Metabolic mechanisms of nitrogen substrate utilisation in three rhizosphere bacterial strains investigated using quantitative proteomics

Nitrogen metabolism in the rhizosphere microbiome plays an important role in mediating plant nutrition, particularly under low inputs of mineral fertilisers. However, there is relatively little mechanistic information about which genes and metabolic pathways are induced by rhizosphere bacterial strains to utilise diverse nitrogen substrates. Here we investigate nitrogen substrate utilisation in three taxonomically diverse bacterial strains previously isolated from Arabidopsis roots. The three strains represent taxa that are consistently detected as core members of the plant microbiome: Pseudomonas, Streptomyces and Rhizobium. We use phenotype microarrays to determine the nitrogen substrate preferences of these strains, and compare the experimental results versus computational simulations of genome-scale metabolic network models obtained with EnsembleFBA. Results show that all three strains exhibit generalistic nitrogen substrate preferences, with substrate utilisation being well predicted by EnsembleFBA. Using label-free quantitative proteomics, we document hundreds of proteins in each strain that exhibit differential abundance values following cultivation on five different nitrogen sources: ammonium, glutamate, lysine, serine and urea. Proteomic data show that the three strains use different metabolic strategies to utilise specific nitrogen sources. One diverging trait appears to be their degree of proteomic flexibility, with Pseudomonas sp. Root9 utilising lysine nutrition via widespread protein-level alterations to its flexible metabolic network, whereas Rhizobium sp. Root491 shows relatively stable proteome composition across diverse nitrogen sources. Our results give new protein-level information about the specific transporters and enzymes induced by diverse rhizosphere bacterial strains to utilise organic nitrogen substrates. Importance Nitrogen is the primary macronutrient required for plant growth. In contemporary agriculture, the vast majority of nitrogen is delivered via mineral fertilisers, which have undesirable environmental consequences such as waterway eutrophication and greenhouse gas production. There is increasing research interest in designing agricultural systems that mimic natural ecosystems, where nitrogen compounds are cycled between plants and soil, with the mineralisation of recalcitrant soil organic-N molecules mediated via microbial metabolism. However, to date there is little mechanistic information about which genes and metabolic pathways are induced by rhizosphere bacterial strains to metabolise organic-N molecules. Here, we use quantitative proteomics to provide new information about the molecular mechanisms utilised by taxonomically diverse rhizosphere bacterial strains to utilise different nitrogen substrates. Furthermore, we generate computational models of bacterial metabolism from a minimal set of experimental information, providing a workflow that can be easily reused to predict nitrogen substrate utilisation in other strains.


Introduction 83
Improved nitrogen management in agricultural systems is crucial for environmental 84 sustainability. Large-scale application of mineral nitrogen fertilisers has extensive off-target 85 effects, such as greenhouse gas production and waterway eutrophication (1). One potential 86 pathway to boost agricultural sustainability involves substituting mineral fertilisers with organic 87 nutrients derived from recycling various waste streams. For low-input agricultural systems to 88 provide sufficient bioavailable nitrogen to meet the demands of plant growth, future crop 89 management practices will need to better incorporate microbial pathways of nitrogen 90 mobilisation (2). One specific suggestion involves engineering the rhizosphere microbiome to 91 promote the mineralisation of organic nitrogen, coupled with engineering of plant root 92 metabolism to release rhizodeposits that recruit beneficial microbial strains (3). However, the 93 ability to manipulate plant-microbe cooperation is limited by an incomplete knowledge of the 94 specific microbial traits involved in root colonisation and nutrient mobilisation (4). 95 96 Nitrogen flows in the rhizosphere are complex, with plants and microbes potentially cooperating 97 but sometimes competing for uptake of diverse nitrogen molecules (5). Legume-Rhizobia 98 symbioses provide an example of cooperation, whereby the majority of the plant's nitrogen 99 nutrition is derived from bacterial fixation of atmospheric N 2 (6). Outside of legumes, it is 100 generally accepted that plants obtain the majority of their nitrogen nutrition from inorganic forms 101 such as NO 3 and NH 4 , whereas microbes are more adept at acquiring more recalcitrant organic 102 nitrogen forms such as proteins and amino acids (7). Therefore, cooperative nutrient transfers can 103 occur when microbes take up soil-bound organic nitrogen, which is subsequently transferred to 104 plants in a mineralised form following microbial lysis or protozoic predation (8). Conversely, 105 competitive flows can occur when microbes immobilise inorganic nitrogen, or when plants take up organic nitrogen (9). Adding further complexity, plant root exudates contain large amounts of 107 organic nitrogen molecules which can serve as carbon and nitrogen substrates for bacterial 108 growth. The rate of amino acid release from plant roots increases under exposure to specific 109 bacterial metabolites (10), but organic nitrogen molecules released via root exudation can also be 110 efficiently re-acquired by the root system (11). 