Evolving phenotypic networks in silico

AI-driven automated discovery tools reveal diverse behavioral competencies of biological networks

Many applications in biomedicine and synthetic bioengineering rely on understanding, mapping, predicting, and controlling the complex behavior of chemical and genetic networks. The emerging field of diverse intelligence investigates the problem-solving capacities of unconventional agents. However, few quantitative tools exist for exploring the competencies of non-conventional systems. Here, we view gene regulatory networks (GRNs) as agents navigating a problem space and develop automated tools to map the robust goal states GRNs can reach despite perturbations. Our contributions include: (1) Adapting curiosity-driven exploration algorithms from AI to discover the range of reachable goal states of GRNs, and (2) Proposing empirical tests inspired by behaviorist approaches to assess their navigation competencies. Our data shows that models inferred from biological data can reach a wide spectrum of steady states, exhibiting various competencies in physiological network dynamics without requiring structural changes in network properties or connectivity. We also explore the applicability of these 'behavioral catalogs' for comparing evolved competencies across biological networks, for designing drug interventions in biomedical contexts and synthetic gene networks for bioengineering. These tools and the emphasis on behavior-shaping open new paths for efficiently exploring the complex behavior of biological networks. For the interactive version of this paper, please visit https://developmentalsystems.org/curious-exploration-of-grn-competencies.

Evolution beats random chance: Performance-dependent network evolution for enhanced computational capacity

The quest to understand structure-function relationships in networks across scientific disciplines has intensified. However, the optimal network architecture remains elusive, particularly for complex information processing. Therefore, we investigate how optimal and specific network structures form to efficiently solve distinct tasks using a framework of performance-dependent network evolution, leveraging reservoir computing principles. Our study demonstrates that task-specific minimal network structures obtained through this framework consistently outperform networks generated by alternative growth strategies and Erdős-Rényi random networks. Evolved networks exhibit unexpected sparsity and adhere to scaling laws in node-density space while showcasing a distinctive asymmetry in input and information readout node distribution. Consequently, we propose a heuristic for quantifying task complexity from performance-dependently evolved networks, offering valuable insights into the evolutionary dynamics of the network structure-function relationship. Our findings advance the fundamental understanding of process-specific network evolution and shed light on the design and optimization of complex information processing mechanisms, notably in machine learning.

Morphogens in the evolution of size, shape and patterning

Much of the striking diversity of life on Earth has arisen from variations in the way that the same molecules and networks operate during development to shape and pattern tissues and organs into different morphologies. However, we still understand very little about the potential for diversification exhibited by different, highly conserved mechanisms during evolution, or, conversely, the constraints that they place on evolution. With the aim of steering the field in new directions, we focus on morphogen-mediated patterning and growth as a case study to demonstrate how conserved developmental mechanisms can adapt during evolution to drive morphological diversification and optimise functionality, and to illustrate how evolution algorithms and computational tools can be used alongside experiments to provide insights into how these conserved mechanisms can evolve. We first introduce key conserved properties of morphogen-driven patterning mechanisms, before summarising comparative studies that exemplify how changes in the spatiotemporal expression and signalling levels of morphogens impact the diversification of organ size, shape and patterning in nature. Finally, we detail how theoretical frameworks can be used in conjunction with experiments to probe the role of morphogen-driven patterning mechanisms in evolution. We conclude that morphogen-mediated patterning is an excellent model system and offers a generally applicable framework to investigate the evolution of developmental mechanisms.

The distribution of fitness effects during adaptive walks using a simple genetic network

The tempo and mode of adaptation depends on the availability of beneficial alleles. Genetic interactions arising from gene networks can restrict this availability. However, the extent to which networks affect adaptation remains largely unknown. Current models of evolution consider additive genotype-phenotype relationships while often ignoring the contribution of gene interactions to phenotypic variance. In this study, we model a quantitative trait as the product of a simple gene regulatory network, the negative autoregulation motif. Using forward-time genetic simulations, we measure adaptive walks towards a phenotypic optimum in both additive and network models. A key expectation from adaptive walk theory is that the distribution of fitness effects of new beneficial mutations is exponential. We found that both models instead harbored distributions with fewer large-effect beneficial alleles than expected. The network model also had a complex and bimodal distribution of fitness effects among all mutations, with a considerable density at deleterious selection coefficients. This behavior is reminiscent of the cost of complexity, where correlations among traits constrain adaptation. Our results suggest that the interactions emerging from genetic networks can generate complex and multimodal distributions of fitness effects.

