Speaker
Description
Drug resistant pathogens are a wide-spread and deadly phenomenon that infect nearly 3 million individuals in the United States each year. If microbial resistance continues to develop at the current rate, bacterial infections are expected to surpass cancer as the leading cause of death worldwide by 2050. Novel approaches to designing therapy that explicitly take into account the adaptive nature of microbial cell populations are desperately needed. In this study, we describe EvoDM, a reinforcement learning system capable of achieving superior population control in a simulated system of evolution. Reinforcement learning is well-studied subfield of machine learning that has been successfully applied to applications ranging from board games and video games to manufacturing automation. Using previously described evolutionary simulation methods that use fitness landscapes to describe selective pressures, we defined an evolutionary "game" for EvoDM to play. We then trained the EvoDM agent using Deep Q learning, a reinforcement learning algorithm well-suited to situations where little is known about the environment. Given access to a panel of simulated drugs with which to treat the simulated population, we demonstrate that EvoDM outperforms two potential treatment paradigms at minimizing the population fitness over time. We also show that EvoDM approaches the performance of the optimal drug cycling policy, computed through backwards induction of a markov-decision process formulation of our system. We also demonstrate how EvoDM performance is affected by modulating the size of the fitness landscape, the degree of epistasis, and the type of input used for training. Crucially, we show that it is possible for EvoDM to learn effective drug cycling protocols using current population fitness as the only training input. These tests will inform future in vitro implementations of our work. EvoDM represents a proof-of-concept for using AI to understand and treat evolving systems in medicine.
