0 & +k_2 & 0 \end{array}.

Then, following Definition 4, we build the directed SR-Graph, represented in Figure 1, from which we derive the R-graph, which has one positive edge since DRijΓjk>0DR_{ij} \Gamma_{jk} > 0 for each species jj. We represent the R-graph in Figure 2.

Directed SR-graph of Example 1, representing Michaelis-Menten kinetics

Figure 1. Directed SR-graph of Example 1, representing Michaelis-Menten kinetics.

The R-graph of Example 1, representing Michaelis-Menten kinetics

Figure 2. The R-graph of Example 1, representing Michaelis-Menten kinetics.

Then, we augment the network adding the two following dummy reactions:

0⟶S0 \longrightarrow S
P⟶P,P \longrightarrow P,

and we verify trivially that the new network has a consistent labeling, which means that output is monotonic with respect to the input. Moreover, since σ(IN)σ(OUT)=+1\sigma(IN)\sigma(OUT)=+1 we can affirm that the output of this CRN, the chemical species P, is positively monotonic with respect to the species S, which means that if we increase the initial concentration of the input S, then the concentration of the output P increases at any time.

As expected, this result is confirmed by simulations. Once we have fixed the chemical rates (k1=0.1k_1=0.1, k2=1000k_2=1000, k3=0.3k_3=0.3), we vary the initial concentration of the input S in the range [100, 2000]. In Figure 3, we show that by increasing the initial concentration of S (denoted [S]0[S]_0), the concentration of P at the steady state (denoted [P]ss[P]_{ss}) increases as well. This actually does not hold only for the steady state: the concentration of P is increased at any time point.

Simulation results of Example 1, representing Michaelis-Menten kinetics

Figure 3. Simulation results of Example 1, representing Michaelis-Menten kinetics. To show how the concentration of the species P is positively monotonic with respect to the species S, we plot on the horizontal axis the initial concentration of S, in a range [100, 2000] and on the vertical axis the concentration of P at the steady state.

Case study

We now apply the result stated in Theorem 5 to the more complex network of ERK signalling pathway.

SIGNALLING PATHWAY EXAMPLE.

A signalling pathway usually consists of enzymatic cascades, having a starting species that triggers the other connected processes. An initial stimulus, perceived by a transductor (a sort of the first messenger), activates the cascade amplifying the signal for the next enzymatic reaction. Many biochemical processes are associated with signalling pathways as protein activation, repression, and expression of genes: anomalies in these processes could give rise to diseases like cancer, diabetes, and others.

One of the most important examples of such processes is the ERK pathway, which is involved in growth, survival, proliferation, and differentiation of cells. We consider the mathematical model of the ERK pathway implemented by Schilling et al. ([27]) and available to the public on the BioModels Database (BIOMD0000000270). It consists of many fast phosphorylation reactions, which spread the signal along the enzymatic cascade. For simplicity, let us focus on a particular portion of the entire pathway, which we will denote as ERK*. We indicate the species and the kinetics rates as originally denoted in the model in ([27]). For simplicity, we refer to the reaction using the notation RiR_i, where ii is the kinetics rate index. The reactions involved are the following:

Raf⇄k19k18PRafMek1⇄k27k21[PRaf]PMek1PMek1⇄k25k23PPMek1,(5)\begin{aligned} \mathrm{Raf} &\mathrel{\mathop{\rightleftarrows}^{k_{18}}_{k_{19}}} \mathrm{PRaf} \\ \mathrm{Mek1} &\mathrel{\mathop{\rightleftarrows}^{k_{21}[\mathrm{PRaf}]}_{k_{27}}} \mathrm{PMek1} \\ \mathrm{PMek1} &\mathrel{\mathop{\rightleftarrows}^{k_{23}}_{k_{25}}} \mathrm{PPMek1}, \tag*{(5)} \end{aligned}

in Table 1 we reported the coefficient rates and the initial conditions of ERK* system.

