Machine learning methods for metabolic pathway prediction.

Area of Science:

• Systems Biology
• Bioinformatics
• Computational Biology

Background:

• Reconstructing an organism's metabolic network from its genome is a key challenge in systems biology.
• Predicting metabolic pathways from annotated genomes using reference databases is a common strategy.

Purpose of the Study:

• To quantitatively validate pathway prediction methods.
• To develop and evaluate machine learning (ML) methods for metabolic pathway prediction.

Main Methods:

• Created a "gold standard" dataset of 5,610 known pathway instances for six organisms.
• Defined 123 pathway features and evaluated their information content.
• Applied various ML methods (naïve Bayes, decision trees, logistic regression, feature selection, ensemble methods) and compared them to the PathoLogic algorithm.

Main Results:

• ML-based prediction methods achieved performance comparable to the PathoLogic algorithm (91.2% accuracy, 0.787 F-measure vs. 91% accuracy, 0.786 F-measure).
• ML methods provide probability outputs for predicted pathways, offering more user information and facilitating filtering.
• The performance of ML methods was found to be comparable to existing algorithms.

Conclusions:

• ML methods for pathway prediction demonstrate performance on par with existing algorithms.
• ML methods offer advantages in extensibility, tunability, and explainability.
• Pathway prediction accuracy is largely constrained by the accurate matching of enzymes to reactions based on genome annotations.