Model checking the evolution of gene regulatory networks
 1.2k Downloads
 2 Citations
Abstract
The behaviour of gene regulatory networks (GRNs) is typically analysed using simulationbased statistical testinglike methods. In this paper, we demonstrate that we can replace this approach by a formal verificationlike method that gives higher assurance and scalability. We focus on Wagner’s weighted GRN model with varying weights, which is used in evolutionary biology. In the model, weight parameters represent the gene interaction strength that may change due to genetic mutations. For a property of interest, we synthesise the constraints over the parameter space that represent the set of GRNs satisfying the property. We experimentally show that our parameter synthesis procedure computes the mutational robustness of GRNs—an important problem of interest in evolutionary biology—more efficiently than the classical simulation method. We specify the property in linear temporal logic. We employ symbolic bounded model checking and SMT solving to compute the space of GRNs that satisfy the property, which amounts to synthesizing a set of linear constraints on the weights.
Keywords
Model Check Linear Temporal Logic Satisfiability Modulo Theory Quantifier Elimination Linear Temporal Logic Formula1 Introduction
Gene regulatory networks (GRNs) are one of the most prevalent and fundamental types of biological networks whose main actors are genes regulating other genes. A topology of a GRN is represented by a graph of interactions among a finite set of genes, where nodes represent genes, and edges denote the type of regulation (activation or repression) between the genes, if any. In [24], Wagner introduced a simple but useful model for GRNs that captures important features of GRNs. In the model, a system state specifies the activity of each gene as a Boolean value. The system is executed in discrete time steps, and all gene values are synchronously and deterministically updated: a gene active at time n affects the value of its neighbouring genes at time \(n+1\). This effect is modelled through two kinds of parameters: threshold parameters assigned to each gene, which specify the strength necessary to sustain the gene’s activity, and weight parameters assigned to pairs of genes, which denote the strength of their directed effect.
Some properties of GRNs can be expressed in linear temporal logic (LTL) (such as reaching a steadystate), where atomic propositions are represented by gene values. A single GRN may or may not satisfy a property of interest. Biologists are often interested in the behavior of populations of GRNs, and in presence of environmental perturbations. For example, the parameters of GRNs from a population may change from one generation to another due to mutations, and the distribution over the different GRNs in a population changes accordingly. We refer to the set of GRNs obtained by varying parameters on a fixed topology as GRN space. For a given population of GRNs instantiated from a GRN space, typical quantities of interest refer to the longrun average behavior. For example, robustness refers to the averaged satisfaction of the property within a population of GRNs, after an extended number of generations. In this context, Wagner’s model of GRN has been used to show that mutational robustness can gradually evolve in GRNs [11], that sexual reproduction can enhance robustness to recombination [1], or to predict the phenotypic effect of mutations. The computational analysis used in these studies relies on explicitly executing GRNs, in the purpose of checking if they satisfy the property. Then, in order to compute the robustness of a population of GRNs, the satisfaction check must be performed repeatedly for many different GRNs. Typically, robustness is estimated by statistically sampling GRNs from the GRN space. In this work, we pursue formal analysis of Wagner’s GRNs which allows to replace this simulationbased statistical testingtype method by a formal verificationtype method that gives higher assurance and scalability.
In this paper, we present a novel method for synthesizing the space of parameters which characterize GRNs that satisfy a given property. These constraints eliminate the need for explicitly executing the GRN to check the satisfaction of the property. Importantly, the synthesized parameter constraints allow to efficiently answer questions that are very difficult or impossible to answer by simulation, e.g. emptiness check or parameter sensitivity analysis. In this work, we chose to demonstrate how the synthesized constraints can be used to compute the robustness of a population of GRNs with respect to genetic mutations. Since constraint evaluation is usually faster than executing a GRN, the constraints precomputation enables faster computation of robustness. To this end, when statistical sampling of the GRNs is employed, higher precision is achieved within the same computational time. Moreover, it sometimes becomes possible to replace the statistical sampling with the exact computation of robustness.
In our method, for a given GRN space and LTL property, we used SMT solving and bounded model checking to generate a set of constraints such that a GRN satisfies the LTL property if and only if its weight parameters satisfy the constraints. The key insight in this method is that the obtained constraints are complex Boolean combinations of linear inequalities. Solving linear constraints has been the focus of both industry and academia for some time. However, the technology for solving linear constraints with Boolean structure, namely SMT solving, has matured only in the last decade [3]. This technology has enabled us to successfully apply an SMT solver to generate the desired constraints.
We have built a tool which computes the constraints for a given GRN space and a property expressed in a fragment of LTL. In order to demonstrate the effectiveness of our method, we computed the robustness of five GRNs listed in [8], and for three GRNs known to exhibit oscillatory behavior. We first synthesised the constraints and then we used them to estimate robustness based on statistical sampling of GRNs from the GRN space. Then, in order to compare the performance with the simulationbased methods, we implemented the approximate computation of robustness, where the satisfiability of the property is verified by executing the GRNs explicitly. The results show that in six out of eight tested networks, the precomputation of constraints provides greater efficiency, performing up to three times faster than the simulation method.
