Using classifiers to understand coarse-grained models and their fidelity with the underlying all-atom systems

Aleksander E P Durumeric; Gregory A Voth

doi:10.1063/5.0146812

Using classifiers to understand coarse-grained models and their fidelity with the underlying all-atom systems

J Chem Phys. 2023 Jun 21;158(23):234101. doi: 10.1063/5.0146812.

Authors

Aleksander E P Durumeric¹, Gregory A Voth¹

Affiliation

¹ Department of Chemistry, Chicago Center for Theoretical Chemistry, James Franck Institute, and Institute for Biophysical Dynamics, The University of Chicago, 5735 S. Ellis Ave., Chicago, Illinois 60637, USA.

PMID: 37318166
DOI: 10.1063/5.0146812

Abstract

Bottom-up coarse-grained (CG) molecular dynamics models are parameterized using complex effective Hamiltonians. These models are typically optimized to approximate high dimensional data from atomistic simulations. However, human validation of these models is often limited to low dimensional statistics that do not necessarily differentiate between the CG model and said atomistic simulations. We propose that classification can be used to variationally estimate high dimensional error and that explainable machine learning can help convey this information to scientists. This approach is demonstrated using Shapley additive explanations and two CG protein models. This framework may also be valuable for ascertaining whether allosteric effects at the atomistic level are accurately propagated to a CG model.