Graph Attention Fusion With Kolmogorov-Arnold Network for Drug-Gene Interaction Prediction

Area of Science:

• Computational biology
• Pharmacogenomics
• Artificial intelligence in drug discovery

Background:

• Predicting drug-gene interactions is crucial for drug development.
• Understanding the influence of attention mechanisms in deep learning models is essential for complex datasets.
• Existing methods may not fully capture the intricate interplay between drug and gene features.

Purpose of the Study:

• To propose a novel deep learning framework, dgKAN, for accurate drug-gene interaction prediction.
• To develop an interpretable model that analyzes the mutual influence of heterogeneous attention in drug-gene relationships.
• To enhance drug discovery by identifying potential therapeutic targets and drug candidates.

Main Methods:

• Constructing subgraphs using dynamic neighbor selection and attention sampling for drug and gene embedding.
• Utilizing a Transformer and Graph Neural Network (TransGNN) module to fuse global and local attention mechanisms.
• Developing an interpretable Kolmogorov-Arnold Network (KAN) with spline functions to model cross-domain information flow.

Main Results:

• dgKAN significantly outperforms existing baseline methods in drug-gene interaction prediction across various datasets.
• The model successfully captures implicit characteristics by analyzing heterogeneous attention in drug-gene relationships.
• The interpretability of the KAN network allows for a deeper understanding of the model's decision-making process.

Conclusions:

• dgKAN represents a significant advancement in deep learning for drug-gene interaction prediction.
• The approach offers valuable insights into the complex relationships between drugs and genes.
• The predicted interactions hold substantial potential for accelerating drug development and improving disease treatment strategies.