Kinase-Kernel Models: Accurate In silico Screening of 4 Million Compounds Across the Entire Human Kinome

Reliable in silico prediction methods promise many advantages over experimental high-throughput screening (HTS): vastly lower time and cost, affinity magnitude estimates, no requirement for a physical sample, and a knowledge-driven exploration of chemical space. For the specific case of kinases, given several hundred experimental IC50 training measurements, the empirically parametrized profile-quantitative structure–activity relationship (profile-QSAR) and surrogate AutoShim methods developed at Novartis can predict IC50 with a reliability approaching experimental HTS. However, in the absence of training data, prediction is much harder. The most common a priori prediction method is docking, which suffers from many limitations: It requires a protein structure, is slow, and cannot predict affinity. Highly accurate profile-QSAR models have now been built for roughly 100 kinases covering most of the kinome. Analyzing correlations among neighboring kinases shows that near neighbors share a high degree of SAR similarity. The novel chemogenomic kinase-kernel method reported here predicts activity for new kinases as a weighted average of predicted activities from profile-QSAR models for nearby neighbor kinases. Three different factors for weighting the neighbors were evaluated: binding site sequence identity to the kinase neighbors, similarity of the training set for each neighbor model to the compound being predicted, and accuracy of each neighbor model. Binding site sequence identity was by far most important, followed by chemical similarity. Model quality had almost no relevance. The median R2 = 0.55 for kinase-kernel interpolations on 25% of the data of each set held out from method optimization for 51 kinase assays, approached the accuracy of median R2 = 0.61 for the trained profile-QSAR predictions on the same held out 25% data of each set, far faster and far more accurate than docking. Validation on the full data sets from 18 additional kinase assays not part of method optimization studies also showed strong performance with median R2 = 0.48. Genetic algorithm optimization of the binding site residues used to compute binding site sequence identity identified 16 privileged residues from a larger set of 46. These 16 are consistent with the kinase selectivity literature and structural biology, further supporting the scientific validity of the approach. A priori kinase-kernel predictions for 4 million compounds were interpolated from 51 existing profile-QSAR models for the remaining >400 novel kinases, totaling 2 billion activity predictions covering the entire kinome. The method has been successfully applied in two therapeutic projects to generate predictions and select compounds for activity testing.