Predicting Drug-Target Interaction Via Self-Supervised Learning.

Area of Science:

• Computational drug discovery
• Graph representation learning
• Bioinformatics

Background:

• Graph representation learning offers new avenues for drug-target interaction (DTI) prediction.
• Current DTI prediction methods face challenges with manual label dependency and vulnerability to attacks.
• Self-supervised learning (SSL) effectively utilizes input data for supervision, offering a promising alternative.

Purpose of the Study:

• To propose SupDTI, a novel SSL-enhanced framework for DTI prediction.
• To leverage heterogeneous networks for comprehensive drug and protein information integration.
• To improve the accuracy and robustness of computational DTI prediction.

Main Methods:

• Developed SupDTI, an end-to-end framework integrating localized and globalized graph convolutions.
• Employed a variational autoencoder to constrain node representations with desired statistical properties.
• Implemented a unified SSL strategy combining contrastive and generative learning modules for enhanced node representations.

Main Results:

• SupDTI effectively captures both local and global node information within heterogeneous networks.
• The unified SSL strategy significantly enhances the quality of node representations.
• Experimental results demonstrate superior DTI prediction performance compared to state-of-the-art methods.

Conclusions:

• SupDTI offers a robust and effective approach to computational DTI prediction.
• SSL integration overcomes limitations of manual labels and enhances model resilience.
• The proposed framework advances the field of computational drug discovery and personalized medicine.