Introduction: Responses to targeted drugs are highly variable across patients and current genomic biomarkers are often ineffective at stratification. Here, we tested the hypothesis that direct quantification of kinase activity markers (using phosphoproteomics) would predict kinase inhibitor efficacy in cancer with greater accuracy than proxies of pathway activation (such as genetic mutations). Experimental procedures We designed an approach (based on computational analysis of phosphoproteomics data) that (i) systematically identifies markers of kinase activity and (ii) measures these markers in primary cancer cells with precision and accuracy. To test the performance of the method, we carried out LC-MS/MS phosphoproteomics analysis of three different cancer cell lines treated with 60 different kinase inhibitors. In parallel, we determined the selectivity profile of the same 60 compounds against 460 kinases. An algorithm was designed to compare the in vitro specificity profiles of these kinase inhibitors with their effects on cellular phosphoproteomes. We then statistically assessed kinase activity enrichment in primary breast tumors from 86 cases and 36 primary acute myeloid leukaemia (AML) blast specimens, by measuring our markers of kinase and signaling activity in these tumors. Models to predict sensitivity to kinase inhibitors were constructed using methods based on multivariate regression (partial least squares) and machine learning (random forest and neural networks). Unpublished data We quantified 22,000 unique protein phosphorylation sites in three cell lines treated with 60 kinase inhibitors each in quadruplicate, resulting in the acquisition of 15.8M quantitative data points. Using our newly developed computational tools, we identified 6,206 kinase activity markers for 106 kinases and 1,508 network edges (kinase-kinase relationships). These markers of kinase activity and network circuitry correlated with the impact that several kinase inhibitors had in reducing the viability of primary AML cases. Machine learning models, that used our markers of kinase activity as input, predicted responses to inhibitors with high accuracy (RMSE <0.15). As an example, our approach was twice more accurate at determining response to midostaurin than the FDA approved FLT3-ITD mutation biomarkers. Quantification of our activity markers in 86 primary breast tumors revealed an anti-correlation between MAPK and PI3K activity markers. Consistent with this observation, MAPK pathway inhibitors decreased the viability of cells with wild-type PIK3CA to a greater extent than those harboring PIK3CA mutations in helical or kinase domains. Our data suggest that absence of PIK3CA mutations, in combination with our MAPK pathway activity markers, should be investigated as a predictive signature for MAPK pathway inhibitors. Conclusion: Our study provides insights into kinase network regulation and represents a unique resource to investigate the relationships between kinase network topology and drug response versus resistance. Crucially, we demonstrate that our approach predicts efficacy of targeted therapies with greater accuracy than mutational analysis.
Citation Format: Pedro Rodriguez Cutillas, Mauran Hijazi, Ryan Smith, Conrad Bessant, David Britton. Chemical phosphoproteomics systematically identifies circuitries of kinase networks in cancer cells and predicts their response to kinase inhibitors [abstract]. In: Proceedings of the AACR-NCI-EORTC International Conference on Molecular Targets and Cancer Therapeutics; 2019 Oct 26-30; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2019;18(12 Suppl):Abstract nr C088. doi:10.1158/1535-7163.TARG-19-C088