Skip to content
Open access

A Linearly Scalable GNN Framework on Drug–Gene Ontology–Disease Tripartite Graphs for Drug–Disease Association Prediction With Positive-Unlabeled Contrastive Learning

2026 · IEEE Access · Vol 14, pp. 167280-167290 · 0 citations · 32 references

Abstract

In drug-disease association prediction, graph neural networks (GNNs) offer a powerful means to maintain linear complexity without compromising high predictive accuracy. Our framework operates on a drug–Gene Ontology (GO)–disease tripartite graph, where GO terms serve as a functional bridge connecting drugs and diseases through shared biological annotations. We extend our preliminary framework by introducing a positive-unlabeled contrastive learning (puCL) objective and a coverage-aware fusion strategy, specifically designed to improve both representation quality and model flexibility. By clustering known associations, the puCL objective refines the latent space geometry and addresses the inherent bias caused by the common assumption that unlabeled pairs are true negatives. To incorporate semantic knowledge, our fusion strategy leverages biomedical language models like PubMedBERT and BioBERT while remaining robust even when textual descriptions are missing for 42% of the compounds. We validated this framework on datasets spanning six years of biomedical progress; notably, it maintains a 99.0% area under the precision-recall curve (AUPRC) despite a 7.4-fold growth in network edges. Our results highlight a “structure-first” principle where topological features remain the primary predictive driver, as confirmed by stratified analysis showing a robust 98.3% AUPRC for drugs lacking any textual metadata. Computational validation of top-ranked predictions identified supporting evidence across three high-burden disease domains, including an active clinical trial for Ketamine in Parkinson Disease. These findings suggest that the proposed framework is a scalable and adaptable solution for handling the rapid expansion of modern biomedical databases.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.