Microbiome Evolution from a New Perspective: Reinforcement Learning Provides an Evolutionary Perspective of Microbial Interactions.

Microorganisms are integral to ecosystems, shaping their environment and material flow by forming a complex network of interacting cells. Predicting the phenotype of microorganisms from their genotype has motivated creation of mathematical models to describe the behavior of microbial communities. While metabolic modeling methods based on flux balance analysis (FBA) offer insights into homogenous and heterogenous microbial community metabolism, it falls short in predicting long-term stability, especially in presence of interacting microbes. We propose "Self-Playing Microbes in Dynamic FBA", a novel reinforcement learning algorithm. It treats microbial metabolism as a decision-making process, enabling microorganisms to adapt metabolic strategies for enhanced fitness in a dynamic context by trial and error in a simulation environment and finding flux regulation policies that stabilize in a microbiome in the presence of other microbes, by relying on first principles of microbial ecology with minimal reliance on pre-determined strategies and experimental observations. Our work demonstrates improved performance over existing methods in various scenarios, such as, metabolite exchange between auxotrophs and secretion of extracellular enzymes by the cells highlighting its biological significance.