DeepReweighting: Reparameterizing Force Field under Explainable Deep Learning Framework

Abstract

The analysis and prediction of biomolecule structures is indispensable in various research fields such as protein engineering, mRNA vaccine research and targeted drug screening. Molecular dynamics (MD) simulation is currently the most commonly used and reliable tool for sampling dynamic conformation ensemble of large biomolecules. However, the force fields commonly used in MD simulations are composed of empirical parameters, and their accuracy are limited by the training set data used for parameterizing the force field.In response to these challenges, we have developed a novel force field optimization strategy based on an explainable deep learning framework, DeepReweighting, for rapid and precise force field re-parameterization and optimization. DeepReweighting demonstrates a significant increase in re-parameterization efficiency compared to traditional Monte Carlo method and exhibits greater robustness. Furthermore, DeepReweighting can rapidly re-parameterize any existing or custom differentiable parameters in the force field, providing a faster and more accurate tool for optimizing and utilizing molecular force fields.