Graph Nets for partial charge prediction

Yuanqing Wang (MSKCC) gave a talk about using Graph Nets for fast prediction of atomic partial charges as a part of OFF webinar series. The preprint is available here.

Abstract: Here we show that Graph Nets — a set of update and aggregate functions that operate on molecular topologies and propagate information thereon — are capable of predicting properties of atoms and molecules which would otherwise require expensive Quantum Mechanics (QM) calculations. This architecture is applied to predict atomic partial charges, which are crucial parameters for Molecular Dynamics (MD) simulations, molecular mechanics calculations, and virtual screening, as they determine the electrostatic contributions to interaction energies. We will also briefly introduce our work-in-progress on using this model for inter-hierarchical multitask learning and on Hypergraph Functional Potential (HGFP) — a fast and flexible machine learning force field.