However, the limitation of the RP method is its inability to extrapolate beyond the range of observed responses. The main objective of incorporating the RP method in the virtual screening process is to rapidly classify unknown compounds based on a small number of readily interpretable descriptors, therefore, for screening compounds. The recursive partition decision tree model was con following structed using a QSAR module of Cerius2 version 4. 10. 17. The splits were scored using the Gini Impurity scoring function, which minimizes the impur ity of the nodes resulting from the split. The tree was set to prune backward through a moderate pruning pro cess, to avoid over splitting. Every node should contain 1% of the samples to qualify for further splits.

The knot value was limited to a threshold of 20 per variable and maximum tree depth was set to 10. The best RP tree was generated with these parameters. Training and test sets of the RP model A total of 225 compounds collected from the literature were classified into two categories, the active class, which includes the compounds having an activity range below or equal to 500 nM, and the inactive class, which covers the activity range of more than 500 nM in the IKKb enzyme inhibition assay. Two dimensional and three dimensional descriptors of Cerius2 were used for the RP tree generation. The descriptors were optimized by means of removing those with constant values and 95% of the zero values, while some of the descriptors were deleted on the basis of the correlation threshold 0. 9.

Totally, 37 descriptors were retained in the RP study that comprised 31 two dimensional and 6 three dimensional descriptors. In the RP study, we defined the activity class column as a dependent variable and the descriptors used as independent variables. A total of 84 compounds were used as an external test set compounds, collected from a different set of pub lished articles, with none of the compounds or similar scaffolds included in the training set. External test set compounds have been reported by two groups. The first set of compounds are derivatives of the imida zothienopyrazine core, with a series of compounds having imidazoquinoxaline synthesized by same group included in training the model. Another set of compounds reported by Chiristoper et al. was synthesized based on the benzimidazole core to specifi cally inhibit IKKu, but instead inhibited IKKb.

The external test sets were combined to serve as an indepen dent test set to asses the generality of the model. Dependent and independent variables were calculated as explained before. Docking procedure The third filter used in the VS scheme was molecular docking. Brefeldin_A To date, there is no crystal structure reported for IKKb. Hence, we modeled the protein based on four other closely related kinase proteins, based on the proce dure of homology modeling detailed elsewhere.