In the benchmarks, we compute mutational robustness with respect to mutation model without selection, described in Sect. 7.2. In this model, the GRNs in a GRNpopulation are assumed to be distributed so that the mutation process from one generation to another is not influenced by evolutionary pressure (there is no priority given to cells which exhibit some selection property). Mathematically, the limiting distribution of such evolutionary process over many generations amounts to the stationary distribution of a discretetime, finite Markov chain, where the state space captures all possible GRN configurations in a GRNspace, and transition probabilities describe the probability that one GRN configuration changes to another one in the next generation; Such stationary distribution can be approximated with complexity polynomial in the size of GRNspace. This model was also used in our related work in [17]. We here extend the approach by introducing a more realistic, but also technically more demanding, model with selection. In this model, a number of individuals not satisfying the selection property dies, while those individuals who meet the property reproduce more, that is, the nature selects for the individuals satisfying the property. We take the selection property to be equal to the property we check robustness for. To this end, the process describing the GRN space over generations is not Markovian. In Sect. 7.3, the principal result is Theorem 2, which shows that robustness for the model with selection equals the second largest eigenvalue of the matrix capturing the transition probabilities of mutations among the GRNs in the GRN space. Moreover, in Theorem 3, we show that, confirming the intuition, in both model with and without selection, robustness measure preserves monotonicity in the sense that larger properties imply larger robustness.
Related Work Formal verification techniques are already used for aiding various aspects of biological research [12, 15, 19, 20, 25]. The robustness of models of biochemical systems with respect to temporal properties has been studied [6, 18, 22]. These approaches differ from ours in that in our work we focus on quantifying robustness with respect to a concrete source of perturbation of the parameters (such as point mutations). In [18] and [22], the authors present methods to evaluate the satisfaction degree of a trace with respect to a temporal formula, but no assumption is made on the distribution of parameters, while in [6], the authors use a symbolic technique to synthesise the parameter space for which the property is satisfied, but robustness is defined in terms of the relative size of the error intervals around some reference parameter values.
This work extends [17] which, to the best of our knowledge, is the first application of formal verification to studying the robustness against evolution of gene regulatory networks. As such, this work opens up a novel application area for the formal verification community. As previously discussed, with respect to related studies in evolutionary biology, our method can offer a higher degree of assurance, more accuracy, and better scalability than the traditional, simulationbased approaches. In addition, while previous works focus on invariant properties, our method allows the study of nontrivial temporal properties that are expressible in LTL, such as bistability or oscillations between gene states.
1.1 Motivating example
Our method automatizes this idea to any given GRN topology and any property specified in LTL. We first translate \(T\) to a labelled transition system, partly shown in Fig. 1b. Then, we apply symbolic model checking to compute the constraints which represent the space of GRNs from \(T\) that satisfy the bistability property.
To illustrate the scalability of our method in comparison with the standard methods, in Fig. 1c, we compare the performance of computing the mutational robustness with and without precomputing the constraints (referred to as evaluation and execution method respectively). We choose a mutation model such that each parameter takes 13 possible values distributed according to the binomial distribution (see Sect. 7 for more details on the mutation model). We estimate the robustness value by statistical sampling of the possible parameter values. For a small number of samples, our method is slower because we spend extra time in computing the constraints. However, more samples may be necessary for achieving the desired precision. As the number of samples increases, our method becomes faster, because each evaluation of the constraints is two times faster than checking bistability by executing GRNindividuals. For \(1.2\times 10^5\) simulations, execution and evaluation methods take same total time, and the robustness value estimated from these many samples lies in the interval (0.8871, 0.8907) with 95 % confidence. Hence, for this GRN, if one needs better precision for the robustness value, our method is preferred.
One may think that for this example, we may compute exact robustness because the number of parameter values is only \(13^6\) (four weights and two inputs). For simplicity of illustration, we chose this example, and we later present examples with a much larger space of parameters, for which exact computation of robustness is infeasible.
2 Preliminaries
In this section, we start by defining a GRN space, which will serve to specify common features for GRNs from the same population. These common features are types of gene interactions (topology), constant parameters (thresholds), and ranges of parameter values that are subject to some environmental perturbation (weights). Then, we formally introduce a model of an individual GRN from the GRN space and temporal logic to express its properties.
2.1 Basic notation
\(\mathbb {R}_{\ge 0}\) (resp. \(\mathbb {Q}_{\ge 0}\)) is the set of nonnegative real (resp. rational) numbers. For \(m < n\), let m..n denote the set of integers from m to n. We denote real vectors by bold face letters. Let \(\mathbf {k}\) be a real vector then let \(k_i\) denote the ith element of \(\mathbf {k}\). With abuse of notation, we treat finite maps with ordered domain as vectors with size of the map. Let S be a set, we denote as \(\mathcal {P}_k(S)\) the set of all subsets of S with cardinality k.
For rationals constants \(k_1,\ldots ,k_n\), a vector of rational variables \(v= (v_1,\ldots ,v_n)\), and a rational constant t, let \( k_1v_1 + \cdots + k_nv_n + t\) denote a linear term. Let \(k_1v_1 + \cdots + k_nv_n + t > 0 \) and \(k_1v_1 + \cdots + k_nv_n + t \ge 0\) be a strict and nonstrict inequality over \(v\) respectively. Let \(\textit{linear}(v)\) be the set of all the (non)strict inequalities over \(v\). Let \(\textit{polyhedra}(v)\) be the set of all the finite conjunctions of the elements of \(\textit{linear}(v)\).
2.2 GRN space
The key characteristics of the behaviour of a GRN are typically summarised by a directed graph where nodes represent genes and edges denote the type of regulation between the genes. A regulation edge is either activation (one gene’s activity increases the activity of the other gene) or repression (one gene’s activity decreases the activity of the other gene) [23]. In Wagner’s model of a GRN, in addition to the activation types between genes, each gene is assigned a threshold and each edge (pair of genes) is assigned a weight. The threshold of a gene models the amount of activation level necessary to sustain activity of the gene. The weight on an edge quantifies the influence of the source gene on destination gene of the edge.
We extend the Wagner’s model by allowing a range of values for weight parameters. We call our model GRN space, denoting that all GRNs instantiated from that space share the same topology, and their parameters fall into given ranges. We assume that each gene always has some minimum level of expression without any external influence. In the model, this constant input is incorporated by a special gene which is always active, and activates all other genes from the network. The weight on the edge between the special gene and some other gene represents the minimum level of activation. The minimal activation is also subject to perturbation.
Definition 1

