Monte-Carlo Tree Search (MCTS) is largely responsible for the improvement not only of many computer games, including Go and General Game Playing (GPP), but also of real-world continuous Markov decision process problems. MCTS initially uses the Upper Confidence bounds applied to Trees (UCT), but the Rapid Action Value Estimation (RAVE) heuristic has rapidly taken over in the discrete and continuous domains. Recently, generalized RAVE (GRAVE) outperformed such heuristics in the discrete domain. This paper is concerned with extending the GRAVE heuristic to continuous action and state spaces (cGRAVE). To enhance its performance, we suggest an action decomposition strategy to break down multidimensional actions into multiple unidimensional actions, and we propose a selective policy based on constraints that bias the playouts and select promising actions in the search tree. The approach is experimentally validated on a real-world biological problem: the goal is to identify the continuous parameters of gene regulatory networks (GRNs).

When modelling a complex biological system, the bottleneck of the process is the determination of parameter values that lead to model dynamics that are compatible with observations. Even for discrete modelling frameworks, this step can be limiting. Here we introduce a representation of the whole family of discrete models that can be associated to a biological system, where states are shared by all models and transitions are labelled by constraints on dynamical parameters. A model checking procedure is defined to handle this new representation. This procedure extracts the conditions on parameter settings that are compatible with a given dynamical property expressed in a temporal logic. We prove the correctness of our model checking procedure and illustrate the advantage of such an approach on some different systems of biological interest.

The hybrid modelling framework of gene regulatory networks (hGRNs) is a functional framework for studying biological systems, taking into account both the structural relationship between genes and the continuous time evolution of gene concentrations. The goal is to identify the variables of such a model, controlling the aggregated experimental observations. A recent study considered this task as a free optimisation problem and concluded that metaheuristics are well suited. The main drawback of this previous approach is that panmictic heuristics converge towards one basin of attraction in the search space, while biologists are interested in finding multiple satisfactory solutions. This paper investigates the problem of multimodality and assesses the effectiveness of cellular genetic algorithms (cGAs) in dealing with the increasing dimensionality and complexity of hGRN models. A comparison with the second variant of covariance matrix self-adaptation strategy with repelling subpopulations (RS-CMSA-ESII), the winner of the CEC'2020 competition for multimodal optimisation (MMO), is made. Results show evidence that cGAs better maintain a diverse set of solutions while giving better quality solutions, making them better suited for this MMO task.

Real-world applications of artificial intelligence often require the decision-maker to choose between multiple optimal solutions at hand before making a final decision. Since there is no guarantee that an increase in the budget or several independent executions will yield different solutions, classic mechanisms are not suitable for identifying multiple solutions. In the context of sequential decision-making problems, Monte Carlo Tree Search (MCTS) is a state-of-the-art online planning algorithm. It is responsible for the improvement of many computer games but also for real-world problems involving continuous action spaces. MCTS has recently been successfully applied to the diverse planning problem in the discrete setting. In this work, we propose different diverse planners based on MCTS (DP-MCTS) to be relevant in the continuous setting. The solution to a diverse planning problem is a Pareto set between diversity and quality of plans. Therefore, we suggest considering a multi-objective setting in which the vectorial reward integrates the diversity measure as an additional objective. In addition, we propose two types of inhibition strategies disregarding the optimal plans to enforce the exploration of the search space during the tree construction. The three different contributions are assessed independently against a diverse multi-armed bandit policy, and the comparison is held on a real-world biological problem involving continuous action and state spaces.

Early derisking decisions in the development of new chemical compounds enable the identification of novel chemical candidates with improved safety profiles. In vivo studies are traditionally conducted in the early assessment of acute oral toxicity of crop protection products to avoid compounds, which are considered "very acutely toxic", with an in vivo lethal dose of 50% (LD50) ≤ 60 mg/kg body weight. Those studies are lengthy and costly and raise ethical concerns, catalyzing the use of nonanimal alternatives. The objective of our analysis was to assess the predictive efficacy of read-across approaches for acute oral toxicity in rats, comparing the use of chemical structure information, in vitro biological data derived from the Cell Painting profiling assay on U2OS cells, or the combination of both. Our findings indicate that the classification of compounds as very acute oral toxic (LD50 ≤ 60 mg/kg) or not is possible using a read-across approach, with chemical structure information, morphological profiles, or a combination of both. When classifying compounds structurally similar to those in the training set, the chemical structure was more predictive (balanced accuracy of 0.82). Conversely, when the compounds to be classified were structurally different from those in the training set, the morphological profiles were more predictive (balanced accuracy of 0.72). Combining the two models allowed for the classification of compounds structurally similar to those in the training set to slightly improve the predictions (balanced accuracy of 0.85).

: In this work, we study the existence of sustained oscillations in the “canonical” repressilator, a basic synthetic circuit of 3 genes leading to sustained oscillations. Previous works mostly used differential equations to study the repressilator. In our work, a pre-existing hybrid modeling framework of gene regulatory networks called HGRN is used to model this system. Compared to differential equations, dynamical properties of HGRNs are easier to prove theoretically due to its lower dynamical complexity. The objective of this work is to find conditions for the existence of sustained oscillations described by separable constraints on parameters. With such separable constraints, each parameter is constrained individually by an interval, which can provide useful information for the design of synthetic circuits. Our two major contributions are the following: firstly, we develop, by using the Poincar´e map, a sufficient and necessary condition for the existence of sustained oscillations; then, based on this condition, we give a method using the range enclosure property of Bernstein coefficients to compute compatible separable constraints. By applying this method, we successfully obtain sets of conditions for the existence of sustained oscillations described as separable constraints.

The biological regulation networks determine the behavior of biological systems and mathematical modeling to a large extent, and is essential to understand how they work, clarify the causal links in order to modify biological functions, design pharmaceutical treatments, etc., for example. This chapter illustrates the major contribution of formal methods for modeling complex biological systems in choosing to focus on regulatory networks. It introduces the first way to select the desired configuration based completely on the logic of Hoare: from a trace observed experimentally and transcribed into a formal language, it is possible to construct the constraints that make this trace possible in the formal model. The chapter presents the TotemBioNet platform dedicated to identifying parameters for the discrete formalization of René Thomas by using the complementarity of the two approaches based on Hoare's logic and temporal logic.

The study of gene regulatory networks (GRNs) allows us to better understand biological systems such as the adaptation of the organism to a disturbance in the environment. Hybrid GRNs (hGRNs) are of interest because they integrate the continuous time evolution in GRN modeling which is convenient in biology. This study focuses on the problem of identifying the variables of hGRN models. In a large-scale case, previous work using constraint-based programming has failed to solve the minimal constraints on such variables which reflect the biological knowledge on the system behavior. In this work, we propose to transform a Constraint Satisfaction Problem (CSP) into a Free Optimization Problem (FOP) by formulating an adequate fitness function and validate the approach on an abstract model of the circadian cycle. We compare several continuous optimization algorithms and show that these first experimental results are in agreement with the specifications coming from biological expertise: evolutionary algorithms are able to identify a solution equivalent to the ones found by continuous constraint solvers.

Ce chapitre démontre que la modélisation des réseaux de régulation biologiques est grandement facilitée par des méthodes symboliques de vérification formelle comme le model-checking et les preuves en logique de Hoare. Il présente une méthodologie complète de modélisation formelle où les approches logiques classiques du génie logiciel s'adaptent remarquablement efficacement au problème majeur de la modélisation des systèmes complexes: l'identification des paramètres.

When designing a biological regulatory network, new information or wet experiments can require adding variables or interactions, inside a previously validated model. They can result in complete reconsiderations of established behaviours. Fortunately, formal methods allow for fully automated verification of properties, and TotemBioNet is an efficient software integrating a collection of formal approaches for regulatory networks. It allowed us to develop a multidisciplinary methodology for designing large dynamical models in an incremental way, including non regression proofs (preservation of important biological properties).

Adaptation of metabolism is a response of many eukaryotic cells to nutrient heterogeneity in the cell microenvironment. One of these adaptations is the shift from respiratory to fermentative metabolism, also called the Warburg/Crabtree effect. It is a response to a very high nutrient increase in the cell microenvironment, even in the presence of oxygen. Understanding whether this metabolic transition can result from basic regulation signals between components of the central carbon metabolism are the the core question of this work. We use an extension of the René Thomas modeling framework for representing the regulations between the main catabolic and anabolic pathways of eukaryotic cells, and formal methods for confronting models with known biological properties in different microenvironments. The formal model of the regulation of eukaryote metabolism defined and validated here reveals the conditions under which this metabolic phenotype switch occurs. It clearly proves that currently known regulating signals within the main components of central carbon metabolism can be sufficient to bring out the Warburg/Crabtree effect. Moreover, this model offers a general perspective of the regulation of the central carbon metabolism that can be used to study other biological questions.

When we model a complex biological system, we try to understand the causality chains that explain the different behaviours observed. However, these observations are often made under experimental conditions which are not necessarily comparable since they depend on the culture medium for example. The construction of a right modelisation therefore depends on our ability to take into account all this information in a single framework.

TotemBioNet is a new software platform to assist the design of qualitative regulatory network models by combining “genetically modified Hoare logic”, temporal logic model checking and optimized enumeration techniques. TotemBioNet is particularly efficient to manage parameter identification, the most critical step of formal modelling. It is also remarkably flexible and efficient to check properties in order to explore biological assumptions. To illustrate this efficacy, we address the classical example of the cell cycle, where the passage from one phase to the next one, called checkpoint , is crucial but is usually a rather fuzzy informal concept. The cyclic behaviour of the cell cycle is specified by temporal logic and the order of individual events inside each phase is explored thanks to quantifiers introduced in Hoare logic. This way, TotemBioNet rapidly suggests a sensible formalization of the notion of checkpoint.

Endocrine disrupting chemicals raise a lot of interest and concern regarding their risk for human health and the environment. They represent a broad variety of natural and synthetic chemicals with different levels of endocrine activity evaluation. In particular, for high production volume chemicals, new methods are required to enable the evaluation of the vast number of chemicals for their potential to alter the endocrine system and prioritize them for deeper characterization. The ToxCast program from the US EPA provides data from high throughput screening assays to develop computational tools aimed at rapid in vitro bioactivity screening and prioritization.Using publicly available data (ToxCast and ToxRef databases), we evaluate whether in vitro assay evaluations could predict in vivo outcomes observed in rat long-term studies for more than 400 chemicals. We focus on effects observed in three endocrine and two sex accessory organs and 42 in vitro assays related to pathways associated with endocrine related toxicity.First, using simple statistical correlation we demonstrate that there is no mutual linear correlation between the selected in vitro assays and any in vivo outcome, with balanced accuracies around 50% for each assay-outcome pair. Then, by applying machine learning to investigate potential non-linear correlations, we show that the combination of different in vitro assays is not correlated with the long-term in vivo effects and cannot help to predict them since balanced accuracies are also around 50%. Moreover, the prediction based on in vitro assays is not better than the one based on classical QSAR methods. This study highlights that the selected in vitro assays do not provide information about in vivo outcomes observed in endocrine and associated organs in long-term rat in vivo studies and stresses the need for the development of in vitro assays that reflect the compounds’ pharmacokinetic properties.

The development of in silico tools able to predict bioactivity and toxicity of chemical substances is a powerful solution envisioned to assess toxicity as early as possible. To enable the development of such tools, the ToxCast program has generated and made publicly available in vitro bioactivity data for thousands of compounds. The goal of the present study is to characterize and explore the data from ToxCast in terms of Machine Learning capability. For this, a large scale analysis on the entire database has been performed to build models to predict bioactivities measured in in vitro assays. Simple classical QSAR algorithms (ANN, SVM, LDA, random forest, and Bayesian) were first applied on the data, and the results of these algorithms suggested that they do not seem to be well-suited for data sets with a high proportion of inactive compounds. The study then showed for the first time that the use of an ensemble method named "Stacked generalization" could improve the model performance on this type of data. Indeed, for 61% of 483 models, the Stacked method led to models with higher performance. Moreover, the combination of this ensemble method with an applicability domain filter allows one to assess the reliability of the predictions for further compound prioritization. In particular we showed that for 50% of the models, the ROC score is better if we do not consider the compounds that are not within the applicability domain.

The study of gene networks allows us to better understand some biological processes such as the adaptation of the organism to a disturbance of the environment. In a discrete modelling framework of gene networks, it has been shown that the Hoare logic can help the modeller to identify the parameters of the model, so that the latter exhibits the observed biological traces. In this paper we present a hybrid modelling of gene networks which pays particular attention to the time spent in each state and we introduce an extension of the Hoare logic in this hybrid framework. The weakest precondition calculus associated with this modified Hoare logic makes it possible to determine the minimal constraints on the dynamic parameters of a gene network from an observed biological trace. These constraints form a continuous CSP that can be solved using the AbSolute continuous solver. The first experimental results show that the obtained solutions are in agreement with the specification of the Hoare triple coming from biological expertise.

The study of gene networks allows us to better understand some biological processes such as the adaptation of the organism to a disturbance of the environment. In a discrete modelling framework of gene networks, it has been shown that the Hoare logic can help the modeller to identify the parameters of the model so that the latter exhibits the observed biological traces. In this paper we present a hybrid modelling of gene networks which pays particular attention to the time spent in each state and we introduce an extension of the Hoare logic in this hybrid case. The weakest precondition calculus associated with this modified Hoare logic makes it possible to determine the minimal constraints on the dynamic parameters of a gene network from an observed biological trace. These constraints form a continuous CSP that we ∗Papier doctorant : Jonathan Behaegel1 est auteur principal. solve using the AbSolute continuous solver. The first experimental results show that the obtained solutions are in agreement with the specification of the Hoare triple coming from biological expertise.

Toxicology studies are subject to several concerns, and they raise the importance of an early detection of the potential for toxicity of chemical compounds which is currently evaluated through in vitro assays assessing their bioactivity, or using costly and ethically questionable in vivo tests on animals. Thus we investigate the prediction of the bioactivity of chemical compounds from their physico-chemical structure, and propose that it be automated using machine learning (ML) techniques based on data from in vitro assessment of several hundred chemical compounds. We provide the results of tests with this approach using several ML techniques, using both a restricted dataset and a larger one. Since the available empirical data is unbalanced, we also use data augmentation techniques to improve the classification accuracy, and present the resulting improvements.

G-Networks and their simplified version known as the Random Neural Network have often been used to classify data. In this paper, we present a use of the Random Neural Network to the early detection of potential of toxicity chemical compounds through the prediction of their bioactivity from the compounds' physico-chemical structure, and propose that it be automated using machine learning (ML) techniques. Specifically the Random Neural Network is shown to be an effective analytical tool to this effect, and the approach is illustrated and compared with several ML techniques.