ZeroBind: a protein-specific zero-shot predictor with subgraph matching for drug-target interactions

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Current drug-target interaction (DTI) prediction models struggle with generalization to new proteins and drugs.
• Accurate DTI prediction is crucial for efficient drug discovery and development.

Purpose of the Study:

• To develop a protein-specific meta-learning framework, ZeroBind, for robust DTI prediction.
• To improve generalization capabilities for predicting interactions involving novel proteins and drugs.

Main Methods:

• ZeroBind employs a meta-learning approach with protein-specific models trained using graph neural networks (GNNs).
• A subgraph information bottleneck (SIB) module identifies informative protein subgraphs (binding pockets).
• Task adaptive self-attention optimizes the contribution of individual protein models for final predictions.

Main Results:

• ZeroBind demonstrates superior performance in DTI prediction compared to existing methods.
• The framework shows significant improvements in predicting interactions for unseen proteins and drugs.
• ZeroBind performs well with fine-tuning on proteins or drugs with limited known binding partners.

Conclusions:

• ZeroBind offers a powerful and generalizable approach for DTI prediction.
• The protein-specific meta-learning strategy effectively addresses the limitations of current DTI prediction models.
• ZeroBind has the potential to accelerate drug discovery by accurately identifying novel drug-target relationships.