\(G=\{g_1,\ldots ,g_{d}\}\) is a finite ordered set of genes,

\(g_{in} \in G\) is the special gene used to model the constant input for all genes,

Open image in new window is the activation relation such that Open image in new window and Open image in new window ,

Open image in new window is the repression relation such that Open image in new window ,

\(t:G\rightarrow \mathbb {Q}\) is the threshold function such that \(\forall g\in G{\setminus }\{g_{in}\} t(g) \ge 0\) and \(t(g_{in}) < 0 \),

Open image in new window is the maximum value of an activation/repression,

Open image in new window assigns a set of possible weight functions to each activation/inhibition relation, so that Open image in new window .
In the following text, if not explicitly specified otherwise, we will be referring to the GRN space Open image in new window .
2.3 GRNindividual
Definition 2
(GRNindividual) A GRNindividual \(\mathcal {G}\) is a pair \((T,w)\), where \(w\in {W}\) is a weight function from the GRN space.
A state \(\mathbf {\sigma }:G\rightarrow \mathbb {B}\) of a GRNindividual \(\mathcal {G}= (T,w)\) denotes the activation state of each gene in terms of a Boolean value. Let \(\varSigma (\mathcal {G})\) denote the set of all states of \(\mathcal {G}\), such that \(\mathbf {\sigma }(g_{in})=true\). In the context where we do not refer to a set of states of a concrete GRNindividual, we denote by \(\varSigma (T)\) the set of all states of any GRN from GRN space \(T\). The GRN model executes in discrete time steps by updating all the activation states synchronously and deterministically according to the following rule: a gene is active at next time if and only if the total influence on that gene, from genes active at current time, surpasses its threshold.
Definition 3
2.4 Temporal properties
A GRN exists in a living organism to exhibit certain behaviors. Here we present a linear temporal logic (LTL) to express the expected behaviors of GRNs.
Definition 4
Note that we did not include next operator in the definition of LTL. This is because a typical GRN does not expect something is to be done in the strictly next cycle.
The algorithm for model checking LTL properties on a GRNindividualalways halts on runs shorter or equal to the number of states \(\varSigma (G)\), because the GRNindividualis deterministic (every infinite path is fully described by its prefix of length \(\varSigma (G)\)). Hence it will suffice to perform bounded model checking for paths of length up to \(\varSigma (G)\).
3 Algorithm for parameter synthesis
In this section we present an algorithm for synthesising the weights’ space corresponding to a given property in linear temporal logic. The method combines LTL model checking [2] and satisfiability modulo theory (SMT) solving [4].
The method operates in two steps. First, we represent any GRNindividualfrom the GRN space with a parametrized transition system. In this system, a transition exists between every two states, and it is labelled by linear constraints, that are necessary and sufficient constraints to enable that transition in a concrete GRNindividual (for example, see Fig. 1b). We say that a run of the parametrized transition system is feasible if the conjunction of all the constraints labelled along the run is satisfiable. Second, we search for all the feasible runs that satisfy the desired LTL property and we record the constraints collected along them. The disjunction of such run constraints fully characterizes the regions of weights which ensure that LTL property holds in the respective GRNindividual.
Definition 5
\(\varPhi (\mathbf {\sigma },\mathbf {\sigma }')\) says that a gene \(g'\) is active in \(\mathbf {\sigma }'\) iff the weighted sum of activation and suppression activity of the regulators of \(g'\) is above its threshold.
A run of \((T, v)\) is a sequence of states \(\mathbf {\sigma }_0, \mathbf {\sigma }_1, \ldots \) such that \(\mathbf {\sigma }_n \in \varSigma (T)\) for all \(n \ge 0\), and \(\varPhi (\mathbf {\sigma }_0,\mathbf {\sigma }_{1}) \wedge \varPhi (\mathbf {\sigma }_1,\mathbf {\sigma }_{2}) \wedge \ldots \) is said to be the run constraint of the run. A run is feasible if its run constraint is satisfiable. We denote by \(\llbracket (T,v)\rrbracket \) the set of feasible traces for \((T,v)\). For a weight function \({w}\), let \(\varPhi (\mathbf {\sigma }, \mathbf {\sigma }')[{w}/v]\) denote the formula obtained by substituting \(v\) by \({w}\) and let \((T, v)[{w}/v] = (\varSigma (T), \varPhi ')\), where \(\varPhi '(\mathbf {\sigma },\mathbf {\sigma }') = \varPhi (\mathbf {\sigma },\mathbf {\sigma }')[{w}/v]\) for each \(\mathbf {\sigma },\mathbf {\sigma }' \in \varSigma (T)\).
In the following text, we refer to the parametrized transition system \((T, v)\) and an LTL property \(\varphi \). Moreover, we denote the run constraint of run \(r=\mathbf {\sigma }_0,\mathbf {\sigma }_1,\ldots \in \llbracket (T,v)\rrbracket \) by \({cons}(r)\).
Lemma 1
For a weight function \({w}\), the set of feasible runs of \((T, v)[{w}/v]\) is equal to the language of \((T,w)\), that is \(\llbracket (T,{w})\rrbracket \).
The proof of the above lemma follows from the definition of the semantics for GRNindividual. Note that the run constraints are conjunctions of linear (non)strict inequalities. Therefore, we may apply efficient SMT solvers to analyze \((T, v)\).
3.1 Constraint generation via model checking
Now our goal is to synthesize the constraints over \(v\) which characterise exactly the set of weight functions \(w\), for which \((T,{w})\) satisfies \(\varphi \). Each feasible run violating \(\varphi \) reports a set of constraints which weight parameters should avoid. Once all runs violating \(\varphi \) are accounted for, the desired region of weights is completely characterized. More explicitly, the desired space of weights is obtained by conjuncting negations of run constraints of all feasible runs that satisfy \(\lnot \varphi \).
Since run constraints are always a conjunction of linear inequalities, goodCons is a conjunction of clauses over linear inequalities. Therefore, we can apply efficient SMT technology to evaluate the condition at line 8. The following theorem states that the algorithm GenCons computes the parameter region which satisfies property \(\varphi \).
Theorem 1
Proof
We use the above presentation of the algorithm for easy readability. However, our implementation of the above algorithm differs significantly from the presentation. We follow the encoding of [7] to encode the path exploration as a boundedmodel checking problem. Further details about implementation are available in Sect. 5. The algorithm has exponential complexity in the size of \(T\). However, one may view the above procedure as the clause learning in SMT solvers, where clauses are learnt when the LTL formula is violated [26]. Similar to SMT solvers, in practice, this algorithm may not suffer from the worstcase complexity.
Example 1
4 Computing robustness
In this section, we present an application of our parameter synthesis algorithm, namely computing robustness of GRNs in presence of mutations. To this end, we formalize GRNpopulationand its robustness. Then, we present a method to compute the robustness using our synthesized parameters.
A GRNpopulationmodels a large number of GRNindividuals with varying weights. All the GRNindividuals are defined over the same GRN space, hence they differ only in their weight functions. The GRNpopulationis characterised by the GRN space \(T\) and a probability distribution over the weight functions. In the experimental section, we will use the range of weights \({W}\) and the distribution \(\pi \) based on the mutation model outlined in Sect. 7.
Definition 6
(GRNpopulation) A GRNpopulation \(\mathcal {Z}\) is a pair \((T,\pi )\), where \(\pi :{W}\rightarrow [0,1]\) is a probability distribution over all weight functions from GRN space \(T\).
We write \(\varphi ({\mathcal {Z}})\in [0,1]\) to denote the expectation that a GRN instantiated from a GRNpopulation \(\mathcal {Z}=(T,\pi )\) satisfies \(\varphi \). The value \(\varphi ({\mathcal {Z}})\) is in the interval [0, 1] and we call it robustness.
Definition 7
The above definition extends that of [11], because it allows for expressing any LTL property as a phenotype, and hence it can capture more complex properties such as oscillatory behaviour.
The mutation model that induces the GRNpopulation we use in our experimental evaluation is shown in Sect. 7.
4.1 Evaluating robustness
Let us suppose we get a GRNpopulation \(\mathcal {Z}= ( T, \pi )\) and LTL property \(\varphi \) as input to compute robustness.
For small size of GRN space \(T\), robustness can be computed by explicitly enumerating all the GRNindividuals from \(T\), and verifying each GRNindividual against \(\varphi \). The probabilities of all satisfying GRNindividuals are added up. However, the exhaustive enumeration of the GRN space is often intractable due to a large range of weight functions \({W}\) in \(T\). In those cases, the robustness is estimated statistically: a number of GRNindividuals are sampled from \(T\) according to the distribution \(\pi \), and the fraction of satisfying GRNindividuals is stored. The sampling experiment is repeated a number of times, and the mean (respectively variance) of the stored values are reported as robustness (respectively precision).
5 Implementation

