A Novel Multi-Scale Graph Neural Network for Metabolic Pathway Prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Metabolic pathway prediction is vital for drug development.
• Accurate classification of compound metabolism aids in understanding drug efficacy and toxicity.
• Existing methods face challenges in capturing complex molecular structures and interactions.

Purpose of the Study:

• To introduce a novel Multi-Scale Graph Neural Network (MSGNN) framework for predicting metabolic pathway classes of compounds.
• To enhance the accuracy and efficiency of metabolic pathway prediction in drug research.
• To provide a robust computational tool for analyzing compound metabolism.

Main Methods:

• Developed a Multi-Scale Graph Neural Network (MSGNN) framework incorporating a subgraph encoder, feature encoder, and global feature processor.
• Implemented a graph augmentation strategy to expand the training dataset.
• Extracted local structural features and atomic characteristics using specialized encoders.
• Integrated pre-trained model information and molecular fingerprints in the global feature processor.

Main Results:

• MSGNN achieved high performance metrics: 98.17% accuracy, 94.18% precision, 94.43% recall, and 94.30% F1-score.
• The proposed MSGNN framework demonstrated superior performance compared to existing similar models.
• Ablation experiments confirmed the essential contribution of each MSGNN module to its overall effectiveness.

Conclusions:

• MSGNN is a highly effective framework for predicting metabolic pathway classes of compounds.
• The model's architecture and strategies contribute to its superior predictive power in drug research.
• MSGNN offers a significant advancement in computational approaches for metabolic pathway analysis.