Abstract
Understanding and predicting the metabolic fate of xenobiotics is essential in early drug discovery stages, as poor ADMET properties are a leading cause of new drug candidates' failure. In silico metabolism modeling offers a way to design safer and more effective compounds. We present Meta(QM), a set of random forest classifiers enhanced with quantum chemical descriptors to predict (i) the occurrence of metabolic reactions (Metaclass(QM)) and (ii) the site of metabolism (Metaspot(QM)). Models were trained on the MetaQSAR database, which contains 3788 expert-curated reactions divided into 3 main categories, 21 classes, and 101 subclasses. The descriptors used to train the models included physicochemical, constitutional, and stereo-electronic features computed at two levels of theory: PM7 (MOPAC 2016) and DFT (B3LYP/6-31G(d)). For Metaclass(QM), the use of DFT descriptors led to improved classification performances by 10% at the class level and 8.6% at the subclass level, compared to PM7 descriptors. In Metaspot(QM), both descriptor sets showed similar performance in SoM prediction across different datasets. DFT descriptors enhance the classification of metabolic reactions, while simpler methods suffice for the prediction of metabolic sites. These findings support the use of quantum descriptors in metabolism modeling workflows, balancing accuracy and computational cost.