Metabolic modeling of zebrafish - Combining robust and stochastic optimization methods with GC-flux and FVA to account for interindividual variation in gene expression
Summary
Introduction: Genome-scale metabolic models (GEMs) predict metabolic fluxes based on genomic information from genes underlying its metabolic network. Van Steijn et al. improved the only available GEM for zebrafish (ZebraGEM) to an updated version (ZebraGEM 2.0), which is now suitable to be used with gene expression data. When comparing groups, it is important to consider the original distribution underlying the gene expression before integrating into ZebraGEM 2.0. This can be taken into account with robust and stochastic optimization methods.
Aim: Our objective is to account for interindividual variation in gene expression by integrating gene expressions into ZebraGEM2.0 based on a robust optimization method or a stochastic optimization method.
Method: Gene expression data were derived from a published dataset of tuberculosis-infected and control zebrafish, RNA was measured at four (4dpi) and five (5dpi) days postfertilization, resulting in four subgroups. Per group and per gene the associated Weibull distribution parameters were calculated. In this study, three methods will be compared. First, the average gene expressions were taken per group as expression variables for integration into ZebraGEM 2.0 by Gene-centric Flux (mean model). Flux Variability Analysis (FVA) was performed and Relative Flux Range Change (RFRC) was calculated to compare conditions at both days. Second, for robust optimization, gene expressions at 5th, 16th and 25th percentile of the gene’s expression distribution were taken as the expression variables. FVA was performed, RFRC was calculated and compared to results of the mean model. Third, for stochastic optimization, per group one hundred scenarios were created with each scenario based on random sampling within each gene’s expression distribution, assuming independence of genes. FVA was performed on all scenarios, and a 10-90% and 25-75% interval of all fluxes per reaction was taken. RFRC was calculated and compared to the mean model.
Results: The mean model showed decreased biomass production for infected larvae at 4dpi and 5dpi. All robust and stochastic models showed different results, none was similar to the mean model. The mean model showed altered histidine metabolism between control and infected groups, whereas in robust models the purine and pantothenate pathways and cofactor biosynthesis were altered.
Conclusion: In this study, random sampling for stochastic optimization did not work properly, therefore it was concluded that the obtained results were invalid. The robust models suggested an altered purine metabolism in infected zebrafish larvae, which is also suggested in literature. The altered histidine metabolism from the mean model and the altered pantothenate pathway and altered cofactor biosynthesis from the robust model require further confirmation by further research. Therefore, it is not clear yet if this robust method predict metabolic changes better compared to integration of average gene expressions.