MLR-Based QSAR Models for Predicting Inhibitory Activity of Reverse Transcriptase by HEPT Derivatives using GETAWAY Descriptors

In this study, quantitative structure-activity relationship (QSAR) models for non-nucleoside reverse transcriptase inhibitors based on 1-[(2-hydroxyethoxy)-methyl]-6-(phenylthio)thymine (HEPT) derivatives were generated. The structures of the compounds and their activities were obtained from the literature. The data set were divided into two sets: training set (N=91) and validating set (N=10). All 3-D structures of these inhibitors were optimized by semi-empirical method, AM1 prior to calculations of 3-D molecular descriptors, GETAWAY. Multiple linear regression (MLR) using stepwise method was applied to determined significant descriptors. Out of 197 GETAWAY descriptors, 4-14 molecular descriptors have significant relationships with the activities (expressed as log (1/EC50)) of HEPT. The MLR method generated 14 models. The predictive power of these models were evaluated internally by applying the following statistical parameters for the training set and test set: root-mean-square error for prediction (RMSE), correlation coefficient (R), squared correlation coefficient (R2), adjusted squared correlation coefficient (R2adj), difference between R2 and R2adj (R2 – R2adj), squared cross-validation correlation coefficient (Q2). External validation was performed by employing Golbraikh and Tropsha criteria. Moreover, residual analysis was performed. Internal validation of Model XX (N = 91) revealed that it has the highest predictive power (RMSE = 0.4288, R = 0.9393, R2 = 0.882, R2adj = 0.8620, R2 - R2adj = 0.0203, Q2 = 0.8317). However, external validation (using the validating set, N=10) showed that Model XII has the highest predictive power (R2 = 0.961, R20 = 0.9565, k = 0.8648, k’ = 0.9800, [R2 - R20] = 0.0066, [R2 - R20] /R2 = 0.0069, R2pred = 0.9481) based on Golbraikh and Tropsha criteria. Residual analysis confirmed that both models are valid.