Metabolic modelling reveals the specialization of secondary replicons for niche adaptation in Sinorhizobium meliloti

The genome of about 10% of bacterial species is divided among two or more large chromosome-sized replicons. The contribution of each replicon to the microbial life cycle (for example, environmental adaptations and/or niche switching) remains unclear. Here we report a genome-scale metabolic model of the legume symbiont Sinorhizobium meliloti that is integrated with carbon utilization data for 1,500 genes with 192 carbon substrates. Growth of S. meliloti is modelled in three ecological niches (bulk soil, rhizosphere and nodule) with a focus on the role of each of its three replicons. We observe clear metabolic differences during growth in the tested ecological niches and an overall reprogramming following niche switching. In silico examination of the inferred fitness of gene deletion mutants suggests that secondary replicons evolved to fulfil a specialized function, particularly host-associated niche adaptation. Thus, genes on secondary replicons might potentially be manipulated to promote or suppress host interactions for biotechnological purposes.

ADPribose diphosphatase Required a dead end metabolite smc03763 DNA (cytosine5)methyltransferase Would produce a dead end metabolite smc03243 Dihydroneopterin dephosphorylase Too general of an annotation in the S. meliloti 1021 genome annotation smc03236 GMP synthase Lacked confidence in the annotation smc00356 LysyltRNA synthetase Lacked confidence in the annotation smc02689 Aminobutyraldehyde dehydrogenase Too general of an annotation in the S. meliloti 1021 genome annotation smc02377 putrescine oxidase Lacked confidence in the annotation smc02263 acetolactate synthase Does not complement auxotrophy of a smc01431 mutation 3 smc01404 Aspartate racemase Would produce a dead end metabolite smc01153 3hydroxyacylCoA dehydratase (3hydroxytetradecanoylCoA) Too general of an annotation in the S. meliloti 1021 genome annotation smc01147 coproporphyrinogen oxidase (O2 required) Lacked confidence in the annotation smc01146 thiamin pyrophosphatase Lacked confidence in the annotation smc00810 Guanosine 3&apos;diphosphate 5&apos;triphosphate Lacked confidence in the annotation smc00808 protoporphyrinogen oxidase (aerobic) Lacked confidence in the annotation smc00410 NADH dehydrogenase Lacked confidence in the annotation smb21586 glutathione synthetase Does not complement mutation of gshB1 4 smb21301 Aminobutyraldehyde dehydrogenase Too general of an annotation in the S. meliloti 1021 genome annotation smb20857 glucose6phosphate isomerase Excluded based on 5 smb20433 ornithine cyclodeaminase Does not complement mutation of ocd 6 smb20115 Dihydroxyacid dehydratase (2,3dihydroxy3methylpentanoate) Does not complement auxotrophy of a smc04045 mutation 3 smb20072 Cytosolic source for myoinositol Lacked confidence in the annotation sma2213 Aminobutyraldehyde dehydrogenase Too general of an annotation in the S. meliloti 1021 genome annotation sma2211 acetolactate synthase Does not complement auxotrophy of a smc01431 mutation 3 sma1871 ornithine cyclodeaminase Does not complement mutation of ocd 6 sma1844 Aminobutyraldehyde dehydrogenase Too general of an annotation in the S. meliloti 1021 genome annotation sma1155 magnesium transport via ABC system Too general of an annotation in the S. meliloti 1021 genome annotation sma0959 3oxoacylacylcarrierprotein reductase (nC10:0) Too general of an annotation in the S. meliloti 1021 genome annotation sma0958 acetolactate synthase Does not complement auxotrophy of a smc01431 mutation 3 sma0956 glutamate1semialdehyde aminotransferase Too general of an annotation in the S. meliloti 1021 genome annotation sma0486 ornithine cyclodeaminase Does not complement mutation of ocd 6 Isoleucine -5.  1 . c The GC content for mRNA was estimated from the overall GC content of S. meliloti 1 . The GC content of tRNA was estimated based on the GC content of the 10 most common codons in S. meliloti 1,7 . The GC content of rRNA was determined based on the rrn loci of S. meliloti 1 . The S11 overall composition of cellular RNA was determined assuming 80% rRNA, 15% tRNA, and 5% mRNA. d The amino acid composition was estimated based on the codon usage of S. meliloti 7 . e The lipid composition used was as previously determined for S. meliloti, with each lipid class represented by a single lipid of the most common lipid size [8][9][10][11] . f A 4:1 ratio of low molecular weight (LMW) to high molecular weight (HMW) succinoglycan was set as previously determined 12 . g These numbers are for when growth was modelled in bulk soil. When growth was simulated in the rhizosphere, the amount of both LMW and HMW succinoglycan was doubled. h Nod factor was included in the biomass composition when growth was modelled in the rhizosphere, but not when growth was modelled in bulk soil.  (7) Values are ratios of all genes in the given group (column) with the gene count given in brackets. Replicon -the replicon on which the transcriptional regulator is located. Entire Genome -the distribution in the entire S. meliloti 1021 genome. Bulk Fit -genes whose deletion results in a fitness decrease in bulk soil. Rhizo Fit -genes whose deletion results in a fitness decrease in the rhizosphere. Nodule Fit -genes whose deletion results in a fitness decrease in the nodule. B2R More and B2R Less-genes associated with more important and less important reactions, respectively, following the bulk soil to rhizosphere transition. R2N More and R2N Lessgenes associated with more important and less important reactions, respectively, following the rhizosphere to nodule transition. Statistically significant biases within each group, compared to the entire genome, was determined using Pearson's Chi Squared tests: * p-value < 0.001.  (1) 0.000 (0) 0.000 (0) 0.000 (0) Values are ratios of all genes in the given group (column) with the gene count given in brackets. Compartment -the pangenome classifications. The distribution is shown for: the entire S. meliloti 1021 genome (entire genome), the model (iGD1575), genes associated with reactions active in bulk soil (Bulk Act), the rhizosphere (Rhizo Act), or the nodule (Nodule Act), and genes contributing to fitness in bulk soil (Bulk Fit), the rhizosphere (Rhizo Fit), or the nodule (Nodule Fit). * Distribution is significantly different (p-value < 0.001) from that of the S. meliloti 1021 genome as determined using Pearson's Chi Squared tests. † Distribution is significantly different (p-value < 0.05) from that of iGD1575 as determined using Pearson's Chi Squared tests.   *If an exchange reaction is not listed, the lower boundary was set to '0', the upper boundary set to '1000', and the reaction did not carry flux in any of the three conditions.

SUPPLEMENTARY NOTES
Supplementary Note 1. Prediction of the genetic basis of D-galacactosamine metabolism. The genetic basis for several of the observed carbon utilization phenotypes has not been previously examined in S. meliloti. By combining our mutant phenotype data with iGD1575, the output of the DuctApe software 16 , the S. meliloti genome annotation 1 , and a previously published ABC transporter induction study 17 , we were able to predict novel carbon catabolic loci. One example compound is D-galactosamine. D-galactosamine was not present in the Phenotype MicroArray TM , but N-acetyl-D-galactosamine, whose catabolism proceeds via Dgalactosamine, was tested. Both the ∆B180 and ∆B181 deletion mutants failed to utilize Nacetyl-D-galactosamine as a carbon source. It was previously shown that two ABC transporters within ∆B180, one encoded by smb21135-smb21138 and the second by smb21216, smb21219-smb21221, are induced by D-galactosamine 17 . The smb21216, smb21219-smb21221 operon also includes a putative sugar isomerase (smb21218) and sugar amine kinase (smb21217). Thus, we hypothesize that the Smb21216, Smb21219-Smb21221 transporter, as well as potentially the Smb21135-Smb21138 transporter, transports D-glucosamine, which is phosphorylated by Smb21217 to D-glucosamine-6-phosphate, and converted to D-tagatose-6-phosphate by Smb21218. D-tagatose-6-phosphate can then be further metabolized via two steps to glyceraldehyde-3-phosphate and dihydroxyacetone phosphate, both which can enter glycolysis. At least one of these two final steps is located within ∆B181 as this mutant cannot grow with Nacetyl-D-glucosamine or D-tagatose. In fact, the smb21373-smb21377 operon, removed in ∆B181, encodes an ABC transporter that we predict transports D-tagatose and that indeed is induced by tagatose 17 , as well as a putative carbohydrate kinase and a putative D-tagatose-1,6bisphosphate aldolase that could complete the D-galactosamine catabolic pathway.
Supplementary Note 2. Validation of the metabolic switching detection during niche adaptation. Flux variability analysis (FVA) was employed to ensure that the observed flux changes between environments were true changes. During FBA, one solution of potentially multiple solutions giving the same optimal flux through the objective function is returned. FVA examines the range of flux that can possibly be carried by each reaction while still maintaining optimal flux through the objective function.
