Computational Biology and Chemistry 56 (2015) 98–108

Contents lists available at ScienceDirect

Computational Biology and Chemistry
journal homepage: www.elsevier.com/locate/compbiolchem

Research Article

On the impact of discreteness and abstractions on modelling noise in
gene regulatory networks
Chiara Bodei a , Luca Bortolussi b,g,h , Davide Chiarugi g,c,∗ , Maria Luisa Guerriero d ,
Alberto Policriti e , Alessandro Romanel f
a

Dip. di Informatica, Università di Pisa, Italy
Dip. di Matematica e Geoscienze, Università di Trieste, Italy
c
Max Planck Institut of Colloids and Interfaces, Potsdam, Germany
d
Systems Biology Ireland, University College Dublin, Ireland
e
Dip. di Matematica e Informatica, Università di Udine, Udine, Italy
f
Centre for Integrative Biology (CIBIO), University of Trento, Italy
g
CNR-ISTI, Pisa, Italy
h
Modelling and Simulation Group, University of Saarland, Campus E 1 3, Saarbruecken, Germany
b

a r t i c l e

i n f o

Article history:
Received 13 August 2014
Received in revised form 26 February 2015
Accepted 7 April 2015
Available online 8 April 2015
Keywords:
Gene regulatory networks
Discrete modelling
Hybrid system
Quasi-steady state approximation
Stochastic noise

a b s t r a c t
In this paper, we explore the impact of different forms of model abstraction and the role of discreteness
on the dynamical behaviour of a simple model of gene regulation where a transcriptional repressor negatively regulates its own expression. We first investigate the relation between a minimal set of parameters
and the system dynamics in a purely discrete stochastic framework, with the twofold purpose of providing an intuitive explanation of the different behavioural patterns exhibited and of identifying the main
sources of noise. Then, we explore the effect of combining hybrid approaches and quasi-steady state
approximations on model behaviour (and simulation time), to understand to what extent dynamics and
quantitative features such as noise intensity can be preserved.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction
Regulating gene expression is a complex work of orchestration,
where the instruments play with improvised variations without
a fixed music sheet. Under this regard, the regulation process,
in which DNA drives the synthesis of cell products such as RNA,
and proteins, can be thought of as a stochastic process. The
amount of RNA and proteins in living cells must be thoroughly
tuned, both to manage effectively housekeeping functions and to
respond promptly to upcoming needs (e.g. to adapt to environmental changes). To this end, gene expression is equipped with
several control mechanisms and strategies that grant both reliability and flexibility in terms of throughput. Nevertheless, when
observed at the single cell level, the amount of molecules involved

∗ Corresponding author at: Max Planck Institute of Colloids and Interfaces, Am
Mühlenberg 1, 14476 Potsdam, Germany. Tel.: +49 (331) 567 9617.
E-mail addresses: chiara@di.unipi.it (C. Bodei), luca@dmi.units.it (L. Bortolussi),
davide.chiarugi@mpikg.mpg.de (D. Chiarugi), maria.guerriero@astrazeneca.com
(M.L. Guerriero), policriti@dimi.uniud.it (A. Policriti), romanel@science.unitn.it
(A. Romanel).
http://dx.doi.org/10.1016/j.compbiolchem.2015.04.004
1476-9271/© 2015 Elsevier Ltd. All rights reserved.

in gene expression and its regulation fluctuates randomly (Raj and
van Oudenaarden, 2009). This stochastic effect at the molecular
level turns out to play important roles in conditioning cell-scale
phenomena, e.g. cellular fate decision making, incomplete penetrance or enhanced fitness through phenotypes variability (Raj and
van Oudenaarden, 2009).
Pioneering works (Novic and Weiner, 1957; Ross et al., 1994)
showed that gene expression in populations of genotypically identical cells (i.e. with the same genetic constitution) is highly variable
even when epigenetic conditions (i.e. the ones that result from
external rather than genetic influence) are kept constant. In Elowitz
et al. (2002), the authors identified such a population variability
and decomposed the extrinsic and intrinsic contributions therein.
Also, in Murphy et al. (2010), it is shown that this variability it is
controllable.
Recent developments in experimental techniques (see Raj and
van Oudenaarden (2009) for a review) have made it possible to
detect and count individual molecules and, therefore, to measure
the amount of mRNA and proteins in single cells. These measurements have clearly shown that the number of mRNA and proteins
can vary significantly from cell to cell. This variability is caused

different parameter combinations lead to a range of qualitatively different dynamics. In such mechanisms. Consequently.. Stamatakis and Zygourakis (2011)... This model has been widely studied (e. This precision comes with a high computational cost.. Computational methods can significantly help investigating the synergistic mechanisms underlying the regulation of gene expression. abstracting away noise and randomness due to stochastic fluctuations. This approach succeeded in gaining insights on the stochasticity of gene expression underlying circadian clocks (Chabot et al. In this context. 1967. Models of biological systems proposed in the literature can vary in terms of the abstraction level used to represent molecular amounts (in this regard a model can be either discrete or continuous) and in terms of the underlying paradigm used for describing the temporal evolution of the system. mostly based on (variants of) Gillespie’s stochastic simulation algorithm (Gillespie. in contrast to the common knowledge. Identifying which strategies come into play in generating or dampening noisy behaviours is even more challenging. by systematically studying the impact of parameters variations on the global dynamics of the system. However.g. As a consequence. having a minor influence on the global noise pattern. even numerical simulation can be computationally very expensive.e. Balázsi et al. obtaining a set of Stochastic Differential Equations (SDEs).g. i. 2000. In particular. the global intensity of the feedback depends on parameters related to every single control point. as in the case of bimodal mRNA distributions generated by long transcriptional bursts. (ii) quantifying the impact of each reaction of the modelled system on the overall dynamics and on noise patterns. Thus. The authors show. These findings have raised new interest in analysing the role of noise in gene expression. At a higher level and on a longer term. In this work.. Marquez Lago and Stelling. Ordinary Differential Equations (ODEs) have been extensively used over the years to describe the behaviour of biological processes. 2008). we consider a simple model of gene regulation in a transcription/translation genetic network. we recall also (Salis and Kaznessis. governing the dynamics of the protein and of the mRNA. 2009) and is studied.. Nevertheless. we study a stochastic model of the negative feedback loop to construct an exact picture of its possible behavioural patterns and of the effects of noise. There are several modelling strategies that can lead to different kinds of computational models. 2010). because this is governed in a non-trivial way by several quantitative parameters. during which mRNA level approaches a new steady state (Raj and van Oudenaarden. Bartholomay. 1958). where a transcriptional repressor negatively regulates its own expression. have been successfully introduced to overcome the modelling limitations of continuous methods. / Computational Biology and Chemistry 56 (2015) 98–108 by the fundamentally stochastic nature of the biochemical events involved in gene expression (Raj and van Oudenaarden. The regulation process includes multiple steps leading from gene transcription to the translation of the resulting mRNA to obtain the encoded protein. In such models.C. The extensively studied regulation paradigm represented by the feedback control strategy can be used to explain the mechanisms controlling gene transcription and translation. Bortolussi and Policriti. 2011. The handy dimension of our model of gene regulation allows us to play with different models and techniques. these models are usually studied resorting to simulation approaches. can be safely approximated in a deterministic fashion. 2009). Characterising the contributions of each single control point in the regulation process of gene expression is a complex task. where population-level mathematical frameworks are introduced and applied. we systematically apply various forms of model abstraction. where several biochemical mechanisms play a role (see Alberts et al. Stamatakis and Zygourakis (2010). On the one hand. Actually. e. we establish a link between the parameter space and the observed temporal patterns. e. continuous methods still fail to properly describe various phenomena arising from stochastic fluctuations in systems involving small copy numbers of molecules (Resat et al. Bodei et al. the diverse behavioural phenotypes. 2001)... 2009).. They provide modellers with powerful and well assessed analysis and simulation techniques. rather directly. approaches based on a discrete and stochastic formulation.e. Raj and van Oudenaarden. (2009). Raj et al. One way to represent noise is to couple a Gaussian noise term to the model equations. with approaches such as the one described in Dauxois et al. in Mantzaris (2007). for example. A compromise between accuracy and efficiency can be obtained by combining discrete and continuous evolution in so-called hybrid approaches (Pahle. Starting from the analysis proposed in Marquez Lago and Stelling (2010). only the most relevant sources of noise will be represented via a fully detailed stochastic description. 1994).g. This allows us to identify those reactions that. On the other hand. 2007) and genetic switches (Becksei et al. 1977). In this way.g. 2006). in some cases. it is becoming clear that noise and stochasticity underlie critical events in cell’s life such as differentiation and decision making (Balázsi et al. In particular.. in order to mitigate the inefficiency of exact stochastic simulation methods. understanding its behaviour is not easy. 2011). Unfortunately. indeed. 2008. 2005. depending on both the particular purposes and on the features of the available data. The overall behaviour emerges from the complex interaction between feedback strength and other parameters of the system. the intensity of noise in this system is analysed in terms of various parameters. because it is a minimal system that explicitly describes the processes of transcription and translation and because it is a basic component of many complex biological systems. 2009.. since the analytic solution of the underlying equation is often infeasible for real size systems. ODE models implicitly assume continuous and deterministic change of concentrations. some authors suggest that random phenotypic switching can represent an efficient mechanism for adapting to fluctuating environments (see. Each step represents a possible control point. with special emphasis on the strength of the negative feedback. Ross et al. where hybrid approaches for stochastic simulation of gene networks have been developed. especially for certain parameter sets. This. However. in Marquez Lago and Stelling (2010). (2002) for a comprehensive review and Lillacci and Khammash (2010) for an application to model selection). the phenotypical variability (i. Despite its apparent simplicity. Hume. Stekel and Jenkins. They also relate the possible different dynamical regimes with the different regions of the parameter space. the variability resulting from the interaction of the genotype with the environment) exhibited by populations of identical organisms can be directly caused by stochasticity at the single cell level. 2010. 2013). such as the ones built upon 99 Continuous-Time Markov Chains (CTMCs) (McQuarrie. in the form of autocatalytic reactions systems. Moreover. 2009. Salis et al. Stochastic systems are formally represented through a chemical master equation and have also been studied. that noise generally increases with feedback strength. The copy numbers of molecules and individual entities in the cell space are discrete and the reactions in which they are involved are stochastic events. we abstract the discrete stochastic dynamics into . makes it difficult to capture qualitatively different outcomes arising from identical initial conditions (e. which can be either deterministic or stochastic. we aim at setting up a systematic strategy for correctly building hybrid models of biochemical systems. That paper demonstrates how the application of engineering principles to the role of feedbacks in a biological context could be misleading. we investigate the model with two goals in mind: (i) identifying the parameters that play a key role in the regulation process.

