Replacement based non-linear data reduction in radial basis function networks QSAR modeling

• Replacement method in combination with RBFN was utilized.
• In spite of data reduction, all variable had collaboration in models.
• Cluster Analysis and K-means clustering were tested in combination with RBFN.
• Three different medicinal-biological data sets were used.
• Use of samples as centers resulted in higher prediction ability.

A combination of radial basis function network (RBFN) and replacement method (RM) is introduced for a better description of quantitative structure activity relationship models (QSAR). RBFN–RM provides a way to choose the informative centers in order to reduce the volume of data and to increase the prediction ability, without eliminating any of the descriptors. This method was applied for predicting the activity of a series of 1-[2-hydroxyethoxy-methyl]-6-(phenylthio) thymine] (HEPT) derivatives, as non-nucleoside reverse transcriptase inhibitors (NNRTIs). Prediction ability of RBFN–RM was compared to combinations of cluster analysis (CA) and K-means clustering with RBFN (RBFN–CA and RBFN–K-means). The Q2 value for RBFN–RM, RBFN–K-means, and RBFN–CA was calculated as 0.9766, 0.7965, and 0.7084, respectively, which showed the merit of RBFN–RM. The method was applied on Selwood and GABA data sets, as well. To check the stability of the RM procedure, for each data set, the models were validated by using different arrangements of calibration and validation sets. Using any of the calibration and validation arrangements for HEPT, Selwood, and GABA data sets the estimated correlation values, r, for calculated versus actual activities in the validation sets were higher than 0.96.