The genome scale model of A. baylyi was iteratively reconstructed following a process depicted in Figure 1. We first built an initial draft model iAbaylyi^v1 using information from the genome annotation, metabolic pathways databases, and the literature. Although facilitated by the automated network reconstruction software PathoLogic 21, this initial reconstruction still required extensive manual curation (see Methods). The draft metabolic network generated by PathoLogic was first inspected to filter out and correct wrongly predicted pathways and reactions, and then completed by reviewing the expert genome annotations and the metabolic information contained in the literature. For instance, specific efforts were dedicated to properly include pathways accounting for the particular degradation capabilities of A. baylyi. Physiological information on A. baylyi was especially helpful to build the set of transport processes, as substrate specificities of transporters are difficult to deduce from genome annotation only. For each metabolite shown to be consumed by A. baylyi we added a corresponding transport reaction to the model. Out of 133 transporters, 23 were initially included in the model using this type of evidence only. The dependency between genes and reactions was modeled using Boolean rules, known as GPR (Gene-Protein-Reaction associations) 22. These rules encode the presence of isozymes or enzymatic complexes for the catalysis of reactions, and predict the effect of genetic perturbations on the activity of reactions. GPR rules were first derived using homology with E. coli enzyme complexes 23 and then completed by manual curation. In order to model the metabolic and energetic demands associated with growth, we introduced a set of intermediary biomass reactions that synthesize generic cell constituents (e.g. protein, DNA, RNA, or lipid) from precursor metabolites, and a global growth reaction consuming them in proportion defined by studies of biomass composition 2425. Energetic parameters required to predict quantitative growth rate using Flux Balance Analysis (FBA) were assumed to be similar to those of E. coli model (see Methods)22. No accurate measurement of A. baylyi growth yields could be used to validate these parameters, however. While such validation would be required to get more accurate predictions of growth yields, the current parameters already provide good approximate values (see Additional file 1 for a sensitivity analysis on these parameters). For the purpose of qualitatively predicting growth ability using Metabolite Producibility analysis (see Figure 2) 26, we designed a reduced list of biomass precursors which are all essential for growth in in vitro conditions. We used this list to predict qualitative growth phenotypes and compare them with those of phenotyping experiments on in vitro environments. In vivo environments may impose harsher conditions requiring additional metabolic responses; this list therefore represents a minimal set of essential precursors that may need to be expanded to properly predict growth phenotypes on more realistic environments 27. The Methods section provides more details on the reconstruction process.

This initial reconstruction process led to the model iAbaylyi^v1 gathering 859 reactions grouped in 7 metabolic categories and 697 distinct metabolites, 109 of which could be transported from the environment. As depicted in Figure 3, the model accounts for all main processes of A. baylyi metabolism, including biosynthetic routes, energy metabolism, and catabolic pathways. Genomic islands of catabolic diversity endow A. baylyi with the ability to degrade a wide variety of soil compounds 19. The metabolic model reflects this nutritional versatility, as 20% of its reactions are dedicated to the catabolism of external compounds. A list of specific compounds that can be degraded by A. baylyi is provided in Table 1.

We used results of large-scale growth phenotyping experiments to perform a first round of model assessment and refinement. Using Biolog assays, we experimentally tested the wild-type strain ability to use 190 distinct metabolites as sole carbon and energy sources (see Methods). Using the model, we predicted the growth phenotypes of the wild-type strain on the corresponding in silico media and compared them to the experimental results.

Out of the 190 screened metabolites, 45 were found to be carbon and energy sources for A. baylyi. This relatively small fraction of carbon sources can be explained by the fact that Biolog microplates are only partially adapted to A. baylyi's biotope: they feature sugars, nucleosides or amino acids but relatively few chemicals originating from plant compounds. iAbaylyi^v1 model predicted 24 of them and missed 21 (see Figure 1). Eight of the missed carbon source metabolites were already present in the model, but with no associated transporter. Amongst them, seven would also be predicted as carbon and energy sources had the corresponding transporters been included. In order to resolve these inconsistencies, we added for each of them a generic transport reaction accounting for A. baylyi's ability to utilize these compounds (see Table 2). Growth on the remaining metabolite (2-ketobutyrate) was contradicted by an additional individual growth experiment.

