Structure-based metabolite function prediction using graph neural networks.

Motivation: Being able to broadly predict the function of novel metabolites based on their structures has applications in systems biology, environmental monitoring, and drug discovery. To date, machine learning models aiming to predict functional characteristics of metabolites have largely been limited in scope to predicting single functions, or only a small number of functions simultaneously. Results: Using the Human Metabolome Database as a source for a wider range of functional annotations, we assess the feasibility of predicting metabolite functions more broadly, as defined by four elements, namely location, role, the process it is involved in, and its physiological effect. We evaluated three graph neural network architectures to predict available functional ontology terms. We compared the graph models with two multilayer perceptron architectures using circular fingerprints and Chemical BiDirectional Encoder Representations from Transformers (ChemBERTa) embeddings. Among the models tested, the graph attention network, incorporating embeddings from the pretrained ChemBERTa model to predict the process metabolites are involved in, achieved the highest performance with a macro F1-score of 0.903 and an area under the precision-recall curve of 0.926. Availability and implementation: The model identified function-associated structural patterns within metabolite families, demonstrating the potential for interpretable prediction of metabolite functions from structural information.