Physics-Informed Bayesian Optimization for Conformational Ensemble Augmentation

Conformational
search is a key part of reaction modeling, molecular
docking, and other fields of computational chemistry, where it is
important to take molecules’ flexibility into account. However,
modern conformational search approaches provide no guarantee that
they did not miss any important conformation. Thus, identifying missing
conformations from an existing ensemble is of broad importance for
computational chemistry. In this paper, we introduce a Bayesian optimization
algorithm for conformational ensemble augmentation, that is, locating
missing conformers in an existing ensemble, which employs Bayesian
optimization with physics-informed torsion-potential-based kernel
function and a novel acquisition function that prioritizes potential
energy surface exploration for increased conformer diversity. The
devised method demonstrates high efficiency on a test set of biologically
relevant molecules.