A total of 199 reactions were considered to change in importance between simulated growth in bulk soil and the rhizosphere based on changes in flux and/or the fitness contribution. Of these, 63 (31.7%) are supported by fitness data while 136 (68.7%) are classified as changing solely on the basis of flux. Of the 136 reactions, the predicted reaction flux during growth in the rhizosphere for 61 of them was outside the FVA range for the reaction when grown in bulk soil. This supports that optimal biomass production in the rhizosphere required that flux through these 61 reactions changed. To further examine the remaining 75 reactions, where the reaction flux in the rhizosphere was within the FVA range of the reaction when grown in bulk soil, the flux through each reaction was individually set to the rhizosphere value, and the flux distribution during growth in bulk soil was monitored. In all but two cases, doing so caused the flux through at least one other reaction (with an average of 9 reactions and standard deviation of 4) to move outside of the rhizosphere FVA range of that reaction. Or in other words, while 61 of the predicted changing reactions between bulk soil and rhizosphere could be artificially set to the rhizosphere flux rate without impairing growth in bulk soil, in all but 2 cases, doing so would S20 require an average of 9 new changes following transition from growth in bulk soil to the rhizosphere. Additionally, when flux through all 61 of the reactions were simultaneously set to the corresponding flux value from the rhizosphere, and growth simulated in bulk soil, the model became unsolvable. Together, these data support that 197 of the 199 reactions (99.0%) predicted to change in importance during transition between growth in bulk soil and the rhizosphere are true changes.
The same procedure was used to confirm the flux changes observed during transition between the rhizosphere and the nodule environments. Of the 451 reactions that were considered changing, 292 (64.7%) were supported by fitness data while 159 (35.3%) are based solely on flux data. Of these 159 reactions, the predicted flux through 57 in the nodule was outside the FVA range for the reaction during simulated growth in the rhizosphere. For all the remaining 102 reactions, individually setting the flux during growth in the rhizosphere to the flux value predicted in the nodule resulted in the flux through at least one (average of 4.5, standard deviation 2) other reaction to move outside the nodule FVA range of that reaction. Additionally, the model becomes unsolvable in the rhizosphere when flux through all 102 reactions are simultaneously set to the corresponding nodule flux value. When considered together, these data support that all of the reactions predicted to change in importance during transition between growth in the rhizosphere and nitrogen fixation in the nodule are true changes.

Supplementary Note 3. Robustness with respect to changes in the nutrients composition and uptake rates.
We explored the effect of random variations in the composition and uptake rates of the nutrients present in each of the simulated ecological niches (bulk soil, rhizosphere, and nodule) on the main outcomes of the model, i.e. predicted growth rates, number of essential and fitness contributing genes and gene pairs. Results obtained (shown in Supplementary Figures 7 and 8) revealed that, overall, conclusions drawn concerning the role of pSymA and pSymB throughout the S. meliloti lifecycle and the growth phenotypes in different conditions are not influenced by such variations. In all environments, the number of essential genes shows very little variation across all the iterations, and the number of essential plus fitness contributing genes on each replicon is also predominately unaffected by the environmental variations. This trend is observed also in the case of essential and fitness contributing gene-pairs (Supplementary Figure 8). A notable exception is represented by the effectiveness of the symbiosis in the nodule (Supplementary Figure 7), which seems to be more dependent on the composition and the utilization rate of the input compounds. However, this higher variation is not surprising as a limited number of nutrients that are used by the bacteria are available to the bacteroid in the nodule; as such, any change in one nutrient is likely to influence flux through the objective function much more.
Overall, the stability of the number of essential/fitness contributing genes across all permutations confirms that the presented results are robust and supports that the conclusions drawn in this work.

Supplementary Note 4. Few biases were detected in the genomic localization of the transcriptional regulators of metabolic genes.
When compared to a recent regulon analysis in S. meliloti 2 , even if not conclusive due to the low representation in terms of gene numbers, a bias was observed for genes associated with reactions classified as less important during either the bulk soil to rhizosphere or the rhizosphere to nodule transitions to be regulated by a pSymB encoded transcription factor (Supplementary Table 4 and Supplementary Data 5). However, S21 while this may be suggestive of a bias for the regulatory machinery associated with niche adaptation to be encoded by pSymB, it may also simply reflect a bias in the dataset as most of the regulated genes were located on pSymB and transcription factors tend to regulated genes on their own replicon 2 .
