Optimizing drug design by merging generative AI with a physics-based active learning framework

Abstract

Machine learning is transforming drug discovery, with generative models (GMs) gaining attention for their ability to design molecules with specific properties. However, GMs often struggle with target engagement, synthetic accessibility, or generalization. To address these, we developed a GM workflow integrating a variational autoencoder with two nested active learning cycles. These iteratively refine their predictions using chemoinformatics and molecular modeling predictors. We tested our workflow on two systems, CDK2 and KRAS, successfully generating diverse, drug-like molecules with high predicted affinity and synthesis accessibility. Notably, we generated novel scaffolds distinct from those known for each target. For CDK2, we synthetized 9 molecules yielding 8 with in vitro activity, including one with nanomolar potency. For KRAS, in silico methods validated by CDK2 assays identified 4 molecules with potential activity. These findings showcase our GM workflow’s ability to explore novel chemical spaces tailored for specific targets, thereby opening new avenues in drug discovery. Generative models show promise in drug discovery by enabling the design of molecules with desired properties, yet often face challenges related to target engagement, synthetic accessibility, and generalization. To address these, the authors developed a workflow combining a variational autoencoder with active learning cycles, generating diverse, drug-like molecules with synthetic feasibility and high predicted affinity for CDK2 and KRAS.