Recent advances in machine learning applications in metabolic engineering

Metabolic engineering encompasses several widely-used strategies, which currently hold a high seat in the field of biotechnology when its potential is manifesting through a plethora of research and commercial products with a strong societal impact. The genomic revolution that occurred almost three decades ago has initiated the generation of large omics-datasets which has helped in gaining a better understanding of cellular behavior. The itinerary of metabolic engineering that has occurred based on these large datasets has allowed researchers to gain detailed insights and a reasonable understanding of the intricacies of biosystems. However, the existing trail-and-error approaches for metabolic engineering are laborious and time-intensive when it comes to the production of target compounds with high yields through genetic manipulations in host organisms. Machine learning (ML) coupled with the available metabolic engineering test instances and omics data brings a comprehensive and multidisciplinary approach that enables scientists to evaluate various parameters for effective strain design. This vast amount of biological data should be standardized through knowledge engineering to train different ML models for providing accurate predictions in gene circuits designing, modification of proteins, optimization of bioprocess parameters for scaling up, and screening of hyper-producing robust cell factories. This review briefs on the premise of ML, followed by mentioning various ML methods and algorithms alongside the numerous omics datasets available to train ML models for predicting metabolic outcomes with high-accuracy. The combinative interplay between the ML algorithms and biological datasets through knowledge engineering have guided the recent advancements in applications such as CRISPR/Cas systems, gene circuits, protein engineering, metabolic pathway reconstruction, and bioprocess engineering. Finally, this review addresses the probable challenges of applying ML in metabolic engineering which will guide the researchers toward novel techniques to overcome the limitations.

Deduction of signaling mechanisms from cellular responses to multiple cues

Cell signaling networks are complex and often incompletely characterized, making it difficult to obtain a comprehensive picture of the mechanisms they encode. Mathematical modeling of these networks provides important clues, but the models themselves are often complex, and it is not always clear how to extract falsifiable predictions. Here we take an inverse approach, using experimental data at the cell level to deduce the minimal signaling network. We focus on cells' response to multiple cues, specifically on the surprising case in which the response is antagonistic: the response to multiple cues is weaker than the response to the individual cues. We systematically build candidate signaling networks one node at a time, using the ubiquitous ingredients of (i) up- or down-regulation, (ii) molecular conversion, or (iii) reversible binding. In each case, our method reveals a minimal, interpretable signaling mechanism that explains the antagonistic response. Our work provides a systematic way to deduce molecular mechanisms from cell-level data.

Geometry of gene regulatory dynamics

Embryonic development leads to the reproducible and ordered appearance of complexity from egg to adult. The successive differentiation of different cell types that elaborate this complexity results from the activity of gene networks and was likened by Waddington to a flow through a landscape in which valleys represent alternative fates. Geometric methods allow the formal representation of such landscapes and codify the types of behaviors that result from systems of differential equations. Results from Smale and coworkers imply that systems encompassing gene network models can be represented as potential gradients with a Riemann metric, justifying the Waddington metaphor. Here, we extend this representation to include parameter dependence and enumerate all three-way cellular decisions realizable by tuning at most two parameters, which can be generalized to include spatial coordinates in a tissue. All diagrams of cell states vs. model parameters are thereby enumerated. We unify a number of standard models for spatial pattern formation by expressing them in potential form (i.e., as topographic elevation). Turing systems appear nonpotential, yet in suitable variables the dynamics are low dimensional and potential. A time-independent embedding recovers the original variables. Lateral inhibition is described by a saddle point with many unstable directions. A model for the patterning of the Drosophila eye appears as relaxation in a bistable potential. Geometric reasoning provides intuitive dynamic models for development that are well adapted to fit time-lapse data.

Why call it developmental bias when it is just development?

The concept of developmental constraints has been central to understand the role of development in morphological evolution. Developmental constraints are classically defined as biases imposed by development on the distribution of morphological variation.

This opinion article argues that the concepts of developmental constraints and developmental biases do not accurately represent the role of development in evolution. The concept of developmental constraints was coined to oppose the view that natural selection is all-capable and to highlight the importance of development for understanding evolution. In the modern synthesis, natural selection was seen as the main factor determining the direction of morphological evolution. For that to be the case, morphological variation needs to be isotropic (i.e. equally possible in all directions). The proponents of the developmental constraint concept argued that development makes that some morphological variation is more likely than other (i.e. variation is not isotropic), and that, thus, development constraints evolution by precluding natural selection from being all-capable.

