In June 2008, it became apparent that two MRs were being assembled by two different research groups 20 (Bumann, unpublished data). Subsequently, a Salmonella reconstruction jamboree was held at the University of Iceland, Reykjavik, from September 5th to 6th, 2008. The jamboree team consisted of over 20 experts in microbiology, proteomics, Salmonella physiology, and computational modeling. Based on the experience with the yeast reconstruction jamboree 17 , a methodology was devised to increase the efficiency of community-based network reconstruction 24 and applied to the Salmonella reconstruction jamboree.

The goal of a network reconstruction jamboree is to provide a 2-D genome annotation that is of higher quality than it may be achieved by bioinformatic analyses alone 24 25 . The objective of this jamboree was to re-evaluate, reconcile, and expand the currently available MRs for S. Typhimurium with a focus on virulence. Furthermore, we aimed to include standard identifiers for reconstruction metabolites, reactions, and genes to facilitate subsequent mapping of 'omics' data. The starting MRs were AJRecon (a variant is published in 20 ) and BRecon (D. Bumann, unpublished data), which were derived from published E. coli MRs, iJR904 26 and iAF1260 27 , respectively, and their contents were modified to account for Salmonella-specific properties; i.e., transport and enzymatic reactions not present in Salmonella were removed and the proteins associated with the reactions were modified to contain proteins present in S. Typhimurium LT2.

The resulting knowledge-base, STM_v1.0 (Table 1; Additional file 2; Additional file 1 Table S2), represents the final product of a community-effort to develop a detailed MR of S. Typhimurium. STM_v1.0 integrates the novel and common features of the starting MRs into a vetted, well-documented consensus knowledge-base, capturing currently available BiGG knowledge about S. Typhimurium. Key features of STM_v1.0 include i) accounting for the periplasm as a compartment between the extracellular space and cytoplasm; ii) Salmonella-specific virulence characteristics, such as iron chelation by salmochelin and serovar Typhimurium LT2 O-antigen production; iii) the possibility to employ the consensus MR as mathematical, predictive model; and iv) comprehensive support data for reactions and associated genes (Additional file 1 Table S2a). Some information was excluded from STM_v1.0, such as the 26 dipeptide and tripeptide transport/digestion reactions that are present in AJRecon, as they represent generic compounds. Accounting for all potential consumable oligopeptides would make computational analysis intractable or unnecessarily difficult. Appropriate oligopeptides may be manually added to STM_v1.0 to represent a specific growth environment. We also attempted to exclude reactions that were included to fit some growth data 28 , but were contrary to other observations 20 29 as was the case for growth with D-aspartic acid 30 as the sole carbon source which requires an unknown transporter and an unknown aspartate racemase 31 .

Additionally, we evaluated the reaction directionality of consensus MR reactions by considering thermodynamic properties of participating metabolites. In the case that a thermodynamic prediction was inconsistent with experimental evidence, the experimental evidence was followed. Thermodynamic predictions are made using the knowledge that is available 45 , and incorrect predictions highlight gaps in our knowledge of biology.

A bacterial MR often includes a biomass reaction that lists all known biomass precursors and their fractional contribution necessary to produce a new bacterial cell in a given environment. The individual biomass constituents of a S. Typhimurium cell have been measured 20 , and adapted for the consensus reconstruction by accounting for the changes in naming and compartments introduced during reconciliation (Additional file 1 Table S3c).

To assess the utility of a mathematical approximation of reality, it is essential to determine the consistency of the model's predictions with real-world benchmarks. In the case of MRs, comparing experimental growth data with predicted biomass production is a commonly employed metric in benchmarking metabolic models 14 . Although biomass production is a commonly employed metric, the results should always be taken with a grain of salt; for instance, it is possible to improve the fitting of a model's predictions to growth data by including enzymatic reactions for which no evidence exists or which are contrary to published experimental observations. The reconstruction committee chose not to include invalidated enzymatic reactions that improved the fit between growth predictions and experimental observations; the failings of the model's predictions highlight areas where knowledge is lacking and experimental undertakings could identify new knowledge.

For S. Typhimurium, there is a wealth of experimental growth data 29 . Overall, we found good agreement between the qualitative growth phenotype predictions and the experimental data (Table 2 Additional file 1 Table S4); with the notable exception of sulfur metabolism where the prediction accuracy was about 40%. As we are becoming increasingly aware of the importance of sulfur-related metabolism in host-pathogen interactions 35 36 37 38 , the deficiencies in our knowledge highlighted by this analysis represent viable targets for experimental enquiry. For the carbon and nitrogen sources accessible by AJRecon our results were comparable, however STM_v1.0 has the ability to metabolize 20 carbon sources and 15 nitrogen sources not accessible to AJRecon. The additional metabolic capabilities of STM_v1.0 are due, in part, to the presence of ~200 additional gene products in STM_v1.0.

