Application of k-means clustering, linear discriminant analysis and multivariate linear regression for the development of a predictive QSAR model on 5-lipoxygenase inhibitors

Abstract

In this work, we performed a quantitative structure activity relationship (QSAR) model for a family of 5-lipoxygenase (5-LOX) inhibitors using k-means clustering and linear discriminant analysis (LDA) for the selection of training and test sets and multivariate linear regression (MLR) for the independent variable selection. With the k-means clustering method, the total set of compounds (58 derivatives of 5-Benzylidene-2-phenylthiazolinones) was divided in two clusters according to a simple discriminant function. We found that piID (conventional bond order ID number) molecular descriptor discriminates correctly 100% of the compounds of each clusters. Thirty different models divided in three series were analyzed and the series with representative training and test sets (series 3) had the most predictive models. The statistical parameters of the best model are Rtrain=0.811 and Rtest=0.801. We found that a rational selection in the setting-up of training and test sets allows to obtain the most predictive models and the random selection is sometimes unsuitable, especially, when the total set of compounds can be classified in different clusters according to structural features.