Attention-Guided Multi-View Contrastive Learning for Predicting Sparse Drug-Gene Associations

Employing deep learning techniques for drug discovery and repurposing necessitates the acceleration of predictions regarding drug-gene interactions. However, the scarcity of experimental support data often constrains the performance and generalization capabilities of existing predictive models. To address this issue, we propose an attention-guided multi-view contrastive learning method (named AMCL) for predicting unidentified drug-gene correlations. Specifically, AMCL integrates multi-scale feature learning and employs a graph convolutional network to extract local topological information. Additionally, it utilizes a kernel function to capture global structural patterns. High-order dependencies are dynamically modeled through the dynamic hypergraph learning module. The model is guided to prioritize the information of densely linked regions in the interactive network and aids in the prediction job by the attention bias mechanism based on the LCA-biased attention. The discriminating capacity of learned embeddings is improved by the cross-view contrastive learning technique, particularly when sparse data is present. The experimental results show that AMCL outperforms other state-of-the-art techniques on the three datasets of DGIdb 5.0, ChEMBL and Guide to Pharmacology. The contributions of each component have been validated through ablation studies, and case studies show AMCL's adaptability in discovering new medications and repositioning existing ones.