Characteristics of the genome-scale metabolic network of B. cenocepacia J2315

The genome-scale reconstructed metabolic model of B. cenocepacia J2315, referred by the conventional naming rules [17] as i KF1028, consists of 859 internal reactions (including transport) and 834 metabolites. The reconstruction accounts for 1,028 genes, covering 14.4% of the 7,116 protein coding genes identified from whole genome sequencing (http://www.ncbi.nlm.nih.gov/genome?term=burkholderia%20cenocepacia%20J2315). The model i KF1028 includes all major pathways required for cell growth and the degradation of common biologically important carbon sources of B. cenocepacia. Apart from these central metabolic pathways, model i KF1028 also includes pathways associated with key metabolic virulence factors, which provides insights into how the system-level metabolic properties affect pathogenicity. For an overview, the properties of the J2315 genome and the reconstructed model i KF1028 were summarized in Table 1. Genome-scale metabolic models have been successfully used to study many pathogenic bacteria, including Staphylococcus aureus[18–20], Acinetobacter baumannii[21], Mycobacterium tuberculosis[22], Salmonella typhimurium[23], and Pseudomonas aeruginosa[24]. A basic comparison between the model i KF1028 and the above five recently published metabolic reconstructions is also illustrated in Table 1. Schematic representation of the metabolic network of B. cenocepacia J2315 with key metabolic virulence factors is shown in Figure 1. Figure 2 enumerates the metabolic pathways included in i KF1028 and, for each pathway, the number of reactions with assigned and non-assigned genes. The high ratio of gene-associated reactions shows that the reconstructed metabolic model of B. cenocepacia J2315 is reliable. [Additional file 1 for i KF1028 and Additional file 2 is in SBML format]

Model	N.A.*	AbyMBEL891	i NJ661	i RR1083	i MO1056	i KF1028
Genome size	2.8 Mb	3.93 Mb	4.4 Mb	4.8 Mb	6.3 Mb	8.1 Mb
Included genes	758	650	661	1,083	1,056	1,028
Total reactions	1,497	891	939	1,087	883	859
Gene-associated reactions (% of total reactions)	1,278	713 (80%)	723 (77%)	1,018 (93.7%)	839 (95%)	832 (96.9%)
Non-gene-associated reactions	219	46	216	69	44	27
Transport reactions	146	130	93	230	133	102
Metabolites	1,431	778	828	744	760	834

Properties of metabolic reconstruction of B. cenocepacia J2315 (iKF1028) were compared with other published metabolic reconstructions of pathogenic microbes, S. aureus N315 (2009) [18] (which is the improvement of the N315 reconstruction by Becker and Palsson 2005 [19], and Heinermann et al. 2005 [20]), A. baumannii AYE (AbyMBEL891) (2010) [21], M. tuberculosis H37Rv (iNJ661) (2007) [22], S. typhimurium LT2 (iRR1083) (2009) [23], and P. aeruginosa PAO1 (iMO1056) (2008) [24]. *: Not available for the reconstruction of S. aureus N315.

Metabolic virulence factors in model i KF1028

The LPS produced by B. cenocepacia J2315 has an important role in both disease aetiology and antibiotic resistance [51, 52]. LPS usually consists of three components: lipid A, core oligosaccharide, and O antigen. Although there were some studies on characterizing the features of LPS in B. cenocepacia, all these studies focused on a certain part/component of LPS. So far, there is no systematic elucidation of the LPS structure and composition specifically for B. cenocepacia strain J2315, nor any global analysis on its biosynthesis process of the LPS. In this study, we depicted the detailed features of the complete LPS structure in B. cenocepacia J2315 by integrating all available reports on LPS. We also reconstructed the LPS-synthesis pathways supplemented with all necessary proteins involved and major metabolic precursors, as illustrated in Figure 3. According to our study, in J2315, each of the three components has a very unique feature. The lipid A portion is modified by an additional Ara4N residue [49, 53], which had been shown to reduce the binding of cationic antibiotics and was proposed as a potential drug target [54]. The inner core oligosaccharide contains an unusual KDO-KO-Ara4N residue instead of the typical KDO-KDO residue [15, 51, 55]. The outer core comprises various polysaccharides including L-glycero-D-manno-heptose, glucose, galactose, quinovosamine, and rhamnose [51]. The O-antigen portion of LPS in J2315 was interrupted by an insertion element in BCAL3125 [47, 56]. These differences might indicate the reason why strain J2315 is of remarkably distinct activity.

