Genome acquisition, quality filtering, and genotyping

Genomes were included from 29 medium-large-scale studies [4,20–46] that comprised diverse KpSC collections (S1 Data). Outbreak investigations were explicitly excluded to reduce clonal oversampling and genomes were de-replicated to minimize sampling biases for overrepresented SLs (see method: Genome dereplication). Reads were acquired from the European Nucleotide Archive using the enaDataGet and enaGroupGet scripts from enaBrowserTools v1.6 [47], trimmed with Trim Galore v0.5.0 using -q 20 option [48], then assembled with Unicycler v0.4.7 with the --keep 0 option [49]. Assemblies failing the established KlebNET-Genomic Surveillance Platform inclusion criteria were excluded (>500 contigs, total length <4,969,898 or >6,132,846 bp). Graphical Fragment Assembly (GFA) files were analyzed to count GFA dead ends using getUnicyclerGraphStats.py from Unicycler [49]

Assemblies were annotated with Bakta v1.1.1 [50] with the ‘--gram –’ option and Bakta database (date accessed: 1/09/2021), while Kleborate v2.0.4 [51] was used to identify species and Sequence Types (STs). Pathogenwatch (https://pathogen.watch/) and BIGSdb (https://bigsdb.pasteur.fr/) [52,53] were used to assign LIN codes [54] and SLs. LIN codes are assigned using core genome MLST profiles that consider variation in 629 core KpSC genes. The third position in the LIN code is used to infer SL membership [54] and results in groupings that correspond to well-recognized deep-branching monophyletic clades within species. Notably, these SL assignments do not rely on de novo cluster analysis, and are stable with the addition of new genomes [55]. Pairwise ANI was calculated using FastANI v1.33 [56] using --fragLen 3000.

Genome dereplication

We noticed that several SLs were vastly overrepresented in the dataset (e.g., due to dense sampling in specific geographies or inclusion of transmission clusters within larger collections). In order to minimize the impact of these sampling biases we de-replicated the dataset using a similar approach to that described in [51]. Initially, popPUNK v2.4.0 [57] was used to finely cluster isolates. The ‘create-db’ function was used with the following options: ‘--sketch-size 1000000 --min-k 15 --max-k 29 --qc-filter prune’. Then the ‘fit-model’ function was used with the following options: ‘dbscan --ranks 1,2,3,5 --graph-weights’. Finally, the ‘fit-model’ function was used again, with the options: ‘refine --graph-weights --unconstrained’. Additionally, the ‘poppunk_visualise’ function was used, with the ‘--distances’ and ‘--previous-clustering’ utilizing the refined model fit, to output a neighbor-joining core tree.

Isolates which shared the same popPUNK cluster, isolation source and country were sorted using newfangled_untangler.py v1.0.4 (https://github.com/bananabenana/newfangled_untangler/) then de-replicated via Assembly-Dereplicator v0.1.0 [58] with the following options: ‘--threshold 0.0003’. Code also available at Figshare (https://doi.org/10.6084/m9.figshare.24503737).

Pan-genome analysis

The 7,835 filtered and dereplicated KpSC genomes were split into species, then Panaroo v1.2.8 [59] was run with the following options: ‘--clean-mode sensitive -a core --aligner mafft --no_clean_edges --core_threshold 0.98 --merge_paralogs --remove-invalid-genes’. Due to computational limits, the 6,676 K. pneumoniae genomes were split into two groups of 3,338 (based on closest genetic distance) and run separately on Panaroo. The output graphs for all species were then merged using the panaroo-merge command with the following options: ‘--merge_paralogs’ to obtain the final pan-genome.

Metabolic gene identification and analysis

To analyze metabolic pathways, the pan_genome_reference.fa file from Panaroo was translated using the transeq tool from EMBOSS v6.6.0 [60] with the following options: “-table 11 -frame 1.” Translated sequences were analyzed using kofam_scan v1.3.0 (https://github.com/takaram/kofam_scan), the command line version of KEGG’s [61] kofamKOALA [62]. KEGG release 94.0 ver. 2020-04-02.

Genes with an e-value of ≤0.001 were retained and analyzed using KEGG Mapper (https://www.genome.jp/kegg/mapper/reconstruct.html) to identify pathway information. Using the BRITE tab, genes which fell under the ‘metabolism’ subheading were used to classify genes as metabolic or nonmetabolic. KEGG ortholog assignments were used to convert the Panaroo gene presence/absence matrix into a metabolic ortholog matrix. This allowed grouping of functional isozymes which would otherwise have an identical KEGG function. Pairwise metabolic ortholog Jaccard similarities were calculated using the vegdist function from R package vegan v2.5-7 [63].

