Inferring dysregulated kinases using kinase-substrate predictions of a machine learning model

Abstract

Protein phosphorylation is a crucial post-translational modification that regulates many signaling pathways and cellular processes. It involves transfer of a phosphate group to specific amino acid residues of substrate proteins. Protein kinases are enzymes that catalyze phosphorylation. Often, when kinases are dysregulated, they trigger abnormal phosphorylation, which alters normal functioning of its substrates. In many human diseases, substrates are abnormally phosphorylated, and it is associated with aberrant kinase activity. ☐ Kinase activity inference methods identify differentially activated upstream kinases from observed differences in phosphorylation abundance of substrates. Typically, kinase activity inference techniques rely on the knowledge of kinase-substrate associations. However, experimentally validated kinase-substrate data is limited and includes less than 7% of human phosphorylation sites. In the last two decades, many tools have been developed to determine kinases of substrates. Most tools are limited by factors such as relying solely on sequence information around phosphorylation sites, offering limited coverage of human kinases, and lacking functional context in their predictions. Hence, in this dissertation, I use machine learning and knowledge graph embedding to predict kinases of human substrates by including functional context of proteins. ☐ Further, I evaluate the applicability of my model’s predictions in kinase activity inference using a kinase perturbation dataset. Finally, I apply my predictions to infer implicated kinases in tumor conditions and validate the results using signatures of cells with kinase-targeted drugs.