CBKG-DTI: Multi-Level Knowledge Distillation and Biomedical Knowledge Graph for Drug-Target Interaction Prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Drug-target interaction (DTI) prediction is crucial for drug discovery.
• Biomedical knowledge graphs offer rich biological system data but require task-specific knowledge extraction.
• Integrating multi-modal information, including structural data, is essential for accurate DTI prediction.

Purpose of the Study:

• To develop a novel framework, CBKG-DTI, for identifying drug-target interactions.
• To distill task-specific knowledge from large-scale biomedical knowledge graphs into a lightweight predictive model.
• To enhance DTI prediction by integrating relational and structural information using a multi-modal approach.

Main Methods:

• Developed a hierarchy-aware knowledge graph embedding (teacher model) to capture semantic hierarchies.
• Constructed a heterogeneous network integrating relational and structural information.
• Employed a heterogeneous graph attention network (student model) with a multi-level distillation mechanism.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.9751.
• Obtained an Area Under the Precision-Recall Curve (AUPR) of 0.6310 under 5-fold cross-validation.
• Demonstrated the superiority of CBKG-DTI in DTI prediction tasks.

Conclusions:

• CBKG-DTI effectively distills task-specific knowledge from complex biomedical knowledge graphs.
• The framework's multi-modal fusion approach significantly improves DTI prediction performance.
• Validated the effectiveness of knowledge distillation in enhancing lightweight DTI prediction models.