111 112 Investigations of how bacteria utilise diverse nitrogen substrates have been documented since the 113 beginning of modern microbiology (12). Ammonium is the preferred nitrogen source for most 114 bacteria, and experimental designs usually include ammonium as a control treatment, to compare 115 against alternative nitrogen sources or starvation treatments (13). Over decades, such studies have 116 provided detailed insight into fundamental physiological mechanisms such as the molecular 117 pathways of bacterial nitrogen assimilation, the perception of nitrogen status, and the response to 118 nitrogen starvation in E. coli (14). However, other bacterial taxa possess different mechanisms 119 for regulating nitrogen metabolism (15,16), with soil bacteria exhibiting extensive diversity 120 regarding their nitrogen substrate preferences and also the metabolic pathways used to metabolise 121 organic nitrogen sources (17,18). Therefore, novel insights into metabolic mechanisms of 122 nitrogen metabolism may be observed by studying nitrogen substrate utilisation in taxonomically 123 diverse bacterial strains isolated from the rhizosphere. 124

125
The rhizosphere microbiome has attracted increasing research attention over the past 20 years. 126 From the results of 16S pyrosequencing studies, it has become increasingly apparent that the 127 rhizosphere hosts a taxonomically diverse bacterial microbiota, which plays an important role in 128 determining plant growth and health (19). Recently, multiple research groups have established 129 large collections of bacterial strains isolated from field-grown plants, which can be used to 130 dissect the functional traits carried out by individual strains, or reassembled into synthetic 131 communities that recapitulate microbiome function (20,21). There is now an opportunity to study 132 these plant-associated microbial strains using high-throughput 'omics techniques, to acquire new 133 insights into the specific molecular mechanisms that confer a selective advantage in the plant-134 associated niche (22). 135 136 Alongside experimental approaches, computational modelling is becoming a widespread 137 approach to investigate microbial metabolism (23). One particularly useful method is the 138 construction of genome-scale metabolic network models, which translate the information 139 encoded in the bacterial genome into a computational formalism that can be analysed with 140 mathematical methods (24). However, curated genome-scale metabolic models are only available 141 for a relatively small set of extensively studied bacterial strains, and generally it is difficult to 142 analyse newly sequenced bacterial strains using computational modelling. This limitation exists 143 because reconstructing a curated genome-scale metabolic network model is a painstaking process 144 that requires extensive manual curation as well as the acquisition of devoted experimental data, 145 particularly regarding biomass composition. Although progress is being made towards automated 146 reconstruction of genome-scale metabolic network models, many challenges still have to be 147 addressed (25). Recently, a method named EnsembleFBA has been proposed as a potential 148 approach to approximate genome-scale metabolic networks for diverse bacterial strains. Instead 149 of relying on the availability of a single manually curated genome-scale model, EnsembleFBA 150 uses the information derived from multiple metabolic networks, which are reconstructed from the 151 same initial draft network and refined through the process of positive and negative gapfilling on 152 randomized sets of growth and non-growth conditions (26). As a proof of concept, it was shown that the EnsembleFBA method achieved greater precision in predicting essential genes than an 154 individual, highly curated model. 155

156
Here we investigate nitrogen metabolism in three taxonomically diverse bacterial strains 157 previously isolated from Arabidopsis roots. We apply a combination of methods, including 158 quantitative proteomics, growth assays, phenotype microarray and EnsembleFBA. With the 159 proteomic data, we were particularly interested in determining the specific proteins that are 160 enriched according to different nitrogen sources, to decipher the metabolic strategies used for 161 nitrogen acquisition across different rhizosphere bacterial strains. In parallel, we applied the 162 EnsembleFBA method to reconstruct and analyse sets of genome-scale metabolic network 163 models for each strain, using the phenotype microarray data for training and testing the model 164 predictions of nitrogen substrate utilisation. 165