medium (M). Therefore. duplications. the faster P’s response. inversions) and can involve different genomic regions (e. defined as ˛ · unbindP . However. ensuring that translation rate does not become too large considering a maximum rate constant of 108 –1010 M−1 s−1 (Zhou and Zhong. 2. hence using a different mechanism to represent repression intensity. hence having potentially different and sometimes unpredictable effects on the expression of the surrounding/overlapping genes.01. we found out. In the following. we identified the following two key parameters: (1) P binding/unbinding ratio ˛ = bindP /unbindP .0166 16. Afterwards. in particular on the behaviour of the bursty protein regime and on the feasibility of the QSSA (see also Section 3. coding).01 XP0 XM0 0. These abstractions are not blindly applied. After a prescreening in which we used several values according to Marquez Lago and Stelling (2010) for stochastic simulations. we define P production rate as the product of P degradation rate and the steady state value for P per M molecules.0166. Similarly to Marquez Lago and Stelling (2010). gene repression.e. and of the interactions between dynamical regimes involved in the model (partly reported in the Supplementary Material. G.g. the authors show that in this network. The gene (G) is transcribed into an mRNA molecule (M). in this work we choose to vary ˛ in the set {0. While in Marquez Lago and Stelling (2010) the authors fix the binding rate and vary the unbinding rate (thus increasing feedback strength by increasing the binding strength). Rate constant Reaction Description prod M prod P deg M deg P bind P unbind P G→G+M M→M+P M →∅ P →∅ G + P → Gb Gb → G + P mRNA production (transcription) protein production (translation) mRNA degradation protein degradation repressor binding repressor unbinding Table 2 Summary of the combinations of parameters considered and their respective initial values for XP and XM . The authors claim that this simple feedback network has a counter-intuitive behaviour: while. Table 1 Biochemical reactions of the self-repressing gene network considered. From a dynamical point of view. Stekel and Jenkins (2008). making it switch from its free active state to an inactive state (Gb).001 and vary degP = ˇ · degM while varying ˇ. it captures how the protein level reacts to fluctuations at the mRNA level: the higher the value of ˇ. 100}. P regulates its own expression by means of a negative feedback loop: P can bind to G.6. In Marquez Lago and Stelling (2010). prodP = degP · Psteady with Psteady equal to 3500. in our work we fix the unbinding rate and vary the binding rate (thus increasing feedback strength by increasing the binding affinity).g. noise increases with feedback strength. We have chosen this value because it allows us to obtain biologically meaningful numbers of proteins for all the parameter combinations.3 and Appendix A in the Supplementary Material). The first parameter ˛ was used in Marquez Lago and Stelling (2010) as an indicator of the strength of the feedback regulation: the higher the value of ˛. XG . 1(a). enhancer. in engineering. and high (H). negative feedbacks are assumed to be a mechanism to decrease noise. but are rather considered only in the regions of the parameter space for which the hypotheses underlying these techniques are (reasonably) satisfied.6 16600 L-H L-M L-L 86000 25 M-H M-M M-L 2700 1 H-H H-M H-L 86 0 First of all. and so noise at the level of XM level will propagate more effectively to XP . where too often in silico techniques based on abstractions or approximations are used off-the-shelf. All reactions follow mass action kinetics and the units of their rate constants are s−1 . is a genetic network composed by the reactions reported in Table 1 and schematically represented in Fig. The second parameter ˇ is the ratio between the half-life of M and the half-life of P. a very detailed phenomenological description of the different behaviours of the model has been carried out. both M and P can be degraded. with a preliminary high level analysis of the stochastic dynamics of M. The model The model we consider. and Gb are denoted by XP . respectively.100 C. we fixed the value for degM to 0. 16600}. These effects may be difficult to model and may introduce a level of variability that goes beyond the aim of this work. Moreover. (2) P/M degradation ratio ˇ = degP /degM . reasonably changing this parameter does not disrupt the observed behaviours. it describes the speed at which production and degradation of the protein reach equilibrium. we ignore copynumber variations (CNVs) and assume to have a single copy of the gene. The reason for this assumption is that CNVs may rise from different structural rearrangements (e. Similarly. 1. This choice has an impact on the dynamics of the network. Bodei et al. The model is a minimal system that describes the self-regulated transcription/translation of a gene into a protein. P.e. The values of the stochastic rate constants we use are based on those from Marquez Lago and Stelling (2010) and their units are s−1 . All reactions are assumed to follow mass action kinetics. ˛ ˇ 100 1 0. and XGb . Appendix A). 1982) (see Appendix B in the Supplementary Material . promoter. / Computational Biology and Chemistry 56 (2015) 98–108 continuous deterministic dynamics for some of the model components. in this section we provide a simpler picture of the possible behavioural patterns of the model. 16. Following on from those results. the speed of the reaction is proportional to the product of the amounts of each reactant and a kinetic constant. XM . the amounts of P. in fact. and ˇ in the set {0. thus obtaining a stochastic hybrid model. by applying quasi-steady state approximations (QSSAs). which in turn is translated into a protein P. The parameter ˛ is varied by fixing the value for unbindP to 1 and modifying bindP . changing parameters in order to explore a large portion of the parameters’ space. This means that XM and XP will be highly correlated for high ˇ. Essentially. that the seemingly counter-intuitive behaviour can be explained by analysing a combination of a minimal set of parameters. and we refer to the combinations of parameters with the names listed in Table 2. we identified the values that proved to be representative in reproducing the complete gamut of all the interesting observable phenotypes. This enables us to explore parameter combinations with a low degradation rate. the stronger the repression and the smaller the number of mRNA molecules on average. without any concern on their applicability and faithfulness. M. instead. Finally. i. described among others in Marquez Lago and Stelling (2010). We believe that the methodology we adopt is a reliable approach for modelling in systems biology. relative to the mRNA. i. we simplify the model structure reducing the number of model components and reactions. In the following. we refer to these three values for ˛ and ˇ as low (L). deletions.

