XenoSoM: A deep learning framework for site of metabolism prediction of xenobiotics

Metabolic biotransformations significantly influence drug efficacy and safety, making early metabolism assessment crucial in drug discovery. While in vivo and in vitro methods remain the gold standard for investigating metabolic properties, they are often costly and time-consuming. Consequently, AI-driven models have emerged as useful tools to predict metabolic reactions, sites of metabolism (SoM), and resulting metabolites. Predicting the SoM is particularly valuable for identifying reactive atoms and anticipating compounds' metabolic profile. In this study, six deep learning architectures (GCN, GIN, GATC, GINE, GATv2 and AttentiveFP) were trained on a curated dataset containing more than 30'000 substrates with annotated SoMs, thus addressing the long-standing scarcity of large, high-quality datasets in computational metabolism prediction. The final selected model, named XenoSoM, employs an AttentiveFP architecture with single-task (ST) and multi-task (MT) variants, both delivering the best performance among the evaluated models, achieving MCC values up to 0.848 and Top-2 accuracies up to 1.00. XenoSoM_ST and XenoSoM_MT were compared against existing tools such as FAME3R and GNN-SOM. The comparison showed that our architectures perform better than FAME3R and GNN-SOM, representing a notable advancement in in silico drug metabolism modeling for both phase I and phase II reactions. Despite the inclusion of proprietary data sources and the lack of enzyme-specific information, XenoSoM represents an important step toward the development of global models for metabolism prediction, leveraging large-scale data integration to improve the identification of metabolic hotspots.