Thirteen carbon source metabolites were unknown to the metabolic model. For two of them, sorbate and tricarballylate, we were able to identify degradation pathways and add them to the model (see Table 2). Sorbate, an unsaturated fatty acid, can be degraded by fatty acids oxidation enzymes, which were already included in the model for the degradation of other fatty acids. Sorbate transport and degradation reactions were therefore added to the model using the same set of genes. Recently, genes coding for tricarballylate transport (tcuC), oxidation to cis-aconitate (tcuA and tcuB), and for a regulatory protein required for tcuABC expression (tcuR) were identified in Salmonella enterica 2829. Highly homologous genes could be found in synteny in A. baylyi: ACIAD1536 (tcuB, 59% identity), ACIAD1537 (tcuA, 76% identity), ACIAD1541 (tcuC, 64% identity), ACIAD1539 (tcuR, 46% identity), and ACIAD1543 (tcuR, 44% identity). Following these clues, we expanded the model by implementing the corresponding transporter and degradation reaction, and annotated the corresponding genes. In four cases, dedicated growth experiments contradicted the Biolog result, weakening the case for further study (see Table 2). Finally, no relevant pathway could be found for the remaining seven unmodeled carbon sources. Further investigations are needed to identify the metabolic processes allowing A. baylyi to exploit these metabolites.

Conversely, only five of the 145 non-carbon source metabolites were wrongly predicted to be carbon sources by the model: 4-hydroxybenzoate, D-fructose, L-arginine, L-ornithine, and D-serine (see Figure 1 and Table 2). Experiments from 15 contradicted the Biolog result on 4-hydroxybenzoate, while additional individual experiments confirmed the Biolog results of the other four.

Interestingly, A. baylyi annotation describes a complete phosphotransferase (PTS) transport system for fructose (ACIAD1990 and ACIAD1993, fruA &fruB) coupled with a 1-phosphofructokinase (ACIAD1992, fruK) leading to fructose-1,6-bisphosphate (see Figure 5). In accordance with the annotation, the model predicts that fructose should be a carbon and energy source, yet this is not observed experimentally. To confirm the ability of the PTS system to transport fructose, we assessed experimentally the growth phenotype of the fructose bisphosphate aldolase (ACIAD1925, fda) knockout mutant (see Figure 5). The ΔACIAD1925 mutant could not be obtained on succinate-supplemented minimal media, reflecting the fact that Fda is required in the gluconeogenesis pathway to provide fructose-1,6-bisphosphate, an essential intermediate for building pentose-phosphates and polysaccharides. The mutant could however be obtained by adding fructose in the medium, showing that fructose could be imported into the cell and converted to fructose-1,6-bisphosphate. The reason why A. baylyi is unable to use fructose as a sole carbon source remains yet to be investigated. Hypothetically, A. baylyi may be unable to use the Embden-Meyerhof-Parnas (EMP) pathway in the glycolytic direction, as it has been observed for the dissimilation of glucose 1930.

As is the case in E. coli, L-ornithine and L-arginine are degraded by A. baylyi using the arginine succinyltransferase (AST) pathway. This pathway allows E. coli to use them as nitrogen sources, but not as carbon sources. Putative explanations include unsuitable regulation and inadequate transport 31. Similar reasons may explain A. baylyi's inability to use L-ornithine and L-arginine as carbon sources.

A. baylyi's genome annotation includes genes for D-serine transport (ACIAD0118 and ACIAD2662, cycA) and D-serine deaminase activity (ACIAD1048 dsdA), which should allow it to use D-serine as a carbon and nitrogen source. The interpretation of this inconsistency is also unclear; a similar unexplained inconsistency was pointed out in a study involving a metabolic model of B. subtilis 4.

Improvements to the model resulted in iAbaylyi^v2, raising predictive accuracy on Biolog-measured phenotypes from 86% to 91% of the growth phenotypes (see Figure 1). Detailed results of the comparison with Biolog results can be found in Additional file 3.

In steps 2 and 3 of the model refinement process, we assessed and improved the model by comparing its predictions to experimentally determined gene essentialities, derived from the ADP1 mutant collection 8 (see Figure 1). Growth phenotypes of all single gene deletion mutants on the corresponding environments were predicted using metabolite producibility analysis (see Figure 2 and Methods). Predicted phenotypes were then compared to the genome-wide gene essentiality results in order to assess the accuracy of the model and to identify inconsistent predictions. Inconsistencies could be either false essential (genes falsely predicted essential by the model) or false dispensable (genes falsely predicted dispensable by the model) predictions. Since these inconsistencies are as many clues that the understanding of A. baylyi's metabolism represented in the model is erroneous or incomplete, we examined them carefully in order to find interpretations and, when needed, refine the model.