XP steady state is also high.95 0.026 H-M 576.649 0.5635 0.701 0.7728 0.3370 0.168 0. one for each parameter combination.0077 24.5245 0.4732 0.967 0.9828 0. In order to outline the possible dynamics.3658 0.1727 0.5245 0.024 H-M 433. moreover.1467 24. 1. A comprehensive summary of population statistics of XP and XM is reported in Table 3 (top panel).1839 0. sampled at time t = 10000 (see footnote 1).027 H-M 441.2987 0.9897 Stochastic model with QSSA on binding/unbinding L-L 85978.4429 1. respectively.199 H-H 3407. by applying a result in (Costa and Dufour.9473 0. the initial XP and XM are 2700 and 1. XP dynamics are slower than XM ’s.751 M-M 3259.0429 0.1. / Computational Biology and Chemistry 56 (2015) 98–108 101 Fig.246 0.128 H-H 2201.274 H-L 84.9569 Hybrid model with QSSA on binding/unbinding L-L 85832.000 s.008 24.4482 7.2427 0. In case of M-L combination P dynamics are still slower than M’s but repression is stronger than before. Analysis of stochastic behaviour We proceed now by building a fully stochastic model from our set of reactions.894 M-H 4503.XM ) Stochastic model with explicit binding and unbinding L-L 85986.524 0. respectively.4504 6.1807 0.7359 0.688 L-H 88768.6555 0.1027 24.2280 0. Table 3 Statistics for different abstractions of the explicit stochastic model.1891 2. computed from 1000 simulated trajectories.445 1. consequently.921 M-H 4369. we are taking the population average at the chosen time 10.261 0.628 0.9847 0. in our context.1970 0. for further details). 2008).e.2315 0.788 M-M 3231.105 24. It is not surprising that this is also the combination with the lowest XP − XM correlation and coefficient of variation for XP .165 H-H 2176. at steady state XP 1 We empirically identified 10. 2006).249 H-L 85.e.5815 0.6452 0.799 M-M 3266.2310 0. 000 and 25. We finally fix prodM to 35 considering an average RNA polymerase transcription rate of 24–79 nucleotides/s and 1100 base pairs as the average size of an mRNA molecule (Vogel and Jensen.2034 0.5176 0.9301 0. for combinations having high ˛.4506 1.752 L-M 86285. respectively.1400 1. Moreover.4492 6. the initial XP and XM are 86.0079 25.0077 24. Moreover. but it can be shown also for the hybrid models.477 0.2052 0.3802 2.2124 0.1527 24..753 0. However.874 0.1479 1. With L-L combination we have that.4538 6.3546 1.908 L-H 85662.962 M-H 4454.37 M-L 2705. 2. Finally.9696 0.6442 0. Arrows indicate the direction of reactions. 2 shows. Fig. and further details are in the Supplementary Material.2178 0. we partition our parameter space into different regions.457 0.1 As listed in Table 2.9487 0.49 M-L 2707.76 0.9803 0.539 L-M 85991.262 0.9689 behaves like a simple birth-death process with very low noise.936 M-H 4440. dot-headed and T-headed lines represent reaction catalysts (i.5105 0.2343 0.9926 0.7833 0.1527 0. XP fluctuates with a lower frequency than XM does.046 0.622 0.797 M-M 3313. as commonly done.7615 0.2001 0. by creating nine different versions of our model.423 0.019 H-M 526.000.1683 0.972 0.76 0.21 H-H 3386.1072 0.457 L-M 86562.2160 0. i.1236 2.0061 2.289 H-L 84.9766 0.058 0.1048 24.5904 0. For combinations having medium ˛.0425 0. This is obvious for the Continuous-Time Markov Chain (CTMC) models.881 M-L 2724. the average XM value is relatively high and.663 0.9133 0.102 24.47 L-M 86693. i.000 s as a limit time that is long enough to allow us to observe the specific steady state dynamics of P and M in all the parameters combinations we considered. in order to undertake a thorough stochastic analysis of its dynamics.1134 0.9408 0. one representative simulated trajectory of XP together with its coefficient of variation (CV) and the correlation between XP and XM at time t = 10.1210 0. characterising ergodicity in terms of a Discrete-Time Markov Chain (DTMC) obtained by sampling trajectories of the Piecewise Deterministic Markov Process (PDMP) at random times.304 1. ˛-ˇ mean(XP ) CV(XP ) mean(XM ) CV(XM ) cor(XP .8783 0. Diagrammatic representation of the gene regulatory model described in this section and of its variants with QSSAs described in Section 4.71 0.14 0.57 0.5913 0.9910 0. 000.196 0.C.5939 0.8464 0. a framework that provides us with all the algorithms needed to run stochastic.29 0.667 L-H 86956. We can say that XP is very close to its deterministic average behaviour.and the right-hand side of the reaction) with. positive and negative roles.8865 0.5999 0.2677 0.55 M-L 2727.9824 Hybrid model with explicit binding and unbinding L-L 85857. respectively. causing XM and XP steady state values .153 0. The simulation tool we use is COPASI (Hoops et al.6356 1.8650 0.1784 0.8841 0.1397 24.7733 0.608 L-H 85935.61 0.7066 0. and deterministic simulations.1759 0.5904 0.6294 0.3392 0.297 0. these two approaches are equivalent as all the models we consider are ergodic. and not the time average along a single trajectory. for combinations having low ˛.6054 0. species which are on both the left.141 1.1114 0.3553 0.8501 0.398 1. Bodei et al. since repression is weak. since P degradation is slower than M degradation. hybrid.04431 0.e.2433 0.3234 0.9803 0.2037 0.99 0. the initial XP and XM are 86 and 0. for each of the nine versions of the model. an mRNA molecule is produced (on average) in 14–46 s.0426 0.6114 0.27 H-L 84.14 25.2507 0. For each model we then set the initial amounts of M and P to the corresponding steady state values and run 1000 stochastic simulations with limit time 10. 1994).4492 1.2287 0.5633 0.

