In silico prediction of metabolic stability for ester-containing molecules: Machine learning and quantum mechanical methods

Carboxylic ester is an important functional group frequently used in the design of pro-drugs and soft-drugs. It is critical to understand the structure-metabolic stability relationships of these types of drugs. This work aims to predict the metabolic stability of ester-containing molecules in human plasma/blood by both machine learning and quantum mechanical methods. A dataset comprising metabolic half-lives with 656 molecules was collected for machine learning models. Three molecular representations (extended-connectivity fingerprint, Chemopy descriptor and Mordred3D descriptor) were used in combination with four machine learning algorithms (LightGBM, support vector machine, random forest, and k-nearest neighborhood). Furthermore, ensemble learning was applied to integrate the predictions of the individual models to achieve improved prediction results. The consensus model reached coefficient of determination values of 0.793 on the test set and 0.695 on the external validation set, respectively. Feature importances of machine learning models were interpreted from SHapley Additive exPlanations, which were consistent with previous esterase-catalyzed hydrolysis reaction mechanism. Moreover, a quantum mechanical model was built to calculate the energy gap of esterase-catalyzed hydrolysis reaction, deriving metabolic stability ranks. Abilities of quantum mechanical model to discriminate relative metabolic stability for molecules in external validation set was compared with machine learning model. Advantages and disadvantages of machine learning and quantum mechanical methods in metabolic stability prediction were discussed. In summary, this work can serve as an in silico high throughput screening tool to accelerate the early development process of pro-drugs and soft-drugs.