Initial concentrationsRates
(1)(2)
Raf=10\mathrm{Raf}=10k18=0.1445k_{18}=0.1445
Praf=0\mathrm{Praf}=0k19=0.37k_{19}=0.37
Mek1=1\mathrm{Mek1}=1k21=0.02k_{21}=0.02
PMek1=0\mathrm{PMek1}=0k23=667.957k_{23}=667.957
PPMek1=0\mathrm{PPMek1}=0k25=0.13k_{25}=0.13
k27=0.07k_{27}=0.07

Table 1. The initial concentrations and the rates of ERK* system.

The species PRaf\mathrm{PRaf}, in the reaction R21R_{21}, is involved as catalyst promoter, which means that its concentration positively influences the production of the species PMek1\mathrm{PMek1}. In our framework, the rate vector for this network is

R(S)=[k18[Raf]−k19[PRaf]k21[Mek1][PRaf]−k27[PMek1]k23[PMek1]−k25[PPMek1]].R(S)= \begin{bmatrix} k_{18}[Raf]-k_{19}[PRaf]\\ k_{21}[Mek1][PRaf]-k_{27}[PMek1]\\ k_{23}[PMek1]-k_{25}[PPMek1] \end{bmatrix}.

We can apply Theorem 5 on the network, considering Raf and PPMek1, respectively, as the input and the output. We augment the network adding the two following dummy reactions:

0⟶Raf0 \longrightarrow\mathrm{Raf}
PPMek1⟶PPMek1.\mathrm{PPMek1} \longrightarrow\mathrm{PPMek1}.

We build the R-graph and we can easily verify that it has a consistent labeling, as we show in Figure 4. Moreover, since σ(IN)σ(OUT)=+1\sigma(IN)\sigma(OUT)=+1, we can conclude that PPMek1 is positively monotonic with respect to Raf. This can be confirmed with simulations, as shown in Figure 5.

R-graph representation of the network 5, augmented by the two dummy reactions

Figure 4. R-graph representation of the network 5, augmented by the two dummy reactions. As we can notice it has a consistent labeling.

Simulation results showing PPMek1 concentration versus initial Raf concentration

Figure 5. Simulation results of CRN 5, representing ERK signalling pathway. To show how the concentration of the species PPMek1 is positively monotonic with respect to the species Raf, we plot on the horizontal axis the initial concentration of Raf, in a range [1, 100] and on the vertical axis the concentration of PPMek1 at the steady state.

Methods

In this section, we illustrate the proofs of the two main results.

Proof of Theorem 3. This proof is basically the same one that appears in [3], but adapted to our slightly different definitions.

If the R-graph has the positive-loop property, then there is a consistent labelling σ\sigma; let Σ∈Rr×r\Sigma\in\mathbb{R}^{r \times r} be the diagonal matrix with Σii=σ(i)\Sigma_{ii}=\sigma(i). The Jacobian of (4) is J=DR⋅ΓJ=DR\cdot\Gamma. The off-diagonal entries of ΣJΣ\Sigma J\Sigma are given by

ΣiiJikΣkk=∑jσ(i)DRijΓjkσ(k),i≠k,\Sigma_{ii}J_{ik}\Sigma_{kk}=\sum_j \sigma(i)DR_{ij}\Gamma_{jk}\sigma(k), \qquad i\ne k,

where all summands are non-negative by the definition of consistent labeling (2). The fact that ΣJΣ\Sigma J\Sigma has non-negative diagonal elements is a sufficient condition for the system to be orthant-monotone (see [30]). □

Proof of Theorem 5. Let eI∈Rse_I \in\mathbb{R}^{s} (resp. eOe_O) be the vector that has an entry 1 in the corresponding to SIS_I (resp. SOS_O) and all other entries equal to 0.

