Investigating ADR mechanisms with knowledge graph mining and explainable AI

Knowledge graphs can help predict drug safety risks: The study successfully used open-access knowledge graphs to identify molecular features that distinguish drugs likely to cause adverse reactions (specifically DILI and SCAR) from those that don't, suggesting this approach could complement traditional clinical trials in drug safety assessment.

Explainability matters for drug safety decisions: Rather than using "black box" AI models, the researchers deliberately chose interpretable methods like Decision Trees and Classification Rules because experts need to understand why a drug is flagged as risky—not just that it is—to validate findings and potentially discover new biological mechanisms.

Expert validation is essential for AI-generated insights: While the AI models achieved good accuracy, experts only agreed that 73% of predictions for DILI (drug-induced liver injury) and 38% for SCAR (severe skin reactions) were actually explanatory for the adverse reactions, highlighting that AI-generated features require human expert review before being considered scientifically meaningful.

AI could accelerate understanding of drug mechanisms: By automatically identifying molecular patterns associated with known adverse reactions, this approach offers a pathway to explain how drugs cause harm in cases where the mechanism is currently unknown—potentially improving drug development and safety monitoring.

Responsible AI requires transparency in healthcare applications: This work demonstrates that for high-stakes decisions like drug safety, AI systems should be designed to show their reasoning through human-readable models rather than optimizing purely for prediction accuracy.