Structure-based metabolite function prediction using graph neural networks.

Area of Science:

• Metabolomics
• Cheminformatics
• Computational Biology

Background:

• Predicting metabolite functions from their structures is essential for various scientific fields, including systems biology, environmental monitoring, and drug discovery.
• Current machine learning models often struggle to predict multiple metabolite functions simultaneously, limiting their broad applicability.

Purpose of the Study:

• To assess the feasibility of broadly predicting metabolite functions, encompassing location, role, process, and physiological effect.
• To evaluate and compare different machine learning architectures for predicting functional ontology terms of metabolites.

Main Methods:

• Utilized the Human Metabolome Database for extensive functional annotations.
• Evaluated three graph neural network (GNN) architectures against two multilayer perceptron (MLP) models.
• Employed circular fingerprints and Chemical BiDirectional Encoder Representations from Transformers (ChemBERTa) embeddings for feature representation.

Main Results:

• The graph attention network (GAT) model, enhanced with ChemBERTa embeddings, demonstrated superior performance in predicting the processes metabolites are involved in.
• Achieved a high macro F1-score of 0.903 and an area under the precision-recall curve of 0.926 for predicting metabolite functions.
• Identified function-associated structural patterns within metabolite families, indicating potential for interpretable predictions.

Conclusions:

• The developed GAT model offers a significant advancement in predicting multiple metabolite functions from structural information.
• This approach holds promise for enhancing interpretability in metabolite function prediction, aiding drug discovery and systems biology.
• The findings underscore the potential of advanced machine learning techniques for comprehensive metabolomic analysis.