for high values of ˛. Hence. the underlying biological mechanism still remains unclear. Even though many complex mechanisms. such dynamics are still very close to a simple birth-death process but. With L-M combination XM fluctuates at steady state around a value of 25. XM fluctuates only between 0 and 1. By further increasing the repression with H-H combination. When ˇ is high. This causes. with higher noise. XM takes very small values).. P fluctuates with a higher frequency than M. our study supports the fact that gene regulation mechanisms themselves (and the . Our exploration of the parameter space is clearly model-specific and cannot be directly generalised to other models of gene regulatory networks. we have that XM and XP start to show a correlation higher than 0. Sampled XP trajectories from the explicit stochastic model computed using the Gibson-Bruck stochastic simulation algorithm (Gibson and Bruck. meaning that its dynamics are noisier. we have that the noise in XM is reflected and amplified at the level of XP .5. hence XM is 0 most of the time. Moreover.. each change in XM is followed by a change in XP . This increase in noise in all combinations having low ˛ can be visually observed also in the XP distributions in Fig. for all three values of ˛. hence it reaches stochastic equilibrium faster than M. Combination H-L generates a strong gene repression that causes XM to fluctuate between 0 and 1. a behaviour that is shown by XP . The coefficient of variation and correlation trends indicated by the arrows and in the panels are reported in detail on the top panel of Table 3. (2005)).102 C. Note that the trajectories are reported for a longer time (t = 50. In particular. involving the biochemical machinery of DNA transcription. S2 it is easy to see that we have three main discrete levels of XP situated around 0. we end up with a bimodal behaviour. to be much lower. As a consequence. 2006). since P degradation is very low. sampled at time t = 10. P dynamics look like M’s with discrete levels properly rescaled. S2. as indicated also by the increment of the coefficients of variation for XM and XP . S2 we can observe that these two levels are 0 and 3500. that is quantitatively equivalent (when scaled on XM ). for a high value of ˇ. This combination of discreteness and possible absence of M is immediately reflected in P dynamics. the time course of P copy number starts to exhibit a multi-modal behaviour. 000 for computing the statistics is valid. XP is able to reach its steady state between successive changes in XM level. XM starts to take very low values with fluctuations between 0 and 4 (see Fig. This is clearly the consequence of the fact that. a behaviour that exhibits bursts. 2000) using different combinations of ˛ and ˇ values. Indeed.98). the correlation between XP and XM to be very high (greater than 0. the effect of repression is even amplified by the almost constant presence of P. 2005) and (especially) in eukaryotes (Raj et al. 000 s) to better visualise the long term trends and to show that the choice of time t = 10. generating a behaviour that is noisier than in L-L. This can be explained by observing that XM fluctuates between 0 and 3 and considering that. Bodei et al. our results clearly illustrate that this kind of thorough exploration is essential when studying genetic networks. XP tends to slowly degrade but displays some small “steps” corresponding to those rare and short intervals in which XM takes value 1. but in this case the coefficient of variation is higher. This is evidently caused by the high degree of discreteness in XM (i. Note that the system alternates periods in which P is degraded to periods in which M is produced in few copies.e. While there is strong experimental evidence supporting transcriptional bursts both in prokaryotes (Golding et al. which mainly alternates between two discrete levels. S1 for instances of XM simulated trajectories). 2. however. With the intermediate combination M-H. in order to understand the role of feedback loops in generating different patterns of time courses of both mRNA and translated proteins. we showed that certain parameter combinations are responsible for bursty temporal patterns. seem to contribute to pulsatile transcription (as suggested in Golding et al. With combination L-H XP fluctuates around the mean value with high level of noise. In this parameter regions. with combination H-M. 7000 and 10. By increasing ˇ to a medium value. 000 s (see footnote 1). / Computational Biology and Chemistry 56 (2015) 98–108 Fig. bursts are emerging because P degradation is fast enough to increase the frequency of the intervals in which XM is equal to 1 and to allow XP to rapidly reach a low amount when XM is back to 0. From Fig. from a methodological point of view. Since the number of P molecules varies in accordance with the number of M ones. In particular. and are computed from 1000 simulated trajectories. described in Section 3. 500. From Fig. and so XP increases. 3500. With M-M combination we can observe that XP starts to present an irregular fluctuating-like behaviour till reaching. In particular.

as rates of discrete jumps can depend on the value of continuous variables. which is a partial differential equation. / Computational Biology and Chemistry 56 (2015) 98–108 relative speed of the different reactions involved) play an important role in determining this pattern of transcription. such as leaping (Gillespie and Petzold. e. 2000) and it usually works well. 2010). We will address this issue in Section 3. due to technical problems in detecting single molecules of mRNA and proteins in the same cell. Evaluating this parameter in individual living cells is not an easy task.1. Model abstractions 3. In particular. Ramaswamy and Sbalzarini. although less accurate as feedback strength increases. Hybrid abstraction In the previous section. where P’s dynamics are much faster than M’s. respectively. one can either use approximate simulation algorithms. Nevertheless this correlation analysis can give us useful insights into the process of gene expression. Modelling each part of the loop following Gillespie’s computational approach. and  and . the discrete process is non-homogeneous in time. 000. at the same time (Raj and van Oudenaarden. 2000. so that noise at the level of P is extremely low. Our case study also shows the importance of modelling without abstracting away any part of the feedback loop in order to reproduce the full spectrum of possible behaviours. (ii) a replacement of the stochastic discrete dynamics with continuous deterministic ones. This also holds for G’s dynamics. 2009). 1993). or adopt some form of model abstraction (see. justifying them by a though a posteriori statistical evaluation. plays a central role in determining the qualitative pattern of dynamical evolution. It is known that this type of approximation is exact in the thermodynamic limit (Gillespie. Crudu et al. especially in the high repression regime. Indeed. Dynamic partitioning.C. Hybrid simulation algorithms for biochemical systems based on PDMPs (Pahle. for small ˇ the approximation is reasonable. while their model can describe both fluctuating and bursty temporal patterns. Continuous variables are subject to continuous evolution. as there are no simply derivable analytic formulae to invoke. Bodei et al. the number of firing P production and degradation events is so large that explicit simulations are severely slowed down. while maintaining discrete dynamics of M and gene repression. 2003. we considered the explicit stochastic model of the simple negative feedback loop. 2006). Here we discuss methods in the latter category. the computational analysis can be difficult to carry out because exact simulation algorithms can be very inefficient under certain parameter configurations. provided that the number of molecules in the system is sufficiently large. where the highest values for XP – XM correlation are obtained when ˇ is high. We will however discuss and use heuristic arguments. instead. in general according to their copy numbers in order to minimise errors caused by the continuous approximation. to justify the use of abstractions in certain regions of the parameter space. In particular. although the latter approach is computationally rather demanding (Trivedi and Kulkarni. These results are consistent with the predictions of our model. whose analysis gives us a precise picture of the patterns of dynamical behaviour and of the effects of noise. from discrete stochastic to continuous deterministic. the rates of mRNA transcription and degradation. thus obtaining a stochastic hybrid model. the amount of M ranges around 25 and P’s slow dynamics have the effect of averaging the noise at the level of M. and (iii) an approximation obtained by the removal of specific model reactions performed by the computation of quasi-steady state values. In general.. 2). This behaviour is different from the one observed by performing stochastic simulations. should be less relevant in this respect if the steady state of XP is sufficiently high. it is worthwhile recalling those obtained by Raj et al. In this case. Evaluating the system of ODEs associated with our model. These processes are described by a set of continuous variables and a set of discrete variables. (2006). it does not account for bimodality in mRNA distribution. while discrete changes happen spontaneously at times determined by exponential distributions. while it fails if noise plays a relevant role in the system dynamics. . which show that mRNA and protein levels are strongly correlated when protein lifetime is short. corresponding to a decrease in the number of M molecules (see Fig. but that this correlation decreases when protein lifetime is long. in this situation. we consider three abstraction approaches: (i) an approximation of the stochastic discrete dynamics with continuous deterministic ones. for instance. in which not all the parameter configurations yield patterns that converge to a steady state (see Fig. focusing both on techniques that abstract the dynamics. When exact simulation is not feasible. noise at the level of M propagates more or less rapidly to P’s dynamics. 2009. localised to some of the components of the model.. 3. Therefore. Mathematically.. This happens. The partitioning rule can be either static or dynamic. The intrinsic noise of P’s dynamics. we can easily verify that the molecule numbers for the different species converge to stable steady states for all parameter configurations considered. This suggests that abstracting away details even in a simple biological model can result in a loss of expressiveness. Another result of our analysis which is worth considering in more detail is the trend of XP – XM correlation. 2). 1993). 103 Establishing the quality of an abstraction a priori is a difficult issue. like quasi-steady state approximation (Rao and Arkin. after a screening of the current parameter set. Peccoud and Ycart (1995) specified a Markovian model of gene regulation considering four parameters:  and . However. as it allows us to study the propagation of fluctuations in mRNA levels to the amount of produced proteins. applied to all model components. From the discussion above. even for this simple genetic network.2. 2009) usually implement some heuristic rule for partitioning species into discrete and continuous ones. Continuous deterministic abstraction The most common approach for dynamics abstraction is to replace the stochastic model with a fully deterministic one based on ODEs. in a study similar to ours. 3. that simplify the model structure reducing the number of reactions and components. in fact. belonging to the class of Piecewise Deterministic Markov Processes (PDMPs) (Davis. a reasonable abstraction of this model would see P as a continuous quantity evolving according to an ODE. is performed during simulation: molecules that at a certain time fall below a specific threshold are treated discretely and stochastically. Among the few experimental results currently available. Then.g. the rates of gene switching from active to inactive state and vice versa. Rathinam et al. and on techniques. as in CTMCs. for low ˛ and high ˇ. it follows that the inherent discreteness of M and G evolution. Gibson and Bruck. PDMPs can be analysed by simulation or even by numerical solution of their master equation. this gives rise to a stochastic hybrid model. instead. thus obtaining a model defined by a set of ordinary differential equations (ODEs). Static partitioning is performed before simulation. 2003). A configuration of parameters where the deterministic approximation of the whole system turns out to be appropriate (with respect to the behaviour of P) corresponds to low ˛ and ˇ (low repression and slow P). and the number of P molecules ranges close to 100. we obtain an exact stochastic model (under the assumption that molecules are well stirred).

