Fast and accurate committor estimation for kinetics simulations

Abstract

Computing long-timescale kinetics of biomolecular processes remains a major challenge for atomistic simulations. A way out is to exploit local kinetic information to construct the global stationary flux across the reaction space. The committor serves as the optimal reaction coordinate for this purpose; however, its calculation is itself highly demanding. Here, we introduce a fast and accurate algorithm for committor estimation by leveraging highly parallelizable short trajectory simulations and analogue prediction. The resulting committor is represented via a neural network ansatz and subsequently coupled with the Milestoning method to predict the mean first passage time at very low computational cost. We demonstrate the robustness and efficiency of this committor-guided Milestoning (CoM) method through examples of increasing complexity.