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

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Drug discovery and repurposing rely on accurate drug-gene interaction predictions.
• Limited experimental data hinders the performance of current predictive models.
• Deep learning offers potential but faces challenges with data scarcity.

Purpose of the Study:

• To develop an advanced deep learning model for predicting drug-gene correlations.
• To overcome data scarcity issues in predicting drug-gene interactions.
• To enhance drug discovery and repurposing through improved prediction accuracy.

Main Methods:

• Proposed an attention-guided multi-view contrastive learning (AMCL) method.
• Integrated multi-scale feature learning, graph convolutional networks, and kernel functions.
• Utilized dynamic hypergraph learning and LCA-biased attention mechanisms for prioritized information extraction.
• Employed cross-view contrastive learning to boost embedding discrimination with sparse data.

Main Results:

• AMCL demonstrated superior performance compared to state-of-the-art methods on three benchmark datasets (DGIdb 5.0, ChEMBL, Guide to Pharmacology).
• Ablation studies validated the effectiveness of individual AMCL components.
• Case studies highlighted AMCL's utility in identifying novel drug candidates and repurposing existing drugs.

Conclusions:

• AMCL effectively addresses data scarcity in drug-gene interaction prediction.
• The proposed method significantly advances the capabilities of deep learning in drug discovery and repurposing.
• AMCL shows promise for accelerating the identification of new therapeutic strategies.