High-throughput virtual screening (HTVS) is a critical tool in the molecular discovery pipeline that aids in narrowing down a massive set of candidate compounds to a much smaller collection to be experimentally validated. The "virtual libraries" commonly evaluated through HTVS continue to grow in size, with several in excess of 108 molecules. As a result, the computational resources required for screening has grown to levels that are both impractical for regular use in early-stage discovery and inaccessible to many academic researchers. Given that a virtual screen generally seeks to identify the top predicted candidate molecules in a virtual library, Bayesian optimization can serve to accelerate this process via the usage of a surrogate structure-property relationship model that guides the selection of future experiments. This talk will describe our efforts to apply Bayesian optimization to structure-based virtual screening as a proxy for protein-ligand binding affinity. We examine the effects of surrogate model architecture, acquisition function, and batch size as applied to protein-ligand docking datasets of varying size. Our studies show significant reductions in the computational costs associated with virtual screening, identifying 87.9% of the top-50000 ligands in a 100M-member library after testing only 2.4M compounds Moreover, we demonstrate that (1) despite the theoretical motivation of uncertainty-aware acquisition functions, greedy acquisition is a consistently reliable strategy; (2) higher capacity surrogate model architectures lead to appreciable increases in optimization performance; and (3) the use of model-guided selection facilitates the identification of the most potent molecules (by docking score) across all datasets tested. Such techniques will mitigate computational costs associated with the screening of increasingly large virtual libraries.

Accelerating high-throughput virtual screening through molecular pool-based active learning

ACS Spring Meeting 2021

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