MARD: Mirror-Augmented Reasoning Distillation for Mechanism-Level Drug-Drug Interaction Prediction
Researchers have developed MARD (Mirror-Augmented Reasoning Distillation), a novel system for predicting mechanism-level drug-drug interactions (DDIs). MARD utilizes a 7B parameter model and incorporates three key training innovations: a single-token KL divergence for direction tags, DPO with programmatic hard negatives, and a leakage-safe retrieval channel. This approach allows for auditable reasoning metrics and surpasses existing methods, including GPT-4o, in accuracy on novel drug pairs while operating at a fraction of the cost. AI
IMPACT Advances drug-drug interaction prediction accuracy and efficiency, potentially improving pharmaceutical research and development.