model variables are partitioned into fast and slow. 2003. as now P exerts a repression also when its value lies between 0 and 1. The idea behind QSSA is that. 3. assuming that the entities involved only in these reactions are at their steady state. since we are approximating only P’s dynamics as continuous. when the number of molecules falls below 10 we switch to a fully stochastic system) and the discrete-to-continuous switching threshold to 20. We stress that the choice of which hybrid or ODE solver to use is crucial. In particular. containing only the slow species. medium ˇ) trajectories of P for three different abstractions. The rate of going from G = 1 to G = 0 is then ˛ · unbindP · XP . The steady .. 2009).3. P’s deterministic dynamics are stiff. the hybrid model outperforms the stochastic simulation in those parameter regions in which P’s dynamics are the bottleneck for stochastic simulation (Table S6). However.e. QSSA of gene dynamics is assumed in the majority of models of genetic networks. We applied dynamic partitioning to our system. in this case. Results for the stochastic model with explicit binding and unbinding can be found in the corresponding panels in Fig. In our system. at the price of reducing the integration step of ODE solvers in order to avoid significant loss in accuracy. in such an extreme. Approximate hybrid strategies can avoid this overhead (Pahle. This supports our initial conjecture that the inherent discrete and stochastic dynamics of G and M are what mainly determine noise modes of protein expression.g. In fact. For large values of ˇ.4 Then. For the simple binding/unbinding mechanism considered in our model. while the rate of going in the other direction is unbindP . while the performance is less appealing for larger feedback strength. 2005) should be used. with some loss of precision for M and large values of ˛. for large values of ˇ). keeping XM and XP fixed. the hybrid model seems to be also able to capture quite accurately the first two moments of the distribution as well as the correlation between XM and XP (Table 3). 2005. Comparison of multi-modal (medium ˛. Quasi-steady state approximation Fig. the two species describing the gene state (G and Gb) and the binding and unbinding reactions can be removed from 3 In hybrid simulation. Then. In all other cases. 2009. we get that Gsteady = 1 with probability 1/(1 + ˛ · XP ). 2007). 3. one can remove these reactions from the model. in fact. and treating P as always continuous can introduce significant errors. we obtain a 277-fold speed-up. Mastny et al. 2. 2006). In terms of simulation time. the ODE integration engine needs to be coupled with an event detection mechanism (Burden and Faires. For a stochastic model. then their dynamics will quickly reach an equilibrium. a reduced system is constructed. 2005). 2005. 2005) that requires to find the roots of non-linear equations to identify the firing time of stochastic events (Pahle. hence can be safely abstracted away. At the quantitative level. and the steady state of fast variables conditional on the slow ones being constant is computed. with rates that depend on the continuous variables. applying QSSA we obtain a nice closed form for the production rate of mRNA. feedback strength is increased. A different abstraction technique consists in reducing the number of model variables and reactions.. for low ˛ and high ˇ. allowing us to study the behaviour of the system for parameter combinations in which stochastic simulation is computationally unfeasible (e. If a non-stiff integrator is employed (such as methods belonging to the explicit Runge-Kutta family Burden and Faires. notice that the execution time of the hybrid model is essentially constant for all parameter sets. In Table 3 and Fig. This is because the stochastic part of the process in the hybrid system is time inhomogeneous.104 C. if a set of reactions acting on one (or more) molecular species is very fast. the hybrid approach is faster for high ˇ. Indeed. the overhead introduced by hybrid simulation3 can overcome the gain in computational efficiency. using the Quasi-Steady State Approximation (QSSA) (Rao and Arkin. Intrinsic fluctuations of the protein due to stochastic production and degradation are less relevant in this respect. In fact. one obtains a steady state distribution for fast variables (assuming ergodicity). Since analytical expressions are rarely obtained in this way. In this case. then no speed-up may be observed at all (data not shown) compared to exact stochastic simulation. These two thresholds have been heuristically selected to avoid unnecessary switching between discrete and stochastic models. 2009). 3 we report the most interesting combinations. Therefore. as the P number is reduced. 4 Conditional on XM and XP . Indeed.. e. P trajectories often approach zero. 2 We do not consider further static partitioning. averaging out the fast variables from the rate functions depending on them according to the previously computed steady state distribution.2 setting the continuous-to-discrete switching threshold to 10 (hence. or using the deterministic steady state distribution of fast variables. In Fig. such averaged rates can be approximated either by stochastic simulation (Mastny and S Liu. 2007). In practice. a stiff solver (Burden and Faires.. i. hence so is the translation frequency and P degradation reactions in the stochastic model. We expect this hybrid scheme to work well in most cases. for instance. with a possible loss of precision for high feedback repression (which reduces the overall number of P molecules). the one obtained from the ODE model. the remaining Markov chain has two states.g. In our model the binding and unbinding of the gene repressor turns out to be a natural candidate for QSSA due to the dynamics of the reaction. as it turned out to be not very accurate for high feedback values (data not shown). high ˇ) and peak-like (high ˛. We can see that the hybrid simulation works very well. The use of separate thresholds helps to avoid too many switches due to noise effects. 3 we compare the behaviour of the hybrid model with dynamic partitioning to the behaviour of the fully stochastic model. Cao et al. / Computational Biology and Chemistry 56 (2015) 98–108 as expected (Crudu et al. and essentially all behaviours of the fully stochastic model are qualitatively captured. Bodei et al. one for G = 1 and one for G = 0. Hoops et al. our hybrid abstraction will improve stochastic simulation only when the number of P production and degradation events dominates the simulation cost.