Supplementary Note 5. The core S. meliloti genome is over-represented among the fitness contributing genes. An advantage of using S. meliloti as a model organism is that many strains have been fully sequenced, facilitating the study of the S. meliloti pangenome. We classified all genes associated with the environmentally variable or fitness promoting reactions as belonging to the 'core', 'accessory', or 'unique' genome (Supplementary Data 5) based on the results of Galardini et al. 2 . Perhaps not surprisingly given that iGD1575 is enriched in the core metabolic processes relative to the entire S. meliloti genome, the core genome was overrepresented in the model (Supplementary Table 5). Remarkably, however, nearly all genes contributing to fitness in either bulk soil, the rhizosphere, or the nodule belonged to the core genome, a clear enrichment relative to the percentage of core genes in iGD1575 overall. This observation highlights that metabolic genes contributing to the fitness of the cell are highly likely to be or become part of the core genome, emphasizing the functional role of core genome as common tool set of genes for a given bacterial species.
Supplementary Note 6. Identification of essential model genes. Currently, no high throughput systematic knock-out studies of S. meliloti exist in the literature, and so the complete set of essential genes in this organism is not known. However, previous work has shown that only two essential genes exist on the pSymB chromid and none are present on pSymA 15,18,19 . In this work we examined a library of pSymB deletion mutants that collectively remove 98% of pSymB 14 for carbon utilization phenotypes using Phenotype MicroArrays TM . We also replicated this experiment in silico using iGD1575. In order to do so, it was necessary to ensure that none of the in silico deletion mutants were lethal. To test this, we used an in silico single gene deletion analysis to identify single copy essential genes in iGD1575 when grown in a minimal medium with sucrose as the carbon source and with thiamine supplementation 18,19 . In this context, an essential gene refers to a gene whose deletion prevents the formation of at least one biomass precursor when grown. This analysis identified a total of 231 single copy essential metabolic genes (Supplementary Table 8), none of which are on pSymA. However, of the 231 essential genes, 216 were on the chromosome while 15 were on pSymB.
Of the 15 genes on pSymB, 12 are involved in the biosynthesis of succinoglycan. While these genes are not truly essential, the inclusion of succinoglycan in the biomass formation reaction resulted in them being considered essential in the model. Two of the essential pSymB genes, wgaG and wgaJ, are predicted to be required for lipopolysaccharide (LPS) biosynthesis. Further examination, through a double deletion analysis and manual screening, revealed another two reactions associated with multiple redundant pSymB genes that are essential for LPS synthesis. All four reactions were required for the synthesis of dTDP-rhamnose, necessary for the production of the O-antigen. Little is known about LPS biosynthesis in S. meliloti, and it is possible that these genes are truly required for complete LPS synthesis. But whereas Sinorhizobium LPS mutants that cannot encorporate rhamnose into their LPS produce a truncated LPS yet survive 20 , the rigidity of the model and encorporation of complete LPS in the biomass reaction meant that such a mutant would fail to produce biomass in silico. Finally, ansB was not surprisingly determined to be essential as it is the only S. meliloti gene predicted to be involved S22 in asparagine biosynthesis, although experimental evidence indicates it is not essential for asparagine biosynthesis and the asparagine biosynthetic pathway in S. meliloti remains unidentified 18 . The two genes on pSymB that are truly essential are a tRNA and a protein involved in ribosomal biogenesis (engA) 15 , and are therefore not present in the model.
When succinoglycan is removed from the biomass reaction, and unknown GPRs are added to the AsnB and the four dTDP-rhamnose synthesis reactions, the simultaneous removal of all pSymA and pSymB genes from the model does not prevent biomass formation. Thus, for the in silico experiments testing the metabolic capacity of iGD1575, the model was modified as described in the previous sentence so that all deletion mutants were viable.

Metabolic network reconstruction.
A draft metabolic model was constructed using the KBase Narrative Interface (www.kbase.us). The S. meliloti 1021 annotated genome was imported from the public KBase database, and a draft metabolic model was reconstructed using the 'Build metabolic model' method. This draft model was gap-filled using the 'Gapfill metabolic model' method to allow biomass formation when grown in a minimal medium containing metal ions, succinate, ammonium, sulphate, phosphate, and biotin. The model was then downloaded, the KBase annotations in the 'Gene_association' field replaced with the S. meliloti 1021 locus tags, and the KBase annotations in the 'Protein_association' field replaced with the S. meliloti 1021 gene names 1 . The KBase biomass objective was removed, new objective functions were formulated (representing biomass formation and symbiosis, as described below) and the model manually gap filled to produce flux through each of the objective functions.
