Application of Parallelized Graph Neural Networks in Predicting Molecular Properties from Large-Scale Chemical Databases

Abstract

A parallel GNN architecture is proposed to solve the problems of large computation and poor scalability in molecular property prediction of large-scale chemical databases. This study uses data-level and model-level parallel methods, distributed training and memory optimization methods to learn graph-structured molecular data. To make GNN training as efficient as possible on heterogeneous datasets on multi-node GPU clusters, we optimize graph batching, adaptive sampling, workload balancing, and communication-efficient synchronization. According to the results of the standard chemical benchmark experiment, the parallelized GNN obtained a great acceleration ratio, a small peak memory usage, and still had a high prediction accuracy compared with the basic method. The comprehensive analysis analyzes how using better partitioning and sampling techniques affects how many conversations and what is needed. In computational chemistry and materials science, these distributed and parallel GNN methods seem to have good large-scale molecular prediction capabilities.