2009. In particular. the dynamics are preserved for low and medium ˛. fixed. The resulting model contains only four reactions and two species. for high ˇ and varying ˛ (with unbinding rate Fig. In this case. we obtain that the two models are very similar 5 If we apply the recipe of stochastic QSSA (Rao and Arkin. 3 reports the most interesting combinations). In this case. If the unbinding rate is reduced then the QSSA assumption on equilibrium of gene repression will cease to be valid. 2007). Note how the model with QSSA severely underestimates noise and exhibits more and much lower peaks w. If the former are much smaller that the latter. consider a variation of the model in which we apply QSSA to both G and P.. In this case. Therefore. In Fig. 2003.r.1. and two reactions. different values of ˛. we fixed the binding rate to bindP = 0. hence we preferred to stick to the simpler deterministic approximation. Practically. it is only one order of magnitude larger). the reader is referred to the web version of this article. Comparison of M trajectories for the explicit stochastic model and the stochastic model with QSSA both on gene repression and protein dynamics. the dynamics of gene repression reach equilibrium (in the stochastic sense) much sooner than M’s. Bodei et al. we have to check if the binding and unbinding rates are a few orders of magnitude larger than the remaining rates. with  = Psteady · XM . In particular. simulating the model we observed that the dynamics are qualitatively similar to those of the explicit model for all parameters and. 2007) consists in comparing the time scales of the reactions that are to be removed by QSSA with the time scales of the remaining reactions. in our model we may apply it to P itself. thus. for high ˇ and different values of ˛ (increasing from top to bottom). and this is indeed the case in our model. ∞ k=0 (1/(1 + ˛k)(1)/(k!)e− k .5 This version of the model is illustrated in Fig. This is confirmed in Fig. For instance. they are in good agreement also from a quantitative point of view. when ˇ is large. In this case. (For interpretation of the references to color in this figure legend. in most cases. However. From a quantitative point of view. the corresponding rates are faster.0166 and varied the unbinding rate accordingly to unbindP = bindP /˛. to get the correct QSSA mRNA production rate we need to compute the steady state distribution of the gene and the protein simultaneously.1). differently from Section 2. However.. with the gene variable G averaged out. 5. The application of QSSA is by no means restricted only to gene dynamics: it can be applied to any variables and sets of reactions that are sufficiently fast (with respect to the species they interact with). we show the results of stochastic simulation of this reduced model.e. 1(c). For example. assuming QSSA on binding and unbinding (also in this case Fig. A typical “rule of thumb” to apply QSSA (Crudu et al.) the model.t. then QSSA is usually safe. very low unbinding rate). if we compare the XM distributions for the full stochastic model and those for this minimal model. and variable unbinding rates (decreasing from left to right).. and then compute the marginal probability that Gsteady = 1. Mastny and S Liu. giving a rate of transcription of the form prodM /(1 + ˛ · prodP /degP · XM ). 2005). as in Section 2. 2010). / Computational Biology and Chemistry 56 (2015) 98–108 105 Fig. 3 and Table 3. Comparison of P trajectories for the explicit stochastic model (continuous blue line) and the QSSA stochastic model (dotted red line) for medium ˇ. the model has only one molecular species left. namely M. and the transcription rate. QSSA is predicted to be valid only for medium to low ˇ. This is not always the case in our system: for large ˇ. this expression coincides with the one that can be obtained from the ODE model. . 5.e. 1(b). 4. Mastny and S Liu. As we can see. becomes prodM /(1 + ˛ · XP ). where we compare the behaviour of the full and the QSSA models for different parameter combinations. we approximate the stochastic QSSA by computing reduced rates from the deterministic system. we are not aware of a closed form solution of the previous summation. we fixed ˇ as medium and varied only ˛ but. we obtain that this probabil- ity equals state probability  for G = 1 can be obtained by solving the balance equation ˛ · unbindP · XP ·  = unbindP (1 − ). as observed also in (Marquez Lago and Stelling. 2005. transcription and M degradation. where XP is replaced by its deterministic steady state value prodP /degP · XM . and simulation results are shown in Fig. Cao et al. from a qualitative point of view. as well. the full stochastic model for large ˛ (i. 4. and is schematically illustrated in Fig. it is known that the QSSA can introduce large errors in the stochastic dynamics (Cao et al.C. i. This can be explained by observing that gene repression influences directly only M and that the dynamics of binding/unbinding reactions are much faster than those of production/degradation of M and. the unbinding rate is comparable to P’s speed (more precisely.

4. or to the particular shape of the density functions. / Computational Biology and Chemistry 56 (2015) 98–108 for low and medium ˛ (Table S5). by treating them as continuous or removing them via QSSA. All the used abstractions are compared in terms of computational cost and of the potential of the underlying model to faithfully reproduce exact behaviours and on the deriving cost/benefit ratio. It is also used as a test for detecting a difference between locations (means or medians). In particular. a discrete treatment of the protein was shown to be necessary for very small numbers in order to avoid significant loss in accuracy. i. in order to carry out numerical analysis. Noise sources The hybrid and QSSA abstractions allowed us to selectively turn on and off the different (internal) noise sources of the system. Its two-sided version can be used to check for a significant difference between two populations. is a classical test for goodness of fit between two samples. collecting interesting . 3. 4. we computed the histogram distance (Cao and Petzold. Different dynamic behaviours: role of parameters A collection of distinct dynamic behaviours emerged. S3. and gene dynamics) and checking the validity of different plausible hypotheses. There could be several reasons for these differences. Also in this case the dynamics are essentially similar to those of the explicit model.1. 3 and Table 3. In this context.106 C. 2010). Subsequently. S4. Moreover. mainly for large values of ˛ and in relation to the hybrid approach. it is the intrinsic discreteness of the system (in terms of low copy numbers of some molecular species) that drives the noise dynamics. We finally investigated the use of QSSAs on gene dynamics and its interaction with hybrid abstraction. we analysed the model quantitatively. recovering known patterns in the accuracy of this approximation also for the hybrid case (Cao et al. 2005). However. Bodei et al. The histogram distance is used in Monte Carlo simulation to calculate the distance between histogram functions that approximate probability density functions of different group of samples. Discussion Aiming at defining a rigourous strategy for building correct hybrid models of biochemical systems. we set up a systematic investigation for evaluating the relative importance of different sources of noise (i. a model describing a negative feedback loop can be safely simplified abstracting away the non-key reactions. protein. It is interesting to notice that gene expression dynamics switching from continuous to oscillating behaviours are common in biological systems and can be used as mechanisms to trigger alternative responses (Walker et al. 2009). we studied the dynamics of a simple and paradigmatic gene regulatory network. However. S2. we were able to identify the reactions which are essential for producing the observed noise patterns. This analysis justified our assumption that feedback repression strength and relative degradation speed are good descriptors of the network behaviour. Finally. i. when the correlation between P and M is close to 1. where only the protein is treated continuously. covering (to the best of our knowledge) the full landscape of behaviours reported in the literature. In Fig. Different dynamic behaviours: emergence and relation to model abstraction Analysing different abstract models allowed us to better clarify which interactions in the model are crucial for distinct behavioural phenotypes to emerge. In this way.2. the different reactions. This can be explained following and combining the considerations we already made for the two abstractions separately.. the tests give different results for some combinations of parameters (with Kolmogorov–Smirnov and Mann–Whitney detecting more significant differences). For instance. 2004) (Tables S1.e. assuming that the two samples come from distributions with the same shape. Hence. low values of both ˛ and ˇ yield the “continuous phenotype”. The results of this analysis allowed us to identify which of the considered reactions can be specified through a deterministic formalisms. we identified two main parameters that govern the system’s behaviour. It is clear that the hybrid abstraction and the QSSA can be combined together (Crudu et al. while high values of ˛ and ˇ correspond to the “multi-modal phenotype”..e. We also identified biologically reasonable ranges for the values of these parameters. 1947) and a Kolmogorov-Smirnov test (Sheskin. mRNA. We showed that while the fully continuous approximation failed to capture relevant behaviours (as expected). 4. relating their validity to the possible values of parameters identified. some statistically significant differences are detected. These three tests can detect different aspects of differences between distributions. the reactions composing the model) in determining the overall behaviour.e. We stress that QSSA on P can be applied only for large values of ˇ. The use of different abstract models also allowed us to identify the key species and interactions that determine the overall behaviour of the system. S2. we show the results of hybrid simulation of the reduced model where QSSA is used to abstract binding and unbinding.4. Some statistical insights Looking at the histograms of the distribution of the amounts of P and M at time t = 10. S5 for the distributions of P) for all the models considered. 2006) and we performed both a Mann-Withney test (Mann and Whitney.3. correlating precise parameter combinations with a corresponding “behavioural phenotype”. we systematically analysed a fully detailed discrete stochastic model of a negative feedback loop. grounding on the fundamental information about the basic “building blocks” of the system (i. 4. Our experimentation showed that the key reactions are those involved in modifying the value of molecular species present in low numbers and that are slow (hence not amenable to QSSA). also comparing the viability of different abstractions. The (two-samples) Kolmogorov-Smirnov test. instead. In summary. while preserving the accuracy of the corresponding fully stochastic model in reproducing the described dynamics. Our methodology of screening and targeting key reactions can be generalised to more complex models. All these tests reported no significant difference between the empirical distributions for most parameter combinations. hence we present the results of all of them in the Supplementary Material. To analyse to what extent noise patterns are correctly reproduced. The Mann-Whitney test is a non-parametric test used to check for a significant statistical dominance between two samples. First.. S3). we considered several possible abstractions of the exact discrete stochastic model. We provided a quantitative analysis of the relative contribution of various parameters to the global dynamics.e. testing the null hypothesis that the two samples come from the same distribution (for the two-sided case). related to the number of samples considered. 000 s (see Figs. worked very well also for moderately low protein numbers. a hybrid approximation scheme. and we were able to link (relative) values of model parameters with observed behaviours. This is confirmed by statistically testing the difference between the empirical distributions of the abstract models and the empirical distribution of the stochastic model. we can observe that the distributions look visually quite similar.

