Benchmarking <i>In Silico</i> Metabolite Prediction Tools against Human Radiolabeled ADME Data for Small-Molecule Drugs

Artificial intelligence-based <i>in silico</i> metabolite prediction tools are increasingly used in drug development, but their performance against high-quality human data remains uncertain. We evaluated four open-access models, including SyGMa, GLORYx, BioTransformer 3.0, MetaPredictor, and MetaTrans, using 11 small-molecule drugs with published human radiolabeled ADME data. Predicted metabolites were compared with experimentally identified human metabolites, and model performance was assessed using recall, precision, balanced accuracy, F1, and Jaccard scores. Results indicated a distinct trade-off between coverage and balanced accuracy: SyGMa provided the broadest metabolic coverage but with low precision; GLORYx and BioTransformer achieved a better balance between coverage and relevance, with BioTransformer showing the best overall performance; MetaPredictor produced fewer metabolites, yet yielded competitive scores for several drugs, despite issues with SMILES interpretability. MetaTrans showed moderate overall performance but also generated incorrect metabolite predictions. None of the models captured the mercapturic acid pathway or predicted metabolite abundances. Overall, current artificial intelligence-based tools can partially reproduce human metabolic profiles but remain insufficient to replace experimental studies, serving best as early-stage screening tools. Further advances in molecular representation, pathway completeness, and data quality are needed to improve next-generation metabolite prediction models.