encoding the bounded model checking (BMC) of the parametrised transition system as a satisfiability problem,

computation of goodCons by quantifier elimination,

minimization of goodCons using linear decision diagrams (LDDs),

sampling of mutants with efficient property checking,
5.1 Encoding the bounded modelchecking problem

\(\llbracket T\rrbracket _k\) is the unrolling of the transition relation induced by \(T\), and

\(\llbracket \lnot \varphi \rrbracket _k\) is the propositional encoding of the bounded semantics of \(\lnot \varphi \).
5.2 Symbolic computation of goodCons
5.3 Minimization of goodCons by efficient representation
We use linear decision diagrams (LDD) [10] for representing formula goodCons. A linear decision diagram is a binary decision diagram where nonterminal nodes are labelled by linear constraints and terminal nodes are labelled by either 0 or 1. A function \(w\) satisfies a nonterminal node if \(w\) satisfies the label and satisfies the left child or if \(w\) does not satisfy the label and satisfies the right child, while it satisfies a terminal node if the label is 1. The LDD produced by the parameter synthesis procedure is an LDD over the variables \(w_{g,g'}\) such that, for all weight functions \(w\in {W}\) of \(T\), \(w\) satisfies the LDD iff \(\llbracket (T,{w})\rrbracket \models \varphi \). The representation of the formula goodCons by means of LDD enjoys the property of being minimal, provided an ordering for the linear constraints over \(w_{g, g'}\). Moreover, LDDs inherit optimizations and variable reordering algorithms from BDDs. We implement LDDs by using the CUDD BDD package, where propositional variables are mapped to linear constraints of the original goodCons formula. In general, the BDD package does not contain information about the theory of linear real arithmetics used by the constraints, hence a satisfying assignment for a BDD, might be unsatisfiable when considering the semantics of the constraints. Our implementation ensures consistency against the theory because of the fact that we represent goodCons using LDD, after checking the models \(\mu \) with the SMT solver. The LDD accumulates models \(\mu \) at line 8 that are ensured to be consistent with the theory, as they are models for \(\psi '\), extracted at line 6.
5.4 Sampling with efficient property checking
The robustness evaluation procedure takes a probability distribution \(\pi :{W}\rightarrow [0,1]\) over the weight functions of \(T\) and computes the expected probability of satisfying property \(\varphi \), as in Def. . The procedure samples a number of weight functions from \(\pi \) and computes the ratio between functions that induce GRNindividuals that satisfy the property and the number of samples.

