Drug–Drug Interaction Prediction: Aligning Evidence, Models, Validation, and Clinical Use

Abstract

Drug–drug interactions (DDIs) are a major source of preventable medication-related harm, but the evidence used to anticipate them spans molecular structure, pharmacology, biomedical text, knowledge graphs, and longitudinal clinical records. Computational studies often treat distinct targets as a single benchmark, even though binary interaction screening, event prediction, relation extraction, and patient-specific risk require different evidence and validation. This critical review introduces a claim-centered framework that aligns the clinical question, evidence source, prediction unit, model inductive bias, validation design, and intended use. Representative methods are compared by what they encode, the assumptions they impose, and their characteristic failure modes rather than by incomparable leaderboard scores. The analysis covers similarity and factorization methods, molecular and relational graph learning, pair-conditioned substructure models, contrastive and pretrained learning, multimodal fusion, and language-based systems. A validation-to-translation ladder then connects leakage-aware splitting, calibration, uncertainty, external testing, mechanistic corroboration, and prospective workflow evaluation. The synthesis indicates that architecture-level gains are conditional on label construction and distribution shift. It also shows that useful explanations require evidence alignment and actionability in addition to visual plausibility. This framework clarifies which claims current DDI benchmarks can support and what additional evidence is needed before predictions can guide prescribing or monitoring.