This article adds to the idea that development is not compatible with the isotropic expectation by arguing that, in fact, it could not be otherwise: there is no actual reason to expect that development could lead to isotropic morphological variation. It is then argued that, since the isotropic expectation is untenable, the role of development in evolution should not be understood as a departure from such an expectation. The role of development in evolution should be described in an exclusively positive way, as the process determining which directions of morphological variation are possible, instead of negatively, as a process precluding the existence of morphological variation we have no actual reason to expect.

This article discusses that this change of perspective is not a mere question of semantics: it leads to a different interpretation of the studies on developmental constraints and to a different research program in evolution and development. This program does not ask whether development constrains evolution. Instead it asks questions such as, for example, how different types of development lead to different types of morphological variation and, together with natural selection, determine the directions in which different lineages evolve.

Biosystems Design by Machine Learning

Biosystems such as enzymes, pathways, and whole cells have been increasingly explored for biotechnological applications. However, the intricate connectivity and resulting complexity of biosystems poses a major hurdle in designing biosystems with desirable features. As -omics and other high throughput technologies have been rapidly developed, the promise of applying machine learning (ML) techniques in biosystems design has started to become a reality. ML models enable the identification of patterns within complicated biological data across multiple scales of analysis and can augment biosystems design applications by predicting new candidates for optimized performance. ML is being used at every stage of biosystems design to help find nonobvious engineering solutions with fewer design iterations. In this review, we first describe commonly used models and modeling paradigms within ML. We then discuss some applications of these models that have already shown success in biotechnological applications. Moreover, we discuss successful applications at all scales of biosystems design, including nucleic acids, genetic circuits, proteins, pathways, genomes, and bioprocesses. Finally, we discuss some limitations of these methods and potential solutions as well as prospects of the combination of ML and biosystems design.

Speeding up anterior-posterior patterning of insects by differential initialization of the gap gene cascade

Recently, it was shown that anterior-posterior patterning genes in the red flour beetle Tribolium castaneum are expressed sequentially in waves. However, in the fruit fly Drosophila melanogaster, an insect with a derived mode of embryogenesis compared to Tribolium, anterior-posterior patterning genes quickly and simultaneously arise as mature gene expression domains that, afterwards, undergo slight posterior-to-anterior shifts. This raises the question of how a fast and simultaneous mode of patterning, like that of Drosophila, could have evolved from a rather slow sequential mode of patterning, like that of Tribolium. In this paper, we propose a mechanism for this evolutionary transition based on a switch from a uniform to a gradient-mediated initialization of the gap gene cascade by maternal Hb. The model is supported by computational analyses and experiments.

Adapting machine-learning algorithms to design gene circuits

BACKGROUND

Gene circuits are important in many aspects of biology, and perform a wide variety of different functions. For example, some circuits oscillate (e.g. the cell cycle), some are bistable (e.g. as cells differentiate), some respond sharply to environmental signals (e.g. ultrasensitivity), and some pattern multicellular tissues (e.g. Turing's model). Often, one starts from a given circuit, and using simulations, asks what functions it can perform. Here we want to do the opposite: starting from a prescribed function, can we find a circuit that executes this function? Whilst simple in principle, this task is challenging from a computational perspective, since gene circuit models are complex systems with many parameters. In this work, we adapted machine-learning algorithms to significantly accelerate gene circuit discovery.

RESULTS

We use gradient-descent optimization algorithms from machine learning to rapidly screen and design gene circuits. With this approach, we found that we could rapidly design circuits capable of executing a range of different functions, including those that: (1) recapitulate important in vivo phenomena, such as oscillators, and (2) perform complex tasks for synthetic biology, such as counting noisy biological events.

CONCLUSIONS

Our computational pipeline will facilitate the systematic study of natural circuits in a range of contexts, and allow the automatic design of circuits for synthetic biology. Our method can be readily applied to biological networks of any type and size, and is provided as an open-source and easy-to-use python module, GeneNet.

In Silico Evolution of Biochemical Log-Response

Numerous biological systems are known to harbor a form of logarithmic behavior, from Weber's law to bacterial chemotaxis. Such a log-response allows for sensitivity to small relative variations of biochemical inputs over a large range of concentration values. Here we use a genetic algorithm to evolve biochemical networks displaying a logarithmic response. A quasi-perfect log-response implemented by the same core network evolves in a convergent way across our different in silico replications. The best network is able to fit a logarithm over 4 orders of magnitude with an accuracy of the order of 1%. At the heart of this network, we show that a logarithmic approximation may be implemented with one single nonlinear interaction, that can be interpreted either as multisite phosphorylations or as a ligand induced multimerization. We provide an analytical explanation for the effect and exhibit constraints on parameters. Biological log-response might thus be easier to implement than usually assumed.

