Agents of Synergy: Patient-Informed Multi-Agent Reinforcement Learning for Safe Drug Combination Design

Mahule Roy, Subhas Roy

Abstract: Prediction of drug dosage synergy is challenged by the vast combinatorial space of drug pairs and the critical trade-off between efficacy and toxicity. We propose a patient-informed, multi-agent reinforcement learning framework that formulates synergy discovery as an active, closed-loop optimization over combination therapies and monotherapies. Unlike static regression models, our approach incorporates patient-specific factors—such as drug clearance and toxicity thresholds—directly into a structured reward function. Three specialized agents—Synergy Scout, Dose Adapter, and Safety Sentinel—coordinate via factorized deep Q-networks to explore the joint dosing space efficiently. Evaluated on over one million drug–patient combinations, our method achieves a validation R2 of 0.913 and 83.2% accuracy on literature-validated synergistic pairs, outperforming DeepSynergy by 7.2× in efficacy and surpassing the best prior multi-agent system by 15% in AUROC. Moreover, the modular architecture provides inherent interpretability, enabling transparent, agent-level explanations of dosing decisions.