Reinforcement Learning Guides Cellular Navigation in Chemical Networks

This paper introduces a novel framework that integrates reinforcement learning principles into chemical reaction networks to model how living systems navigate complex environments. Specifically, it reinterprets the phototaxis of unicellular algae, traditionally seen as a mechanistic stimulus-response, as a subjective, information-driven sensorimotor process. The core idea links a Partially Observable Markov Decision Process (POMDP) with biochemical reaction dynamics. The model allows a cell to update its internal state based on noisy observations, balancing light orientation with exploratory reorientation. These internal dynamics are implemented using Chemical-Reaction-Network Ordinary Differential Equations (CRN-ODEs). By applying Inverse Reinforcement Learning to experimental Chlamydomonas trajectories, the researchers inferred the underlying behavioral objective, showing that the model accurately reproduces empirical alignment-to-light distributions. This work suggests that behaviors like run-tumble alternation emerge as a strategy for information acquisition, resolving sensory ambiguity through intracellular biochemical networks.