Best use of atomistic MD simulations for generalized molecular models

Determine effective strategies for leveraging atomistic molecular dynamics simulation data to develop generalized machine learning models that improve prediction accuracy for novel molecular structures, beyond approaches that rely solely on static structures.

Background

The paper highlights that many successful molecular representation learning approaches are trained on static structures of drug-like compounds and proteins. While such models can achieve strong performance, they may miss dynamical aspects of binding interactions captured by molecular dynamics (MD) simulations.

The authors explicitly state that integrating MD into model development to build more generalized representations is an unresolved challenge, motivating their SurGBSA approach as one step toward addressing this broader question.