Abstract PR12: A model of HGF and EGF dual ligand stimulation in ACHN cells constructed with the aid of a novel model merging heuristic

Abstract

Abstract The investigation of context-specific perturbations of cell signaling pathways requires the construction of large scale network models of multiple interacting pathways. A key issue in the production of these models is the process of model calibration, sometimes referred to as regression or optimization, during which parameters of a model are systematically adjusted to improve the model's predictions on a set of training data. We have developed a novel heuristic, termed “model merging,” to exploit parameter information from simpler models in the construction of new complex models. Using this merging heuristic, we developed a model of HGF and EGF stimulation of the MAP kinase cascade in ACHN cells, a renal cancer cell line. We constructed the model of HGF and EGF stimulation by first constructing smaller models of stimulation by each separate ligand. We use the single ligand models to inform the training of the dual ligand model by applying the model merging heuristic. Our model accurately reflects the time courses of pMET, pEGFR, and pMAPK under single ligand and dual ligand stimulation over two-hour time courses. It also captures several previously unmodeled features of MET and EGFR pathway crosstalk. The goal of model merging is to resolve parameter differences while maintaining the predictive quality of the previously known models. By the nature of model calibration, there are discrepancies between parameter estimates of the single ligand models whenever a parameter appears in both models. We design a scoring system for model predictions that takes into account the discrepancies between the parameter estimates. To resolve these discrepancies, our model merging heuristic iteratively adjusts the scoring system to create a series of optimization problems. The solutions to the series of problems converge to a resolved parametrization for the general model while maintaining the high quality of predictions of the component models. A key advantage of our heuristic over other model calibration algorithms is its ability to identify context-specific differences in parametrizations during model calibration. This is possible because the heuristic does not enforce any context-specific or context-free assumptions. Rather, it tries to resolve differences between multiple parameter sets by searching for a single consensus parameter set. If the training data precludes such a consensus for some of the parameters, then the heuristic can identify these parameters and simultaneously resolve the other parameters for which there is a consensus. In this way, the heuristic identifies parameters that are context-specific. This major advantage of our model merging method was integral to our construction of a model of HGF and EGF stimulation of the MAP Kinase cascade in ACHN cells. We identify context-specific parameters that describe a previously known, receptor-specific difference in phospho-MAPK negative feedback to the adapter protein GAB1. Identification of context-specific parametrizations also demonstrated that a prior model of EGF stimulated MAP kinase signaling employed an overly simplified mechanism of phospho-MAPK negative feedback to the protein SOS. Using this more detailed mechanism, we are now able to better interrogate the role of MAPK negative feedback in the more general multipathway setting. We also investigated the generalizability of our model merging heuristic by applying it repeatedly to a simplified version of our dual ligand model. We demonstrate that applying model merging in combination with Nelder-Mead calibration generates a significantly more robust estimation of parameters than Nelder-Mead calibration alone. Model merging thus produces a higher likelihood of finding a global best model parametrization instead of a local best parametrization. These results highlight the potential for our model merging technique to enable the creation of multipathway models with high mechanistic detail with reduced computational cost. This proffered talk is also presented as Poster A10. Citation Format: Andrew L. Matteson. A model of HGF and EGF dual ligand stimulation in ACHN cells constructed with the aid of a novel model merging heuristic [abstract]. In: Proceedings of the AACR Special Conference on Chemical Systems Biology: Assembling and Interrogating Computational Models of the Cancer Cell by Chemical Perturbations; 2012 Jun 27-30; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2012;72(13 Suppl):Abstract nr PR12.