ClinicalFocal loss improves GNN drug-drug interaction prediction accuracy by 19.3 percentage points

Researchers integrated ClinicalFocal loss into a relation-aware graph convolutional network to enhance the prediction of polypharmacy side effects, addressing the limitation of standard binary cross-entropy which treats all examples equally.

• Accuracy increased from 0.699 to 0.892 and F1 score from 0.700 to 0.894 on the TWOSIDES dataset.
• AUROC rose from 0.766 to 0.914, while AUCPR improved from 0.714 to 0.860.
• The false-negative rate dropped significantly from 29.8% to 9.1%, and specificity increased from 69.6% to 87.5%.
• Overall classification error decreased by 64.1% relative, achieving 90.9% recall for observed interaction triples.

The study concludes that asymmetric focal optimization serves as a direct, tunable lever for improving graph-based drug-drug interaction prediction without requiring modifications to the underlying model architecture.