A two-stage framework with contrastive representation learning for drug-disease association prediction under positive-unlabeled learning

Predicting drug-disease associations (DDA) is severely hindered by extreme data sparsity and the positive-unlabeled (PU) nature of biomedical datasets, where unobserved pairs are often misidentified as true negatives. To prevent negative label contamination and information leakage commonly observed in end-to-end models, we propose a two-stage framework that decouples representation learning from link verification. The first stage performs offline contrastive representation learning on a heterogeneous graph via a dual-view graph attention mechanism and node-level InfoNCE loss, learning discriminative embeddings without adjacency reconstruction or premature negative assumptions. The second stage applies subgraph embedding-based link prediction (SEAL) verification by extracting k-hop enclosing subgraphs with target-edge removal. By combining a negative sampling strategy under the PU setting that integrates easy and hard negatives, the framework enables robust training under open-world settings. Experimental results under ten-fold cross-validation demonstrate that the proposed method achieves an AUC of 0.9406 and an AUPR of 0.9461 on the benchmark Cdataset, outperforming several competitive baselines and confirming its effectiveness under extreme sparsity.