Machine learning guided weighted ensemble for rare event simulation in biophysics.

Abstract

Studying the kinetics and thermodynamics of rare events such as protein folding, protein-ligand binding, and biomolecular self-assembly through atomistic molecular dynamics simulation is a challenge due to the long timescale involved. Weighted ensemble is a path sampling technique, where multiple trajectories are initiated and simulated in parallel, and as the configurational space is being explored, the trajectories propagating towards the target state are duplicated with appropriate redistribution of weights, and the trajectories which do not are discontinued. A progress coordinate is a descriptor which can capture the transitions in a biophysical process and can also distinguish the metastable states in a system. In this study, we present a Weighted Ensemble approach, where we use a progress coordinate trained using machine learning to calculate the mean first passage times of the rare events. From our method we were able to construct the free-energy landscape and estimate rate constants for the transitions in a small protein with reasonable accuracy. This method will potentially facilitate the exploration of configuration space of complex biomolecular processes in the future.