The rate and stoichiometry matrix of the augmented network are

DR^=1…rIN[DR0]OUTkeOT,Γ^=1…rINOUT[ΓeI0].\widehat{DR} = \begin{array}{cc} &1\ldots r\\ IN&\left[\begin{array}{c}DR\\0\end{array}\right]\\ OUT&k e_O^T \end{array}, \qquad \widehat{\Gamma} = \begin{array}{cc} &1\ldots r\quad IN\quad OUT\\ &\left[\begin{array}{ccc}\Gamma&e_I&0\end{array}\right] \end{array}.

and its Jacobian is

J^=DR^⋅Γ^=1…rINOUT1…r[JDReI0]IN[000]OUT[eOTΓ00].\widehat{J}=\widehat{DR}\cdot\widehat{\Gamma}= \begin{array}{cc} &1\ldots r\quad IN\quad OUT\\ 1\ldots r&\left[\begin{array}{ccc}J&DRe_I&0\end{array}\right]\\ IN&\left[\begin{array}{ccc}0&0&0\end{array}\right]\\ OUT&\left[\begin{array}{ccc}e_O^T\Gamma&0&0\end{array}\right] \end{array}.

Let σ\sigma be a consistent labeling for the augmented network, and assume without loss of generality (up to replacing σ\sigma with −σ-\sigma) that σ(IN)=1\sigma(IN)=1.

Hence ΣDReI≥0\Sigma DRe_I\geq0, σ(OUT)eOTΓΣ≥0\sigma(OUT)e_O^T\Gamma\Sigma\geq0, and ΣJΣ\Sigma J\Sigma has non-negative diagonal entries.

We now take two initial values S0S^0 and S‾0=S0+δeI\overline{S}^0=S^0+\delta e_I with δ>0\delta>0, and aim to prove that σ(OUT)(S‾O(t)−SO(t))≥0\sigma(OUT)(\overline{S}_O(t)-S_O(t))\geq0 for each t≥0t\geq0; indeed, this is the input-output monotonicity statement that we need to prove.

Define the dynamical system

{ddty(t)=R(S0+zeI+Γx(t)),ddtz(t)=0.(6)\begin{cases} \frac{d}{dt}y(t)=R(S^0+ze_I+\Gamma x(t)),\\ \frac{d}{dt}z(t)=0. \tag*{(6)} \end{cases}

Let

J˘=yzy[JDReI00]z,Σ˘=yzy[Σ001].\breve{J}= \begin{array}{cc} &y\quad z\\ y&\left[\begin{array}{cc}J&DRe_I\\0&0\end{array}\right]\\ z& \end{array}, \qquad \breve{\Sigma}= \begin{array}{cc} &y\quad z\\ y&\left[\begin{array}{cc}\Sigma&0\\0&1\end{array}\right]. \end{array}

It is simple to verify that J˘\breve{J} is the Jacobian of (6), and that Σ˘J˘Σ˘\breve{\Sigma}\breve{J}\breve{\Sigma} has non-negative off-diagonal, hence the system (6) is orthant-monotone by the same result in Smith [30] that we have used in the proof of {xref}Theorem 3`.

Direct verification shows that the solution of this system with initial value y(0)=0y(0)=0, z(0)=0z(0)=0 is y(t)=x(t)y(t)=x(t), z(t)=0z(t)=0, whereas the solution y‾(t)\overline{y}(t), z‾(t)\overline{z}(t) with initial value y‾(0)=0\overline{y}(0)=0, z‾(0)=δ\overline{z}(0)=\delta is y‾(t)=x‾(t)\overline{y}(t)=\overline{x}(t), z‾(t)=δ\overline{z}(t)=\delta, where the quantity x‾(t)\overline{x}(t) is defined analogously to (4) but with initial value S‾0\overline{S}^0.

We have

[0δ]=Σ˘[y‾(0)z‾(0)]≥Σ˘[y(0)z(0)]=[00].\begin{bmatrix}0\\\delta\end{bmatrix} =\breve{\Sigma} \begin{bmatrix}\overline{y}(0)\\\overline{z}(0)\end{bmatrix} \geq \breve{\Sigma} \begin{bmatrix}y(0)\\z(0)\end{bmatrix} = \begin{bmatrix}0\\0\end{bmatrix}.

By the orthant-monotonicity of (6), this implies

Σ˘[y‾(t)z‾(t)]≥Σ˘[y(t)z(t)]for all t≥0,\breve{\Sigma} \begin{bmatrix}\overline{y}(t)\\\overline{z}(t)\end{bmatrix} \geq \breve{\Sigma} \begin{bmatrix}y(t)\\z(t)\end{bmatrix} \qquad\text{for all }t\geq0,

and this means in particular that Σx‾(t)≥Σx(t)\Sigma\overline{x}(t)\geq\Sigma x(t).