In the first method, which we will call execution method, each sampled GRNindividualis verified by explicitly executing each run of the GRNindividualfrom each initial state and checking if each run satisfies \(\varphi \);

In the second method, which we will call evaluation method, the constraints are first precomputed, and each sampled function is verified by evaluating it in the produced LDD.
6 Experimental results
We ran our tool on a set of GRNs from literature.
We ran the robustness computation by the evaluation and execution methods (the methods are described in Sect. 4.1). In order to obtain robustness and to estimate the precision, we computed the mean of 100 experiments, each containing a number of samples ranging from \(10^3\) to \(10^6\). The total computation time in the execution methods linearly depends on the number of samples used. The total computation time in the evaluation method depends linearly on the number of samples, but initially needs time to compute the constraints. Technically, the time needed to compute robustness by execution method is \(t_{ex} = k_{ex}p\), and the time needed to compute robustness by evaluation approach \(t_{ev} = k_{ev}p+t_{c}\), where p represents the total number of samples used, \(t_c\) is the time to compute the constraints, and \(k_{ex}\) (resp. \(k_{ev}\)) is the time needed to verify the property by evaluation (resp. execution). We used linear regression to estimate the parameters \(k_{ex}\) and \(k_{ev}\), and we present the ratio \(\frac{k_{ex}}{k_{ev}}\) in topleft position of Fig. 4. The results indicate that on six out of eight tested networks, evaluation is more efficient than execution. For some networks, such as osc7, the time for computing the constraints is large, and the gain in performance becomes visible only once the number of samples is larger than \(10^6\). On the other hand, for the networks qi and ncc performance are worse, namely evaluating the constraints is more expensive than finding a counterexample for the property by simulation. In fact, due to their high connectivity, the produced constraints is very big, with respect to the number of possible simulations, as the first is worse case exponential in the number of regulatory relations and the second is worse case exponential in the number of genes.
The network osc25 has been designed to demonstrate the drastic advantage of our approach on a network with 25 genes and for a property with robustness value 1. The model checking approach has a precomputation time of about 3.5 s, and an evaluation time of about 0.95 s for \(10^6\) samples. The simulation approach times out with any number of samples. The reason is that the property is valid, hence the solver does not find any counterexample. The resulting LDD will consist of the node 1 only, hence every sample is evaluated in constant time. On the other hand, as the simulator cannot test for the validity of the property in advance, it has to simulate the network for each out of \(2^{25}\) initial states. The osc25 example shows that the strength of our method stands in its logic engine, and that it is very wellsuited for networks that are likely to be robust, as its approach consists of searching for and enumerating counter examples.
This experimental evaluations shows that our method is preferrable to the simulation method when networks are sparse and when logic reasoning can play a role in reducing the computational effort.
7 GRNpopulationaccording to a mutation model
In this Section, we first present the mutation model without selection, which defines the GRNpopulation we use in our experimental evaluation. Then we show another, more realistic mutation model with selection, which accounts for evolutionary pressure towards a given property.
Genetic mutations refer to events of changing base pairs in the DNA sequence of the genes and they may disturb the regular functioning of the host cell. Such mutations affect the result of the weight function \(w \in {W}\), inducing a range defined by the GRN space.
Evolutionary biologists are interested in how a population of cells evolves over generations, and, more specifically, in a distribution over phenotypes for an extended number of generations. In our model of GRNpopulation \(\mathcal {Z}=(T,\pi )\), the distribution \(\pi \) represents such limiting distribution over the phenotypes \(T\). Depending on whether population is subject to evolutionary pressure selecting for some property, two variants of GRNpopulation arise.
In the following, we first show a model of mutation, and then how the distribution \(\pi \) is derived in both the model without selection and in the model with selection.
7.1 Mutation model
Each gene, a part of DNA which can be transcribed to a protein, is said to be active when the protein encoded by \(g\) is present in the cell. The transcription process is initiated when a transcription factor binds to the promoter region of gene \(g\). A mutation in the promoter region may affect the binding affinity to the transcription factor, and, consequently, that gene’s weight; We assume here that the maximal weight of gene \(g\), that is, \(w(g)\), is achieved for a single sequence of nucleotides, and that it will linearly decrease with the number of mutated nucleotides (where ‘mutated’ refers to being different to the sequence with maximal weight).
A possible way of interpreting the mutation model used in this work is the following. Imagine an experiment which starts with a population of cells, each having a promoter of gene \(g\) in configuration \(\mathbf {a}_0=(a_1,\ldots ,a_{l})\in \{\mathtt {A},\mathtt {T},\mathtt {C},\mathtt {G}\}^{l}\), that is some sequence of \(l\) nucleotides. A mutation in the promoter region may affect the binding affinity to the transcription factor, and, consequently, the gene expression level; We assume that the maximal weight \(w\) is achieved for a single sequence of nucleotides, \({\mathbf {a}}\), and that it will linearly decrease with the increasing number of mismatching nucleotides. This is clearly an oversimplification, since different mutations at the same site will have different effects in the binding energy and not all sites contribute in the same way to the total binding energy. However, it correlates with what is seen experimentally, where it is generally found that there is one “preferred” sequence, and mutations from that sequence decrease the total binding energy. As such, this assumption simplifies the modelling immensely without compromising too much of the biological detail.