B. cenocepacia strains possess multiple quorum sensing systems, which regulate the expression of versatile virulence determinants, such as biofilm formation and motility. Strain J2315 owns the ability to synthesize and recognize three types of chemical signals used for cell-to-cell communication: N-acylhomoserine lactones (AHLs), 4-quinolones (4Qs), and the DSF-like molecule cis-2-dodecenoic acid (BDSF). Two AHLs-based QS systems have been found in J2315, namely CciIR and CepIR [25, 57], which can both produce N-hexanoyl-L-homoserine lactone (C6-HSL) and N-octanoyl-L-homoserine lactone (C8-HSL) signals using acyl side chain (Hexanoyl-ACP and Octanoyl-ACP, respectively) and S-adenosyl-methionine (SAM) as precursors [58]. The CepIR system is conserved in all species of the Bcc. The CciIR system is encoded within a pathogenicity island, designated as the B. cenocepacia island (cci), which was the first time that cell-signalling genes were found on a genomic island [59]. The 4Qs-based signal, the 2-heptyl-4-quinolone (HHQ), is produced by B. cenocepacia strains [26]. HHQ is the precursor of 2-heptyl-3-hydroxy-4-quinolone (PQS) [60] and its synthesis requires four proteins: PqsA, PqsB, PqsC, and PqsD. It had been reported that the exported HHQ from B. cenocepacia can be recognized by Pseudomonas aeruginosa within which HHQ is converted into PQS which is one of the QS signals for P. aeruginosa[26], highlighting the possibility of inter-species communication during the CF co-infection caused by P. aeruginosa and B. cenocepacia. BDSF is a newly discovered signal molecule produced by B. cenocepacia[28]. The synthesis of BDSF requires the gene BCAM0581 [29].

The synthesis pathway of rhamnolipids was also reconstructed in i KF1028. Although there has not been any report demonstrating that B. cenocepacia can produce rhamnolipid, Dubeau et al demonstrated that Burkholderia thailandensis has the orthologs of rhlA, rhlB, and rhlC, which are responsible for the biosynthesis of rhamnolipids in P. aeruginosa[61]. By protein similarity search against the UniProt database, proteins coded by genes BCAM2340, BCAM2338, and BCAM2336 in B. cenocepacia J2315 were identified as highly similar in sequence to rhlA, rhlB, and rhlC in both B. thailandensis (with BLAST E value of 1E-121, 1E-173, and 1E-108, respectively) and P. aeruginosa (with BLAST E value of 3E-60, 7E-98, and 1E-67, respectively). This facilitates us to hypothesize that B. cenocepacia can potentially generate rhamnolipids. Further experimental investigations are needed.

Model validation and gap-filling using phenotype data

Of the remaining 55 carbon sources tested, 14 were indirectly compared with the model due to the missing knowledge of whether the transport mechanisms of these compounds exist in J2315 or not. Initially, all those 14 carbon sources showed a no-growth phenotype both in silico and in the BIOLOG assays. Then we made the assumption that each of these carbon sources could be transported into the cell (i.e. to function as intracellular compounds) and re-tested whether the in silico model can grow on each of them. The results showed that 11 of the 14 carbon sources enabled i KF1028 to grow after applying the above assumption. This supports the hypothesis that J2315 lacks of transporters for all those 11 carbon sources, even though their catabolic pathways are complete. For the rest 3 carbon sources, the agreement between the in silico results and BIOLOG assays remained.

