Drug-drug interactions (DDIs) remain a major source of preventable harm, and many clinically important mechanisms are still unknown. Existing models either rely on pharmacologic knowledge graphs (KGs), which fail on unseen drugs, or on electronic health records (EHRs), which are noisy, temporal, and site-dependent. We introduce, to our knowledge, the first system that conditions KG relation scoring on patient-level EHR context and distills that reasoning into an EHR-only model for zero-shot inference. A fusion "Teacher" learns mechanism-specific relations for drug pairs represented in both sources, while a distilled "Student" generalizes to new or rarely used drugs without KG access at inference. Both operate under a shared ontology (set) of pharmacologic mechanisms (drug relations) to produce interpretable, auditable alerts rather than opaque risk scores. Trained on a multi-institution EHR corpus paired with a curated DrugBank DDI graph, and evaluated using a a clinically aligned, decision-focused protocol with leakage-safe negatives that avoid artificially easy pairs, the system maintains precision across multi-institutuion test data, produces mechanism-specific, clinically consistent predictions, reduces false alerts (higher precision) at comparable overall detection performance (F1), and misses fewer true interactions compared to prior methods. Case studies further show zero-shot identification of clinically recognized CYP-mediated and pharmacodynamic mechanisms for drugs absent from the KG, supporting real-world use in clinical decision support and pharmacovigilance.

The method needs to search through paths for long-term information, it's like to find conflict facts How to deal with the conflict is not mentioned. In Figure 1, we have "search with only macro stage" ( Random, Reinforce, Bayes) and "with only We will elaborate more on this in Sec. Some recent work on graph alignment were not included in the comparison. The code of VR-GNN [Y e et al. 2019] is not publicly available. The search cost still takes tens of hours.

Scoring functions, which measure the plausibility of triples, have become the crux of knowledge graph embedding (KGE). Plenty of scoring functions, targeting at capturing different kinds of relations in KGs, have been designed by experts in recent years. However, as relations can exhibit intricate patterns that are hard to infer before training, none of them can consistently perform the best on existing benchmark tasks. AutoSF has shown the significance of using automated machine learning (AutoML) to design KG- dependent scoring functions. In this paper, we propose AutoSF+ as an extension of AutoSF. First, we improve the search algorithm with the evolutionary search, which can better explore the search space. Second, we evaluate AutoSF+ on the recently developed benchmark OGB. Besides, we apply AutoSF+ to the new task, i.e., entity classification, to show that it can improve the task beyond KG completion.