167
We studied nitrogen metabolism in three taxonomically diverse bacterial strains isolated from 168 roots of field-grown Arabidopsis: Pseudomonas sp. Root9, Streptomyces sp. Root66D1 and 169 Rhizobium sp. Root491. Strains were previously isolated in Bai et al (20), and the three strains 170 chosen here correspond to taxa that were repeatedly observed to be highly abundant in the 171 microbiome of field-grown Arabidopsis plants (20,27,28 showing an accuracy in predicting growth in about 78% of cases for the three strains 185 (Supplementary Table S2). However, there is a relatively poor correlation between the proxy 186 values of metabolic activity predicted by the models versus the experimental measurements, with 187 a comparison of percentile rank between the datasets yielding r 2 values between 0.23 and 0.5 188 across the three strains (Supplementary Figure S2). Interestingly, the accuracy of the model prediction seems to vary across different molecular classes, with good concordance for amino 190 acids but poor concordance for nitrogen bases ( Figure 1 The main aim of this study was to define systems-level differences in cellular proteome 208 composition in three rhizosphere bacterial strains cultivated on five different nitrogen sources. 209 Therefore, bacteria were cultivated on the same nitrogen sources shown in Figure 2 (ammonium, 210 glutamate, lysine, serine and urea), cells were harvested during the exponential growth phase, and 211 cellular protein composition analysed using label-free quantitative proteomics. A numerical 212 summary of protein IDs is shown in Table 1, a visual overview of the derived results is shown in 213  Supplementary Table S4.  216   217 Comparing protein composition across the three strains, it seems that Pseudomonas sp. Root9 218 exhibits more protein-level flexibility compared to the other two strains. This is evident in the 219 PCAs and heatmaps presented in Figure 3, which show that lysine treatment of Pseudomonas sp. 220 Root9 elicits a large proteomic remodelling compared to the other four nitrogen treatments, 221 characterised by hundreds of differentially expressed proteins. In contrast, we see that Rhizobium 222 sp. Root491 exhibits a degree of proteomic homeostasis across the different nitrogen treatments, 223 as shown by the closer clustering of the PCA data points and the lower number of differentially 224 expressed proteins in this strain. 225

226
Comparing across the five different nitrogen sources, we see that each individual nitrogen source 227 seems to elicit a differential proteomic impact in the three different strains. For example, lysine 228 nutrition elicits large-scale changes in the proteome of Pseudomonas sp. Root9, yet relatively few 229 proteomic changes in the other two studied strains. In both Streptomyces sp. Root66D1 and 230 Rhizobium sp. Root491, urea nutrition elicited no proteomic changes compared to ammonium, 231 whereas in Pseudomonas sp. Root9 there were over 100 proteins with differential abundance 232 values between ammonium versus urea (Supplementary Figures S4-S6). 233 234

Orthologous proteins and metabolic pathways modulated by nitrogen nutrition 235
To allow inter-strain comparisons of the label-free quantitative proteomic data acquired from the 236 three taxonomically diverse rhizosphere bacterial strains, we utilised cross-species gene 237 annotation via KEGG orthologues (32). We selected individual proteins that represent the 495 238 KEGG orthologues which were detected in all five treatments across all three strains, and 239 visualise the abundance of these representative orthologues using a heatmap and PCAs in Figure  240 4, with numerical data provided in Supplementary Table S5. As can be seen in Figure 4A and 4B, 241 the samples group together according to the three bacterial strains rather than the five nitrogen 242 sources. This indicates that the baseline differences in strain-specific proteome composition are 243 much greater than any treatment-induced differences elicited by nitrogen nutrition. In Figure 4C  244 we plot a PCA of these 495 KEGG orthologues when protein abundance in the four organic 245 nitrogen sources is normalised versus the inorganic nitrogen source ammonium. This shows that 246 lysine nutrition in Pseudomonas sp. Root9 elicits a proteomic response that is qualitatively 247 different compared to the strain-medium combinations profiled in this study. 