Perspectives and applications of machine learning for evolutionary developmental biology

Evolutionary Developmental Biology (Evo-Devo) is an ever-expanding field that aims to understand how development was modulated by the evolutionary process. In this sense, "omic" studies emerged as a powerful ally to unravel the molecular mechanisms underlying development. In this scenario, bioinformatics tools become necessary to analyze the growing amount of information. Among computational approaches, machine learning stands out as a promising field to generate knowledge and trace new research perspectives for bioinformatics. In this review, we aim to expose the current advances of machine learning applied to evolution and development. We draw clear perspectives and argue how evolution impacted machine learning techniques.

Around the clock: gradient shape and noise impact the evolution of oscillatory segmentation dynamics

BACKGROUND

Segmentation, the subdivision of the major body axis into repeated elements, is considered one of the major evolutionary innovations in bilaterian animals. In all three segmented animal clades, the predominant segmentation mechanism is sequential segmentation, where segments are generated one by one in anterior-posterior order from a posterior undifferentiated zone. In vertebrates and arthropods, sequential segmentation is thought to arise from a clock-and-wavefront-type mechanism, where oscillations in the posterior growth zone are transformed into a segmental prepattern in the anterior by a receding wavefront. Previous evo-devo simulation studies have demonstrated that this segmentation type repeatedly arises, supporting the idea of parallel evolutionary origins in these animal clades. Sequential segmentation has been studied most extensively in vertebrates, where travelling waves have been observed that reflect the slowing down of oscillations prior to their cessation and where these oscillations involve a highly complex regulatory network. It is currently unclear under which conditions this oscillator complexity and slowing should be expected to evolve, how they are related and to what extent similar properties should be expected for sequential segmentation in other animal species.

RESULTS

To investigate these questions, we extend a previously developed computational model for the evolution of segmentation. We vary the slope of the posterior morphogen gradient and the strength of gene expression noise. We find that compared to a shallow gradient, a steep morphogen gradient allows for faster evolution and evolved oscillator networks are simpler. Furthermore, under steep gradients, damped oscillators often evolve, whereas shallow gradients appear to require persistent oscillators which are regularly accompanied by travelling waves, indicative of a frequency gradient. We show that gene expression noise increases the likelihood of evolving persistent oscillators under steep gradients and of evolving frequency gradients under shallow gradients. Surprisingly, we find that the evolutions of oscillator complexity and travelling waves are not correlated, suggesting that these properties may have evolved separately.

CONCLUSIONS

Based on our findings, we suggest that travelling waves may have evolved in response to shallow morphogen gradients and gene expression noise. These two factors may thus also be responsible for the observed differences between different species within both the arthropod and chordate phyla.

Synthetic circuits reveal how mechanisms of gene regulatory networks constrain evolution

Phenotypic variation is the raw material of adaptive Darwinian evolution. The phenotypic variation found in organismal development is biased towards certain phenotypes, but the molecular mechanisms behind such biases are still poorly understood. Gene regulatory networks have been proposed as one cause of constrained phenotypic variation. However, most pertinent evidence is theoretical rather than experimental. Here, we study evolutionary biases in two synthetic gene regulatory circuits expressed in Escherichia coli that produce a gene expression stripe-a pivotal pattern in embryonic development. The two parental circuits produce the same phenotype, but create it through different regulatory mechanisms. We show that mutations cause distinct novel phenotypes in the two networks and use a combination of experimental measurements, mathematical modelling and DNA sequencing to understand why mutations bring forth only some but not other novel gene expression phenotypes. Our results reveal that the regulatory mechanisms of networks restrict the possible phenotypic variation upon mutation. Consequently, seemingly equivalent networks can indeed be distinct in how they constrain the outcome of further evolution.

