Interpretable and Adaptive Graph Contrastive Learning With Information Sharing for Biomedical Link Prediction

The identification of unobserved links in drug-related biomedical networks is essential for various drug discovery applications, which is also beneficial for both disease diagnosis and treatment through exploring the underlying molecular mechanisms. However, existing solutions face significant challenges due to three main limitations, i.e., (1) lack of interpretability to provide comprehensive and reliable insights, (2) insufficient robustness and flexibility in cold-start scenarios, and (3) inadequate interaction and sharing of multi-view information. In light of this, we propose DrugXAS, an interpretable and adaptive cross-view contrastive learning framework with information sharing for biomedical link prediction. Specifically, DrugXAS has three distinctive characteristics for addressing these challenges. To solve the first problem, we propose an attention-aware augmentation scheme to provide understandable explanations of intrinsic mechanisms. To deal with the second challenge, we propose an adaptive graph updater and neighborhood sampler, which select proper neighbors according to the feedbacks from the model to improve aggregation ability. To tackle the third issue, an information sharing module with diffusion loss is proposed to incorporate chemical structures into heterogeneous relational semantics and facilitate the contrast process. Empirically, extensive experiments on seven benchmark datasets involving multi-type tasks demonstrate that DrugXAS outperforms the state-of-the-art methods in terms of precision, robustness, and interpretability.