MIA-QSAR Coupled to Different Regression Methods for the Modeling of Antimalarial Activities of 2-aziridinyl and 2,3-bis-(aziridinyl)-1,4-naphtoquinonyl Sulfate and Acylate Derivatives

Abstract

The antimalarial activities of a series of 2-aziridinyl and 2,3-bis-(aziridinyl)-1,4-naphtoquinonyl sulfate and acylate derivatives have been modeled using multivariate image analysis (MIA) descriptors. The two-dimensional chemical structures correlated reasonably well with dependent variables (Y block) through partial least squares--PLS (for the unfolded data) and multilinear partial least squares--N-PLS (for the three-way array). However, the use of PCA-ranking as variable selection method and least-squares support vector machines (LS-SVM) as regression method improved significantly the prediction ability of the model. All models were validated through leave-one-out and leave-25%-out cross-validations, as well as by means of a Y-randomization test, and demonstrated advantages in prediction performance over an existing model, in which descriptors related to physicochemical and geometric properties of molecules were used to derive multiple linear regression (MLR) and artificial neural networks (ANN) based models. Accounting for non-linearity seems to be an important task for the QSAR modeling of bioactivities of the studied antimalarial compounds.