The concentrations of the output species SO(t)S_O(t), S‾O(t)\overline{S}_O(t) at time tt with the two initial conditions S0S^0, S‾0\overline{S}^0 are given by

SO(t)=SO0+eOT⋅Γx(t),S‾O(t)=S‾O0+eOT⋅Γx‾(t),S_O(t)=S_O^0+e_O^T\cdot\Gamma x(t),\qquad \overline{S}_O(t)=\overline{S}_O^0+e_O^T\cdot\Gamma\overline{x}(t),

respectively (SO0=S‾O0S_O^0=\overline{S}_O^0 because SO≠SIS_O\ne S_I), hence

σ(OUT)(S‾O(t)−SO(t))=σ(OUT)eOTΓΣ⏟≥0Σ(x‾(t)−x(t))⏟≥0≥0.\sigma(OUT)(\overline{S}_O(t)-S_O(t)) =\underbrace{\sigma(OUT)e_O^T\Gamma\Sigma}_{\geq0} \underbrace{\Sigma(\overline{x}(t)-x(t))}_{\geq0} \geq0.

which completes the proof.

Discussion and Conclusions

In this paper, we proposed a new notion of monotonicity, namely the Input-Output monotonicity, for which two species, considered as input and output of the network, are monotonic if the variation of the initial concentration of the input implies a monotonic variation in the concentration of the output. Moreover, we established a new sufficient condition based on the R-graph that guarantees that such monotonicity property holds. We showed how this new notion of monotonicity can have a great practical impact. Indeed, monotonicity assessment can drastically reduce the number of simulations necessary to study the dynamical behaviour of a chemical reaction network under uncertain initial conditions. This can be very useful in several different cases, for example when initial concentrations are not exactly known or when the focus is on studying the effects of perturbations of initial concentrations (as in the case of robustness). The proposed method allows us to study complex chemical reaction networks, performing a preliminary analysis of the system dynamics, in order to investigate different biological properties, such as robustness or sensitivity to perturbations.

We have shown the application of our approach to the small example of Michaelis-Menten kinetics and to the quite complex model of the ERK signaling pathway [20].

In order to apply our sufficient condition on larger and more complex network, different approaches of model reduction can be applied. For instance, in [19] the authors show how to remove particular nodes of the network preserving the steady state fluxes of the system. Moreover, for analysing large-scale biochemical models, we can also simplify the model using the common approach of the separation of timescale [14], which allows us to consider the reaction networks as divided into processes having different timescales leading to an approximate – but accurate – version of the original model.

In addition, in [8], the authors propose an innovative approach based on machine learning on graphs to predict whether a biological system is robust studying its topological features. As future work, we intend to apply a similar method to predict the monotonicity of a system.

AUTHOR’S CONTRIBUTIONS

Conceptualization & Formal analysis: R. Gori, P. Milazzo, L. Nasti, F. Poloni. Simulations & Experiments: L. Nasti. Writing original draft: R. Gori, P. Milazzo, L. Nasti, F. Poloni. Writing - review & editing: R. Gori, P. Milazzo, L. Nasti, F. Poloni.

FUNDING

R. Gori, P. Milazzo, L. Nasti are partially supported by the University of Pisa’s project PRA_2020_26 “Metodi informatici integrati per la biomedica”. F. Poloni is partially supported by the University of Pisa’s project PRA_2020_61 “Analisi di reti complesse: dalla teoria alle applicazioni” and by GNCS/INDAM (Istituto Nazionale di Alta Matematica).

