MVCL-DTI: Predicting Drug-Target Interactions Using a Multiview Contrastive Learning Model on a Heterogeneous Graph

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Accurate prediction of drug-target interactions (DTIs) is crucial for accelerating drug discovery and repurposing.
• Synthesizing information from varied biological subnetworks presents a significant challenge in DTI prediction.

Purpose of the Study:

• To introduce MVCL-DTI, a novel heterogeneous graph-based model for predicting DTIs.
• To effectively integrate and adaptively weight information from multiple biological data views.

Main Methods:

• MVCL-DTI integrates neighbor, meta-path, and diffusion views to capture semantic features.
• An attention-based contrastive learning approach and a multiview attention-weighted fusion module are employed.
• The model's robustness was tested using benchmark datasets under various conditions, including hard negative sampling and masking known DTIs.

Main Results:

• MVCL-DTI demonstrated strong robust generalization across diverse testing scenarios.
• The model successfully predicted novel DTIs, with specific application to COVID-19-related drugs.
• Feature visualization identified critical elements like Gene Ontology and substituent nodes influencing DTI prediction.

Conclusions:

• MVCL-DTI offers an effective framework for DTI prediction by leveraging heterogeneous graph information.
• The model's robust performance and practical applicability, particularly in antiviral drug research, are highlighted.
• Understanding the contribution of specific features and initialization strategies can further optimize DTI prediction.