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

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Medicinal chemistry and molecular design

Background:

• Machine learning, particularly generative models (GMs), shows promise for designing molecules with desired properties in drug discovery.
• Existing GMs face challenges in achieving target engagement, synthetic accessibility, and robust generalization.
• Addressing these limitations is crucial for advancing AI-driven drug design.

Purpose of the Study:

• To develop and validate an improved generative model workflow for drug discovery.
• To enhance molecule design by integrating active learning with chemoinformatics and molecular modeling predictors.
• To generate novel, drug-like molecules with high predicted affinity and synthetic accessibility for specific biological targets.

Main Methods:

• Developed a generative model workflow combining a variational autoencoder with two nested active learning cycles.
• Iteratively refined molecular predictions using chemoinformatics and molecular modeling tools.
• Applied the workflow to design molecules targeting Cyclin-Dependent Kinase 2 (CDK2) and KRAS proteins.

Main Results:

• Successfully generated diverse, drug-like molecules with high predicted affinity and synthesis accessibility for both CDK2 and KRAS.
• Discovered novel molecular scaffolds distinct from known inhibitors for each target.
• Synthesized 9 molecules for CDK2, with 8 showing in vitro activity (one with nanomolar potency); identified 4 potential KRAS inhibitors.

Conclusions:

• The developed generative model workflow effectively explores novel chemical spaces tailored for specific targets.
• The workflow demonstrates significant potential for accelerating the discovery of potent and synthesizable drug candidates.
• This approach opens new avenues for AI-driven drug discovery, overcoming limitations of traditional generative models.