Data-driven enhanced sampling of mechanistic pathways.

Abstract

The mechanisms of molecular processes can be characterized by following the minimum free energy pathway (MFEP) on the underlying multidimensional conformational landscapes. Despite recent advancements in enhanced sampling algorithms, obtaining a converged high-dimensional molecular free energy landscape remains a considerable challenge. To circumvent this issue, we employ a deep multitask learning algorithm that integrates deep neural networks with the established enhanced sampling method of well-tempered metadynamics to iteratively learn the MFEP between reactant and product conformations, without the knowledge of the underlying free energy landscape. Our approach improves upon existing pathway exploration algorithms by following a simpler protocol, thereby eliminating the need to identify intermediate structures along a guess path. From the learned pathway, an automatic reconstruction of a mechanistic fingerprint can be performed by following the sequence of events in the molecular process, allowing for a direct characterization of the molecular mechanism. We demonstrate applications of our algorithm to prototypical chemical reactions, protein folding, and ligand-receptor binding problems. Due to its low computational cost and overall simplicity, this framework is expected to find widespread applications in elucidating molecular mechanisms at all-atom resolution.