DynHeter-DTA: Dynamic Heterogeneous Graph Representation for Drug-Target Binding Affinity Prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Drug-target affinity (DTA) prediction is crucial for drug efficacy and safety assessment.
• Existing deep learning models struggle to fully capture complex drug-receptor interactions.
• Limitations exist in current approaches for modeling intricate drug-drug, protein-protein, and drug-protein relationships.

Purpose of the Study:

• To propose DynHeter-DTA, a dynamic heterogeneous graph model for enhanced DTA prediction.
• To leverage complex interaction networks for adaptive learning of optimal graph structures.
• To improve the expressive power and generalization performance in DTA forecasting.

Main Methods:

• Dynamic heterogeneous graph construction with adaptive connection strengths for drug-drug, protein-protein, and drug-protein pairs.
• Utilizing Graph Isomorphism Networks (GIN) and Self-Attention Graph Pooling (SAGPooling) to handle imbalanced node quantities (protein vs. drug).
• Implementing a two-layer approach for data processing and model design to optimize graph structure and prediction efficiency.

Main Results:

• DynHeter-DTA demonstrates superior performance on public datasets (Davis, KIBA, Human) compared to existing models.
• The dynamic heterogeneous graph structure significantly enhances model expressiveness and generalization.
• The proposed GIN and SAGPooling approach improves prediction efficiency and accuracy.

Conclusions:

• DynHeter-DTA offers an innovative solution for accurate drug-target affinity prediction.
• The model effectively captures complex interactions by adaptively learning graph structures.
• This approach advances computational methods in drug discovery and development.