As the catabolism of the remaining 41 carbon sources out of 55 has not been well studied and information regarding their role in the cell could not be found in public resources, these 41 carbon sources cannot be analyzed in our model. (Complete comparison results with BIOLOG assays are supplied in the Additional file 3)

Model-driven refinement of genome annotation

The first type of refinement was to re-annotate genes in i KF1028 - based on literature evidence and BLAST searches - that have been improperly annotated. An example is the gene BCAL1281, which was annotated in both the Burkholderia Genome Database and KEGG as a "hypothetical protein", but that was now reassigned coding for an "ornithine N-acyltransferase". It was reported that the outer membrane of "B. cepacia" [62] possesses unusual polar lipids, ornithine amide lipids (OL) [63, 64]. In addition, the protein OlsB is required for the first step of OL biosynthesis [65]. By BLAST searches of OlsB against the UNIPROT database, the gene BCAL1281 of B. cenocepacia J2315 was identified with high similarity.

Another type of annotation refinement was based on gap analyses, which pinpointed reactions for which the gene products involved were missing. For instance, we identified a missing reaction that should be catalyzed by IspE and that takes part in the biosynthesis of polyprenyl-PP, a necessary precursor of the ubiquinone biosynthesis. The IspE encoding genes for other strains of B. cenocepacia (AU1054, HI2424, MC0-3) could be identified in the Burkholderia Genome Database and KEGG. By querying GeneDB, we found that the genomic location from 872938 to 873820, named BCAL0802, was not assigned any function in the above two databases. A BLAST search of BCAL0802 against the UNIPROT database revealed a perfect match with IspE from other B. cenocepacia strains. Consequently, BCAL0802 is annotated as a 4-diphosphocytidyl-2-C-methyl-D-erythritol kinase and this missing reaction was supplemented into the model i KF1028. This example exemplifies how the reconstruction process can drive the reconciliation of isolated data from different biological databases.

BIOLOG substrate utilization assays have already been successfully used on the refinement of metabolic reconstructions [24]. It is an efficient approach for gap analysis and refinement of genome annotation. For example, in our study we can highlight the case for the substrate D-galactose, associated in BIOLOG assays with a growth phenotype. From that result we inferred that J2315 should contain a transport mechanism for D-galactose. By homology searches of MglA, MglB, and MglC, which are galactose-binding proteins conveying galactose into the cell, gene BCAL1431 (mglB), BCAL1432 (mglA), and BCAL1433 (mglC) were identified and they had been annotated as a putative sugar ABC transporter ATP-binding protein, a putative ribose ABC transport system, and a putative sugar transport system permease protein respectively. The ordering of mglB, mglA, and mglC is consistent with a) previous studies indicating that the mgl operon contains three genes and the genes are transcribed in the order of mglB, mglA, and mglC [66] and b) mglA and mglC being located downstream of mglB [67]. As a result, the annotations of BCAL1431-1433 were refined to account for the galactose transport. In total, 7 genes were reannotated based on BIOLOG assays.

Gene essentiality analysis

The term 'essential gene' means a gene for which knockout is lethal (i.e. no biomass yield) under certain conditions (e.g. glucose minimal medium) [68]. Identification of essential genes is helpful to understand the basic functions required for survival and it is an efficient way to discover novel targets for new antimicrobial therapies. Here in this study, i KF1028 was used as a framework to predict computationally identified essential genes in B. cenocepacia J2315 on both M9 minimal medium and synthetic CF medium (SCFM). About 19% (192) and 15% (154) of the 1,028 metabolic genes included in i KF1028 were predicted to be essential on M9 and SCFM media, respectively. There are more genes predicted as essentials on M9 than on SCFM, which indicates the influence of the living environment on the bacterium. 147 overlapping predictions were on both media. These essential genes are unequally located on the three chromosomes and most of the essential genes are located on chromosome 1 (Figure 4a). This result agrees with known features of the J2315 genome: chromosome 1 contains a higher proportion of coding sequence (CDS) involved in central metabolism and other house-keeping functions, whereas chromosomes 2 and 3 contain a greater proportion of CDS encoding accessory functions [13].

