(613f) Assessing the Importance of Parametric Uncertainty on Flux Balance Analysis

Flux balance analysis (FBA) and associated techniques operating on stoichiometric genome-scale metabolic models have a central role in quantifying metabolic flows and constraining feasible phenotypes. At the heart of these methods lie two important assumptions: (i) the biomass composition neither changes in response to growth conditions nor environmental/genetic perturbations, and (ii) all metabolites incur production and consumption terms that always equal one another. Despite the stringency of these two assumptions, FBA has been shown to be surprisingly robust at predicting cellular phenotypes. In this paper, we formally assess the impact of these two assumptions on FBA results by quantifying how uncertainty in (i) biomass equation coefficients, and (ii) departures from steady-state due to temporal fluctuations propagate to FBA results. The straightforward implementation of Monte-Carlo simulation or chance-constraint programming will yield inaccurate results as in case (i) it is necessary to re-weight the biomass equation so as the molecular weight remains equal to one, whereas in case (ii) metabolite pools must be conserved even under temporally varying conditions. This implies that in both cases conditional sampling of parameter values must be carried out so as the above conditions are satisfied. Our results confirm the importance of standardizing biomass molecular weight to 1 g/mmol. We find that although on average only 4.7% of the uncertainty in the biomass equation coefficients propagates to the biomass yield calculation, not re-weighing the biomass equation magnifies this value to 14.7%. Interestingly, uncertainty in growth-associated ATP maintenance has the greatest effect on biomass yield (28.7%). Furthermore, FBA subject to metabolite pool dynamics was performed while accounting for metabolite concentration coupling relationships and allowing for perturbations in metabolite steady-states. Overall, uncertainty-to-variation of FBA predictions reported here can be interpreted as a cautionary tale and the framework presented in this work can be used for applications where temporal or parameter variability in FBA needs to be assessed.