Asymmetric Focal Loss Improves Graph Neural Network Prediction of Drug-Pair-Associated Polypharmacy Side Effects

Abstract

Accurate prediction of drug-pair-associated polypharmacy side effects remains an important challenge in computational pharmacovigilance. Standard binary cross-entropy (BCE) may provide insufficient emphasis on difficult positive examples. We evaluated ClinicalFocal, an asymmetric focal-loss function that assigns different focusing exponents and class weights to observed positive and generated negative drug-side-effect-drug triples, within a relation-aware graph convolutional network. On TWOSIDES, five-fold cross-validation showed that ClinicalFocal increased accuracy from 0.699 to 0.892 (+19.3 percentage points), F1 score from 0.700 to 0.894 (+19.4 percentage points), AUROC from 0.766 to 0.914, and AUCPR from 0.714 to 0.860. Sensitivity increased from 0.702 to 0.909, while overall classification error decreased from 30.1% to 10.8%, a 64.1% relative reduction. Brier score decreased from 0.183 to 0.147, a 19.6% relative reduction, indicating improved probability-level predictive accuracy. After independent 100-trial loss-specific optimization, ClinicalFocal remained superior across all five folds. These findings support loss-function design as an important determinant of graph-based polypharmacy side-effect prediction.