AI-Based Prediction of Drug-Drug Interactions

Abstract

Drug-drug interactions (DDIs) -- pharmacological events arising when two or more drugs are co-administered and the pharmacokinetic or pharmacodynamic effect of one drug is altered by the presence of another -- represent a major preventable cause of adverse drug reactions, contributing to an estimated 125,000 deaths annually in the United States and accounting for 7% of all hospital admissions attributable to medication errors in Europe. The combinatorial scale of the DDI problem makes experimental characterisation impractical: for a formulary of 2,000 drugs, there are approximately 2 million possible pairwise combinations to test, of which only a small fraction have been systematically evaluated in clinical pharmacokinetic interaction studies. Artificial intelligence approaches -- graph neural networks operating on drug-target interaction networks, transformer models encoding chemical structure, knowledge graph embeddings, and ensemble methods combining multiple evidence streams -- offer scalable computational DDI prediction that can prioritise the highest-risk combinations for experimental validation and provide real-time clinical decision support at prescribing. This benchmarking study evaluates 12 AI-based DDI prediction methods against a gold-standard validation set of 247 confirmed DDIs from DrugBank; KEGG Drug; and published clinical pharmacokinetic studies, assessing AUC-ROC, precision-recall, DDI type classification accuracy, and prospective clinical alert actionability. Graph neural network approaches (MHCADDI; SSI-DDI; MIRACLE) achieve the highest DDI prediction AUC (0.91 +- 0.04). A DDI-AI Clinical Utility Score (DAICUS) integrating predictive accuracy, DDI type specificity, false positive burden, and EHR integration readiness is developed. The DAICUS-optimal ensemble achieves AUC 0.94, F1 = 0.87, and reduces clinically actionable alert false positives by 41.4% vs. rule-based systems -- a meaningful reduction in alert fatigue with maintained safety.