Predictions Of Human Kinome-small Molecule Interactions

Protein kinase is one of the largest gene families with various functions. Since protein kinases play important roles in regulating cell signaling pathways and other cellular processes, many diseases are closely related to protein kinases. In fact, protein kinases are the second largest drug targets except GPCR. Thus, screening of small molecule inhibitors has always been the hot spot in drug discovery. In order to discover multi-targeted drugs, high-throughput kinase profiling based on the whole human kinome is becoming the mainstream. However, the costs of kinome-wide experimental profiling are very large. In order to reduce such costs and increase the success rate, it is interesting to develop in silico systems for predicting human kinome-small molecule interactions. In such a way, the binding affinities of small molecules with all kinases in the kinome could be predicted, and thereby a small number of molecule candidates would be selected for experimental validation with corresponding kinases.In this thesis, known crystal structures of 113 human kinases were first optimized as the templates for modeling catalytical domains of kinases. In the beginning of the modeling procedure, the known structures were refined:the coordinates of missing atoms in the crystal structures were added; and large-scale refinements were applied to optimize the structures. Then, based on multiple sequence alignment of whole kinome, template(s) with the greatest sequence identity were selected for homology modeling of about 400 kinases without structures, and therefore about 500 structures of human kinase catalytic domains were obtained. To testify the reliability of our modeling procedure, we took the same procedure to predict structures for the kinases with known structures, by excluding the control crystal structure itself as the template, and compared the predicted models with the crystal structures. Results showed that 88 of the 113 predicted models have Ca-RMSDs (root mean square deviations) lower than 3 A. This indicates that the used procedure is reliable and better than usual homology modeling approaches.Taking the obtained models as an assay panel of receptors and quantitatively describing the binding affinities of small molecules with kinases by using the AutoDock Vina energy function, a molecular docking system for predicting the small molecule-human kinome interactions was constructed. To testify its reliability,38 small molecule inhibitors in clinical trials were profiled against the whole kinome, and the calculated binding free energies were compared with published experimental data. The standard error of the deviations between the predicted and experimental values is～1.7 kcal/mol, and 88% of the deviations were found to fall within the scope of 2.7 kcal/mol, i.e., the standard error of Autodock Vina energy model when applied to original testing data sets. Meanwhile, taking dissociation constant Kd= 100 nM and=1μM of as the selection thresholds, we screened out those small molecule-kinase interaction pairs with coincident Kd values, and comparison with experimental data indicated that the prediction accuracies for two thresholds were about 23% and 56%, respectively. These results suggest that the constructed system could offer practical predictions for experimental studies such as discovery of multi-targeted inhibitors, small molecule-based blockage of certain kinase-related signaling pathways, etc.