248 249 Our next step was to analyse which specific KEGG pathways were modulated according to 250 nitrogen treatment in the three strains. In Figure 5, we show the results of Fisher's exact test to 251 determine whether the constituent proteins of 30 KEGG pathways exhibited altered abundance 252 profiles in the 10 pairwise comparisons between different nitrogen sources. Numerical data for all 253 126 tested pathways compared is provided in Supplementary Table S6. Looking at the specific 254 pathways modulated by nitrogen nutrition across the three strains, it seems that Rhizobium sp. 255 Root491 undergoes fewer alterations to KEGG pathways related to metabolism, but instead 256 exhibits extensive modulation to pathways related to environmental processing and motility. For 257 Pseudomonas sp. Root9 and Streptomyces sp. Root66D1, we see that many of the pairwise 258 comparisons are characterised by widespread modulation to all KEGG pathways, indicating that 259 extensive proteome remodelling has taken place between the different nitrogen sources. 260 Next, we compared the metabolic flux distributions outputted from EnsembleFBA versus the 262 differentially expressed proteins identified in the quantitative proteomic datasets (Supplementary  263   Tables S7-S13). To visualise how nitrogen source affects protein abundance and computationally 264 predicted fluxes, we used the Interactive Pathway Explorer to map KEGG orthologues and 265 reactions onto the KEGG map 'Metabolic Pathways' (33). Visualisations for each of the three 266 amino acid treatments (glutamate, lysine and serine) in pairwise comparisons versus ammonium 267 were produced for both the proteomic data (Supplemental Figure S7) and also the computational 268 modelling data (Supplemental Figure S8). Overall, it is evident that a similar set of metabolic 269 pathways have been mapped in both the experimental and computational approaches, with good 270 coverage of glycolysis, TCA cycle, and amino acid metabolism. However, there is relatively little 271 concordance between the differentially regulated metabolic steps identified by the proteomics 272 data versus the differentially regulated fluxes outputted by EnsembleFBA. For instance, the 273 proteomic data show that lysine nutrition elicits significant modifications to lipid metabolism in 274 Pseudomonas sp. Root9, whereas many of the reaction steps in lipid metabolism are absent from 275 the EnsembleFBA flux distributions. This difference could derive from a known limitations of 276 genome-scale modelling approaches such as EnsembleFBA, because we used a generic biomass 277 function to construct the models, which does not account for variations in bacterial lipid 278 composition between genotypes and treatments (34). Therefore, improved model accuracy 279 probably requires condition-specific measurement of microbial biomass composition. 280 281

Proteins correlated to the PII protein of the nitrogen stress response 282
Analysing the quantitative proteomics data, we noticed that the different nitrogen sources often 283 elicited changes in the abundance of proteins involved in the well-characterised nitrogen stress 284 response, such as GlnK (PII protein), amtB (ammonium transporter) and GlnA (glutamine synthetase) (14). Therefore, we postulated that our dataset may allow us to discover new proteins 286 that are regulatory targets of the nitrogen stress response in less studied bacterial taxa. We first 287 analysed the abundance of PII, a well characterised protein of the nitrogen stress response that 288 exhibited significantly different abundance values between certain nitrogen treatments in all three 289 strains ( Figure 6A). Next, we assessed which other proteins in the dataset were correlated to PII 290 in terms of protein abundance, by plotting their correlation against PII on the x-axis and the slope 291 of this correlation on the y-axis ( Figure 6B, numerical data in Supplementary Table S14). These 292 analyses show that Rhizobium sp. Root491 shows the highest nitrogen stress response under these 293 nitrogen treatments, with all three amino acid treatments leading to dramatic increases in the 294 abundance of the PII protein, and also with many more proteins positively correlated to PII 295 abundance in Rhizobium sp. Root491 compared to the other two strains. Looking at the identity 296 of proteins whose abundance was correlated to PII in Rhizobium sp. Root491, we see that 10 297 proteins controlled by the exo operon that conduct the synthesis and export of extracellular 298 polysaccharides are positively correlated to PII abundance (Supplementary Table S14). 299 Analogous findings have been reported via genetic manipulation of V. vulnificus and S. meliloti, 300 with knockout of nitrogen stress response elements NtrC and NtrX resulting in reduced 301 production of extracellular polysaccharides (35, 36). In Pseudomonas sp. Root9, the data point 302 that exhibits a strong negative correlation to PII is an NADP-dependent glutamate dehydrogenase 303 (Supplementary Table S14), previously shown to be a target of NtrC-driven transcriptional 304 repression in P. putida (37). 305

Discussion 307
