ML-DTI: Mutual Learning Mechanism for Interpretable Drug-Target Interaction Prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) offers potential for identifying drug-target interactions (DTIs) but faces challenges in interpretability.
• Existing DL methods often treat drug and target encoders independently, neglecting their inherent relationships.

Purpose of the Study:

• To develop a DL model for DTI identification that enhances interpretability and addresses the independent encoding of drugs and targets.
• To propose a mutual learning mechanism to bridge the gap between drug and target encoders in DTI prediction.

Main Methods:

• A mutual learning mechanism was integrated between drug and target encoders using multi-head attention and position-aware attention.
• The DTI problem was approached from a global perspective, incorporating mutual learning layers.
• Neural attention mechanisms were employed for enhanced model visualization and analysis.

Main Results:

• The proposed method showed similar performance to baseline models in random split settings.
• Significant improvements in predictive performance were observed in orphan-target and orphan-drug split settings, indicating better generalization.
• The method demonstrated enhanced generalization and interpretation capabilities for DTI modeling.

Conclusions:

• The proposed mutual learning mechanism effectively bridges drug and target encoders, improving DTI prediction.
• The attention-based approach enhances model interpretability, facilitating analysis of drug-target relationships.
• This study advances DL applications in DTI identification, particularly for challenging datasets with orphan entities.