Integrating Local Runtime Geometry and Global Equivariant Channels for 3D Molecular Representation Learning

Abstract

Equivariant graph neural networks have greatly advanced the development of machine learning molecular force fields, achieving molecular dynamics simulations with near quantum chemical accuracy. As an efficient vector-scalar interactive equivariant model, ViSNet relies on runtime geometric calculation (RGC) to implicitly extract local geometric information such as bond angles, dihedral angles, and improper angles with linear complexity, achieving excellent performance on small and medium-sized molecular datasets. However, native ViSNet only relies on local neighborhood message passing within a cutoff radius, meaning long-range interactions can only propagate indirectly by stacking multiple network layers. In large molecules and flexible protein systems, this easily leads to vanishing gradients and over-smoothing of representations, limiting its modeling capacity for long-range conformational spaces. This paper proposes G-ViSNet, which introduces an equivariant global scalar-vector distributing-aggregating mechanism into every ViS-MP block of ViSNet to construct a low-cost global information interaction pathway. Experiments conducted on the MD17 small molecule benchmark and MD22 large molecule benchmark datasets show that G-ViSNet exhibits no performance degradation on small molecules and significantly improves the prediction accuracy of energy and forces for large molecules. G-ViSNet balances local geometric feature extraction and long-range information modeling, providing an efficient solution for machine learning molecular dynamics simulations of flexible biological macromolecules.