GraphsformerCPI: Graph Transformer for Compound-Protein Interaction Prediction.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Predicting compound-protein interactions (CPI) is crucial for drug design.
• Growing data necessitates efficient and interpretable prediction models.
• Existing deep learning methods often lack transparency.

Purpose of the Study:

• To introduce GraphsformerCPI, an end-to-end deep learning framework for improved CPI prediction.
• To enhance model interpretability in compound-protein interaction prediction.
• To leverage spatial structures and attention mechanisms for deep molecular representations.

Main Methods:

• GraphsformerCPI treats compounds and proteins as structured node sequences.
• Utilizes structure-enhanced self-attention for integrating molecular features.
• Employs a dual-attention mechanism for extracting atom-residue relational features.
• Extends Transformer capabilities to spatial structures for enhanced learning.

Main Results:

• GraphsformerCPI outperforms baseline models on classification CPI datasets.
• Achieves competitive performance on regression CPI datasets.
• Demonstrates significant improvements in AUC, precision, and recall on benchmark datasets.
• Shows notable gains in Concordance index (CI) and mean squared error (MSE) on the KIBA dataset.
• Molecular docking reveals insights into binding mechanisms and interactions.

Conclusions:

• GraphsformerCPI offers superior performance and interpretability in CPI prediction.
• The framework provides practical significance for drug design and discovery.
• Identifies key molecular components and enhances understanding of binding mechanisms.
• Advances the field of interpretable deep learning for biological applications.