SF-Rx: A Multioutput Deep Neural Network-Based Framework Predicting Drug–Drug Interaction under Realistic Conditions for Safe Prescription

Drug-drug interaction (DDI) can compromise therapeutic efficacy and cause detrimental effects in polypharmacy. Computational prediction of DDI has emerged as an alternative approach to time-consuming clinical experiments for investigating potential drug interactions, yet reliable prediction remains challenging. We present SF-Rx (Safe Prescription), a DDI predictive framework that incorporates structural similarity profiles with pharmacokinetic (PK) and pharmacodynamic (PD) features to predict severity, types, and directionality. Our study employs a scaffold-based cross-validation strategy for paired drugs and enables a realistic evaluation of model performance while quantifying prediction uncertainty. The implementation of federated learning across multiple DDI data sets improves model generalization and overcomes limited chemical diversity in single-source data sets. Our framework provides a promising approach for developing a reliable DDI prediction model under real-world scenarios, potentially improving patient safety in multidrug treatments.