A.. (Hybrid) automata and (stochastic) programs: the hybrid automata lattice of a stochastic program.. U.. N. 212 (1). but very difficult to realise without a systematic modus operandi. 1993. Garland Science. her current affiliation is AstraZeneca.. . Parameter estimation and model selection in computational biology. B.doi. B. G. Johnson. L. E.. 2007. Kummer. Luitel. J. Lewis.J. 221 (1). Gillespie.E.. G. 104 (9). 2007. 2340–2361. 6 (3). 1742–1750.. 014116. Accuracy limitations and the measurement of errors in the stochastic simulation of chemically reacting systems. Weiner.. 1249–1252. Probab...A. Theor.. Efficient exact simulation of chemical systems with many species and many channels. 18 (1). 1947. Enzyme induction as an all-or-none phenomenon..J. Proc. O. J. Numerical simulation for biochemical kinetics. 2010).F. A. Thomson Brooks/Cole. Haseltine. F. e1000696.. 38 (8). Hybrid semantics of stochastic programs with dynamic reconfiguration. Swain. 43 (7). Stability and Ergodicity of Piecewise Deterministic Markov Processes.A.B.. Gillespie.P05). Zawilski. following the analysis steps presented here. Furthermore. The chemical Langevin equation. 194107. Walter.. 2010. 79. Numerical Analysis. Paulsson.. Petzold. Xu. the key points for the regulation dynamics. 297–306. MIT Press. In this setting. Nucleic Acids Res. The slow-scale stochastic simulation algorithm. approximate stochastic and hybrid approaches. Rev. 92 (12). Ann. In: Ch. Chem. Cox. 1957.. Pedraza. On a more general level. • large genetic networks may be studied by exploiting an extensive analysis of the simpler modules composing them... P.. J. Counter-intuitive stochastic behavior of simple gene circuits with negative feedback. Balázsi. K. Radulescu. 2009. G.C. 1958. Biol. Cao. Hume. Golding. Khammash. Dufour. Whitney. Collins. Marquez Lago. followed by an extensive computational analysis.D. Popul... E. C. 413–478.. Phys. L. This approach should produce a classification of relevance/weight of each point in the overall evaluation of the chosen level of abstraction: a classification certainly useful from a biological point of view.... 2013.. Exact stochastic simulation of coupled chemical reactions. 2006. J. M. in the online version. Biophys. 127 (9). J. A. van Oudenaarden. 2002. J. 910–925. • a complete and detailed landscape of parameter values must be determined and tested for each proposed model. at the same time. 53–64. L. 553–566. J.S. 2712–2726.. SIAM J. D. 2011. Davis. 1025–1036. P.. Furthermore.. Policriti.. Biophys. J. Stochastic gene expression in a single cell. D. 2007. An interesting extension of this work would be the definition of a (semi)-automatic procedure which. we precisely analysed the impact of abstractions on the viability of model analysis. Costa.. J. Murphy. 81 (25). 123. 761–798. Crudu. L. M.D. 20 (3). Biol. W. Two classes of quasi-steady-state model reductions for stochastic kinetics... P.. Nonlinear Soft Matter Phys. S.V. Phys. Adams..R. 98 (9). S. T. D. 2008.. 2005. van Oudenaarden.T.L. 1977.H.. We believe that our methodology. Petzold. Phys. By systematically comparing the exact model with its abstractions.. J. A. R.. i. Natl. Phys.org/10. Sahle. Mendes.004 References Alberts.1016/j. Chabot. Positive feedback in eukaryotic gene networks: cell differentiation by graded binary response conversion. D. Logic Comput. Debussche. 2005. Bioinformatics 22 (24). • it is important to identify the “hot points” in control structures. S.. Cao.B.T. A.. 2000. Rawlings. R. Nested stochastic simulation algorithm for chemical kinetic systems with disparate rates. W.T. 2010. 48 (2).. 2009. F. 2010. Stochastic gene expression out-of-steady-state in the cyanobacterial circadian clock. 2001.. Siggia. J. Phys. S. N.L.. A. Lee. Bodei et al. I. J. the noise properties of simple feedback mechanisms of genetic networks and a precise characterisation of the validity of hybrid and QSSAs can be used to construct accurate abstract models in a modular way. A. Serrano. J. Di Patti. 2002. Math.. 158–180. Policriti. Chem.. / Computational Biology and Chemistry 56 (2015) 98–108 insights on its dynamics. Phys. Vanden-Eijnden. at http://dx. System Modelling in Cellular Biology. M.. 50–60.. J.E.. Pahle. Y.R. 20 (10). McQuarrie. 175–190. Cell. Nested stochastic simulation algorithms for chemical kinetic systems with multiple time scales. 222–234.F. Math.. 1183–1186. Fanelli.M.. 1–29. Gillespie. Bartholomay. the methodology proposed here for the study of feedback loops is based on the following assumptions and considerations: • fully discrete stochastic models are the most reliable and can be used as “benchmarks” for the evaluation of alternative models.R. 036112. Roberts. Y. R. Markovian modelling of gene-product synthesis.. 10 (1). 2528–2535. Real-time kinetics of gene activity in individual bacteria. Mann. J. Elowitz. E. Cambridge. Gibson. Collins. Bortolussi. Novic. we highlighted the key interactions that determine the different behavioural phenotypes exhibited by the model... Bull. Stochastic approach to chemical kinetics. 122 (1). Burden. D.J... B. 1053–1077. 2005. M. J. pp. Hybrid stochastic simplifications for multiscale gene networks. Nature 450 (7173). Phys. Mastny... H.compbiolchem. Bruck.. This would in principle allow to select for abstractions exhibiting a dynamics that closely relates to the expected dynamics of the real system under consideration.. Acknowledgements Davide Chiarugi acknowledges the support of the Flagship “InterOmics” project (PB. O. Brief Bioinform. 2006. Biol. our approach can be extended to a thorough analysis of more complex genetic networks. Probability in transcriptional regulation and its implications for leukocyte differentiation and inducible gene expression.. 3. Chapman & Hall. Mole. Tuning and controlling gene expression noise in synthetic gene networks. 2005. Lillacci. for each control point. Ycart. it is crucial to investigate to which extent the different approximations capture the emerging behaviours. i. pp. • the best trade-off between computational costs and biological faithfulness is to be found in hybrid models preserving such “hot points”. Pahle. Dauxois. J. Cell 144 (6). 47.. it reduces the overall computational effort compared to a more blind and extensive exploration of the parameter space. Such methodology should proceed with an attempt to (semi-)automate parameters and (especially) “hot points” search exploiting modularity and it would have as ultimate goal the identification of the “right” level of abstraction of a network. Liu. Biophys.. Raff.T. McKane. E. D. Sci. 63–76.A. 094106. J. Becksei. Hoops. Biol..... our results can be considered an extension of the ones presented in (Marquez Lago and Stelling. Markov Models and Optimization. J. Chem. 1876–1889. 4271–4288. L. 2323–2328. From single-cell genetic architecture to cell population dynamics: quantitatively decomposing the effects of different population heterogeneity sources for a genetic network with positive feedback architecture.. Chem. 1995.. J.e.. J. In: Proceedings of CompMod’09. Gillespie. with respect to the fully discrete stochastic model. Cell 123 (6). E: Stat. Phys. Chem. D. 1967. A.V. PLoS Comput. 2000. In this sense. it would be also interesting to investigate to what extent the proposed methodology can be extended to preserve specific classes of behaviours. Comput.. COPASI: a COmplex PAthway SImulator. in this respect.e. EMBO J.. Acad. Wang. Vanden-Eijnden. M. possibly using deductive arguments. D. 6–24. 3067–3074. J.. allows us to gain a deep understanding of the model behaviour and.. BMC Syst. J. M. D.04.. Appl... A. T. Faires. Gauges. Maria Luisa Guerriero was partially supported by Science Foundation Ireland research grant SFI 13/IF/B2792.. 107 Appendix A.. 2000.. E. A. Petzold. 23 (4).L. Blood 96 (7). 4 (3). Supplementary data Supplementary data associated with this article can be found. X.C.M. Stelling. 2009. 2015. 2009. On a test of whether one of two random variables is stochastically larger than the other. Peccoud. supported by the Italian MIUR and CNR organizations. Liu. will aid in building hybrid models of biological networks. P. Chem.J.. 2007. Cellular decision making and biological noise: from microbes to mammals. A. L.. Control Optim. Biochemical simulations: stochastic. M. a preliminary computational screening for identifying the key structural parameters and the regions in which different abstractions are applicable.. J. Enhanced stochastic oscillations in autocatalytic reactions. UK. Levine. Comput.. Balázsi. 113 (1). U. 2006. E. D.. J.. Stat. A. Phys.. Simus.M. L. Alessandro Romanel’s contribution was partially supported by the ABSTRACTCELL ANR-Chair of Excellence. A. Singhal. 89. Stochastic models for chemical reactions.B.. Bortolussi. Séraphin. Science 297 (5584). In particular.. Mantzaris. K.

