Benchmark dataset for training machine learning models to predict the pathway involvement of metabolites

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Metabolomics

Background:

• Metabolic pathway analysis is crucial for understanding biochemical processes, but many identified metabolites lack defined pathway involvement in public datasets.
• Existing machine learning models for predicting metabolite pathway involvement rely on outdated and flawed datasets, hindering accurate metabolic interpretation.
• The Kyoto Encyclopedia of Genes and Genomes (KEGG) is a key resource, but its metabolite data requires rigorous curation for computational applications.

Approach:

• Developed a new, reproducible benchmark dataset for metabolite pathway prediction using KEGG data, ensuring data integrity and updateability.
• Implemented an atom coloring methodology for feature representation.
• Trained and compared Random Forest, XGBoost, and multilayer perceptron with autoencoder models using the curated dataset.

Key Points:

• The new benchmark dataset addresses limitations of previous KEGG-based datasets, including over 1500 duplicate entries.
• XGBoost binary classification models demonstrated superior performance in predicting metabolite pathway involvement across 11 KEGG pathway categories.
• Achieved a high weighted average F1-score of 0.8180 and Matthews correlation coefficient of 0.7933 across 1000 cross-validation folds.

Conclusions:

• The developed dataset and XGBoost models significantly enhance the accuracy of predicting metabolite pathway involvement.
• This work provides a robust computational framework for advancing metabolic pathway interpretation and discovery.
• Facilitates better understanding of metabolic networks by assigning unknown metabolites to their respective pathways.