Gene network dynamics controlling keratinocyte migration

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Translation of large‐scale data into a coherent model that allows one to simulate, predict and control cellular behavior is far from being resolved. Assuming that long‐term cellular behavior is reflected in the gene expression kinetics, we infer a dynamic gene regulatory network from time‐series measurements of DNA microarray data of hepatocyte growth factor‐induced migration of primary human keratinocytes. Transferring the obtained interactions to the level of signaling pathways, we predict in silico and verify in vitro the necessary and sufficient time‐ordered events that control migration. We show that pulse‐like activation of the proto‐oncogene receptor Met triggers a responsive state, whereas time sequential activation of EGF‐R is required to initiate and maintain migration. Context information for enhancing, delaying or stopping migration is provided by the activity of the protein kinase A signaling pathway. Our study reveals the complex orchestration of multiple pathways controlling cell migration. ### Synopsis So far it is unclear how to translate large‐scale ‘omics’ data into a coherent model of cellular regulation that allows one to simulate, predict and control cellular behavior. The functional components that lead to a cellular decision have been debated with respect to requiring a controlled interplay of protein signaling and gene regulation. Combining these complex processes into one model is inherently difficult: protein signaling pathways and gene regulation are feedback‐entangled processes, yet occurring on different timescales in the range of minutes and hours, respectively. This makes it practically impossible to experimentally observe all required cellular components at a sufficiently high sampling rate. Here, we propose a complexity reduction approach for the above problem based on a timescale separation of cellular events. The slaving principle states that the long‐term macroscopic behavior of a system is governed by its slowest evolving variables ([Haken, 2004][1]). Hence, the long‐term phenotypic response of a cell can be expressed in terms of its slowest evolving functional elements. Postulating that a cell reaches a decision on a timescale of hours, its phenotype should be controlled by the slow protein turnover rates. Assuming a proportionality between protein and mRNA concentrations, information on cellular processes lasting hours or days should be reflected in the cellular gene expression dynamics, which are experimentally accessible by time‐resolved microarray experiments. We studied the mechanisms of keratinocyte cell migration in a heterologous co‐culture combining primary human keratinocytes and genetically defined murine fibroblasts as an in vitro model for cell migration in the context of wound healing ([Szabowski et al , 2000][2]). The dermal‐derived hepatocyte growth factor (HGF) is involved in the regulation of migration via the activation of the Met receptor, a proto‐oncogene commonly mutated in metastasizing epithelial cells ([Birchmeier et al ., 2003][3]). However, from a system control point of view, it is not known how the cell produces a sustained and context‐dependent migratory response upon stimulation of Met. We stimulated keratinocytes with HGF and performed DNA microarray experiments at 1, 2, 3, 4, 6 and 8 h after stimulation. We identified a putative set of genes responsible for the decision to migrate, by ranking the gene kinetics according to their mean and peak fold expression. Interestingly, the rank score distribution of genes displays a long tail at high rank scores ([Figure 1B][4]), likely indicating that only a small set of genes mediate the decision to migrate. Genes that are strongly regulated upon the migratory HGF stimulus show a weak or even a downregulated response upon FGF‐7 stimulation, a cytokine mediating proliferation ([Figure 1C][4]). Subsequent filtering according to their biological function further narrowed down candidate genes for in silico modeling. In the next step, a dynamic gene regulatory network was inferred from the kinetics of the identified genes using a continuous time recurrent neural network as the model function ([Beer, 1995][5]). In our model, all genes can interact, time‐delayed, with each other, including themselves via a sigmoid activation function. The impact of the HGF‐induced protein signaling on gene expression is taken into account via a time‐varying input function for each gene. Gene expression kinetics was fitted using a genetic algorithm combined with a search for robust system solutions based on the maximal Lyapunov exponent evaluation ([Rosenstein et al , 1993][6]). Network parameters were obtained from averaging over 250 out of 5000 fittest and most robust solutions. For the complete workflow, see [Figure 1A][4]. The inferred model allowed predictions on the gene network topology and its long‐term dynamic behavior. Network topology was shown to be robust with increasing network size. Starting with the top‐ranked genes and growing the network in size from three to nine genes, each time repeating the inverse modeling workflow, qualitatively conserved the topology of the smallest three‐gene network. This core consists of the highest ranked genes considered for modeling, namely the transcription factor egr3 , akap12 and ptgs2 . Moreover, it could be shown that the regulatory effect of genes not in the core network is weak to negligible, as the external input and the higher ranked genes control their dynamics. Hence, a network of nine genes sufficed to include most genes having a regulatory effect on the cellular decision toward migration. Model verification was performed experimentally on the protein level, with the assumption that (i) protein levels follow mRNA change over time and (ii) fast protein signaling events can be...

This publication has 58 references indexed in Scilit: