Learning probabilistic models of biological systems using active inference with belief propagation

Abstract

In this work, the normative framework of active inference is integrated with belief propagation for inverting a probabilistic causal model using data generated from planned interactions between a Bayesian modeling agent and a biological system. Thompson sampling of parameter distributions is used to estimate the free energy of the expected future when beliefs about beliefs are rolled over a planning horizon. Learning a probabilistic model for maximizing biomass production in the well-known Baker’s yeast example is used as an ex-ample. The prior parameter distributions in the system model of a fed-batch cultivation are updated as new observations are obtained. Planned action sequences aim to excite the yeast metabolism by introducing changes in the feed rate of two nutrients (glucose and nitrogen). Results obtained demonstrate that by maximizing the model evidence, the proposed approach constraints biological system dynamics to relevant trajectories for improved parametric precision in the preferred region of physiological states that favor biomass productivity.