Drug-target interaction prediction based on graph convolutional autoencoder with dynamic weighting residual GCN

Area of Science:

• Computational biology
• Bioinformatics
• Network science

Background:

• Drug-target interaction (DTI) prediction is crucial for drug discovery and repurposing.
• Network-based methods, particularly Graph Convolutional Networks (GCNs), are effective for DTI prediction.
• Existing shallow GCNs struggle to extract higher-level semantic information and lack effective training guidance.

Purpose of the Study:

• To propose a novel graph convolutional autoencoder model, DDGAE, for enhanced DTI prediction.
• To improve the representation capabilities of models for heterogeneous DTI networks.
• To enhance the learning efficiency, performance, and stability of DTI prediction models.

Main Methods:

• Developed a Dynamic Weighting Residual Graph Convolutional Network (DWR-GCN) module for improved representation.
• Implemented a dual self-supervised joint training mechanism to boost learning efficiency.
• Integrated DWR-GCN with a graph convolutional autoencoder within the DDGAE framework.

Main Results:

• The proposed DDGAE model demonstrates superior performance in DTI prediction.
• The DWR-GCN module effectively enhances the representation capability for heterogeneous DTI networks.
• The dual self-supervised training mechanism improves overall model learning performance and stability.

Conclusions:

• DDGAE significantly outperforms state-of-the-art (SOTA) models in DTI prediction tasks.
• The proposed method achieves optimal performance and demonstrates reliability through case studies.
• DDGAE offers a robust and effective approach for advancing DTI prediction.