DiffMeta-RL: Reinforcement Learning-Guided Graph Diffusion for Metabolically Stable Molecular Generation.

Area of Science:

• * Computational Chemistry
• * Artificial Intelligence in Drug Discovery
• * Pharmacology

Background:

• * Optimizing ADMET properties, especially metabolic stability, is crucial in drug discovery.
• * Cytochrome P450 (CYP450) enzymes significantly influence drug metabolism, clearance, and potential drug-drug interactions.
• * Unfavorable CYP450 interactions contribute to high rates of clinical drug failure.

Purpose of the Study:

• * To develop an AI-driven method for optimizing drug molecules with improved metabolic stability.
• * To create a controllable molecular generation framework that minimizes CYP450 liability.
• * To enhance the efficiency and success rate of early-stage drug discovery.

Main Methods:

• * Developed DiffMeta-RL, a discrete graph diffusion model incorporating reinforcement learning.
• * Implemented a streamlined reinforcement learning approach by directly optimizing the diffusion model as the policy.
• * Integrated MetaCYP, a CYP450 inhibition predictor, into the reward function for guided molecular generation.

Main Results:

• * DiffMeta-RL demonstrated superior performance in validity, uniqueness, and novelty compared to existing models.
• * Generated proton pump inhibitor (PPI) candidates showed favorable predicted binding affinity.
• * Designed PPIs exhibited reduced susceptibility to CYP2C19 and CYP3A4 metabolism.

Conclusions:

• * DiffMeta-RL offers a powerful, controllable approach for designing drug candidates with enhanced metabolic stability.
• * The integration of CYP450 inhibition prediction directly into the generative process streamlines optimization.
• * This AI framework holds significant potential for advancing drug discovery and reducing attrition rates.