Model-based selection of the robust JAK-STAT activation mechanism

JAK-STAT pathway family is a principal signaling mechanism in eukaryotic cells. Evolutionary conserved roles of this mechanism include control over fundamental processes such as cell growth or apoptosis. Deregulation of the JAK-STAT signaling is frequently associated with cancerogenesis. JAK-STAT pathways become hyper-activated in many human tumors. Therefore, components of these pathways are an attractive target for drugs, which design requires as adequate models as possible. Although, in principle, JAK-STAT signaling is relatively simple, the ambiguities in a receptor activation prevent a clear explanation of the underlying molecular mechanism. Here, we compare four variants of a computational model of the JAK1/2-STAT1 signaling pathway. These variants capture known, basic discrepancies in the mechanism of activation of a cytokine receptor, in the context of all key components of the pathway. We carry out a comparative analysis using mass action kinetics. The investigated differences are so marginal that all models satisfy a goodness of fit criteria to the extent that the state of the art Bayesian model selection (BMS) method fails to significantly promote one model. Therefore, we comparatively investigate changes in a robustness of the JAK1/2-STAT1 pathway variants using the global sensitivity analysis method (GSA), complemented with the identifiability analysis (IA). Both BMS and GSA are used to analyze the models for the varying parameter values. We found out that, both BMS and GSA, narrowed down to the receptor activation component, slightly promote the least complex model. Further, insightful, comprehensive GSA, motivated by the concept of robustness, allowed us to show that the precise order of reactions of a ligand binding and a receptor dimerization is not as important as the on-membrane pre-assembly of the dimers in the absence of ligand. The main value of this work is an evaluation of the usefulness of different model selection methods in a frequently encountered, but not much discussed case of a model of a considerable size, which has several variants differing at peripheries. In such situation, all considered variants can reach nearly perfect agreement with respect to their numerical simulations results and, most often, the sufficient experimental data to test against is not available. We argue that in such an adverse setting, the GSA and IA, although not directly corresponding to the model selection problem, can be more informative than the representative, generalizability-based approaches to this task. An additional insight into how the responsibility for the network dynamics spreads among model parameters, enables more conscious, expert-mediated choice of the preferred model.