Machine Learning based Predictive Model for Screening Mycobacterium Tuberculosis Transcriptional Regulatory Protein Inhibitors from High-Throughput Screening Dataset

In view of the essential role played by dosRS in the survival of Mycobacterium in the infected granuloma cells, dosRS transcriptional regulatory proteins were considered as a validated target for high throughput screening (HTS). However, the cost and time factor involved in screening large compound libraries are an important hurdle in identifying lead compounds. Therefore, the use of computational machine learning techniques to build a predictive model for screening putative drug-like molecule has gained significance. In this regard, a target-based predictive model using machine learning approaches was built to develop fast and efficient virtual screening procedures to screen anti-dosRS molecules. In the present study, we have used various structural and physiochemical attributes of compounds from HTS dataset to train and build a chemoinformatics predictive model based on four state-of-art supervised classifiers (Random forest, SMO, J48, and Naïve Bayes). The trained model was applied to test dataset for validating the robustness, accuracy, and sensitivity of the predictive model in screening active antidosRS molecules. The Cost-Sensitive Classifier (CSC) with Random Forest (RF) algorithm based predictive model showed a high sensitivity (100%) and specificity (83.13%) to identify active and inactive molecules, respectively from assay dataset (ID: 1159583). CSC-RF proved to more robust and efficient in classifying active molecule from an imbalanced dataset with highest Balancing Classification Rate (BCR) (91.57%) and maximum Area under the Curve (AUC) value (0.999). Keywords—Mycobacterium; dosRS-transcriptional regulatory proteins; High Throughput Screening (HTS); virtual screening; machine learning algorithms; classification; predictive chemoinformatics model