Druggability Assessment in TRAPP using Machine Learning Approaches

Abstract

Accurate protein druggability predictions are important for the selection of drug targets in the early stages of drug discovery. Due to the flexible nature of proteins, the druggability of a binding pocket may vary due to conformational changes. We have therefore developed two statistical models, a logistic regression model (TRAPP-LR) and a convolutional neural network model (TRAPP-CNN), for predicting druggability and how it varies with changes in the spatial and physicochemical properties of a binding pocket. These models are integrated into TRAPP (TRAnsient Pockets in Proteins), a tool for the analysis of binding pocket variations along a protein motion trajectory. The models, which were trained on publicly available and self-augmented data sets, show equivalent or superior performance to existing methods on test sets of protein crystal structures, and have sufficient sensitivity to identify potentially druggable protein conformations in trajectories from molecular dynamics simulations. Visualization of the evidence for the decisions of the models in TRAPP facilitates identification of the factors affecting the druggability of protein binding pockets.