We classified refinements into three categories according to the model component that was modified: GPR, NETWORK or BIOMASS (see Figure 2). These three components model different kinds of biological processes which contribute to determining the growth phenotype of mutant strains (see Methods). The GPR component, consisting of the GPR Boolean rules, computes the effect of the genetic perturbation on the activity of reactions in the model. The NETWORK component, the actual network of reactions, models the metabolic conversion capabilities of the organism. Finally, the BIOMASS component, consisting of the list of metabolites required for growth, models the biomass precursor requirements of the organism.

The overall refinement process led to the final model iAbaylyi^v4 gathering 774 genes, 875 reactions and 701 metabolites (see Figure 1). iAbaylyi^v4 integrates all refinements resulting from the three experimental datasets introduced in this work. Accordingly, its predictions are consistent with the experimental results in 91% of the cases for dataset 1, 94% of the cases for dataset 2, and 94% of the cases for dataset 3. Compared with iAbaylyi^v1, it was expanded by 19 reactions and 2 genes, while 3 reactions and 16 genes were removed in the refinement process (see Figure 1, Model corrections).

In order to facilitate the exploration of A. baylyi metabolism using the genome scale model, we created NemoStudio 20 (Combe et al, in preparation), a web interface combining a simulation layer for the model with AcinetoCyc, A. baylyi Pathway-Genome Database 21. NemoStudio gathers data on functional genomics annotations, metabolic reactions and pathways, and experimental mutant phenotyping results within a single interface. Additionally, it allows performing phenotype predictions using the constraint-based model.

AcinetoCyc gathers information on the metabolic network of A. baylyi and is used to display interactive metabolic maps. After its initial automated construction using PathoLogic 21, AcinetoCyc has been undergoing constant curation. It includes all metabolic reactions present in the model.

NemoStudio integrates the latest version of A. baylyi metabolic model, iAbaylyi^v4. Growth phenotype predictions can be performed for any set of environmental conditions and genetic perturbations of this study. We implemented both Flux Balance Analysis (FBA) and Metabolite Producibility methods to predict growth phenotypes (see Methods). When performed on sets of environmental conditions and sets of gene deletions, prediction results are displayed in a table format in parallel to the actual experimental results. Predictions can thus be readily compared with the experimental observations. Furthermore, predicted and experimental phenotypes are both displayed on AcinetoCyc metabolic maps, and conversely gene deletions can be directly set from these metabolic maps (see Figure 9). When performed for a single environment and a single genetic perturbation, FBA predicts an optimal flux distribution towards biomass production; these fluxes are both displayed in a table and on AcinetoCyc metabolic pathways.

The availability of this resource as a web interface makes it easily usable by scientists interested in A. baylyi metabolism. Compared with previous web-based software for genome-scale metabolic modeling 27, the A. baylyi NemoStudio interface provides better interactivity, direct visualization of results on metabolic maps and integrated comparison with experimental data. By interfacing as much as possible results deriving from systems level analyses with experimental data of various forms, it allows the simultaneous exploitation of both information types.

The initial reconstruction of the metabolic network was carried out using data provided by (i) the genome expert annotation 19, (ii) the BioCyc metabolic pathway database automatically generated from these annotations 21 and (iii) various literature resources on biochemistry, including textbooks, reviews and journal publications (see Additional file 2). The genome annotation was downloaded from the MaGe interface 5051 and used as input of the Pathway Tools software 21 in order to generate a BioCyc automatic reconstruction of the metabolic network. The predicted pathways were classified into 7 metabolic categories (central metabolism, nucleotide metabolism, amino acids metabolism, lipid & cell wall metabolism, degradation pathways, cofactor biosynthesis, transport) and examined manually before being included in the model. In order to meet the requirements of the modeling framework the mass balance and reversibility of the reactions were checked.

