Search papers, labs, and topics across Lattice.
This paper introduces MARD, a novel pipeline for mechanism-level drug-drug interaction (DDI) prediction that employs a structured taxonomy and innovative training techniques to enhance accuracy in identifying specific pharmacological interactions. By implementing a 7B reasoning model that utilizes a single-token KL divergence and leakage-safe retrieval channels, the authors achieve a significant accuracy improvement over existing systems, particularly in handling novel drug pairs. The results indicate that MARD not only surpasses the best baseline by 13.9 percentage points but also demonstrates an anti-memorisation effect, suggesting that the model's success is rooted in structured reasoning rather than mere memorisation of frequently encountered drug pairs.
MARD's accuracy improves on novel drug pairs by 13.9 percentage points, showcasing the power of structured pharmacological reasoning over memorisation.
Mechanism-level drug-drug interaction (DDI) prediction requires identifying which enzyme or pharmacodynamic axis is implicated, in which direction, and with which evidence -- not merely whether two drugs interact. We introduce a reproducible mechanism-level DDI labelling and evaluation protocol with a structured 7-family/147-subtype taxonomy, leakage-safe cold-split protocols, and auditable reasoning metrics for evaluating pharmacological prediction beyond flat interaction classification. We propose a pipeline that produces a 7B reasoning MARD (Mirror-Augmented Reasoning Distillation), combining three training innovations: a single-token KL divergence on direction tag that ties the model's prediction, per-loss PRM-weighted DPO with programmatic hard negatives, and a leakage-safe mechanism-aware retrieval channel. Process-reward step labels are automatically verifiable against DrugBank-structured fields, requiring no human or LLM judges. On the April-2026 DrugBank release, our MARD-7B is the only system in a 32-system comparison whose accuracy survives drug-pair novelty, beating the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% of frontier API cost. Further analysis reveals an anti-memorisation signature where accuracy improves on rarely seen drugs, suggesting that gain comes from structured pharmacological reasoning rather than drug-frequency memorisation. We release corpus, DDI-PRM, retrieval index, and training code.
