Data-driven extraction of human kinase-substrate relationships from omics datasets.

Phosphorylation forms an important part of the signalling system that cells use for decision making and regulation of processes such as cell division and differentiation. In human, >90% of identified phosphosites don't have annotations regarding the relevant upstream kinase. At the same time around 30% of kinases (as annotated in Uniprot) have no known target. This knowledge gap stresses the need to make large scale, data-driven computational predictions. In this study, we have created a machine learning-based model to derive a probabilistic kinase-substrate network from omics datasets. Our methodology displays improved performance compared to other state-of-the-art kinase-substrate prediction methods and provides predictions for more kinases. Importantly, it better captures new experimentally-identified kinase-substrate relationships. It can therefore allow the improved prioritisation of kinase-substrate pairs for illuminating the dark human cell signalling space. Our model is integrated into a web server, SELPHI, to allow unbiased analysis of phosphoproteomics data, facilitating the design of downstream experiments to uncover mechanisms of signal transduction across conditions and cellular contexts.