Critical assessment ofE. coligenome-scale metabolic model with high-throughput mutant fitness data

Abstract

TheE. coligenome-scale metabolic model (GEM) is a gold standard for the simulation of cellular metabolism. Experimental validation of model predictions is essential to pinpoint model uncertainty and ensure continued development of accurate models. Here we assessed the accuracy of theE. coliGEM using published mutant fitness data for the growth of gene knockout mutants across thousands of genes and 25 different carbon sources. We explored the progress of theE. coliGEM versions over time and further investigated errors in the latest version of the model (iML1515). We observed that model size is increasing while prediction accuracy is decreasing. We identified several adjustments that improve model accuracy – the addition of vitamins/cofactors and re-assignment of reaction reversibility and isoenzyme gene to reaction mapping. Furthermore, we applied a machine learning approach which identified hydrogen ion exchange and central metabolism branch points as important determinants of model accuracy. Continued integration of experimental data to validate GEMs will improve predictive modeling of the mapping from genotype to metabolic phenotype inE. coliand beyond.SynopsisE. coligenome-scale metabolic model flux balance analysis (FBA) prediction accuracy was quantified with published experimental data assaying gene knockout mutant growth across different carbon sources. Insights into model development trends and sources of inaccuracy were revealed.Model representational power (size) has been increasing over time, while accuracy has been decreasing.Adding vitamins/cofactors to the model environment and re-assigning reaction reversibility and isoenzyme gene-to-reaction mapping improves correspondence between model predictions and experimental data.Machine learning reveals hydrogen ion exchange and central metabolism branch points as important features in the determination of model accuracy.

Accurate prediction of in vivo protein abundances by coupling constraint-based modelling and machine learning

“…The reasons for the lower predictive performance for yeast in comparison to E. coli could be due to several factors: First, from a modelling perspective, there is more and better-curated knowledge about the metabolism of E. coli compared to that of S. cerevisiae (Bernstein et al, 2023). In addition, the metabolism of yeast is compartmentalized and includes mechanisms for enzyme regulation not accounted for in pcGEMs of any organism (e.g., organellar enzyme pools, post-translational modifications, diffusion effects).…”

Integration of graph neural networks and genome-scale metabolic models for predicting gene essentiality

“…This phenomenon could arise from various sources, such as the data imbalance that in our case favors essential labels, or because predicting non-essential genes is intrinsically more challenging than essential ones [4]. Our approach illustrates the potential of exploring new ways of combining traditional tools such as Flux Balance Analysis with modern data-driven approaches, and adds to the growing body of literature at the interface of genome-scale metabolic modeling with machine learning [40,47].…”

“…Flux Balance Analysis has shown good prediction accuracy for gene essentiality in the E. coli bacterium [31] and other model microbes, but predictions for eukaryotes and higherorder organisms have produced mixed results [18,22]. The quality of FBA predictions have also been shown to vary strongly across published models as well as the performance metrics employed to quantify prediction accuracy [4]. An often overlooked limitation of the FBA approach is the tacit assumption that the metabolism of deletion strains optimizes the same objective as the wild type.…”

Metabolic interaction models recapitulate leaf microbiota ecology

“…Our metabolic modeling predicted interaction outcomes on the basis of carbon sourceutilization capabilities. Further curation that integrates additional properties such as vitamin auxotrophies and storage (66), as well as metabolic shifts occurring from gene regulation (67,68), is likely to improve the quantitative predictive power of this framework (33,69). Moreover, although our predictions based on metabolic mechanisms are informative of key aspects underlying the assembly of plantassociated microbiomes, they do not consider additional factors that are known to shape leaf communities, such as signaling molecules, antagonistic interactions, or host immunity (5,(70)(71)(72)(73).…”