Reinforcement learning in chemical networks explains algae's curiosity-driven phototaxis

Living systems navigate noisy environments with incomplete sensory data—a challenge traditionally modeled with simple stimulus–response rules. In a new paper accepted at ALIFE 2026, Tang et al. reframe phototaxis in unicellular algae as an information-driven, curiosity-based exploration process. They propose a framework that links a Partially Observable Markov Decision Process (POMDP) with biochemical reaction dynamics via Chemical-Reaction-Network Ordinary Differential Equations (CRN-ODEs). The cell’s hidden environmental variables are inferred through a memoryless Bayesian update after each observation, balancing light-seeking with exploratory reorientation. A biophysical photoreception process and a polynomial bound on information gain make the model chemically computable.

To validate their approach, the team applied Inverse Reinforcement Learning (IRL) to 30 experimental trajectories of Chlamydomonas algae. The inferred behavioral objective closely matches observed phototactic motion, and the resulting dynamics perform comparably to standard Stochastic Simulation Algorithm (SSA) baselines. Critically, the model reveals that the classic run-tumble alternation is not just a mechanical reflex but an active information-acquisition strategy: tumbling reorients the cell to sample new sensory configurations and resolve ambiguity. This demonstrates how intracellular biochemical networks can support adaptive, information-seeking behavior, bridging machine learning, biophysics, and minimal cognition.