Stephan C. Schürer, Steven M. Muskal

Large corpora of kinase small molecule inhibitor data are accessible to public sector research from thousands of journal article and patent publications. These data have been generated employing a wide variety of assay methodologies and experimental procedures by numerous laboratories. Here we ask the question how applicable these heterogeneous data sets are to predict kinase activities and which characteristics of the data sets contribute to their utility. We accessed almost 500,000 molecules from the Kinase Knowledge Base (KKB) and after rigorous aggregation and standardization generated over 180 distinct data sets covering all major groups of the human kinome. To assess the value of the data sets, we generated hundreds of classification and regression models. Their rigorous cross-validation and characterization demonstrated highly predictive classification and quantitative models for the majority of kinase targets if a minimum required number of active compounds or structure-activity data points were available. We then applied the best classifiers to compounds most recently profiled in the NIH Library of Integrated Network-based Cellular Signatures (LINCS) program and found good agreement of profiling results with predicted activities. Our results indicate that, although heterogeneous in nature, the publically accessible data sets are exceedingly valuable and well suited to develop highly accurate predictors for practical Kinome-wide virtual screening applications and to complement experimental kinase profiling.