Reversibility of the reactions was determined from literature evidence when available or based on simple thermodynamic considerations 52. Proton translocation efficiencies of reactions of the respiratory chain were assumed to be similar to those of E. coli 53. Resulting P/O ratio can range between 0.5 to 2, depending on the types of cytochrome oxidase and NADH dehydrogenase that are used. Reactions using generic compounds (for example a nitrile or a polymer of undetermined length) were instantiated with defined representative metabolites. In this respect, polymeric pathways were expanded into chains of specific reactions. Large polymeric molecules such as the acyl carrier protein (ACP) or tRNAs were included in the model when they were involved as substrate cofactors of biochemical reactions. Their specific synthesis was not considered in the model. Dependency between reactions and genes were coded by Gene-Protein-Reaction (GPR) Boolean relationships (see below). Using the Cyclone interface to BioCyc 54, we implemented a simple method based on gene homologies between Escherichia coli and Acinetobacter baylyi to infer enzyme complexes and find AND Boolean associations between genes. Information from the literature was used to close gaps in the metabolic pathways, include pathways specific to A. baylyi that were unknown to the metabolic databases, and check the predicted pathways, for instance for the specificity of the cofactors. Physiological information derived from the literature 155556575859 was used together with genome annotation tools, e.g. TransportDB 60, to add transport reactions in the model. A generic transport reaction was added to the model for each metabolite shown to be utilized by A. baylyi. A fixed biomass composition was chosen according to data found in the literature for strains growing on standard media (see Additional file 4). This biomass composition was used to build the reduced list of essential biomass precursors and derive a biomass reaction for Flux Balance Analyses (see below). To help properly account for all metabolic requirements associated with growth, we decomposed the biomass reaction into a set of intermediary biomass reactions synthesizing generic cell constituents (e.g. protein, DNA, RNA, or lipid) from precursor metabolites and a global growth reaction consuming them according to the chosen biomass composition. See Additional file 4 for details on these reactions.

The metabolic model is composed of three components, namely GPR, NETWORK and BIOMASS. The GPR component models the dependency between genes and reactions using Boolean functions usually called gene-protein-reaction (GPR) associations 22). For each reaction, a Boolean rule encodes how genes are related to the activity. Genes that are required together are linked with an AND relation while isofunctional genes are linked with an OR relation. The set of GPR associations yields the set of potentially active reactions given the set of available genes.

The NETWORK component models the metabolic network using the constraint-based modeling framework 3. This framework describes the distributions of reaction fluxes that are compatible with constraints that derive from basic physical assumptions or specific biological information. They are usually formulated as linear constraints, which allow to explore the fluxes solution space using linear programming tools. The main constraint is imposed by the steady-state assumption, represented by the matrix equation:

S·v = 0

where S is the stoichiometric matrix of the metabolic network and ν the vector of reaction fluxes. The stoichiometric matrix is a matrix of size (m × n) where m is the number of metabolites and n the number of reactions. Each element S_i,j of the matrix represents the relative stoichiometric coefficient of metabolite i in reaction j. Additional constraints on the fluxes, such as irreversibility and capacity constraints, are imposed by inequalities in the form:

ν_{lb, i} ≤ ν_i ≤ ν_{ub, i}

where ν_lb,i and ν_ub,i are respectively the lower and upper bounds of the flux of reaction i.

Environmental conditions are applied to the model by constraining the exchange fluxes of extracellular metabolites. Exchange fluxes are sink reactions allowing to control the input or output of metabolites in the model. They are constrained to 0 ≤ ν_i ≤ ∞ for metabolites absent from the medium and -∞ ≤ ν_i ≤ ∞ for metabolites present in the medium, except for limiting nutrients for which a maximum uptake rate is chosen (-ν_uptake ≤ ν_i ≤ ∞). When simulating the metabolic network of a knockout mutant, the activity of each reaction is determined by evaluating its GPR association according to the set of removed genes. Fluxes of the inactivated reactions are constrained to be equal to zero.

The BIOMASS component models the essential metabolic requirements for growth. It consists of a list of metabolites that are considered to be essential biomass precursors. Growth phenotype is therefore determined by checking their producibility 26. To do so, the steady-state constraints for the essential biomass precursors are changed to strict producibility constraints:

{ S i n t e r n a l ⋅ ν = 0 S b i o m a s s p r e c u r s o r s ⋅ ν ≥ ε ν l b , i ≤ ν i ≤ ν u b , i MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaWaaiqaaeaafaqaaeWabaaabaGaem4uam1aaSbaaSqaaiabdMgaPjabd6gaUjabdsha0jabdwgaLjabdkhaYjabd6gaUjabdggaHjabdYgaSbqabaGccqGHflY1cqaH9oGBcqGH9aqpcqaIWaamaeaacqWGtbWudaWgaaWcbaGaemOyaiMaemyAaKMaem4Ba8MaemyBa0MaemyyaeMaem4CamNaem4CamNaemiCaaNaemOCaiNaemyzauMaem4yamMaemyDauNaemOCaiNaem4CamNaem4Ba8MaemOCaiNaem4CamhabeaakiabgwSixlabe27aUjabgwMiZkabew7aLbqaaiabe27aUnaaBaaaleaacqWGSbaBcqWGIbGycqGGSaalcqWGPbqAaeqaaOGaeyizImQaeqyVd42aaSbaaSqaaiabdMgaPbqabaGccqGHKjYOcqaH9oGBdaWgaaWcbaGaemyDauNaemOyaiMaeiilaWIaemyAaKgabeaaaaaakiaawUhaaaaa@7494@