. Theor. C. Raj. 383–387. Stochastic chemical kinetics and the quasi-steady state assumption: application to the Gillespie algorithm. Vargas. Salis. D. K. A.L. van Oudenaarden. Raj. Lightman.M. Resat. Kinetic modeling of biological systems. Kaznessis. Biol. J. Jenkins. 3rd ed. D. 2003. G.. J. A...J.. 2009. 913–918. D. Hume.. 1982. 1994.S. . J. Salis. I. Bacteriol.A. A partial-propensity variant of the compositionrejection stochastic simulation algorithm for chemical reaction networks. BMC Bioinform. 2010. Annu. Methods Mol. pp. Biochem.Y.. A. BMC Bioinform. Petzold. A. Tsaneva-Atanasova.. Browne. Trivedi.. Y. 41–61. Walker.. Rathinam. 216–226. 1994.. 2009. Neuroendocrinol. Tyagi. 2003.. K. 176 (10). Jensen. Ramaswamy. 266 (1). Petzold.. Sheskin.F.. Stamatakis.J. 044102. 132 (4). Stochastic mRNA synthesis in mammalian cells.. 38. 6. The RNA chain elongation rate in Escherichia coli depends on the growth rate... Transcription of individual genes in eukaryotic cells occurs randomly and infrequently. Zygourakis... D. Variability in gene expression underlies incomplete penetrance. e309. Nature 463 (7283).. Gillespie.. 72 (2). K. Accurate hybrid stochastic simulation of a system of coupled chemical or biochemical reactions. D. Zygourakis. R.L. Phys. Deterministic and stochastic population level simulations of an artificial lac operon genetic network. M. S. Phys. L. Immunol. A. Biol... V. Ross. 4 (10). 2010. 311–335.-Z. D. A mathematical and computational approach for integrating the major sources of cell population heterogeneity. Biol. Rifkin...V. Cao.. 541 (10). 2008.. Terry. 2010.G. Stamatakis. McArdle. van Oudenaarden. 2006.T. Vogel. 2011. 12... Biophys. 2807–2813. 118 (11).. 2008. Bodei et al. BMC Syst. M. H. L. Chapman & Hall. / Computational Biology and Chemistry 56 (2015) 98–108 Raj.F. Andersen. D. Stiffness in stochastic chemically reacting systems: the implicit tau-leaping method.. Eur.. Handbook of Parametric and Nonparametric Statistical Procedures. 4999–5010. 1993. 1226–1238. 122 (5). 255–270. Y. 93.. Single-molecule approaches to stochastic gene expression. Encoding and decoding mechanisms of pulsatile hormone secretion.A. H. Sotiropoulos.J.R. Peskin.J.P. 22 (12)..P. 128 (2-3).. J. I. 177–185. S. Raj.. or chance: stochastic gene expression and its consequences. K. 301. M. Phys. M.. Stekel.. PLoS Biol.. A. Kaznessis. Arkin. Journal of Chemical Physics 119 (24).S. V. Rev. 24–31. C. Kulkarni. J. Pettigrew.. Zhong.. 2004. U. S. 54103. Tranchina. J. Diffusion-controlled reactions of enzymes. Cell 135 (2). nurture. Strong negative self regulation of prokaryotic transcription factors increases the intrinsic noise of protein expression.R. Rao.. E. 12784–12794. H. A.. Chem. S. Chem. J... Y. Multiscale Hy3S: hybrid stochastic simulation for supercomputers. W. van Oudenaarden.-Q.. FSPNs: fluid stochastic Petri nets. Zhou... 2006. J. Cell Biol. C. A.. 2005.A. Chem. 7.N.108 C. J. C. Nature. Sbalzarini. Armstrong... In: Application and Theory of Petri Nets. 2010. 2.. K.