Understanding cell populations sharing information through the environment, as reinforcement learning

Abstract

Collective migration is a phenomenon observed in various biological systems, where the cooperation of multiple cells leads to complex functions beyond individual capabilities, such as in immunity and development. A distinctive example is cell populations that not only ascend attractant gradient originating from targets, such as damaged tissue, but also actively modify the gradient, through their own production and degradation. While the optimality of single-cell information processing has been extensively studied, the optimality of the collective information processing that includes gradient sensing and gradient generation, remains underexplored. In this study, we formulated a cell population that produces and degrades an attractant while exploring the environment as an agent population performing distributed reinforcement learning. We demonstrated the existence of optimal couplings between gradient sensing and gradient generation, showing that the optimal gradient generation qualitatively differs depending on whether the gradient sensing is logarithmic or linear. The derived dynamics have a structure homogeneous to the Keller-Segel model, suggesting that cell populations might be learning. Additionally, we showed that the distributed information processing structure of the agent population enables a proportion of the population to robustly accumulate at the target. Our results provide a quantitative foundation for understanding the collective information processing mediated by attractants in extracellular environments.