Graph neural networks for molecular dynamics simulations

Abstract

Graph neural networks (GNNs) are emerging as powerful tools for advancing molecular dynamics (MD) simulations, providing data-driven frameworks to complement traditional physics-based approaches. By representing atoms and their interactions as graphs, GNNs naturally encode chemical and structural information, enabling accurate neural network force fields trained on quantum data, automated discovery of collective variables for enhanced sampling, and efficient prediction of atomic forces to extend simulation timescales. Beyond driving MD, GNNs facilitate the analysis of high-dimensional trajectories, offering interpretable insights through attention mechanisms or transferable embeddings. Applications such as protein–DNA assembly, pretrained featurizers, and cryptic pocket discovery illustrate the breadth of GNNs, underscoring their potential to transform biomolecular simulations and accelerate mechanistic and translational discoveries.