We assume discrete generations and that mutations happen only during replication. When the DNA is passed from the mother to the daughter cell, a nucleotide \(a_i\) can mutate to a different value with probability p, and, consequently, it can remain the same with probability \(1p\). Furthermore, we assume that the mutations occur only in the promoter region of genes and that the mutation events are independent within the whole genome, and independent of the history of mutations, as is typically assumed in biology.
7.1.1 Modeling mutations in a nucleotide
7.1.2 Modeling mutations for a single gene
Let us assume \(l\) to be the length of the DNA binding region of a gene^{2} \(g\), and \(P^{id}\) to specify the probabilities of single point mutations occurring in it. Moreover, we assume that the maximal weight of gene \(g\), i.e. \(w(g)\), is achieved for a single sequence of nucleotides, and that the weight will linearly decrease with the number of mutated nucleotides (where ‘mutated’ refers to being different to the sequence achieving the maximal weight). Therefore, if k out of \(l\) nucleotides in \(g\) are mutated, the weight of \(g\) becomes \(w(g)(1\frac{k}{l})\).
If the whole promoter sequence of one gene is modelled by a string \(\mathbf {a}\in \{\mathtt {A},\mathtt {T},\mathtt {C},\mathtt {G}\}^{l}\), its changes over generations are captured by an ldimensional random process of type \((X_n^1,\ldots ,X_n^l)(\mathbf {a})\in \{\mathbf {0},\mathbf {1}\}^l\). Mutation events \(X_n^1\), \(X_n^2\),..., \(X_n^l\) are assumed to happen independently across the genome, and independently of the history of mutations. Then, a random process \(M_n:=X_n^1+\cdots +X_n^l\), such that \(M_n=k\in \{0,\ldots ,l\}\), denotes that the configuration at generation n differs from the optimal configuration in exactly k points. The following lemma defines the values of the transition matrix of the process \(\{M_n\}\).
Lemma 2
Proof
It suffices to observe that u represents the number of mutated nucleotides which remain mutated at time \((n+1)\), and \((ju)\) is the number of unmated nucleotides which become mutated at time \((n+1)\). The existence and convergence of the stationary trivially follows from that \(\{M_n\}\) is regular (irreducible and aperiodic). Each process \(\{X_n^i\}\) at a stationary behaves as a Bernoulli process, with probability of being mutated equal to \(\beta =\lim _{n\rightarrow \infty } P(X_n^i=1)\), and a probability of remaining unmutated \(1\beta \). The claim follows as the sum of l independent Bernoulli processes is binomially distributed. The special case follows because matrix \(\hat{P}^{id}\) has a unique stationary distribution (0.25, 0.75) (denoting by a and b its coordinates, we obtain equations \(a(1p)+b\frac{p}{3}=a\) and \(ap+b(1\frac{p}{3})=b\), which can be satisfied only if \(a=\frac{b}{3}\)).
Moreover, \(\{M_n\}\) converges to a unique stationary distribution, which is a binomial, with success probability \(\beta =\lim _{n\rightarrow \infty } P(X_n^i=\mathbf {1})\). In a special case Open image in new window , the stationary distribution is \(\beta =\frac{3}{4}\), independently of the value of p.
7.1.3 Modeling mutations in a GRN
We assume mutations happen independently across the genome, therefore if the genome is in the configuration \(\mathbf k =(k_1,\ldots ,k_d)\), the probability that the next generation will be in the configuration \(\mathbf k '=(k_1',\ldots ,k_d')\) will be a product of transition probabilities from \(k_i\) to \(k_i'\), for \(i=1,\ldots ,d\). In other words, each sequence \(\mathbf k =(k_1,\ldots ,k_d)\in 0..l_1\times \cdots \times 0..l_{d}\) defines one weight function \(w_\mathbf{k }\), with \(w_\mathbf{k }(g_i)=w(g_i)(1\frac{k_i}{l_i})\). The possible weight configurations of a GRN are defined accordingly.
Definition 8
7.2 GRNpopulationin a model without selection
Definition 9
Equation (7) tells that the probability of observing weight w in generation \(i+1\) equals the total probability of mutating from any possible weight \(w'\) in generation i to weight w.
Lemma 3
The sequence defined in (7) converges to a unique distribution \(\pi (w_\mathbf{k }) = \prod _{i=1}^d{l_i\atopwithdelims ()k_i}\beta _i^{k_i}(1\beta _i)^{d_ik_i}\), where \(\beta _i = \lim _{n\rightarrow \infty } P(X_n^i=1)\).
Proof
Let \({\mathbf {T}}\) be a matrix notation for transition function \(\mathcal{T}\), given some fixed ordering among the weight function. The matrix \({\mathbf {T}}\) is obviously regular, because, from (5), each two states have a nonzero transition probability.
7.3 GRNpopulationin a model with selection
In a model with selection, upon mutations occur, only the individuals satisfying the selection property survive. However, the total population size remains constant, meaning that, even though a number of individuals dies, those individuals who meet the property reproduce more, that is, the nature selects for the individuals satisfying the property. We take the selection property to be equal to the property we check robustness for. In this context, the robustness is the ratio of surviving individuals upon mutations occur on the longterm.
With abuse of notation, let \(\varphi ({W})\subseteq {W}\) denote the subset of surviving weight configurations: \(\varphi ({W}) = \{w\in {W}\mid \varphi (w) = 1\}\).
Definition 10
