A simple clustering technique to improve QSAR model selection and predictivity: application to a receptor independent 4D-QSAR analysis of cyclic urea derived inhibitors of HIV-1 protease

A training set of 50 tetrahydropyrimidine-2-one based inhibitors of HIV-1 protease, for which the -log K(i) values were measured, was used to construct receptor independent 4D-QSAR models. A novel clustering technique was employed to facilitate and improve model selection as well as test set predictions. Following the manifold model theory, five unique models were chosen by the clustering algorithm (q(2) = 0.81-0.84). The models were used to map the atom type morphology of the inhibitor binding site of HIV-1 protease as well as to predict the potencies (-log K(i)) of 10 test set compounds. The rank-difference correlation coefficient was used to evaluate the quality of the test set predictions, which was improved from 0.39 to 0.68 when the clustering technique was applied. The set of five models, collectively, identify the important binding characteristics of the HIV protease receptor site. This study demonstrates that the selected simple clustering technique provides a discrete algorithm for model selection, as well as improving the quality of test set, or unknown, compound prediction as determined by the rank-difference correlation coefficient.