To combat the rise in antibiotic-resistant pathogens, it is crucial to identify new drug targets. Genes or sets of genes that are essential for growth are potential drug targets. To identify novel drug targets in STM_v1.0, we performed single and double gene deletion studies. We identified 201 essential genes in M9/glc, 144 of which were also essential in LB (Additional file 1 Table S5a). The synthetic lethal gene pair simulations were performed using only genes that were found to be non-essential in the condition of interest (Additional file 1 Table S6). In M9/glc, there were 87 synthetic lethal gene-pairs comprised of 102 unique genes. For E. coli, Suthers et al. 39 predicted 86 synthetic lethal gene-pairs, however, there were only 83 unique genes involved. In LB, there were 56 synthetic lethal gene-pairs comprised of 76 unique genes. Interestingly, 10 of LB synthetic lethal genes were also essential in M9/glc and were members of 12 of the LB synthetic lethal gene-pairs. The very small fraction of essential synthetic lethal gene pairs (< 100 synthetic lethalities out of >500,000 possibilities - assuming approx. 1000 non-essential metabolic genes) emphasizes the robustness of S. Typhimurium's metabolic network, which has previously been noted 5 .

Our observed, very small number of synthetic lethal pairs in STM_v1.0 indicates that antimicrobial regimens may need to target more than two elements to be effective. Unfortunately, it will take less time for a pathogen to evolve a solution to a conjoint two-target antimicrobial strategy compared to a single-target strategy. To reduce the probability of a pathogen evolving resistance to a conjoint two-target strategy, it may be plausible to employ a combination of two-target strategies. Although a combination approach may be suitable for dealing with antibiotic resistance, there are potential shortcomings associated with clearance and toxicity because all the components of a regimen must reach a target at a specific time with the requisite concentrations. Despite these difficulties, multi-component, multi-target drugs are becoming standard therapeutics for complex diseases, including cancer, diabetes, and infectious diseases 40 . Experimental identification and characterization of therapeutic strategies that require multiple targets for effectiveness is a resource intensive undertaking (e.g., creating over 500,000 double mutant strains). An in silico approach using an MR, such as STM_v1.0, could be implemented to prioritize the experiments by indicating which multi-target therapies would adversely affect the pathogen's metabolic capabilities.

As mentioned above, the synthetic gene deletion analysis yielded 56 synthetic lethal gene pairs disrupting growth of S. Typhimurium in silico. We grouped these gene pairs based on different criteria to assess their potential value as multi-drug targets (Figure 2). It is notable that five gene pairs are between protein complexes while a further three gene pairs are between genes involved in the same pathway - this indicates the presence of a layer of 'redundancy' for the enzyme or pathway that confers protection against a single-target therapy. Moreover, three of the genes involved in gene pairs are known to be essential for virulence, but not for growth, and have known inhibitors based on BRENDA 41 . This structured overview of in silico synthetic lethal gene pairs identified numerous candidate drug targets many of which have known inhibitors. In subsequent studies, these model-generated hypotheses need to be tested and validated.

Additional gene products shown to play a central role in virulence yet are not essential for growth in laboratory conditions or do not have an unequivocal functional annotation represent additional therapeutic targets. These gene products could serve as potential points for manipulating host metabolism 38 , could be essential for metabolism in the host environment (e.g., Salmonella-containing vacuoles are nutrient poor) 42 , and will represent an energy and materials demand when creating integrated metabolic and expression reconstructions 8 9 . Recent examples of relevant gene products that have not been annotated but are crucial for virulence include gene products STM3117-STM3120 43 ; as the metabolic functions of unannotated genes are elucidated, they will need to be incorporated into future revisions of the MR.

The starting reconstructions, AJRecon and BRecon, were built on scaffolds derived from published E. coli MRS. AJRecon is a pre-publication version of iRR1083 20 , and was based on iJR904 26 . For its scaffold, BRecon (Bumann, unpublished) employed iAF1260 27 - a direct descendent of iJR904. The two reconstructions, differ in content due to: (1) different components being targeted for manual curation (e.g., BRecon extended Fe chelation and AJRecon extended lipid production), and (2) differences in E. coli MRs that were used as comparative genomics scaffolds for initializing the Salmonella MRs (e.g., iAF1260 accounted for the periplasm whereas its ancestor did not).