The probability of seeing weight w in generation \(i+1\) is the probability of transition from any \(w'\) in generation i, but first normalised over all surviving configurations \(w''\). The above definition is inspired by a similar definition presented in [11], i.e., the probability of a singly mutated GRN satisfying the property.
The next theorem states how to compute the robustness in the model with selection. See Fig. 6 for illustration.
Theorem 2
Let \(\mathbf {T}\in [0,1]^{(0..l_1\times \cdots \times 0..l_{d})^2}\) be the matrix encoding the transition probability \(\mathcal{T}\) between two mutated weights, and let \(\varvec{\pi }\) be the (row) vector notation for the converging distribution function \(\pi \). Let \(\tilde{\mathbf {T}}\in [0,1]^{\varphi ({W})\times \varphi ({W})}\) (resp. \(\tilde{\varvec{\pi }}\)) be the restriction of \(\mathbf {T}\) (resp. \(\varvec{\pi }\)) to the components in \(\varphi ({W})\). Then, \(\tilde{\varvec{\pi }}\) is the eigenvector for \(\tilde{\mathbf {T}}\), with the eigenvalue \(\Vert \tilde{\varvec{\pi }}\tilde{\mathbf {T}}\Vert \). Moreover, this eigenvalue is exactly robustness, \(\varphi ( \mathcal {Z})\).
Proof
Finally, \(\Vert \tilde{\varvec{\pi }} \tilde{\mathbf {T}}\Vert \) is the matrix form of \(\varphi ( \mathcal {Z})\) in Dfn. 7, which proves the second part of the claim.
Finally, we show that robustness measure preserves monotonicity in the sense that larger properties imply larger robustness in both model with and without selection.
Theorem 3
Let \(\varphi \) and \(\varphi '\) be two properties such that for all \(w\in {W}\), \(\varphi (w) \Rightarrow \varphi '(w)\), that is, the parameter space where \(\varphi \) holds is fully contained in the space where \(\varphi '\) holds. Then, \(\varphi ( \mathcal {Z})\le \varphi '(\mathcal {Z})\), that is, the robustness with respect to property \(\varphi \) is not larger than the robustness with respect to property \(\varphi '\), both in model without selection and model with selection with respect to property \(\varphi \) (resp. \(\varphi '\)).
Proof
In case of model without selection, the claim trivially holds. For model with selection, by Theorem 2, it suffices to show that the eigenvalue of \(\tilde{\mathbf {T}}(\varphi )\) is not larger than the eigenvalue of \(\tilde{\mathbf {T}}(\varphi ')\). By the assumption, \(\tilde{\mathbf {T}}(\varphi )\) can be obtained from \(\tilde{\mathbf {T}}(\varphi ')\) by deleting rows and columns corresponding to those states that do not satisfy \(\varphi \). By instantiating Gelfand’s formula for the 1norm, the claim follows.
The following terminology and results can be found in any standard linear algebra textbook.
Definition 11
A matrix \(M\in {\mathbf R}^{m\times n}\) is called positive (resp. nonnegative), written \(M>0\) (resp. \(M\ge 0\)), if all entries of M are strictly positive (resp. nonnegative).
Definition 12
A square matrix \(M\in {\mathbf R}^{n\times n}\) is regular, if there exists \(k>0\), such that \(M^{k}\ge 0\).
Theorem 4
 1.
M has an eigenvalue \(\hat{\lambda }\in {\mathbf R}_{>0}\), such that for all other eigenvalues \(\lambda \), we have \(\lambda <\hat{\lambda }\),
 2.
the right and lefteigenvectors associated to eigenvalue \(\hat{\lambda }\) are the only eigenvectors with all real, positive entries.
 3.
the eigenvalue \(\hat{\lambda }\) has multiplicity one.
Proof
(Sketch) We only show the proof for (1). If M has a spectral radius \(\rho (M)\), then \(M'=\frac{M}{\rho (M)}\) has an eigenvector on the unit circle, and all other eigenvalues of \(M'\) are smaller or equal to 1. Assume \(r=\rho (M')\ne 1\). Then, there exists \(m\in {\mathbb N}\), such that \(Re(r^m)<0\). Notice that the matrix \({M'}^m\epsilon I\) has all positive entries for some small enough \(\epsilon >0\). Moreover, by Gelfand’s formula (\(\lim _{k\rightarrow \infty } \Vert A^k\Vert ^{1/k}=\rho (A)\) for any matrix norm \(\Vert \cdot \Vert \)), it follows that \(\rho ({M'}^m\epsilon I)\le \rho ({M'}^m)\le {\rho (M')}^m=1\). On the other hand, since \(({M'}^m\epsilon I)v=(r^m\epsilon )v\) for any vector v, it follows that \(r^m\epsilon \) is an eigenvalue for \({M'}^m\epsilon I\), and consequently \(\rho ({M'}^m\epsilon I)>\Vert r^m\epsilon \Vert =1\). Contradiction.
Hence the spectral radius of M is achieved for a positive real eigenvalue.
Theorem 5
Proof
(Sketch) As M is nonnegative and regular, there exists \(m>0\) such that \(g^m(v_0)>0\). The convergence can be shown from the Jordan canonical form and by subsequently applying the power iteration approach. Let \(v_0\) be the initial distribution, and let \(M=VJV^{1}\) be the Jordan canonical form of M. The vector \(v_0\) can be written as a linear combination of the rows of \(V^{1}\) (generalized right eigenvectors): \(v_0=c_1v_1+\cdots +c_nv_n\), where \(c_1\ne 0\) is nonnegative eigenvector \(v_1\) corresponding to the dominant eigenvalue.
Lemma 4
A regular stochastic square matrix has a unique stationary distribution and it converges to it.
Proof
Since all row sums are equal to 1, the column vector of only ones is an eigenvector, and the corresponding eigenvalue is 1. Hence, there exists a positive left eigenvector \(\pi \) for eigenvalue 1. Then, \(\nu =\frac{\pi }{\pi }\) is a stationary distribution. The proof of uniqueness and convergence follow as in the proof for Theorems 4 and 5.
8 Conclusion and discussion
In this paper, we pursued formal analysis of Wagner’s GRN model, which allows symbolic reasoning about the behavior of GRNs under parameter perturbations. More precisely, for a given space of GRNs and a property specified in LTL, we have synthesized the space of parameters for which the concrete, individual GRN from a given space satisfies the property. The resulting space of parameters is represented by complex linear inequalities. In our analysis, we have encoded a bounded modelchecking search into a satisfiability problem, and we used efficient SMT solvers to find the desired constraints. We demonstrated that these constraints can be used to efficiently compute the mutational robustness of populations of GRNs. Our results have shown the cases in which the computation can be three times faster than the standard (simulation) techniques employed in computational biology.