To assess the predictive potential of the model, we compared the in silico essential genes predicted on SCFM with experimental essentiality data for P. aeruginosa PAO1 and P. aeruginosa PA14 [69, 70] since there is no experimental gene essentiality data available for B. cenocepacia. As both P. aeruginosa and B. cenocepacia are CF pathogens, and B. cenocepacia was historically classified under the genus Pseudomonas[71], it is possible that partial similarity exists between them. SCFM was chosen due to its similarity to the nutritional composition of sputum from CF patients. The common set of the essential genes from two P. aeruginosa strains was chosen for comparison in order to reduce the effect of strain-dependent variation. Totally, there are 294 in silico essential metabolic and non-metabolic genes of B. cenocepacia J2315 with high similarity to the common set of P. aeruginosa by BLAST searches, of which 91 in vivo essential genes are present in i KF1028. Genes in the in vivo essential set but not in the i KF1028 were assumed to be involved either in non-metabolic functions or in accessory functions of metabolism. A total of 55% (50) of the in silico predicted essential genes agreed with the in vivo essential genes based on gene homology with P. aeruginosa (Figure 4b, refer to Additional file 4 for detailed information about essential genes).

Based on gene essentiality analysis, the genome-scale metabolic model of B. cenocepacia was further refined. For example, BCAL0660 and BCAL3421, which were homologous genes encoding protein AccC according to their annotation, were originally both included in the model with the gene-protein-reaction (GPR) relationship of "BCAL0660 or BCAL3421". Through in silico gene deletion study, BCAL3421 is identified as non-essential, which is inconsistent with the in vivo essential gene results. Such discrepancy was subject to further analysis. AccA, AccB, AccC, and AccD were four subunits of Acetyl-CoA Carboxylase (ACC) catalyzing the first step of fatty acid biosynthesis [72]. The gene accB, of which the locus in J2315 is BCAL3420, is frequently adjacent to the gene accC and both genes are cotranscribed and form an operon together [73, 74]. Furthermore, BCAL3420 and BCAL3421 shows greater than 2-fold expression for J2315 under CF conditions versus soil conditions, yet BCAL0660 shows an opposite result [75]. Taken together, BCAL0660 was excluded from model i KF1028. Further studies are necessary to validate the function of BCAL0660.

Identification of essential enzymes and potential drug targets

Essential enzyme/protein refers to a gene product (catalyzing the relevant reactions) for which individual deletion (i.e. imposing the fluxes through these reactions to zero) is lethal under certain conditions. Through FBA using i KF1028,we could obtain a collection of essential enzymes (protein), based on which 45 essential enzymes were identified as potential drug targets and supported by experimental evidences from literatures. There are 39 of them which were also predicted as drug targets for P. aeruginosa PAO1 [76]. All the 39 targets are nonhomologous to human protein sequences and thus could serve as potential candidate antibiotic drug targets for CF patients infected by both B. cenocepacia J2315 and P. aeruginosa PAO1. Among 39 targets, there are 9 targets, namely AccA, AccB, AccC, AccD, MurA, FolP, PhoA, RibE, and RibH, which have approved drugs in the DrugBank database [77].

Reconstruction of the metabolic network

The reconstruction process for B. cenocepacia J2315 is illustrated in Figure 5. The process followed the procedure described previously [81]. The reconstruction was carried out on ToBiN (Toolbox for Biochemical Networks, http://www.lifewizz.com), which was first mentioned in the paper [82]. ToBiN is a modular platform for metabolic modelling and the structural analysis of networks. It consists of a collection of open-source computational tools. Sets of reactions can be uploaded in the platform via a web interface, merged with already existing sets, and the resulting stoichiometric matrix is then processed by the server as a FBA problem. The linear solver that ToBiN used is the Clp (Coin-or linear programming), an open-source linear programming solver written in C++ and is part of the COIN-OR (Computational Infrastructure for Operations Research) project (http://www.coin-or.org). The platform works in a similar way as the COBRA toolbox with the main difference that, by being web-based, it permits users to adopt a more efficient and collaborative workflow.

An initial draft reconstruction was derived from the annotated genome of Burkholderia cenocepacia J2315 available at the Burkholderia Genome Database (http://www.burkholderia.com). To link annotated genes to proteins and proteins to reactions, biological databases such as KEGG, GeneDB, UniProt, BRENDA, Transport Classification Database (TCDB), and TransportDB were used [83–88]. Manual curations were performed to establish gene-protein-reaction (GPR) associations, which connect genetic data to reactions in the metabolic network and allow for subsequent exploration of metabolic phenotypes using genetic perturbations.

