Asymmetric Focal Loss Improves Graph Neural Network Prediction of Drug-Drug Interactions

Jul 16, 2026

Graph neural networks can predict polypharmacy side effects, yet standard binary cross-entropy training treats easy and difficult examples similarly, risking the omission of clinically significant drug-drug interactions. This work tests whether an asymmetric focal loss objective improves multi-relational DDI prediction by emphasizing difficult positive interactions. ClinicalFocal loss is integrated into a relation-aware graph convolutional network using molecular fingerprints, physicochemical descriptors, and learned embeddings, then evaluated on TWOSIDES with five-fold cross-validation against a binary cross-entropy baseline.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages14

Reading Time35 min

Highlights

Main findings

Implications

Abstract

Introduction

Asymmetric Focal Loss Improves Graph Neural Network Prediction of Drug-Drug Interactions Faranak Hatami1, Mousa Moradi2, * 1 Department of Chemistry, University of Illinois Chicago, Chicago, IL 60607, USA; [fhatami@uic.edu](fhatami@uic.edu) 2 Harvard Ophthalmology AI Lab, Schepens Eye Research Institute of Massachusetts Eye and Ear, Harvard Medical School, Boston, MA, USA; [mmoradi2@meei.harvard.edu](mmoradi2@meei.harvard.edu) * [Correspondence: mmoradi2@meei.harvard.edu](Correspondence: mmoradi2@meei.harvard.edu) Highlights What are the main findings? • ClinicalFocal loss improved accuracy, F1 score, AUROC, and AUCPR with superior convergence stability compared with binary cross-entropy. • The asymmetric objective achieved 90.9% recall, reduced the false-negative rate from 29.8% to 9.1%, and increased specificity from 69.6% to 87.5% . What are the implications of the main findings? • Emphasizing difficult positive interactions improves DDI prediction without architectural changes. • ClinicalFocal loss enables safety-oriented DDI screening; external validation and calibration assessment are needed. Abstract Background: Graph neural networks improve computational prediction of polypharmacy side effects, but standard binary cross-entropy training allocates equal capacity to wellclassified and difficult examples, potentially missing clinically significant interactions. We evaluated whether an asymmetric focal objective could improve multi-relational drugdrug interaction (DDI) prediction by emphasizing difficult positive interactions. Methods: ClinicalFocal loss was integrated into a relation-aware graph convolutional network using molecular fingerprints, physicochemical descriptors, and learned embeddings. The model was evaluated on TWOSIDES using five-fold cross-validation with identical experimental conditions (architecture, features, data partitions, hyperparameters, and random seeds) for ClinicalFocal loss and binary cross-entropy baseline. Results: ClinicalFocal loss increased accuracy from 0.699 to 0.892 (+19.3 percentage points) and F1 score from 0.700 to 0.894 (+19.4 percentage points) . AUROC increased from 0.766 to 0.914, and AUCPR increased from 0.714 to 0.860. The falsenegative rate decreased from 29.8% to 9.1%, while specificity increased from 69.6% to 87.5% . Overall classification error decreased from 30.1% to 10.8%, corresponding to a 64.1% relative reduction. Improvements were consistent across all five folds. Conclusions: Asymmetric focal optimization improved classification and ranking performance while achieving 90.9% recall for observed interaction triples, without modifying the underlying architecture. Loss-function design is a direct, tunable lever for improving graph-based DDI prediction. Keywords: drug-drug interactions; polypharmacy; graph neural networks; focal loss; patient safety 1. Introduction The simultaneous use of multiple medications (polypharmacy) is increasingly common in managing older adults and patients with multiple chronic conditions [1,2] . Adverse drug events account for approximately 110,000 deaths annually in the United States [3], making them a leading cause of preventable harm, yet the number of possible drug combinations far exceeds what can be prospectively evaluated in clinical trials [4] . Adverse drug events therefore remain an important source of emergency care and hospitalization, particularly among older adults and patients receiving complex medication regimens [1,2,5]. As medication use increases, identifying potentially harmful drug-drug interactions (DDIs) before they produce clinically significant consequences becomes a critical computational and clinical challenge. Because exhaustive experimental assessment of all possible drug combinations is infeasible, computational methods have emerged as a practical strategy for detecting and prioritizing potential DDIs [6] . Graph-based machine-learning methods are especially suitable for this pro

FAQs

What problem does the paper address in drug-drug interaction prediction?

Standard binary cross-entropy training does not focus learning on difficult positive interactions, which can reduce the ability to detect clinically significant DDI signals.

How does ClinicalFocal loss change the training objective?

ClinicalFocal loss is an asymmetric focal objective that emphasizes difficult positive interactions, improving both classification and ranking without changing the underlying architecture.

What performance improvements are reported over the baseline?

ClinicalFocal loss increases accuracy and F1, improves AUROC and AUCPR, reduces the false-negative rate, and raises specificity, with consistent gains across five folds.

Jul 16, 2026

Upload by Evangeline