Sampling Function-Related Metastable States of Proteins With DASH

Abstract

Rational discovery of function-specific protein modulators as well as activity-enhanced engineering proteins underscore the need to identify function-related metastable states (FMSs) of proteins. However, current experimental and computational methods struggle to generate these states directly from their native state (NS), likely because the NS [->] FMS transition is non-spontaneous. To address this challenge, we introduce Deep learning guided Adaptive sampling with seed Selection and Hopping (DASH), integrating both deep learning and physical functions to guide molecular dynamics (MD) simulation towards FMS. DASH successfully sampled NS [->] FMS transitions in 18 cases across two tasks: protein activation and cryptic allosteric site opening. DASH combined with secondary-structure collective variables is further able to sample folding of disordered regions. Compared to existing methods, DASH demonstrates superior performance while requiring comparable simulation time. Crucially, we applied DASH to sample new conformations of proteins, by which revealing new folding states, previously unknown allosteric sites, and potential activators. These predictions have been further verified in wet-lab experiments and crystal structures determinations. Collectively, our framework provides a robust strategy for function-specific pharmacy research and could accelerate future drug discovery efforts.