Manual curation and further expansion of the model through the inclusion of additional reactions and gene-protein-reaction associations (GPRs) was performed in several main stages. Where possible, 'Unknown' GPRs were replaced with genes from the S. meliloti 1021 genome, and when supported by published experimental data, incorrect GPRs and reactions were removed from the draft model. We next identified and included additional metabolic and transport genes in the S. meliloti 1021 genome annotation. Following this, the genes present in the draft model were compared to the list of genes included in the existing S. meliloti metabolic model iHZ565 21 , and the majority of the additional genes and associated reactions from iHZ565 were added to our draft model. However, 31 genes were not transferred from iHZ565 (Supplementary Table 1) as the annotation was extremely general and we lacked confidence in the true substrates/products of the reaction, experimental data was inconsistent with their inclusion, or the reaction produced a dead-end metabolite (a metabolite produced or consumed by only a single reaction, meaning that the reaction will never be active during flux balance analysis). An iterative gap-filling procedure was then employed to reconcile the predictions of 'growth' or 'no growth' with various carbon and nitrogen substrate with the known ability or inability of S. meliloti 1021 to grow on these substrates. Most of the experimental growth data came from a previous Phenotype MicroArray TM (Biolog) experiment 22 , with a few substrate taken from additional literature sources [23][24][25][26] . Finally, a library of large S. meliloti deletion mutants was screened for carbon utilization defects using the Omnilog Phenotype MicroArray TM system (Biolog), as described below. In silico predictions were compared with the experimental results, and were necessary and possible, discrepancies were fixed by further refinement and expansion of the metabolic model. Where possible, GPRs for transport reactions were added based on published mutation or induction studies; otherwise, transport reactions were added as a diffusion reaction with no associated GPR. Support for manually added transporters and metabolic reactions came from experimental evidence, review articles, and the KEGG and BioCyc databases [3][4][5][6]8,9,17,22,23, .
The final S. meliloti model was termed iGD1575 in accordance with the nomenclature standard 66  Metabolic modelling. The ability of the model to produce flux through the specified objective functions was examined using flux balance analysis (FBA). Simulations were performed in Matlab R2015a (Mathworks), using scripts from the Cobra Toolbox 67 and the Gurobi 6.0.1 solver (www.gurobi.com). Single gene deletion and double gene deletion analyses were performed using methods in the Cobra Toolbox. Iteratively removing each reaction from the model and then examining the effect with FBA determined the essentiality of each individual reaction. Metabolic capacity was determined in silico by iteratively providing the model with a unique carbon or nitrogen source and then observing the ability of the model to produce biomass with FBA. The predicted effect of the large-scale genome deletions on the growth phenotype (either growth or no growth) was addressed by simultaneously removing all reactions dependent on the deleted genes from the model and then running the in silico phenotype microarray experiment. For this analysis, exopolysaccharide was removed from the objective function and an unknown GPR added to the other four essential reactions dependent on pSymB genes (see Supplementary Note 6 for additional details) in order to allow all deletion mutants to grow in the minimal media.
Evaluation of robustness with respect to changes in the nutrients composition and uptake rates. Simulation of growth/symbiosis in 1000 randomly generated media for each environment was performed as follows. For each of the 1000 iterations, a random variation in the allowable uptake rate for each of the nutrients in the medium was introduced, with the allowable variation set to a value 50% greater or lower than the original one (e.g. if the original uptake rate was 1 mmol g -1 h -1 , in each of the 1000 conditions, the uptake rate was randomly set to a speed between 0.5 to 1.5 mmol g -1 h -1 ). Additionally, further noise was introduced by randomly removing, at each iteration, two nutrients from the niche simulated nutrients set and restoring them for the following iteration. FBA was then used, at each iteration, to evaluate i) the predicted growth rate in each of the iterations for each of the three environments, ii) the variations in the number of essential genes, and iii) the variations in replicon specific essential plus fitness contributing genes. Due to time constraints, when determining the variation in the number of replicon specific essential/fitness-contributing gene pairs related to random changes in the nutrients composition, the number of iterations was reduced to 100.

Flux visualization.
Metabolic networks were visualized with the online tool iPath 2.0 68 . Where possible, KEGG IDs were associated with each reaction based on comparison with the Seed Reference list, and these reactions mapped to the corresponding reaction, if one existed, in the 'Metabolic Pathway' map of iPath.
