MARD: Mirror-Augmented Reasoning Distillation for Mechanism-Level Drug-Drug Interaction Prediction

The article introduces MARD (Mirror-Augmented Reasoning Distillation), a 7 billion parameter model designed for mechanism-level drug-drug interaction (DDI) prediction, which outperforms 32 other systems in accuracy, particularly on novel drug pairs. Key innovations include a single-token KL divergence for direction tagging, PRM-weighted DPO training with hard negatives, and a mechanism-aware retrieval channel, all contributing to a model that requires no human or LLM judges for verification. This advancement is significant for practitioners as it enhances pharmacological reasoning capabilities in DDI prediction while maintaining low operational costs.