Reply to “Do genome-scale models need exact solvers or clearer standards?”

I n their Correspondence entitled, “Do genome-scale models need exact solvers or clearer standards?”, Ebrahim et al (2015) suggest an unnecessary dichotomy. They discuss the findings of our paper, “An exact arithmetic toolbox for a consistent and reproducible structural analysis of metabolic network models” (Chindelevitch et al, 2014), and suggest that our work highlights the need for better model encoding standards. Moreover, the authors dispute our claims that multiple previously published metabolic network models are unable to produce growth when analyzed with an exact arithmetic approach. They attribute discrepancies between their findings and ours solely to a misinterpretation of the formatting conventions used to encode these models. The authors conclude that genomescale metabolic network models need better standards, rather than the improvements in accuracy obtained with exact arithmetic. We argue here that improved standards and exact arithmetic are complementary advances that both benefit this field. Thus, the answer to the question posed by Ebrahim et al (2015) is “both.” In this response, we acknowledge the discrepancies in model interpretation between our approach and that of Ebrahim et al (2015), but maintain the key conclusions of our original study. Namely, a number of published metabolic network models are unable to exhibit growth even when our interpretation of these models is identical to that of Ebrahim et al (2015). We attribute the remaining differences between the results of our original study and their study to significant changes made to the models since our results were initially published. Indeed, our MONGOOSE tool provides a model verification platform, which will continue to be useful in identifying errors in model functionality, helping curators to fix them. Additionally, we demonstrate on a specific real-model example that exact arithmetic can change the results of the analysis of genome-scale metabolic network models. We conclude that exact arithmetic remains an important tool for the verification and analysis of metabolic network models. The original parser for models in SBML format used in our study interprets some boundary metabolites as subject to flux balance constraints; as the authors point out, this is contrary to the tacit convention in the field. However, we find blockage in many of the same genome-scale models, even when we interpret them according to this convention. Specifically, we initially reported that 16 out of 39 SBML models (the currently preferred format) had a blocked biomass reaction when only flux balance and irreversibility constraints are taken into consideration (Chindelevitch et al, 2014). After re-analysis with the parsing interpretations used by Ebrahim et al (2015), we find that 8 SBML published models still have a blocked biomass reaction under these conditions (Dataset EV1). Notably, all 10 highly curated SBML models reported in the BiGG database, which includes 3 models we previously reported as blocked, EC3 (iAF1260), HP2 (iIT341), and MT1 (iNJ661), exhibit flux through the biomass reaction after reanalysis (Dataset EV1). Combined with the 28 out of 50 non-SBML models (typically provided as an Excel spreadsheet) that we reported as blocked (Chindelevitch et al, 2014), a total of 36 out of 89 genome-scale models (rather than the original 44 out of 89) still exhibit the problem of having a blocked biomass reaction, which, as we already pointed out in the original paper, can be easily corrected in a systematic way by the MONGOOSE toolbox. In order to understand the outstanding discrepancies between the two analyses of SBML models, we performed a comparison between the inputs used in our approach and that of Ebrahim et al (2015). We found that the original models we analyzed were modified in Ebrahim et al (2015) in one of two ways. First, the source files of many of the models have been altered—beyond the 9 models Ebrahim et al (2015) explicitly state to have modified in consultation with their creators, 10 other models appear to be different between the original versions we analyzed and those analyzed now by Ebrahim et al (2015). All model differences, as well as the scripts necessary to identify them, are included as Dataset EV2. Of the 8 SBML models that still have a blocked biomass reaction in our analysis, the sourcefile modifications account for four. Second, a number of models have been modified algorithmically during processing, as described in the worksheets provided by Ebrahim et al (2015), including 2 of the 8 SBML models we still find to have a blocked biomass reaction. Of the remaining two models, one is correctly identified as blocked, and the last one is absent from Ebrahim et al’s analysis (see Dataset EV3 for details). Note, that while there are also