While computing mutational robustness is one of the applications of our synthesized constraints, the constraints allow to efficiently answer many other questions that are very difficult or impossible to answer by executing the sampled GRNs. In our future work, we aim to work on further applications of our method, such as parameter sensitivity analysis for Wagner’s model. Moreover, we plan to work on the method for exact computation of robustness by applying point counting algorithms [5].
The Wagner’s model of GRN is maybe the simplest dynamical model of a GRN—there are many ways to add expressiveness to it: for example, by incorporating multistate expression level of genes, nondeterminism, asynchronous updates, stochasticity. We are planning to study these variations and chart the territory of applicability of our method.
Footnotes
 1.
The artifact is available at http://pub.ist.ac.at/~mgiacobbe/grnmc.tar.gz.
 2.
More precisely the length of the promoter of gene.
Notes
Acknowledgments
Open access funding provided by Institute of Science and Technology (IST Austria).
References
 1.Azevedo, R.B.R., Lohaus, R., Srinivasan, S., Dang, K.K., Burch, C.L.: Sexual reproduction selects for robustness and negative epistasis in artificial gene networks. Nature 440(7080), 87–90 (2006)CrossRefGoogle Scholar
 2.Baier, C., Katoen, J.P.: Principles of Model Checking. The MIT Press, Cambridge, Massachusetts, London (2008)MATHGoogle Scholar
 3.Barrett, C., Deters, M., de Moura, L., Oliveras, A., Stump, A.: 6 years of SMTCOMP. J. Autom. Reason. 50(3), 243–277 (2013)CrossRefGoogle Scholar
 4.Barrett, C.W., Sebastiani, R., Seshia, S.A., Tinelli, C.: Satisfiability modulo theories. Handb. Satisf. 185, 825–885 (2009)Google Scholar
 5.Barvinok, A., Pommersheim, J.E.: An algorithmic theory of lattice points in polyhedra. New Perspect. Algebraic Combin. 38, 91–147 (1999)MathSciNetMATHGoogle Scholar
 6.Batt, G., Yordanov, B., Weiss, R., Belta, C.: Robustness analysis and tuning of synthetic gene networks. Bioinformatics 23(18), 2415–2422 (2007)CrossRefGoogle Scholar
 7.Biere, A., Cimatti, A., Clarke, E.M., Strichman, O., Zhu, Y.: Bounded model checking. Adv. Comput. 58, 117–148 (2003)CrossRefGoogle Scholar
 8.Cardelli, L.: Morphisms of reaction networks that couple structure to function. BMC Syst. Bio. 8(1), 84 (2014)CrossRefGoogle Scholar
 9.Cardelli, L., CsikászNagy, A.: The cell cycle switch computes approximate majority. Sci. Rep. 2, 1–9 (2012)CrossRefGoogle Scholar
 10.Chaki, S., Gurfinkel, A., Strichman, O.: Decision diagrams for linear arithmetic. In: Formal Methods in ComputerAided Design, 2009. FMCAD 2009, pp. 53–60. IEEE (2009)Google Scholar
 11.Ciliberti, S., Martin, O.C., Wagner, A.: Robustness can evolve gradually in complex regulatory gene networks with varying topology. PLoS Comput. Biol. 3(2), 164–173 (2007)CrossRefMathSciNetGoogle Scholar
 12.Danos, V., Laneve, C.: Formal molecular biology. Theor. Comput. Sci. 325(1), 69–110 (2004)CrossRefMathSciNetMATHGoogle Scholar
 13.de Moura, L.M., Bjørner, N.: Z3: an efficient SMT solver. In: TACAS (2008)Google Scholar
 14.Elowitz, M.B., Leibler, S.: A synthetic oscillatory network of transcriptional regulators. Nature 403(6767), 335–338 (2000)CrossRefGoogle Scholar
 15.Fisher, J., Henzinger, T.A.: Executable cell biology. Nature Biotechnol. 25(11), 1239–1249 (2007)CrossRefGoogle Scholar
 16.Gardner, T.S., Cantor, C.R., Collins, J.J.: Construction of a genetic toggle switch in Escherichia coli. Nature 403(6767), 339–342 (2000)CrossRefGoogle Scholar
 17.Giacobbe, M., Guet, C.C., Gupta, A., Henzinger, T.A., Paixao, T., Petrov, T.: Model checking gene regulatory networks. In: Baier, C., Tinelli, C. (eds.) Tools and Algorithms for the Construction and Analysis of Systems, pp. 469–483. Springer, Berlin, Heidelberg (2015)Google Scholar
 18.Hafner, M., Petrov, T., Lu, J., Koeppl, H.: Rational design of robust biomolecular circuits: from specification to parameters. In: Koeppl, H., Setti, G., di Bernardo, M., Densmore, D. (eds.) Design and Analysis of Biomolecular Circuits, pp. 253–279. Springer, New York (2011)CrossRefGoogle Scholar
 19.Jha, S.K., Clarke, E.M., Langmead, C.J., Legay, A., Platzer, A., Zuliani, P.: A Bayesian approach to model checking biological systems. In: Degano, P., Gorrieri, R. (eds.) Computational Methods in Systems Biology, pp. 218–234. Springer, Berlin, Heidelberg (2009)CrossRefGoogle Scholar
 20.Kwiatkowska, M., Norman, G., Parker, D.: Using probabilistic model checking in systems biology. ACM SIGMETRICS Perform. Eval. Rev. 35(4), 14–21 (2008)CrossRefGoogle Scholar
 21.Lahiri, S.K., Nieuwenhuis, R., Oliveras, A.: SMT techniques for fast predicate abstraction. In: Ball, T., Jones, R.B. (eds.) Computer Aided Verification, pp. 424–437. Springer, Berlin, Heidelberg (2006)CrossRefGoogle Scholar
 22.Rizk, A., Batt, G., Fages, F., Soliman, S.: A general computational method for robustness analysis with applications to synthetic gene networks. Bioinformatics 25(12), i169–i178 (2009)CrossRefGoogle Scholar
 23.Schlitt, T., Brazma, A.: Current approaches to gene regulatory network modelling. BMC Bioinform. 8(Suppl 6), S9 (2007)CrossRefGoogle Scholar
 24.Wagner, A.: Does evolutionary plasticity evolve? Evolution 50(3), 1008–1023 (1996)CrossRefGoogle Scholar
 25.Yordanov, B., Wintersteiger, C.M., Hamadi, Y., Kugler, H.: SMTbased analysis of biological computation. In: Brat, G., Rungta, N., Venet, A. (eds.) NASA Formal Methods, pp. 78–92. Springer, Berlin, Heidelberg (2013)CrossRefGoogle Scholar
 26.Zhang, L., Madigan, C.F., Moskewicz, M.H., Malik, S.: Efficient conflict driven learning in a boolean satisfiability solver. In: Proceedings of the 2001 IEEE/ACM International Conference on ComputerAided Design, pp. 279–285. IEEE Press (2001)Google Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.