Development of QSAR Models Using Singular Value Decomposition Method: A Case Study for Predicting Anti-HIV-1 and Anti-HCV Biological Activities

Abstract

This study performed a detailed approach derived by coupling singular value decomposition (SVD) with multiple linear regression (MLR) methods on the performance and predictive capability of the quantitative structure-activity relationship (QSAR). The study was carried out on two different datasets of 128 HIV-1 attachment inhibitors and 115 HCV analogs. For both datasets, the structure of each compound was represented by suitable molecular descriptors. Then, the two datasets were divided into training and test sets employing the Kennard-Stone procedure (K-S). Both MLR and SVD-MLR models were developed to link the structure of the studied compounds to their reported biological activities. The selected models were subjected to the internal leave-one-out cross-validation method, and their predictive abilities were evaluated using the external test set. The developed SVD-MLR models were robust and reliable with an external determination coefficient (R_test^2) of 0.9755 and a mean-square error (MSE) of 0.0205, as well as an R_test^2 of 0.9179 and MSE of 0.0298 for the HCV and the HIV set, respectively. In return, this model could be developed to predict the activities of a non-seen extra set of organic molecules for the purpose of either virtual screening or lead/hit optimization.