The performance of machine learning (ML) models for drug discovery often depends on the amount of data available for training. In the case of modeling ab initio properties, many costly quantum chemistry calculations may be required. We report a dataset of nearly 50,000 drug-like structures and ab initio properties including high-quality atomic electrostatic multipoles. Using a distributed, high-throughput workflow, all calculations were performed using over 6,000 processors in less than 24 hours, demonstrating a capacity to rapidly develop problem-specific datasets.<br/><br/>The message passing neural network (MPNN) framework is one of the most popular and promising ML architectures in chemistry, although it currently has incompatibilities with the prediction of directional atomic properties. We propose a modified Cartesian MPNN (CMPNN) suitable for predicting directional properties, including atom-centered multipoles, an essential component of <i>ab initio</i> force fields. The efficacy of this model is illustrated using the newly developed dataset. We show that the CMPNN accurately predicts atom-centered charges, dipoles, and quadrupoles, and that errors in the predicted atomic multipoles have a negligible effect on multipole-multipole electrostatic energies. The CMPNN is accurate enough to capture conformational changes of a molecule's electrostatic potential. These results suggest the possibility of dynamically predicting atomic multipoles and other directional properties throughout simulations in which they might exhibit strong conformational dependence.

High-throughput quantum chemistry dataset generation for directional atomic properties

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.