Knowledge of the dominant tautomer for a drug molecule is a critical component of drug discovery. The tautomeric state of a molecule affects its solubility, drug binding, permeability and virtually all other ADMET and physicochemical properties. As such, a method to predict the preferred tautomer would be a valuable tool in drug discovery and various cheminformatics applications. Consequently, an artificial neural network ensemble (ANNE) model has been developed to score and rank multiple candidate tautomers of a molecule to identify the preferred tautomer in aqueous solution. The model has been trained on ~1500 molecules spanning ~7000 tautomers. The training data was obtained from literature compilations (e.g., Tautobase<sup>1</sup>) as well as additional data collected directly from the literature. The training algorithm used a special version of our ADMET Modeler™ program with a softmax activation function along with a cross-entropy loss function for correctly classifying the dominant tautomer among all the candidates for a given molecule. The model achieves high accuracy predicting the preferred tautomer with better than 80% accuracy for both training and test sets with most misses being “close calls”. Details of the training algorithm along with interesting examples of both correct and incorrect predictions will be provided.

Building a machine learning model for tautomer prediction

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.