φ-evo: A program to evolve phenotypic models of biological networks

Molecular networks are at the core of most cellular decisions, but are often difficult to comprehend. Reverse engineering of network architecture from their functions has proved fruitful to classify and predict the structure and function of molecular networks, suggesting new experimental tests and biological predictions. We present φ-evo, an open-source program to evolve in silico phenotypic networks performing a given biological function. We include implementations for evolution of biochemical adaptation, adaptive sorting for immune recognition, metazoan development (somitogenesis, hox patterning), as well as Pareto evolution. We detail the program architecture based on C, Python 3, and a Jupyter interface for project configuration and network analysis. We illustrate the predictive power of φ-evo by first recovering the asymmetrical structure of the lac operon regulation from an objective function with symmetrical constraints. Second, we use the problem of hox-like embryonic patterning to show how a single effective fitness can emerge from multi-objective (Pareto) evolution. φ-evo provides an efficient approach and user-friendly interface for the phenotypic prediction of networks and the numerical study of evolution itself.

Designing synthetic networks in silico: a generalised evolutionary algorithm approach

Background

Evolution has led to the development of biological networks that are shaped by environmental signals. Elucidating, understanding and then reconstructing important network motifs is one of the principal aims of Systems & Synthetic Biology. Consequently, previous research has focused on finding optimal network structures and reaction rates that respond to pulses or produce stable oscillations. In this work we present a generalised in silico evolutionary algorithm that simultaneously finds network structures and reaction rates (genotypes) that can satisfy multiple defined objectives (phenotypes).

Results

The key step to our approach is to translate a schema/binary-based description of biological networks into systems of ordinary differential equations (ODEs). The ODEs can then be solved numerically to provide dynamic information about an evolved networks functionality. Initially we benchmark algorithm performance by finding optimal networks that can recapitulate concentration time-series data and perform parameter optimisation on oscillatory dynamics of the Repressilator. We go on to show the utility of our algorithm by finding new designs for robust synthetic oscillators, and by performing multi-objective optimisation to find a set of oscillators and feed-forward loops that are optimal at balancing different system properties. In sum, our results not only confirm and build on previous observations but we also provide new designs of synthetic oscillators for experimental construction.

Conclusions

In this work we have presented and tested an evolutionary algorithm that can design a biological network to produce desired output. Given that previous designs of synthetic networks have been limited to subregions of network- and parameter-space, the use of our evolutionary optimisation algorithm will enable Synthetic Biologists to construct new systems with the potential to display a wider range of complex responses.

Electronic supplementary material

The online version of this article (doi:10.1186/s12918-017-0499-9) contains supplementary material, which is available to authorized users.

Untangling the Hairball: Fitness-Based Asymptotic Reduction of Biological Networks

Complex mathematical models of interaction networks are routinely used for prediction in systems biology. However, it is difficult to reconcile network complexities with a formal understanding of their behavior. Here, we propose a simple procedure (called ϕ¯) to reduce biological models to functional submodules, using statistical mechanics of complex systems combined with a fitness-based approach inspired by in silico evolution. The ϕ¯ algorithm works by putting parameters or combination of parameters to some asymptotic limit, while keeping (or slightly improving) the model performance, and requires parameter symmetry breaking for more complex models. We illustrate ϕ¯ on biochemical adaptation and on different models of immune recognition by T cells. An intractable model of immune recognition with close to a hundred individual transition rates is reduced to a simple two-parameter model. The ϕ¯ algorithm extracts three different mechanisms for early immune recognition, and automatically discovers similar functional modules in different models of the same process, allowing for model classification and comparison. Our procedure can be applied to biological networks based on rate equations using a fitness function that quantifies phenotypic performance.

Phenotypic spandrel: absolute discrimination and ligand antagonism

We consider the general problem of sensitive and specific discrimination between biochemical species. An important instance is immune discrimination between self and not-self, where it is also observed experimentally that ligands just below the discrimination threshold negatively impact response, a phenomenon called antagonism. We characterize mathematically the generic properties of such discrimination, first relating it to biochemical adaptation. Then, based on basic biochemical rules, we establish that, surprisingly, antagonism is a generic consequence of any strictly specific discrimination made independently from ligand concentration. Thus antagonism constitutes a 'phenotypic spandrel': a phenotype existing as a necessary by-product of another phenotype. We exhibit a simple analytic model of discrimination displaying antagonism, where antagonism strength is linear in distance from the detection threshold. This contrasts with traditional proofreading based models where antagonism vanishes far from threshold and thus displays an inverted hierarchy of antagonism compared to simpler models. The phenotypic spandrel studied here is expected to structure many decision pathways such as immune detection mediated by TCRs and FCϵRIs, as well as endocrine signalling/disruption.

