Overlapping structure features selection in linear and non-linear QSAR

Abstract

Multiple linear regressions (MLR) and support vector machine (SVM) supervised linear and non-linear quantitative structure–activity relationship (QSAR) models were developed for a dataset of (81) N–N-disubstituted trifluoro-3-amino-2-propanol derivatives. QSAR dataset was divided into training set (64) and tests (17) to facilitate external validation. QSAR models derived from MLR and SVM yielded appreciable internal and external predictability. Though SVM models were found statistically fit, MLR models attained more effective and consistent predictability of training and test set. It is important to mention that molecular descriptors (structural information) selected in linear and non-linear QSAR models belong to same categories and code for same structural properties. These different descriptors coding for same structural properties are identified and discussed as overlapping structure features in linear and non-linear QSAR models. Molecular descriptors deduced in relation with biological responses are EEig09d, R3u+, P1P1C6, EPS0 and nCb- in MLR aided linear QSAR models and EEig07d, R6u+, P2C6, G1p and Mor12m in SVM aided non-linear. QSAR models are found reliable for further optimization of N–N-disubstituted trifluoro-3-amino-2-propanol CETP inhibitory activity. Concluding remarks include selection of overlapping structure features in linear (MLR) and non-linear (SVM) models and utility of statistical approaches in QSAR. These overlapping features may reveal underlying structural properties which convert a linear relationship to non-linear and better understandings of bio-chemical aspects of QSAR models to medicinal chemists.