Abstract

The Human Pregnane X Receptor (hPXR) is a regulator of drug metabolising enzymes (DME) and efflux transporters (ET). The prediction of hPXR activators and non-activators has pharmaceutical importance to predict the multiple drug resistance (MDR) and drug-drug interactions (DDI). In this study, we developed and validated the computational prediction models to classify hPXR activators and non-activators. We employed four machine learning methods support vector machine (SVM), k-nearest neighbour (k-NN), random forest (RF) and naïve bayesian (NB). These methods were used to develop molecular and fingerprint based descriptors for the prediction of hPXR activators and non-activators. Total 529 molecules consitsting of 317 activators and 212 non-activators were used for model development. The overall prediction accuracy of models was 69% to 99% to classify hPXR activators and nonactivators using RDkit descriptors. In case of 5 and 10-fold cross validation the prediction accuracy for training set is 74% to 82% and 79% to 83% for hPXR activators respectively and 50% to 62% and 49% to 65% non-activators, respectively. The external test prediction is between 59% to 73% for hPXR activators and 55% to 68% for hPXR non-activators. In addition, consensus models were developed in which the best model shows overall 75% to 83% accuracy for fingerprint and RDkit descriptors, respectively. The best developed model will be utilized for the prediction of hPXR activators and non-activators.

Keywords: hPXR activators, descriptors, machine learning, Support Vector Machine, Random Forest, k-Nearest Neighbour, Naive Bayesian.