Metabolite-GEN: integrated prediction of sites of metabolism, reaction class, and metabolite structures using graph edit networks

Introduction Predicting the metabolic fate of drug candidates remains a central challenge in pharmaceutical research, with direct implications for efficacy, safety, and pharmacokinetic optimization. Methods We present Metabolite‐GEN, an integrated graph neural network framework that predicts SOM and cytochrome P450–mediated reaction class from molecular graph representations, then uses these predictions to enumerate and rerank feasible metabolite structures. The framework was evaluated by five-fold scaffold-stratified cross‐validation on 578 parent–metabolite pairs spanning 12 CYP-mediated reaction classes from the MetXBioDB database, followed by a single evaluation on a scaffold‐disjoint held-out test set of 144 molecules. Results In cross‐validation the model achieved mean SOM atom‐ranking top‐1/top‐3/top-5 accuracies of 56.9% ± 6.7%, 78.6% ± 3.2% and 87.5% ± 1.8% (MRR 0.698), reaction class accuracy 37.2% ± 11.4% (weighted F1 0.365), and metabolite structure top‐1/top‐3/top‐5 accuracies of 56.6% ± 4.1%, 73.4% ± 2.4% and 79.1% ± 3.7% (MRR 0.653). On the test set, SOM accuracies were 54.9%, 79.2% and 86.8% (MRR 0.688), reaction class accuracy 31.9% (weighted F1 0.323, macro F1 0.271), and metabolite structure accuracies 53.5%, 72.2% and 75.0% (MRR 0.626). Discussion The framework returns explicit metabolite structures rather than atom‐level scores, and the candidate pool for a typical compound is short enough to inspect directly, so the output can be acted on without further filtering—prioritizing compounds for metabolic‐stability screening, flagging reactive metabolite liabilities, and focusing experimental metabolite identification.