HARMONIQ: A QSAR-Guided Reinforcement Learning Framework for the Design of Small-Molecule MAO-B Inhibitors Targeting Alzheimer’s Disease

The design of small molecule inhibitors for monoamine oxidase B (MAO-B), a critical therapeutic target for Alzheimer’s disease, poses significant challenges for improving potency and druglikeness. We introduce HARMONIQ, a deep reinforcement learning framework with quantitative structure-activity relationship modeling that targets MAO-B for molecular generation. By integrating sequence-based and fragment-based drug generation, HARMONIQ designs molecules that are both diverse and synthesizable. The inhibitors generated by HARMONIQ had an average docking score of -12.06 kcal/mol, a 24.5% improvement over inhibitors generated by the baseline de novo model. The top HARMONIQ-generated molecule was synthesized and experimentally validated via fluorometric testing alongside 2 lead compounds from the Enamine REAL database. All 3 compounds have inhibitory potency against MAO-B and are promising drug candidates for Alzheimer’s disease.