A deep learning architecture for metabolic pathway prediction.

Area of Science:

• Computational biology
• Cheminformatics
• Machine learning

Background:

• Predicting metabolic pathways is crucial for designing new molecules and understanding cellular processes.
• Existing methods often require extensive feature engineering or struggle with complex molecular representations.

Purpose of the Study:

• To develop an automated machine learning framework for predicting metabolic pathway classes of biochemical compounds.
• To improve the accuracy and efficiency of metabolic pathway prediction compared to existing methods.

Main Methods:

• A hybrid approach combining graph convolutional networks (GCNs) for molecular shape feature extraction from SMILES representations.
• Utilizing a random forest classifier to predict metabolic pathway classes based on extracted features.

Main Results:

• Achieved 95.16% accuracy in predicting single metabolic pathway classes, significantly outperforming competing methods (84.92% or less).
• Demonstrated high accuracy (97.61%) for multi-label classification of compounds with mixed pathway memberships.
• Showcased the ability to predict global physicochemical features from extracted molecular shape features.

Conclusions:

• The proposed hybrid machine learning framework effectively predicts metabolic pathway classes by automatically extracting molecular shape features.
• This approach offers a significant advancement in the accuracy and automation of metabolic pathway prediction, with potential applications in drug discovery and synthetic biology.