Computational docking is an invaluable tool in drug discovery pipelines that enables the rapid screening of ultra-large compound libraries. Previous studies have utilized docking to identify novel anti-inflammatory and antibiotic drug candidates in chemical libraries containing 1.4B and 100M compounds, respectively. However, workflows of these magnitudes are expensive and impractical as standard practice. We and others have recently investigated the application of surrogate model-guided optimization techniques to docking-based high-throughput virtual screening (HTVS) and shown marked increases in sample efficiencies. In a retrospective study of a library with 100M compounds, we identify 94.8% of the top-50000 compounds after screening only 2.4% of the library. However, a practical challenge remains in the limited availability of open-source software that implement these algorithmic advancements in a fully pipelined manner. Tools exist that execute the individual steps of a docking-based optimization, but tying these tools together requires continual monitoring of job progress or proficiency with HPC job schedulers. This talk will describe our group's recent efforts developing two open-source software tools to support model-accelerated virtual screening: pyscreener and MolPAL. Pyscreener enables python-based calls to popular computational docking programs, such as the AutoDock Vina family and DOCK6. It allows users to calculate the docking scores of molecules directly from their respective SMILES strings, serving as a simple black-box oracle for screening or molecular optimization tasks. With pyscreener, users can rapidly test out different docking programs or protocols with minimal changes to their code. Moreover, the first-class support for task distribution facilitates the seamless scaling from local to distributed hardware setups. MolPAL (molecular pool-based active learning) uses Bayesian optimization to accelerate HTVS in a fully end-to-end manner. The software is able to leverage large resource allocations to iteratively (1) dock batches of millions of molecules, (2) train machine learning models on datasets containing millions of samples, and (3) perform inference on 100M+ samples. Additionally, it is agnostic as to the underlying system architecture, running the same on AWS as on a SLURM-managed HPC cluster. We describe the design, use, and scalability of these software tools in addition to ongoing improvements to further decrease computational costs.

Open source software tools to accelerate docking-based molecular discovery

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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