A novel hybrid framework for metabolic pathways prediction based on the graph attention network

Area of Science:

• Pharmacology and Cheminformatics
• Computational Drug Discovery
• Machine Learning in Chemistry

Background:

• Understanding drug metabolic pathways is crucial for predicting absorption, distribution, metabolism, and excretion (ADME).
• A compound's structure and composition dictate its involvement in specific metabolic pathways.
• Accurate metabolic pathway prediction aids in optimizing drug design and development.

Purpose of the Study:

• To develop a novel hybrid framework, HFGAT, for predicting metabolic pathway classes of drug compounds.
• To leverage both global and local compound characteristics for improved prediction accuracy.
• To provide a robust computational tool for drug discovery and development.

Main Methods:

• Developed a hybrid framework (HFGAT) integrating a graph attention network (GAT) with other components.
• Employed a two-branch feature extraction layer to capture global and local compound features.
• Utilized a fully connected layer to integrate features and output predicted metabolic pathway categories.

Main Results:

• HFGAT demonstrated superior performance in multi-class classification compared to six other methods, including GCN.
• Deep learning methods, including HFGAT and GCN, outperformed traditional machine learning approaches.
• HFGAT achieved higher scores on 8 out of 11 pathway categories than the GCN-based method.

Conclusions:

• The HFGAT framework effectively utilizes global and local compound information for accurate metabolic pathway prediction.
• The integration of GAT in HFGAT enhances focus on critical substructures for prediction.
• HFGAT offers a promising computational approach for drug discovery, encompassing diverse metabolic reactions.