ZeroBind

ZeroBind predicts drug-target interactions using a protein-specific meta-learning framework that leverages graph neural networks, subgraph matching, and a Subgraph Information Bottleneck (SIB) to identify binding-relevant substructures in protein and molecular graphs.

Key Features:

• Meta-Learning Framework: Trains protein-specific models as individual tasks to improve generalization to unseen proteins and drugs.
• Graph Neural Networks (GNNs): Learns embeddings from protein graphs and molecular graphs to capture structural information for interaction prediction.
• Subgraph Information Bottleneck (SIB) Module: Uses a weakly supervised SIB module to identify and compress maximally informative subgraphs in protein graphs, highlighting potential binding pockets.
• Task Adaptive Self-Attention Module: Applies a task-adaptive self-attention mechanism to learn the relative importance of protein-specific tasks and weight their contributions to final predictions.
• Subgraph Matching: Employs subgraph matching techniques to compare and align binding-relevant substructures between proteins and molecules.

Scientific Applications:

• Drug-target interaction prediction: Predicts interactions for novel or underrepresented proteins and drugs using structural and task-specific information.
• Binding site identification: Identifies potential binding pockets and binding-relevant substructures within protein structures.
• Drug discovery support: Prioritizes candidate drug-target pairs and potential therapeutic targets based on structural evidence.
• Generalization to unseen interactions: Enables prediction of interactions involving unseen proteins and drugs through meta-learning and task-adaptive weighting.

Methodology:

Meta-training protein-specific models using GNNs to generate embeddings from structural data; SIB module identifies and compresses informative subgraphs representing potential binding pockets; task-adaptive self-attention learns task importance; subgraph matching compares binding-relevant substructures between proteins and molecules.