In silico environmental representations. In silico representations of the nutritional composition of the rhizosphere and bulk soil were constructed from data available in the literature. For both the rhizosphere and bulk soil in silico representations, ammonium and nitrate were included at a one to one ratio, and sufficient ammonium, nitrate, phosphate, sulphate, metal ions, and gases were included so that these compounds were not growth rate limiting.
The sugar content of the in silico rhizosphere was primarily set according to the average monosaccharide composition of pea (Pisum sativum) and cowpea (Vigna unguiculata) root mucilage 69,70 . Sucrose, raffinose, and stachyose were added to the list of sugars, with the ratio of these three sugars set according to their approximate ratio in alfalfa (Medicago sativa) root exudate 71 . The total amount of sucrose, raffinose, and stachyose was arbitrarily set so that 60% of the total glucose was within these sugars, and the amounts of free glucose and galactose were reduced accordingly. The organic acids included in the rhizosphere, and their relative ratios, was S25 based on their prevalence in unstressed alfalfa root exudate 72 . The total ratio of organic acids to sugars in legume root exudate was not found; however, organic acids were ~ 10, 4, and 2 fold more prevalent than sugars in the root exudates of tomato (Lycopersicon esculentum), cucumber (Cucumis sativus), and sweet pepper (Capsicum annum), respectively 73 . Therefore, the total amount of organic acid in the rhizosphere was set at a molar ratio two fold greater than the total carbohydrates. The amino acid content of the rhizosphere was primarily based on the amount of each amino acid present in pea root exudate when grown in quartz sand 74 . To this, hydroxyproline was added according to the serine to hydroxyproline ratio in pea root mucilage 69 . The carbohydrate to amino acid ratio in legume root exudate was not found; however, an approximate carbohydrate to protein ratio of four to one, or greater, was observed in three rice varieties (Oryza sativa) 75 . Therefore, the total amount of amino acids in the rhizosphere was set at a molar ratio four fold less than the total carbohydrates.
The molar ratio of sugars in the in silico bulk soil representation was set as determined previously 76 . Unlike in the rhizosphere condition, sucrose, raffinose, and stachyose were not added to the bulk soil as they appear to be largely absent 71,77 . The concentration of organic acids appears to be quite low in bulk soil, with the organic acid content of the rhizosphere likely at least 50-fold higher than that of bulk soil 78 . Therefore, the three organic acids included in the rhizosphere formulation were also included in bulk soil, but at 2% the concentration. The dominant amino acid content of bulk soil was set as the average of two previously analyzed soil samples 79 . These amino acids accounted for 66.9% of the amino acids, with the remaining 33.1% split evenly between the non-modified amino acids not displayed in the reference 79 as they never exceeded 5% of the total amino acid population. The total carbohydrate content of the bulk soil was set ten fold higher than the total amino acid content, as determined previously 79 .
The nutritional environment within the root nodule was set as used previously for constraint based modelling of iHZ565 21 . The upper and lower bounds for the exchange reactions in all three environments are listed in Supplementary Table 6.
Phenotype MicroArray™ analysis. S. meliloti RmP110 13 , a derivative of S. meliloti 1021 in which a frameshift mutation within pstC was fixed, was used as the wild type reference strain. All deletion mutant strains were described previously 14,15 , and consist of large deletions of the pSymB replicon that each span ~ 40 to 370 kilobase pairs. Phenotype MicroArray™ experiments were performed largely as described previously 22 , using Biolog plates PM1 and PM2A. Strains were initially grown at 30°C on LBmc agar supplemented with CoCl 2 18 . To begin the Phenotype MicroArray™ analysis, colonies were picked up with sterile cotton swabs and resuspended in 0.8% NaCl to a cell density of 81% turbidity (OD 600 ~ 0.1) as measured with a Biolog turbidimeter. 2 mL of each suspension was diluted in 22 mL of carbon free M9 minimal medium 18 containing 240 µL of the redox dye MixA 100x (Biolog), and 100 µL of the final mixture was added to each well of the Biolog plates. The one exception was S. meliloti RmP2754 (∆B180), which grew very poorly when inoculated directly from the agar plate. Therefore, this strain and a second replicate of S. meliloti RmP110 were pregrown in liquid M9glucose, washed twice with 0.8% saline, and diluted in 0.8% saline to a turbidity of 81%. This cellular suspension was then treated as described above and used to inoculate the PM plates. All PM plates were incubated at 30°C in an OmniLog plate reader, which measured reduction of the dye every 15 minutes for 120 hours.