Efficient Fisher Information Computation and Policy Search in Sampled Stochastic Chemical Reaction Networks through Deep Learning

Markov jump processes constitute the central class of Chemical Reaction Network models used to account for the intrinsic stochasticity observed in the dynamics of molecular species abundance throughout Molecular Biology. These models are specified in a parametric form, and their identification requires the use of inference procedures, and in particular the estimation of the Fisher Information. Here, a fast and accurate computation method is introduced in the case of partial observations at discrete time points, based on the use of a Mixture Density Network. We also demonstrate how this Neural Network can be used to perform fast policy search. The efficiency of these approaches is illustrated on a set of examples, and is compared to that of the current state-of-the-art.