Computer-Aided Drug Design: An Innovative Tool for Modeling
We have applied a retrosynthetic program to determine the scaffold and R-group chemical space seen within a library of known kinase inhibitors and non-kinase drug-like molecules. Comparison of the differences quickly revealed that kinase inhibitors are distinct in several chemical fragment and physical properties. We then applied these descriptors in a multivariable quantitative structure-activity relationship (QSAR) model with the goal to distinguish kinase inhibitors from non-kinase drug-like molecules. This model is heuristic in that it was trained over a dataset of 258 known kinase inhibitors and 230 non-kinase drug molecules. The final model recognized 98% of the training set as being kinase inhibitors and had a false positive rate of 15%. This trait for false positives was accepted out of a desire to maintain diversity and not miss possible good kinase inhibitors for screening. The model was validated by reserving a portion of the datasets as test sets, which were not included in the QSAR model building stage. This was done repetitively for different percentiles of the total dataset population. It was seen that model recognition and false positive were only slightly damaged well down to a 70% reserve (30% dataset used for QSAR model training while 70% used for reserve test set). Beyond 70%, the QSAR models were inconsistent, signifying that the training sets were inadequately diverse to represent the greater reserve test sets. We applied this model to evaluate the commercial kinase libraries available from Asinex, BioFocus, ChemDiv and LifeChemicals to facilitate purchase decisions for compounds for HTS for lead compounds. We observed that there are significant differences in populations of recognizable kinase inhibitors across the vendors analyzed, with BioFocus showing the greatest population of kinase like molecules.