Incorporating Phenotype Data to Improve Parental Strain Flux Estimations

Genome-scale constraint-based models have been widely used to predict extracellular and intracellular fluxes. Many of these methods, such as Flux Balance Analysis (FBA), depend upon the assumption of maximum growth (or some other objective function). When cells have not evolved to maximize growth or where other processes (e.g., regulation) limit metabolic fluxes, these methods may fail to accurately describe the flux distribution in a cell. Methods have been developed which predict a perturbed cellular state from a previously estimated parental state; however, these methods (e.g. Minimization of Metabolic Adjustment) are heavily dependent upon the accuracy of the parental state flux distribution.

In order to improve the accuracy of the parental state flux prediction, various methods have been developed which incorporate various types of experimental data including extracellular flux, transcriptomic, and proteomic data. However, extracellular fluxes are not sufficiently constraining to substantially improve flux predictions, gene and protein expression levels do not strongly correlate with flux or help improve flux estimations, and intracellular fluxes are difficult and costly to measure using ¹³C metabolic flux analysis. Therefore, we propose a new method called Relative Phenotypes for Parental Strain estimation (REPPS) that uses multiple sets of extracellular flux measurements. By utilizing extracellular data for a number of perturbed reference strains (e.g. single gene knockouts), as well as the parental strain (i.e. wildtype), we can substantially improve the descriptive flux estimation of the parental strain as well as flux predictions in subsequent mutant strains derived from the parental strain. For example, using REPPS in Escherichia coli with data from five single gene knockout reference strains, the mean squared error between model-estimated and measured fluxes is 53% less than the equivalent parsimonious FBA prediction for the parental strain. This improved parental flux estimate further improves predictions of fluxes in derived strains, where using MOMA with the REPPS estimated parental flux distribution yielded a 58% decrease in mean squared error between predicted and measured fluxes, compared to using MOMA with the pFBA predicted parental flux distribution.

REPPS is also powerful in its ability to be less affected by omissions in a model. Since REPPS attempts to fit a wide variety of experimental data, it is less sensitive to missing or incorrect reactions that may be in a model. Therefore, REPPS can also be used for lesser studied organisms whose models are less developed or for organisms where optimal growth is not their principal objective.

These improvements in accuracy with REPPS result in both better descriptive flux distributions in the parent strain, as well as better prediction of derived-strain phenotypes and fluxes. The improvement in descriptive modelling can improve our understanding of the key mechanisms underlying observed behaviors in a strain. Furthermore, the improved predictive capabilities can reduce the number of experiments necessary to create a chemical production strain, thus improving our ability to engineer strains with desired phenotypes.