Predicting drug-target interactions based on multivariate information fusion and graph contrast learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Traditional drug-target interaction (DTI) prediction methods are slow and expensive.
• Machine learning offers a faster alternative, but data sparsity and insufficient node representations hinder accuracy.
• Existing methods often neglect topological similarities in integrating node information.

Purpose of the Study:

• To develop an advanced model, MGCLDTI, for accurate DTI prediction.
• To address the challenges of data sparsity and inadequate node embeddings in DTI prediction.
• To leverage multi-view information and graph contrastive learning for improved DTI identification.

Main Methods:

• Utilized DeepWalk for global topological representation extraction from a heterogeneous graph (drugs, targets, diseases).
• Implemented a densification strategy to mitigate noise from sparse DTI matrices.
• Applied a graph contrastive learning (GCL) model with node masking to enhance local awareness and optimize embeddings.
• Employed the LightGBM algorithm for final DTI score prediction.

Main Results:

• MGCLDTI demonstrated superior predictive performance compared to state-of-the-art methods.
• Ablation studies confirmed the significant contribution of each model component.
• Case studies validated MGCLDTI's accuracy in identifying potential drug-target interactions.

Conclusions:

• MGCLDTI effectively predicts drug-target interactions by integrating multi-view data and employing advanced graph learning techniques.
• The model successfully addresses data sparsity and enhances node representation learning for improved DTI prediction.
• MGCLDTI shows significant potential for accelerating drug discovery and medical research.