DualKG-DC: A Drug-Centric Dual-Layer Knowledge Graph Framework for Drug Combination Prediction

Most existing approaches to drug combination discovery are disease-centered, aiming to identify drug pairs for specific diseases. Complementarily, a drug-centered strategy starts from known drug combinations and explores new therapeutic indications, facilitating translational applications by leveraging combinations with established safety profiles. Here, we introduce DualKG-DC, a drug centered computational framework that provides a complementary perspective by identifying potential disease indications for a given drug combination. The dual layer knowledge graph architecture, which is pretrained on a foundation biomedical knowledge graph and subsequently refined on a task specific drug combination subgraph, may reduce reliance on large, labeled datasets by leveraging existing knowledge on drug targets, biological pathways, and observed phenotypic effects. In systematic benchmarking against three state-of-the-art models, DualKG-DC outperformed all comparison models, achieving an average Hits@10 of 0.48, MRR of 0.30, AUROC of 0.99, and AUPRC of 0.31. Notably, in cold start scenarios, DualKG-DC outperformed baseline methods in predicting indications for unseen drug combinations, achieving superior results with an average Hits@10 score of 0.32, an MRR of 0.18, an AUROC of 0.98, and an AUPRC of 0.23. These results highlight DualKG-DC as an effective platform for systematically discovering therapeutic opportunities of drug combinations. By leveraging a dual-layer architecture, the model enables effective knowledge transfer, enhancing predictive performance and robustness, particularly for previously unseen drug combinations.