Differential nitrogen treatments are a classical experimental manipulation in microbiology, but 308 the majority of molecular knowledge about bacterial nitrogen metabolism has been acquired in E. 309 coli (14). To deepen our knowledge of nitrogen metabolism in the rhizosphere microbiome, this 310 study analyses nitrogen substrate utilisation in three taxonomically diverse bacterial strains 311 previously isolated from field-grown Arabidopsis roots (20). The three strains represent taxa that 312 are consistently detected as core members of the plant microbiome: Pseudomonas, Streptomyces 313 and Rhizobium (21). Using label-free quantitative proteomics, we document hundreds of proteins 314 in each strain that exhibit differential abundance values between nitrogen sources. To enable 315 protein-level comparisons between these taxonomically diverse strains, we integrate the 316 identified proteins using KEGG Orthologues, and map the differential expression of orthologous 317 proteins onto metabolic maps to determine which specific metabolic pathways are modulated by 318 nitrogen source at the protein level. We also determine novel proteins linked to the nitrogen stress 319 response in these three strains, by investigating which proteins display abundance values that are 320 positively and negatively correlated to the PII signal transduction protein. Furthermore, we 321 integrate experimental data with computational models, using the EnsembleFBA method to test 322 how accurately metabolic phenotypes can be computationally predicted from a minimal set of 323 experimental data. Our results show that the three strains exhibit diverse metabolic responses to 324 different nitrogen nutrition regimes, with a summary of key results presented in Supplementary 325 Table S15. show that urea is a relatively poor sole nitrogen source for plant growth (42). Although plants can 387 uptake urea to some degree, a large proportion of the nitrogen delivered via urea fertilisers must 388 first undergo hydrolysis by microbial metabolism before it can subsequently contribute to plant 389 nutrition (43). Therefore, urea metabolism in the rhizosphere microbiome is a potential target for 390 improving agricultural nitrogen use efficiency. In our work, we show that all three tested strains 391 can grow rapidly on urea as a sole nitrogen source. However, the proteomic impact of urea 392 nutrition differed widely between the three strains, with Streptomyces sp. Root66D1 and 393 Rhizobium sp. Root491 both showing zero proteins that were differentially expressed between 394 ammonium versus urea treatment, whereas this comparison in Pseudomonas sp. Root9 elicited 395 126 differentially expressed proteins. The urease enzyme that converts urea to ammonium is 396 required under normal conditions for catabolism of purine and arginine, and is increasingly 397 expressed under nitrogen stress as a nutrient salvage mechanism. In our dataset, all three strains 398 exhibit high expression of urease subunits under all conditions tested, and our investigations of 399 the nitrogen stress response showed that many urease subunits are tightly correlated to PII 400 expression. For all three strains, we see that at least one amino acid treatment actually elicits a 401 higher urease expression compared to urea nutrition. This suggests that urease abundance is not the limiting factor for utilisation of urea as a sole nitrogen source, and that other mechanisms may 403 explain urea-induced proteome remodelling in Pseudomonas sp. Root9. Inspecting the data, we 404 see many transporter proteins are differentially expressed in Pseudomonas sp. Root9 under urea 405 versus ammonium nutrition, which may be involved in urea uptake or the excretion of urea-406 derived waste products. In comparison, the transport machineries of both Streptomyces sp. 407 Root66D1 and Rhizobium sp. Root491 seem to already be primed for urea uptake when cultivated 408 on ammonium. Future studies could investigate how to optimally coordinate urea transport and 409 metabolism between plants and rhizosphere microbes to deliver higher nitrogen use efficiency 410 from urea fertilisers. 411 412 Many microbial strains have been labelled as plant growth promoting, but there is relatively little 413 knowledge about the genes and mechanisms that underpin this trait (44). In previous work, 414 Rhizobium sp. Root491 was characterised as a plant growth promoting bacterium by its ability to 415 increase Arabidopsis root length in co-cultivation experiments (45). Furthermore, 416 exometabolomics profiling has shown that Rhizobium sp. Root491 can consume a wide variety of 417 plant-derived metabolites as carbon substrates (46). Here, we show that Rhizobium sp. Root491 418 exhibits fast growth on a variety of nitrogen sources, that its set of ABC transporters exhibit 419 differential abundance values in response to nitrogen source, and also that amino acid nutrition 420 induces the expression of multiple proteins involved in the production of extracellular 421 polysaccharides. When combined with previous observations of Rhizobium sp. Root491, we can 422 begin to characterise the functional traits possessed by this strain that contribute to plant growth 423 promotion, such as: recruitment to the rhizosphere via the consumption of plant root metabolites, 424 adherence to the root surface via biofilm production in the presence of plant-derived amino acids, 425 and the potential for mineralisation of diverse nitrogen molecules to fuel plant nutrition.
Potentially, future studies could predict whether other rhizosphere strains can also promote plant 427 growth via similar mechanisms, by investigating genetic similarities with Rhizobium sp. Root491. 428 Also, future work could investigate whether plant genotypes differ in their ability to attract 429 growth-promoting strains to the rhizosphere, and how to design synthetic microbial communities 430 that combine multiple growth-promoting strains. 431 432 There is increasing interest in combining experimental and computational approaches to analyse 433 microbial metabolism, with the long-term goal of quantitatively predicting the behaviour of 434 microbial communities (47). Metabolic modelling is rapidly progressing as a powerful 435 computational tool to explore the metabolic capacities of bacteria. However, the main limitation 436 that prevents modelling approaches from being applied to diverse bacterial strains is the need to 437 obtain a highly curated genome-scale metabolic model for each strain of interest. This process of 438 model curation still requires a significant amount of manual inspection and relies heavily on 439 accurate genome annotation (25). In the present study, we used EnsembleFBA (26) to produce 440 metabolic models for three diverse bacterial strains using a minimal set of experimental 441 information. We compared the derived models versus experimental data by assessing how 442 accurately they can predict growth phenotypes and proteome remodelling across different 443 nitrogen sources. This showed that EnsembleFBA gives relatively accurate predictions of 444 nitrogen substrate utilisation, with binary phenotypes (growth versus no growth) correctly 445 predicted in around 80% of cases. However, there was only an intermediate correlation between However, proteomic measurements showed that the strains deploy different metabolic strategies 472 to utilise specific nitrogen sources. One diverging trait appears to be their degree of proteomic 473 flexibility, with Pseudomonas sp. Root9 utilising lysine via widespread protein-level alterations to its flexible metabolic network. In contrast, Rhizobium sp. Root491 shows relatively stable 475 proteome composition across diverse nitrogen sources, characterised by minimal alterations to 476 central metabolism but differential abundance of many transport proteins. In addition, we 477 document a large set of functionally uncharacterised proteins that display differential abundance 478 values in response to nitrogen source, with functional annotations being particularly unclear in 479 Gram-positive Streptomyces sp. Root66D1. These proteins are potentially important for nitrogen 480 metabolism in the rhizosphere, and could be the targets of future functional study. Our results 481 could inform the selection of high-performing strains in synthetic microbial communities 482 designed to mediate plant nitrogen nutrition under lower inputs of mineral fertilisers. 483

Bacterial strains 486
Bacterial strains used in this study were Pseudomonas sp. Root9 (NCBI Taxonomy ID:  For phenotype microarrays using PM3B (Biolog), 12 ml of inoculant was prepared comprising 10 501 mL of 1.2× IF-0 (Biolog), 1.2 mL of 500 mM glucose, 600 uL of bacterial suspension (as 502 prepared above), 120 uL of Redox Dye D (Biolog) and 80 uL of sterile water. Next, 100 uL of 503 this inoculant (starting OD 600 of 0.05) was loaded into each well of the phenotype microarray, 504 which was transferred to a plate reader (Tecan Infinite Pro 100) and incubated at 28° C for 72 h 505 with shaking (30 sec continuous orbital shaking followed by 9:30 min stationary, shaking 506 amplitude 3 mm). Tetrazolium reduction at A 590 was measured once per 10 min cycle, without 507 correcting for path length, and derived curves were fitted to a logistic equation using the Growthcurver program (51). For each well in every assay, background was subtracted by 509 subtracting the value of the negative control (well A1) from each time point. In our hands, 510 guanosine (well F7) gave a very high background reading and was excluded from the analysis. 511 Wells were considered growth-positive if the carrying capacity (k) of the logistic fit was greater 512 than A 590 of 0.1 in at least two of the three independent biological replicates. Next, area under the 513 curve (AUC) values for all growth-positive wells were z-score normalised within each strain, and 514 the average value of the three replicate assays was calculated. These averaged z-score values 515 were divided into quartiles, so data presented in Fig 1 represent five possible growth intensities, 516 ranging from 0 (no growth) to 4 (highest AUC quartile). 517 518

