Development of a minimal kinase ensemble receptor (MKER) for surrogate AutoShim

The target-tailored 3-D virtual screening (VS) method "Surrogate AutoShim" adds pharmacophoric shims to a 16-kinase crystal structure "Universal Kinase Ensemble Receptor" (UKER) to generate highly predictive, target-customized docking models. Predocking a corporate archive of millions of compounds into the 16-structure ensemble takes months. However, since the 16 UKER structures are always the same, docking need only be done once. The predocked results are then "shimmed" to reproduce experimental training data for any number of additional kinases far more accurately than conventional docking. Training new kinase models and predicting activity for millions of predocked compounds against dozens of kinases takes only hours. However reducing the predocking time would make the method even more advantageous. Sequential Floating Forward Search (SFFS) was employed to rationally identify a reduced subset using only 8 of the 16 structures, a "Minimal Kinase Ensemble Receptor" (MKER) that preserved the predictive accuracy for 20 kinase models. Furthermore, a performance evaluation of this subset on an extended set of 52 kinase targets and 100,000 compounds showed statistical model performance comparable to the original UKER. The MKER has halved the time for predocking large databases of internal and commercial compounds. For ad hoc virtual libraries, where predocking is not possible, 2- or 3-kinases "Approximate Kinase Ensemble Receptors" (AKER) were also identified with only a modest loss of prediction accuracy.