Architecture of a minimal signaling pathway explains the T-cell response to a 1 million-fold variation in antigen affinity and dose

T cells must respond differently to antigens of varying affinity presented at different doses. Previous attempts to map peptide MHC (pMHC) affinity onto T-cell responses have produced inconsistent patterns of responses, preventing formulations of canonical models of T-cell signaling. Here, a systematic analysis of T-cell responses to 1 million-fold variations in both pMHC affinity and dose produced bell-shaped dose-response curves and different optimal pMHC affinities at different pMHC doses. Using sequential model rejection/identification algorithms, we identified a unique, minimal model of cellular signaling incorporating kinetic proofreading with limited signaling coupled to an incoherent feed-forward loop (KPL-IFF) that reproduces these observations. We show that the KPL-IFF model correctly predicts the T-cell response to antigen copresentation. Our work offers a general approach for studying cellular signaling that does not require full details of biochemical pathways.

Positional information from oscillatory phase shifts : insights from in silico evolution

Complex cellular decisions are based on temporal dynamics of pathways, including genetic oscillators. In development, recent works on vertebrae formation have suggested that relative phase of genetic oscillators encode positional information, including differentiation front defining vertebrae positions. Precise mechanisms for this are still unknown. Here, we use computational evolution to find gene network topologies that can compute the phase difference between oscillators and convert it into a decoder morphogen concentration. Two types of networks are discovered, based on symmetry properties of the decoder gene. So called asymmetric networks are studied, and two submodules are identified converting phase information into an amplitude variable. Those networks naturally display a 'shock' for a well defined phase difference, that can be used to define a wavefront of differentiation. We show how implementation of these ideas reproduce experimental features of vertebrate segmentation.

Predicting Ancestral Segmentation Phenotypes from Drosophila to Anopheles Using In Silico Evolution

Molecular evolution is an established technique for inferring gene homology but regulatory DNA turns over so rapidly that inference of ancestral networks is often impossible. In silico evolution is used to compute the most parsimonious path in regulatory space for anterior-posterior patterning linking two Dipterian species. The expression pattern of gap genes has evolved between Drosophila (fly) and Anopheles (mosquito), yet one of their targets, eve, has remained invariant. Our model predicts that stripe 5 in fly disappears and a new posterior stripe is created in mosquito, thus eve stripe modules 3+7 and 4+6 in fly are homologous to 3+6 and 4+5 in mosquito. We can place Clogmia on this evolutionary pathway and it shares the mosquito homologies. To account for the evolution of the other pair-rule genes in the posterior we have to assume that the ancestral Dipterian utilized a dynamic method to phase those genes in relation to eve.

The last common ancestor of the fruit fly (Drosophila) and mosquito (Anopheles) lived more than 200 Million years ago. Can we use available data on insects alive today to infer what their ancestor looked like? In this manuscript, we focus on early embryonic development, when stripes of genetic expression appear and define the location of insect segments (“segmentation”). We use an evolutionary algorithm to reconstruct and predict dynamics of genes controlling stripes in the last common ancestor of fly and mosquito. We predict a new and different combinatorial logic of stripe formation in mosquito compared to fly, which is fully consistent with development of intermediate species such as moth-fly (Clogmia). Our simulations further suggest that the dynamics of gene expression in this last common ancestor were similar to other insects, such as wasps (Nasonia). Our method illustrates how computational methods inspired by machine learning and non-linear physics can be used to infer gene dynamics in species that disappeared millions of years ago.

BioJazz: in silico evolution of cellular networks with unbounded complexity using rule-based modeling

Systems biologists aim to decipher the structure and dynamics of signaling and regulatory networks underpinning cellular responses; synthetic biologists can use this insight to alter existing networks or engineer de novo ones. Both tasks will benefit from an understanding of which structural and dynamic features of networks can emerge from evolutionary processes, through which intermediary steps these arise, and whether they embody general design principles. As natural evolution at the level of network dynamics is difficult to study, in silico evolution of network models can provide important insights. However, current tools used for in silico evolution of network dynamics are limited to ad hoc computer simulations and models. Here we introduce BioJazz, an extendable, user-friendly tool for simulating the evolution of dynamic biochemical networks. Unlike previous tools for in silico evolution, BioJazz allows for the evolution of cellular networks with unbounded complexity by combining rule-based modeling with an encoding of networks that is akin to a genome. We show that BioJazz can be used to implement biologically realistic selective pressures and allows exploration of the space of network architectures and dynamics that implement prescribed physiological functions. BioJazz is provided as an open-source tool to facilitate its further development and use. Source code and user manuals are available at: http://oss-lab.github.io/biojazz and http://osslab.lifesci.warwick.ac.uk/BioJazz.aspx.