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

Area of Science:

• Pharmacology
• Computational Chemistry
• Drug Development

Background:

• Artificial intelligence (AI)-driven *in silico* metabolite prediction is vital in modern drug development.
• Assessing the accuracy of these AI tools against high-quality human data is crucial but remains challenging.

Purpose of the Study:

• To evaluate the performance of four open-access AI metabolite prediction models (SyGMa, GLORYx, BioTransformer 3.0, MetaPredictor, MetaTrans).
• To compare AI-predicted metabolites against experimentally identified human metabolites using established performance metrics.

Main Methods:

• Utilized published human radiolabeled Absorption, Distribution, Metabolism, and Excretion (ADME) data for 11 small-molecule drugs.
• Assessed model performance using recall, precision, balanced accuracy, F1, and Jaccard scores.
• Compared predicted metabolites with experimentally identified human metabolites.

Main Results:

• A trade-off exists between metabolic coverage and prediction accuracy.
• BioTransformer 3.0 demonstrated the best overall performance, balancing coverage and relevance.
• No models predicted metabolite abundances or captured the mercapturic acid pathway.

Conclusions:

• Current AI metabolite prediction tools offer partial insights and are best suited for early-stage drug development screening.
• Further advancements in AI, including molecular representation and pathway completeness, are necessary for improved accuracy.