Active Learning to Select the Most Suitable Reagents and One-Step Organic Chemistry Reactions for Prioritizing Target-Specific Hits from Ultralarge Chemical Spaces

Designing chemically novel and synthesizable ligands from the largest possible chemical space is a major issue in modern drug discovery to identify early hits that are easily amenable to medicinal chemistry optimization. Starting from the sole three-dimensional structure of a protein binding site, we herewith describe a fully automated active learning protocol to propose the commercial chemical reagents and one-step organic chemistry reactions necessary to enumerate target-specific primary hits from ultralarge chemical spaces. When applied in different scenarios (single transform and multiple transforms) addressing chemical spaces of various sizes (from 670 million to 4.5 billion compounds), the method was able to recover up to 98% of virtual hits discovered by an exhaustive docking-based approach while scanning only 5% of the full chemical space. It is therefore applicable to the structure-based screening of trillion-sized chemical spaces at a very high throughput with minimal computational resources.