Systematic inference of regulation by protein kinases finds surprising level of transcription factor deactivation

Gene expression is controlled by pathways of regulatory factors often involving the activity of protein kinases on transcription factor proteins. Despite this well established mechanism, the number of well described pathways that include the regulatory role of protein kinases on transcription factors is surprisingly scarce in eukaryotes. To address this, PhosTF was developed to infer functional regulatory interactions and pathways in both simulated and real biological networks, based on linear cyclic causal models with latent variables. GeneNetWeaverPhos, an extension of GeneNetWeaver, was developed to allow the simulation of perturbations in known networks that included the activity of protein kinases and phosphatases on gene regulation. Over 2000 genome-wide gene expression profiles, where the loss or gain of regulatory genes could be observed to perturb gene regulation, were then used to infer the existence of regulatory interactions, and their mode of regulation in the budding yeast Saccharomyces cerevisiae. Despite the additional complexity, our inference performed comparably to the best methods that inferred transcription factor regulation assessed in the DREAM4 challenge on similar simulated networks. Inference on integrated genome-scale data sets for yeast identified [~]8800 protein kinase/phosphatase-transcription factor interactions and [~]6500 interactions among protein kinases and/or phosphatases. Both types of regulatory predictions captured statistically significant numbers of known interactions of their type. Surprisingly, kinases and phosphatases regulated transcription factors by a negative mode or regulation (deactivation) in over 70% of the predictions. Author summaryIn this work we addressed the challenging problem of inferring regulation by protein kinases and phosphatases via their activity on transcription factors. Although many protein kinase activity predictors have been developed for classes of protein kinases on specific amino acids within target sequences, our approach (PhosTF) provides predictions of regulatory activity for specific protein kinases and phosphatases on specific transcription factors. Our inference approach achieves this using the functional output observed in gene expression data of gene knock out stains, along with known transcription factor regulatory interactions. We formulated and tested a model for inference of regulation as well as a model for simulation of genes expression, transcription and translation. The simulation was used for in-silico validation of the inference method, which performed comparably to the best performers on simpler inference problem in the DREAM4 competition. The inference method was then applied to yeast expression data, with significant validation by known kinase/phosphatase interactions. Over 15300 novel regulatory interactions were predicted, suggesting that kinase activity provided a surprising level of repression of gene expression, either through the deactivation of activators or the activation of repressors.