Metabolic models and computational simulations 519
The EnsembleFBA workflow from Biggs and Papin (26) was adapted to analyse the three studied 520 bacterial strains. Scripts were implemented either in Matlab (Mathworks) as the original code, or 521 adapted for Python (Python Software Foundation). Briefly, genomes were downloaded from 522 NCBI (52) and uploaded to KBase (25), where genome re-annotation and draft metabolic model 523 reconstruction was performed. Outputted draft networks were downloaded and used as inputs for 524 the EnsembleFBA workflow. Also inputted to Ensemble FBA were the composition of the 525 Biolog media, and the experimentally derived growth matrices obtained from PM3B phenotype 526 microarray. Next, 50 metabolic networks were generated for each strain, with each network being 527 trained on 26 nitrogen substrates that supported growth and 11 nitrogen substrates that didn't 528 support growth, in order to perform positive and negative gapfilling. Compounds present on the 529 phenotype microarray but not found in the ModelSEED database (24) were excluded, and a 530 second set of simulations excluding the five N-sources used for proteomics experiments were 531 also obtained for unbiased integration with the proteomics datasets. To evaluate the performance 532 of EnsembleFBA for predicting growth on the different N-sources, its accuracy, precision and 533 recall were compared to randomly generated predictions, after masking the conditions used to 534 gapfill the individual networks to avoid bias. Metabolic activity on a given nitrogen source was 535 estimated as the average growth rate obtained with EnsembleFBA, and weighted according to the 536 fraction of networks in the ensemble that predicted growth. Metabolic fluxes through specific 537 reactions were estimated by averaging the reaction flux for each reaction across all the networks 538 in the ensemble, and weighted according to the fraction of networks where the reaction was 539 active. To visualise up-or down-regulated metabolic fluxes in metabolic pathway maps, 540 metabolic fluxes obtained by simulating growth on Glutamate, Serine or Lysine were compared 541 versus Ammonium, and filtered for reactions with log2 fold change greater than 1. 542 543

Cultivation on individual N-sources for growth assays and proteomic analysis 544
For growth assays on individual N-sources, media were based on M9 formulation (53), with 545 nutrient concentrations of: 50 mM glucose, 24 mM Na 2 HPO 4 , 11 mM KH 2 PO 4 , 4 mM NaCl, 350 546 μM MgSO 4 , 100 μM CaCl 2 , 50 μM Fe-EDTA, 50 μM H 3 BO 3 , 10 μM MnCl 2 , 1.75 μM ZnCl 2 , 1 547 uM KI, 800 nM Na 2 MoO 4 , 500 nM CuCl 2 , 100 nM CoCl 2 . To this, one nitrogen source was 548 added at 5 mM elemental-N (ie: 5 mM of ammonium, glutamate and serine, or 2.5 mM of urea 549 and lysine). For growth assays, 20 μL of bacterial suspension (as prepared above) was inoculated 550 into 380 μL of growth medium (starting OD 600 of 0.05), in individual wells of a sterile 48-well 551 plate (Corning). These plates were then transferred to a plate reader (Tecan Infinite Pro 100) and 552 incubated at 28° C for 48 h with shaking (3 min continuous orbital shaking followed by 7 min 553 stationary, shaking amplitude 3 mm). Culture density at OD 600 was measured once per 10 min 554 cycle, without correcting for path length. To obtain quantitative growth metrics, a logistic 555 equation was fitted to measured growth curves using the Growthcurver program (51). To collect samples for proteomics, cultivation was identical, except that bacterial cells were harvested 557 during the exponential growth phase. Harvest involved pooling of four duplicate wells (total of 558 1.6 mL culture), followed by centrifugation at 10,000× g for 3 min at 4° C. Supernatant was 559 discarded, and cell pellets were rinsed twice with 900 uL of 4° C PBS via centrifugation at 560 10,000x g for 3 min at 4° C. Rinsed cell pellets were then flash-frozen and stored at -80° C. 561 562 Proteomic sample preparation 563 Cellular protein was extracted using protocols modified from Tanca et al (54) as well as Wessel 564 and Flugge (55). To frozen cell pellets, 250 uL of lysis buffer (5% SDS, 100 mM DTT, 100 mM 565 Tris pH 7.5) was added, along with ~100 uL of acid-washed glass beads (1 mm diameter). 