Poster: Ensemble Methods for ADR Prediction

Danika Gupta, Bharath Ramsundar

Abstract: Predicting Adverse Drug Reactions (ADRs) is crucial for preventing health damage and reducing costs, currently exceeding $30 billion annually in the US. ADRs are a leading cause of death, ranking fourth nationally and sixth globally pre-pandemic. Despite rigorous testing, ADR-related deaths continue to rise. Our study assesses an ADR prediction ensemble pipeline combining drug chemistry, side effect interrelationships, targets, and indications. Results show that an ensemble of multi-task classification neural networks and knowledge graph embeddings predicts ADRs across all 27 MedDRA drug reaction groups, reducing false negatives by 2.4x. A case study of 10 diverse drugs from the holdout set revealed that our model forecasts over 88% of all FDA Adverse Drug Events Reporting System (FAERS) incidences by reaction group since 1968 for all 10 drugs, with a success rate over 95% for 7. Our ensemble approach also enables new information to be added throughout the drug lifecycle via additions to the knowledge graph, creating the potential for better ADR predictions over time.