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

Modern drug discovery faces significant challenges in optimizing compounds for favorable ADMET properties, notably metabolic stability, as rapid metabolism impacts bioavailability, half-life, and drug-drug interactions. Central to these issues are cytochrome P450 (CYP450) enzymes, which metabolize most approved drugs, with unintended inhibition or accelerated clearance contributing substantially to clinical failures. To address this, we developed DiffMeta-RL, a discrete graph diffusion model enhanced with reinforcement learning, enabling controllable optimization of pharmacological properties. Unlike conventional actor-critic frameworks that require a separate policy network, DiffMeta-RL introduces a streamlined approach by directly optimizing the generative diffusion model itself as the policy, embedding reward-guided updates into the denoising process. Additionally, we proposed MetaCYP, a highly accurate CYP450 inhibition predictor, integrating it into DiffMeta-RL's reward function to guide molecular generation toward reduced CYP450 liability and improved metabolic stability. In benchmark evaluations, DiffMeta-RL outperforms existing models in validity, uniqueness, and novelty. Specifically applied to designing proton pump inhibitors (PPIs) targeting gastric H⁺/K⁺-ATPase, DiffMeta-RL successfully generated candidates exhibiting favorable predicted binding affinity and reduced susceptibility to CYP2C19 and CYP3A4 metabolism, highlighting its practical potential for advancing drug discovery efforts.