Many important biological processes, such as the formation of biomolecular condensates by liquid-liquid phase separation, involve interactions between a large number of both folded and disordered proteins. They usually occur on a time scale that is beyond the capability of current  atomic molecular simulations. Studying these processes using simulations requires a coarse-grained (CG) force field for both folded and disordered proteins. The transferability of the CG force field is essential to make accurate predictions from simulations. Here we introduce an efficient machine learning method for learning a CG force field for both folded and disordered proteins. Two strategies are used to maximize the CG force field’s transferability. First, it combines information from both atomic molecular simulations and experimental data. Second, the CG force field is simultaneously optimized on a large number of both folded and disordered proteins. As a result, the learned CG force field can match conformational ensembles from atomic molecular simulations as closely as possible while being able to reproduce experimental data for both folded and disordered proteins. We expect that the learned CG force field will be useful for studying and predictive understanding of long-timescale biological processes involving both folded and disordered proteins. In addition to learning CG force fields, the machine learning method used in this work will also be useful for constructing classical force fields using data from quantum mechanics calculations.

Machine learning a coarse-grained force field for both folded and disordered proteins

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.