References

  1. [1]Alur, R., Courcoubetis, C., Henzinger, T. A., & Ho, P.-H. 1992, in Hybrid systems, ed. R. L. Grossman, A. Nerode, A. P. Ravn, & H. Rischel, 209–229
  2. [2]Angeli, D., De Leenheer, P., & Sontag, E. 2010, Journal of mathematical biology, 61, 581
  3. [3]Angeli, D., De Leenheer, P., & Sontag, E. D. 2006, in Decision and Control, 2006 45th IEEE Conference, IEEE, 7–12
  4. [4]Angeli, D., De Leenheer, P., & Sontag, E. D. 2007, Mathematical biosciences, 210, 598
  5. [5]Barbuti, R., Maggiolo-Schettini, A., Milazzo, P., Pardini, G., & Tesei, L. 2011, Natural Computing, 10, 3
  6. [6]Behinaein, B., Rudie, K., & Sangrar, W. 2014, in Electrical and Computer Engineering (CCECE), 2014 IEEE 27th Canadian Conference on, ed. IEEE
  7. [7]Bernini, A., Brodo, L., Degano, P., Falaschi, M., & Hermith, D. 2018, Natural Computing, 17, 345
  8. [8]Bove, P., Micheli, A., Milazzo, P., & Podda, M. 2020, in 13th International conference on bioinformatics models, methods and algorithms (BIOINFORMATICS 2020), SCITEPRESS, 32–43
  9. [9]Danos, V., Feret, J., Fontana, W., & Krivine, J. 2008, in International Workshop on Verification, Model Checking, and Abstract Interpretation, ed. Springer, 83–97
  10. [10]Dresch, J. M., Liu, X., Arnosti, D. N., & Ay, A. 2010, BMC systems biology, 4, 142
  11. [11]Feinberg, M. 1987, Chemical Engineering Science, 42, 2229
  12. [12]Gori, R., Milazzo, P., & Nasti, L. 2019, in 10th International conference on bioinformatics models, methods and algorithms (BIOINFORMATICS 2019), ed. SciTePress, 250–257
  13. [13]Henzinger, T. A. 2000, in Verification of Digital and Hybrid Systems, ed. Springer, Vol. 736, 265–292DOI
  14. [14]Ingalls, B. P. 2013, Mathematical modeling in systems biology: an introduction (MIT press)
  15. [15]Iooss, B., & Lemaître, P. 2015, in Uncertainty management in simulation-optimization of complex systems, ed. Springer, 101–122
  16. [16]Kitano, H. 2002, Systems biology: towards systems-level understanding of biological systems. Kitano H ed, Foundations of Systems Biology, MIT Press, CambridgeDOI
  17. [18]Koch, I. 2010, Molecular Informatics, 29, 838
  18. [19]Küken, A., Wendering, P., Langary, D., & Nikolovski, Z. 2021, bioRxiv, doi: 10.1101/2021.07.23.457585
  19. [20]Kwang-Hyun, C., Sung-Young, S., Hyun-Woo, K., et al. 2003, in International Conference on Computational Methods in Systems Biology, ed. Springer, 127–141
  20. [22]Murata, T. 1989, Proceedings of the IEEE, 77, 541
  21. [23]Nasti, L. 2020, PhD thesis, Università di Pisa, Dipartimento di Informatica
  22. [24]Nasti, L., Gori, R., & Milazzo, P. 2018, in Federation of International Conferences on Software Technologies: Applications and Foundations, Springer, 81–97
  23. [25]Regev, A., Silverman, W., & Shapiro, E. 2000, in Biocomputing 2001, ed. W. Scientific, 459–470
  24. [26]Rizk, A., Batt, G., Fages, F., & Soliman, S. 2011, Theoretical Computer Science, 412, 2827
  25. [27]Schilling, M., Maiwald, T., Hengl, S., et al. 2009, Molecular systems biology, 5
  26. [28]Shinar, G., Alon, U., & Feinberg, M. 2009, SIAM Journal on Applied Mathematics, 69, 977
  27. [29]Shinar, G., & Feinberg, M. 2010, Science, 327, 1389
  28. [30]Smith, H. L. 1995, Mathematical Surveys and Monographs, Vol. 41, Monotone dynamical systems (American Mathematical Society, Providence, RI), x+174
  29. [31]Zi, Z. 2011, IET systems biology, 5, 336

Paper details

vlnc:W_spbfqnkr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents