Parametrisation of hybrid Gene Regulatory Networks using Artificial Evolution and Reinforcement Learning

Genetic regulatory network (GRN) modelling aims at studying and understanding the molecular mechanisms that enable the organism to perform essential functions ranging from metabolism to environmental disturbance adaptation. Studying the dynamics of these systems opens new perspectives with crucial applications in fundamental biology, pharmacology, medicine, or chronotherapy for instance, which tries to choose the best time of day to administer the medication in order to limit the side effects while preserving the therapeutic effects. The bottleneckin the modelling approach is the search for model parametrisations that are consistent with the available biological knowledge, often represented as irregularly spaced time series of observable events. In this work, we have chosen hybrid frameworks (hGRN), which add to the discrete ones the time spent in each of the discrete states leading to the search of piecewise linear trajectories in a multidimensional space. These trajectories, each governed by a parametrisation, must respect the observed biological data, which have been previously and manually interpreted as a set of constraints based on Hoare logic [2]. A previous work logically applied continuous Constraint Satisfaction Problem (CSP) solvers to identify all valid trajectories [1], but faced difficulties in extracting solutions, especially when the number of genes exceeded 3 genes. Here, we first consider the determination of valid parameterisations as an unconstrained optimisation problem and solve it by comparing different metaheuristics. In a second step, we treat this problem as a sequential decision problem using reinforcement learning with Monte Carlo Tree Search (MCTS) algorithms. The two approaches are summed up in the next sections.