Prediction of multi-relational drug-gene interaction via Dynamic hyperGraph Contrastive Learning

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Drug-gene interaction prediction is vital for drug discovery tasks like repurposing and lead discovery.
• Existing methods often focus on binding interactions and struggle with sparse, noisy data common in drug discovery.
• Graph neural networks show promise but degrade in performance with insufficient training data.

Purpose of the Study:

• To propose a novel Dynamic hyperGraph Contrastive Learning (DGCL) framework to address data sparsity and noise in drug-gene interaction prediction.
• To effectively model both local and global relationships between drugs and genes for improved prediction accuracy.

Main Methods:

• Utilized graph convolutions to extract explicit local relationships between drugs and genes.
• Employed dynamic hypergraph structure learning and hypergraph message passing for global information aggregation.
• Integrated a self-augmented contrastive learning component to refine hypergraph structure and enhance drug/gene representations.

Main Results:

• DGCL outperformed eight state-of-the-art methods across three datasets.
• Achieved a significant 7.6% performance improvement on the DGIdb dataset.
• Demonstrated robustness in alleviating issues related to data sparsity and over-smoothing.

Conclusions:

• The proposed DGCL framework effectively captures complex drug-gene relationships, overcoming limitations of existing methods.
• DGCL offers a robust solution for drug-gene interaction prediction in practical, data-scarce scenarios.
• This approach advances computational drug discovery by enabling more accurate predictions from limited interaction data.