A Hybrid Bottom-Up and Data-Driven Machine Learning Approach for Accurate Coarse-Graining of Large Molecular Complexes.

Abstract

Bottom-up coarse-graining refers to the development of low-resolution simulation models that are thermodynamically consistent with certain distributions from fully atomistic simulations. Force-matching and relative entropy minimization represent two major, frequently applied methods that allow to develop such bottom-up models. Nevertheless, atomistic simulations can often provide only limited sampling of the phase space. For bottom-up coarse-graining, these limitations may result in overfitting of the atomistic reference data, especially for large molecular complexes, where the learning may be agnostic of the actual affinities between binding partners. As a solution to this problem, we devise a data-driven machine learning hybrid coarse-graining concept that represents a regularized version of the relative entropy minimization approach. We demonstrate that this new approach allows one to develop coarse-grained models for molecular complexes that reproduce the targeted binding affinity but also describe the underlying complex structure accurately. The trained models therefore show diverse behavior as they can undergo frequent unbinding and binding events and are also transferable for simulating entire protein lattices, e.g., for a virus capsid.