Comparative Analysis of Yeast Metabolic Network Models Highlights Progress, Opportunities for Metabolic Reconstruction

We have compared 12 genome-scale models of the Saccharomyces cerevisiae metabolic network published since 2003 to evaluate progress in reconstruction of the yeast metabolic network. We compared the genomic coverage, overlap of annotated metabolites, predictive ability for single gene essentiality with a selection of model parameters, and biomass production predictions in simulated nutrient-limited conditions. We have also compared pairwise gene knockout essentiality predictions for 10 of these models. We found that varying approaches to model scope and annotation reflected the involvement of multiple research groups in model development; that single-gene essentiality predictions were affected by simulated medium, objective function, and the reference list of essential genes; and that predictive ability for single-gene essentiality did not correlate well with predictive ability for our reference list of synthetic lethal gene interactions (R = 0.159). We conclude that the reconstruction of the yeast metabolic network is indeed gradually improving through the iterative process of model development, and there remains great opportunity for advancing our understanding of biology through continued efforts to reconstruct the full biochemical reaction network that constitutes yeast metabolism. Additionally, we suggest that there is opportunity for refining the process of deriving a metabolic model from a metabolic network reconstruction to facilitate mechanistic investigation and discovery. This comparative study lays the groundwork for developing improved tools and formalized methods to quantitatively assess metabolic network reconstructions independently of any particular model application, which will facilitate ongoing efforts to advance our understanding of the relationship between genotype and cellular phenotype.Author Summary: Scientists have been mapping the chemical reactions cells use to grow and manage waste since before enzymes were first identified more than 150 years ago. The model yeast Saccharomyces cerevisiae has one of the most extensively studied metabolic networks, including at least 25 metabolic network models published since 2003. If iterative model improvement refines the metabolic network map, we would expect eventual convergence to a full, accurate metabolic network reconstruction. In this study, we looked for evidence of such convergence through comparative analysis of 12 genome-scale yeast models. We conducted simulations and evaluated model features such as predictive accuracy, genomic coverage and the included metabolites and reactions. We found that no single metric for evaluating models can adequately summarize important aspects of model quality. In some cases, we observed tradeoffs between model predictive accuracy and network coverage. We found evidence of incremental changes to the network reconstruction, but not marked shifts in model predictive ability or other metrics clearly arising from changes to the network alone. This work has broader implications to computational reconstruction of metabolic networks for any organism, and suggests that there is opportunity for refocusing the model building process to better support mapping cellular metabolic networks.