(570d) Discriminating Significant From Insignificant Model Parameters: The Case of a Dynamic CHO Cell Model

Kinetic modeling for metabolic networks, formulated as a set of ordinary differential equations for intracellular species concentrations, provides the ability to simulate the dynamic behavior of cellular metabolism. Such models aim to predict cellular response to various external stimuli, allowing an investigator to develop a detailed fundamental understanding of the phenomena studied. Investigators often choose to include more than the necessary model details rather than risk the error of including less than necessary details. However, this increases the number of parameters that need to be identified from experimental data and introduces a substantial mathematical limitation in the identification of the important parameters. As more details are added to the model, the increased number of parameters implies an increased amount of experimental data. Most importantly, it is not clear whether all the model details and the corresponding parameters are necessary for a desired set of model predictions. The present paper presents a computationally efficient methodology to identify the model parameters that are significant for the model predictions and thus distinguish them from the insignificant ones. 

The proposed approach is inspired by the classic Design of Experiments (DOE) techniques, performed in silico using a preliminary model. We start by defining the possible range of each of the unknown model parameters and designing a set of experiments that are simulated through the preliminary model. Utilizing analysis of variance (ANOVA) and response surface model (RSM) linear regression tools we develop a simplified nonlinear meta-model in which only the significant parameters are retained. We applied this methodology to a dynamic model of Chinese hamster ovary (CHO) cell metabolism [Nolan et al. 2010]. This model, comprising 51 parameters and 34 reaction fluxes, was able to provide a reliable preliminary prediction of the effects of fed-batch process variables such as temperature shift, specific productivity, and nutrient concentrations. A D-optimal design of experiments was used to sample the parameters across their ranges, and a RSM was obtained with antibody flux as the output. Investigating linear, linearly interactive, and quadratic RSMs, we efficiently eliminated approximately 90% of the terms as being insignificant, shedding light on the importance of each of the 51 original model parameters in the predictions of the metabolic model.