From Monoliths to Pharmacists-at-Scale: Patient-Aware Multi-Agent Reasoning Tames Million-Dimensional Discovery

Keywords: Multi-Agent Reinforcement Learning (MARL), Drug Synergy Prediction, Clinical Dose Optimization, Patient-Aware AI, Multi-agent LLM System

Abstract: Drug synergy prediction is constrained by vast combinatorial spaces, costly validation, and the trade-off between efficacy and toxicity. We introduce a patient-aware, reinforcement-learning-augmented multi-agent system that re-imagines discovery as an active, closed-loop search over both drug pairs and individual pharmacology. Where traditional QSAR and even recent deep-learning baselines treat synergy as a static regression problem and thus plateau at dataset-wide RMSE near 0.06, our environment embeds patient-specific clearance, BSA, and toxicity thresholds directly into the reward. A factorized set of agents—Synergy Scout, Dose Adapter, and Safety Sentinel—explore the joint space via distributed deep Q-networks with prioritized replay, while an ensemble of analysts continuously recalibrates predictions against clinical outcomes. Across more than one million drug–patient combinations, this design delivers a validation R² of 0.913 and an 83.2% accuracy on literature-validated pairs, translating to a 722% efficacy gain over DeepSynergy and a 15% AUROC lift over the best prior multi-agent framework. The resulting system is not only more accurate but also intrinsically interpretable, providing transparent rationales that monolithic pipelines cannot.

Submission Number: 224