After the initial reconstruction was generated, gaps in metabolic pathways necessary to produce biomass components and key virulence factors were filled by cautious literature mining, BIOLOG substrates utilization assays, and BLAST searches on homology and protein sequence similarity analyses [89, 90]. The genome annotation was refined as consequence of the gap-filling and model extension process.

Flux Balance Analysis (FBA) was carried out throughout this study to explore the metabolic capabilities of i KF1028 under various environments. In addition to minimal medium, synthetic cystic fibrosis medium (SCFM) representing the physical living environment during CF infection was simulated in silico to investigate the metabolic flux distribution in a CF-like condition.

Biomass composition

Flux balance analysis

Flux balance analysis (FBA) is an algorithm based on linear programming (LP) and on the assumption that the represented metabolic network is in steady-state (i.e. all the intracellular metabolite concentrations are constant). Being a LP problem, FBA also requires the selection of an objective function and of whether the value for that same function should be maximized or minimized. FBA is usually used to compute the optimal growth yield (the maximized objective function) based on the assumption that the evolutionary fitness of the organism depends on growth alone and, consequently, the implicit regulatory mechanism are organized to permit the theoretical maximal growth. If the system of equations (stoichiometric matrix which represents the metabolic network) is feasible, the algorithm generates an optimal flux distribution for that same network, taking into account the imposed thermodynamic constraints (reaction directionality) and limits on substrate uptake rates. The mathematical description is as follows:

Where S is a stoichiometric matrix containing i rows representing metabolites and j columns representing reactions, v is a vector of all reaction fluxes, v_min and v_max are imposed lower and upper bounds on flux v _j respectively, and c ^T is a vector of coefficients for each reaction that is to be maximized.

In silico media composition

Two different living environments were simulated in silico for strain J2315: M9 minimal medium [94], which contains PO₄^3-, SO₄^2-, NH₄⁺, H⁺, Fe²⁺, K⁺, Mg²⁺, Na⁺, H₂O, and thiamine, with glucose or other BIOLOG substrates as sole carbon source; and synthetic CF sputum medium (SCFM) [95] representing the nutrient conditions inside a host-cell during CF infection. Details of the simulated SCFM composition are provided in the supplemental material [Additional file 6].

BIOLOG assay

To validate the model and estimate the metabolic capabilities of strain J2315, BIOLOG assay was performed by using various carbon sources for strain cultivation [16]. The BIOLOG assay was carried out in triplicates using Biolog GN2 MicroPlates (Biolog, Inc.), which can test the ability of a microorganism to oxidize a panel of 95 different carbon sources simultaneously. The procedure for using the MicroPlates was according to the manufacturer's specification. The strain J2315 was obtained from DSMZ GmbH (DSMZ 16553, equivalent to LMG 16656 as which strain J2315 has been deposited in the BCCM/LMG Bacteria Collection). The strain was cultured overnight in CASO agar plate. Then the bacteria were swabbed from the plate surface and suspended in GN/GP inoculating fluid (Biolog, Inc.) and 150 μl of the suspension was transferred to each well of the GN2 MicroPlate. The MicroPlates were incubated at 30°C for 48 hours and were read by a microplate reader at 24 and 48 hours and analyzed with the Biolog MicroLog3 4.20 software (Biolog, Inc.). A comparison between the BIOLOG results and in silico predictions is provided in the supplemental material. [Additional file 3]

Gene and enzyme essentiality

FBA can be used to interpret genetic modification, such as gene deletion and enzyme inhibition, and subsequently make comprehensive in silico predictions on gene and enzyme essentiality [96]. To assess the essentiality of a gene, its GPR is checked for a unique relation with the associated reaction(s). If the gene is necessary to the reaction, the reaction flux will be constrained to zero and a solution for the maximal growth yield is searched. The deleted gene is predicated to be essential if, as consequence of that added constraint, the value of the objective function (growth yield) changes to zero. The deletion of every gene accounted in the model was simulated for growth on minimal medium with glucose as sole carbon source, and on SCFM. Similarly, an enzyme is considered essential if, by constraining to zero the flux on every associated reaction that has no alternative means of catalysis, the value for the growth yield changes to zero. The essentiality of every enzyme accounted in the model was analysed for growth on SCFM.