where S_internal is the stoichiometric matrix without the biomass precursors, S_{biomass precursors} the stoichiometric matrix restricted to the biomass precursors and ε a vector of small reals, taken as 10^-3. Linear programming tools are used to query for a flux distribution fulfilling this set of constraints. If a flux distribution could be found, the model predicted growth, otherwise it predicted no growth.

In order to assess quantitative growth defects, Flux Balance Analyses (FBA) were performed 3. A biomass reaction was introduced in the model to quantitatively account for the respective contributions of constituent metabolites in the biomass composition (see Additional file 4). Using linear programming, the flux through this reaction was maximized under all constraints, representing the maximal growth rate achievable by the model. Energetic parameters, including growth associated (GAM) and non growth associated (NGAM) maintenance fluxes, were assumed to be similar to those of E. coli model 22. We chose to set NGAM to a constant ATP hydrolysis flux of 10 mmol/h/gDW and GAM to a value of 40 mmol/gDW of ATP in the growth reaction. In all simulations, upper bounds of nutrient exchange fluxes were set to 10 mmol/h/gDW for carbon sources and 100 mmol/h/gDW for other nutrients (see Additional file 2).

Model simulations were performed within FluxAnalyzer 61 and MATLAB^® (The MathWorks Inc., Natick, MA) using the YALMIP optimization toolbox 62 and MOSEK optimization solver (Mosek ApS, Copenhagen, Denmark).

The metabolic model is available both as Excel and SBML files (see Additional files 2 and 5) and will be submitted to the Biomodels.net repository 63. Whenever possible, cross-references for the model reactions and species to AcinetoCyc 20, KEGG 64 and BiGG 65 databases are provided.

The model is accessible through the NemoStudio web interface 20. NemoStudio supports growth phenotype predictions, and comparison to experimental results, as well as browsing of model pathways through an interface with AcinetoCyc 20.

Growth phenotyping experiments of A. baylyi were performed by Biolog, Inc. (Hayward, CA) following experimental procedures described in 66. Basically, growth of wild-type strains of A. baylyi was monitored in PM1 and PM2 microplates containing a defined minimal medium supplemented with 190 distinct carbon sources. The Biolog quantitative growth measures were discretized to yield growth/no-growth qualitative phenotypes by choosing thresholds based on the negative growth control measures and previously known growth phenotypes for A. baylyi. Growth phenotypes that were inconsistent with model predictions were checked by examining results from previous work 15, or retesting them individually. Detailed results of Biolog experiments are provided in Additional file 3.

Detailed experimental protocol for the growth phenotyping of the mutant strains is described in 8. Basically, using 96-wells plates, the mutant strains were grown in liquid MA minimal media (31 mM Na2HPO4, 25 mM KH2PO4, 18 mM NH4Cl, 41 μM nitrilotriacetic acid, 2 mM MgSO4, 0.45 mM CaCl2, 3 μM FeCl3, 1 μM MnCl2, 1 μM ZnCl2, 0.3 μM (CrCl3, H3BO3, CoCl2, CuCl2, NiCl2, Na2NoO4, Na2SeO3)) supplemented with 25 mM of carbon sources. Succinate/urea medium was composed of MA minimal medium without NH4Cl supplemented with 25 mM of succinate and 20 mM of urea. Absorbance at 600 nm of 24 h cultures was measured to monitor growth. Experiments were performed in duplicates. Measures with discrepant repeats or with weak precultures were discarded from the analyses. Repeats were filtered according to the following rule: a measure was kept if either (1) both repeats were under the growth threshold or (2) the relative difference between the repeats was lower than 50% of the highest value. A threshold of a tenth of the mean absorbance was chosen to classify the mutants in growth or no growth categories. This threshold was chosen particularly low in order to consider as essential only mutants with marked fitness defect.