There are three essential phases for community-based MR development: (1) preparation, (2) jamboree, and (3) reconstruction finalization 24 . The preparation and finalization phases are carried out by a small contingent of researchers, whereas, the collective knowledge of the community is harnessed during the jamboree. In the preparation phase, the two MRs were compared as described below in terms of metabolites, reactions, and gene-protein-reaction associations (GPRs). Overlapping content between both original MRs was directly moved into the consensus MR (Additional file 1 Table S1). Discrepancies in the listed three areas were presented to the jamboree team, which was split into three groups: metabolite curation, reaction curation group, and GPR curation group. The metabolites group curated the list of all metabolites present in either original MR for i) protonation state of metabolites at physiological pH, ii) missing metabolite identifiers: KEGGID, PubChemID, ChEBI ID, and iii) comparison of neutral formulae in reconstruction and metabolite databases. The reaction group was responsible for identifying evidence for orphan reactions in either original MR with and without a KEGG reaction ID. Reactions without a KEGG ID had to be extensively audited as there were no database evidences for the correctness of the reaction mechanisms. The GPR group had to resolve the discrepancies in GPR assignments using genome databases and literature. Each team evaluates their problem set based on evidence within the consensus MR and available resources (literature, databases, and annotations). Items that are not adequately addressed during the jamboree are subject to extensive manual curation during the MR finalization phase. The finalization phase includes: (1) manual curation, (2) benchmarking the consensus MR against experimentally-derived phenotypic data, and (3) MR dissemination. The consensus MR is expected to be maintained, updated and expanded in subsequent reconstruction jamborees.

Reconciling multiple MRs requires that the MRs' contents employ a common nomenclature so that the contents may be compared. For this work, we employed the KEGG database 44 as the source of common identifiers (Figure 1); although all of the reactions and metabolites in KEGG may not be accurate or complete, KEGG has the benefit of being an extensive, freely accessible resource used by the broader biological community. The complete consensus reconstruction can be found in Additional file 1 Table S6 and in Additional file 2 as an SBML file.

Thermodynamic directionality for each reaction was calculated as described in 45 . Briefly, assuming a temperature of at 310.15 K, intracellular pH of 7.7, extracellular/periplasmic pH of 7.0, and a concentration range of 0.01-20 mM, we calculated upper and lower bounds on transformed reaction Gibbs energy, and assigned reaction directionality accordingly. Transport reactions were not subject to thermodynamic consistency analysis as there is still uncertainty associated with the directionality prediction of transmembrane transport.

The conversion of a reconstruction into a mathematical model has been described in detail elsewhere 14 . The unit of reaction fluxes was defined as mmol/g_DW/hr.

Flux balance analysis 46 was employed to assess the STM_v1.0 model's ability to correctly predict biomass production in a variety of limiting conditions. The accuracy of the model was assessed by comparing the predictions to benchmarks drawn from experimental data 20 29 . In this assessment, there are four possible observations: (1) STM_v1.0 model correctly predicts growth (G/G), (2) STM_v1.0 model incorrectly predicts growth (G/NG), (3) STM_v1.0 model correctly predicts no growth (NG/NG), and (4) STM_v1.0 model incorrectly predicts no growth (NG/G). For a prediction to be counted as a true positive (G/G) or true negative (NG/NG), the prediction needed to match one or more experimental observations. The predictions were first compared with the Biolog phenotype microarray (PM) data http://www.biolog.com. False positive predictions (G/NG) and false negative predictions (NG/G) were then compared with the data from Gutnick et al. 29 and references cited in Ragunathan et al. 20 . For limiting conditions not represented in the PM, predictions were only compared with data from Gutnick et al. 29 or cited in Ragunathan et al. 20 .

The gene deletion studies were performed by converting STM_v1.0 into a stoichiometric model and performing flux balance analysis 46 . For each gene, or gene pair, the associated reaction(s) were disabled (v_{min, i} = v_{max, i} = 0 mmol.gDW^-1.hr^-1) and the ability of the model to produce biomass was assessed, i.e., the biomass reaction was chosen as the objective function and maximized.

All simulations were performed using the COBRA Toolbox v2.0 47 using Matlab (Mathworks, Inc) as the programming environment, and Tomlab (TomOpt, Inc) as the linear programming solver.