Site of Metabolism Prediction Based on ab initio Derived Atom Representations

Machine learning models for site of metabolism (SoM) prediction offer the ability to identify metabolic soft spots in low‐molecular‐weight drug molecules at low computational cost and enable data‐based reactivity prediction. SoM prediction is an atom classification problem. Successful construction of machine learning models requires atom representations that capture the reactivity‐determining features of a potential reaction site. We have developed a descriptor scheme that characterizes an atom's steric and electronic environment and its relative location in the molecular structure. The partial charge distributions were obtained from fast quantum mechanical calculations. We successfully trained machine learning classifiers on curated cytochrome P450 metabolism data. The models based on the new atom descriptors showed sustained accuracy for retrospective analyses of metabolism optimization campaigns and lead optimization projects from Bayer Pharmaceuticals. The results obtained demonstrate the practicality of quantum‐chemistry‐supported machine learning models for hit‐to‐lead optimization. Metabolism learned from quantum chemistry: Machine learning models were trained to predict sites of metabolism in drug molecules. This approach was successfully evaluated against lead molecules from current drug discovery campaigns in industry.