566 Samples were then incubated for 10 min on an orbital mixer at 95° C with 1500 rpm shaking, 567 then at -80° C for 10 min, then bead-beaten (Bead Ruptor 24, Omni International) at 5 ms-1 for 568 10 min. Next, samples were again incubated at -80° C for 10 min, then again incubated for 10 569 min on an orbital mixer at 95° C with 1500 rpm shaking, then again bead-beaten at 5 ms-1 for 10 570 min. Finally, samples were centrifuged at 20,000x g for 10 min at RT, and 200 uL of supernatant 571 was transferred to a new tube. Protein was then precipitated via the addition of 800 uL MeOH, 572 500 uL H 2 O, and 200 uL chloroform followed by centrifugation at 10,000x g for 5 min at 4° C. 573 The upper aqueous phase was removed and discarded, then 700 uL MeOH was added to the 574 lower organic phase and samples were centrifuged at 20,0000x g for 10 min at 4° C. Protein 575 pellets were then rinsed twice with -20° C acetone via centriguation at 20,0000x g for 10 min at 576 4d C, before being air-dried at RT for 15 min. Dried protein pellets were then stored at -80° C. To 577 solubilise protein pellets, 40 uL of solubilisation buffer (8 M urea, 50 mM TEAB, 5 mM DTT) 578 was added, and samples were incubated on an orbital mixer at 28° C for 1 h with 350 rpm 579 mixing. Next, CAA was added to a final concentration of 30 mM, and samples were incubated on an orbital mixer at 28° C for 30 min with 350 rpm mixing in darkness. To quantify protein 581 concentration, an aliquot of the protein extract was taken and diluted 1:8 in water, then a 582 Bradford assay was performed on the diluted protein samples using BSA as standard. Next, 40 ug 583 of protein extract was transferred to a new tube and incubated with 0.8 ug Lys-C for 2 h at 37d C 584 with 350 rpm shaking. Samples were then diluted 1:8 in TEAB, 0.8 ug of trypsin was added, and 585 samples were incubated overnight at 37° C. Next day, samples were acidified by adding formic 586 acid to a final concentration of 1%. Peptides were then cleaned up via SPE using SDB-RP stage 587 tips. Following elution from stage tips, peptides were dried down in a vacuum centrifuge and 588 stored at -80° C. 589 590

Mass spectrometry 591
Digested peptides were analysed on a QExactive Plus mass spectrometer (Thermo Scientific) 592 coupled to an EASY nLC 1000 UPLC (Thermo Scientific). Dried peptides were resolubilised in 593 solvent A (0.1% formic acid), and loaded onto an in-house packed C18 column (50 cm × 75 µm 594 I.D., filled with 2.7 µm Poroshell 120, (Agilent)). Following loading, samples were eluted from 595 the C18 column with solvent B (0.1% formic acid in 80% acetonitrile) using a 2.5 h gradient, 596 comprising: linear increase from 4-27% B over 120 min, 27-50% B over 19 min, followed by 597 column washing and equilibration. Flow rate was at 250 nL/min. Data-dependent acquisition was 598 used to acquire MS/MS data, whereby the 10 most abundant ions (charges 2-5) in the survey 599 spectrum were subjected to HCD fragmentation. MS scans were acquired from 300 to 1750 m/z 600 at a resolution of 70,000, while MS/MS scans were acquired at a resolution of 17,500. Following 601 fragmentation, precursor ions were dynamically excluded for 25 s. 602 603  proteins (DEPs) for each of the three strains. To define DEPs, protein abundance in one condition 835 was compared to its abundance in the other four conditions. If in any of these 10 comparisons, a 836 protein has a log2FC > 1 and a BH-p-value < 0.05, then it is considered a DEP. Only DEPs that 837 were detected in at least three replicates for all five nitrogen treatments are included in the 838 heatmaps. Rows were clustered using Pearson's correlation coefficient. rhizosphere bacterial strains. Different letters above data series indicate p<0.05 following two-864 way ANOVA and Tukey's HSD test. B: Plots to highlight proteins that are positively or 865 negatively correlated to PII according to their abundance values across five nitrogen treatments. 866 Y-displays the slope of the linear fit (z-score normalised) between protein abundance versus the abundance of PII protein, and X-axis displays correlation between protein abundance versus PII 868 abundance. If a protein has a correlation higher than 0.75 and a slope higher than 2, it is deemed 869 positively correlated, whereas if a protein has a correlation lower than 0.75 and a slope lower 870 than -2, it is deemed negatively correlated to PII. 871 872   showing the correlation between predicted metabolic activity (EnsembleFBA) versus measured 889 metabolic activity (EnsembleFBA) for 81 nitrogen substrates across three bacterial strains. D: