Accelerating Virtual Directed Evolution of Proteins via Reinforcement Learning

With the advancement of machine learning methods, the protein fitness landscape can be predicted, providing reliable guidance in the selection of advantageous mutations for the directed evolution of proteins. However, the potential multiple mutational variants derived from the simple combination of a limited number of advantageous single mutations may not represent superior choices. Moreover, the exploration and selection of the astronomical number of multiple mutational variants remain a highly challenging task. In this study, we introduce a virtual directed evolution pipeline, RelaVDEP, for the rapid identification of mutational variants with explicit enhancement in the desired property of the target protein. By adapting and fine-tuning a pre-trained fitness predictor to improve sequence-based protein functional prediction and by designing a model-based reinforcement learning framework to efficiently explore the vast combinatorial space of protein mutations, this pipeline is capable of effectively accelerating the directed evolution process for a broad spectrum of proteins with versatile functional profiles. According to a series of experimental validations, the diversified mutational variants identified by our method exhibit notable improvements in desirable protein functional properties. In particular, by integrating RelaVDEP with active learning, we successfully optimized the sequence of a PETase, enhancing its catalytic activity through previously unknown mutations.