MARD-7B: A Drug-Drug Interaction Prediction System Based on Mirror-Enhanced Reasoning Distillation

MARD-7B is a small language model with 7 billion parameters. Using mirror-enhanced reasoning distillation technology, it outperforms GPT-4o in drug-drug interaction (DDI) prediction tasks while only costing about 1% of the latter's inference cost. The system adopts a structured classification system of 7 families and 147 subtypes to achieve mechanism-level DDI prediction, which can identify specific biological mechanisms, predict interaction directions, and provide verifiable evidence chains.

Drug-drug interaction (DDI) prediction is a core challenge in the field of drug safety. Traditional methods only determine whether an interaction exists but cannot answer key questions such as mechanism, affected enzymes/pathways, and interaction direction. Mechanism-level DDI prediction requires models to have deep pharmacological reasoning capabilities, combining professional pharmaceutical knowledge with interpretable reasoning processes.

System Architecture

MARD uses a multi-stage training framework consisting of four core phases:

• Data Construction: Build structured corpus based on DrugBank, using 7-family/147-subtype classification with directional information, and three segmentation strategies to ensure rigorous evaluation.
• Teacher Generation and PRM Training: Heterogeneous three-teacher ensemble generates reasoning trajectories, and a reward model (PRM) screens trajectories during training.
• Student Model Training: The 7B-parameter student model undergoes supervised fine-tuning (mirror-symmetric KL loss) and PRM-weighted DPO optimization.
• Evaluation and Validation: Introduce novel reasoning metrics such as Mechanism Fidelity Score (MFS).

Core Technologies

• Single-Token KL Directional Constraint: Ensures consistency in direction prediction.
• PRM-Weighted DPO and Programmatic Hard Negative Samples: Improve the model's discrimination ability.
• Leakage-Preventing Mechanism-Aware Retrieval Channel: Avoids data leakage in cold-start evaluation.
• Automatic Verifiable Reasoning Metrics: Validated based on DrugBank fields to reduce evaluation costs.

In the comparative experiment of 32 systems using the April 2026 version of DrugBank, MARD-7B performed outstandingly:

• 13.9 percentage points higher than the best baseline model
• 6.7 percentage points higher than GPT-4o
• Inference cost is only about 1% of GPT-4o's

The model exhibits an "anti-memorization" feature, with higher accuracy for rare drugs, indicating that its performance comes from structured pharmacological reasoning rather than memorization of high-frequency drugs.

Case Study

Taking Voriconazole (DB00582) and Axitinib (DB06626) as examples, the reasoning process of MARD-7B is: Input Processing → Retrieval Enhancement → Multi-Step Reasoning → Structured Output (predicts "PK Metabolism/Metabolism/A to B/Inhibition" with a confidence of 0.85), and each step has verifiable references.

Open-Source and Reproducibility

• The code is open-sourced under the MIT License, and the model weights and datasets are licensed under CC BY-NC 4.0 (for non-commercial research).
• Provide reproducibility guides, configuration files, and scripts to support various deployment scenarios.
• Original DrugBank data needs to be obtained by users with their own authorization; the codebase includes data reconstruction processes and hash verification.

MARD-7B verifies the potential of small specialized models in vertical domains: the 7-billion-parameter model outperforms general-purpose large models and has low deployment costs through carefully designed distillation strategies and structured data. Implications for medical AI: High-quality professional datasets, verifiable reasoning metrics, and targeted distillation technologies are more practically valuable than simply expanding model size, providing a solution that balances performance and efficiency for drug safety monitoring and clinical decision support.