Twilight [64] was used to classify distribution of metabolic orthologs among KpSC taxa and 48 common K. pneumoniae SLs (n ≥ 15 genomes each, defined by LIN codes) using the -s 5 and -s 20 commands, respectively. ‘Core’: ≥95% prevalence. ‘Intermediate’: <95% and >15% prevalence. ‘Rare’: ≤15% and >0% prevalence. ‘Absent’: 0% prevalence.

Metabolic modeling

Metabolic models were generated via Bactabolize v1.0.1 [65] using the draft_model command with the following options: “--media_type m9 --atmosphere_type aerobic --min_coverage 25 --min_pident 80.” The KpSC-pan v2.0 model [66] was used as the input reference. One-hundred and sixty of 7,835 models (2.0%) required gap-filling to enable simulation of growth on M9 minimal media with glucose (median: 1 reaction added, range: 1–23), using flux balance analysis to optimize the most recent version of the biomass objective function “BIOMASS_Core_Feb2022.” We have previously shown that models gap-filled to simulate growth in minimal media plus glucose can accurately predict a range of other substrate growth phenotypes [65].

Flux balance analysis was used to predict growth phenotypes for all possible carbon, nitrogen, phosphorus, and sulfur sources supported by the reference model as sole sources of carbon, nitrogen, phosphorus, or sulfur in M9 minimal media defined in silico, in both aerobic and anaerobic conditions. These analyses were performed using Bactabolize’s fba command with the following option: “--fba_spec_name m9.” All growth lower bounds and default substrate sources used in this analysis are available at https://github.com/kelwyres/Bactabolize. Models were considered capable of simulating growth when the optimized biomass value was ≥0.0001.

Pairwise Jaccard similarities and population distribution of positive growth phenotypes were calculated as described above for metabolic orthologs.

To evaluate co-occurrence of metabolic traits and control for population structure, genomes were randomly subsampled to a maximum of 10 per SL (n = 2,427 genomes), then predicted aerobic growth phenotypes analyzed using Coinfinder v1.2.0 [67] with the --associate option.

In vitro growth experiments

To validate substrate usage predictions from the metabolic models, 13 isolates (S2 Data) were selected from our in-house collections [4,26,68] and tested in triplicate on each of nine substrates in minimal media, aerobic conditions. Two experiments were performed: i) endpoint OD₆₀₀ at 24- and 48-hours; ii) 3-day substrate coaxing to induce gene expression [69] for predicted false-negative growth results in the endpoint experiment. Isolates were coaxed onto the testing substrate via sub-culturing into progressively lower amounts of D-glycerol, which all isolates can use as a carbon source (10 mM on subculture 1, 5 mM on subculture 2, 0 mM on subculture 3).

Nine growth substrates were tested: seven as sole sources of carbon (methanol, L-hydroxyproline, xylitol, butyrate, acetoacetate, D-glycerol, and D-glucose), one as a sole source of nitrogen (allantoin) and one as a sole source of sulfur (methanesulfonate).

For endpoint experiments, isolates were grown overnight in 2 mL M9 media + 10 mM D-glycerol. M9 media was prepared by using 1X M9 Minimal Salts (Sigma) and adding 2 mM MgSO₄ and 0.1 mM CaCl₂ after autoclaving. 1 mL of culture was pelleted at 8,000 x g for 10 min, supernatant discarded, then centrifugally washed using 1 mL 0.9% NaCl to remove excess carbon/nitrogen/sulfur. Isolates were then diluted 1:100 using 0.9% NaCl and 5 µL sub-cultured into 96-well plates (CLS3603, Corning) containing 200 µL M9 + substrate. Substrates were tested at the following concentrations: 400 mM methanol, 20 mM L-hydroxyproline, 20 mM xylitol, 20 mM sodium butyrate, 20 mM acetoacetate, 20 mM D-glucose, 10 mM D-glycerol 20 mM methanesulfonate, 60 mM allantoin, all at pH 7.0 and 0.2 µm filter sterilized. For allantoin as a nitrogen source, custom nitrogen-negative M9 was prepared (34 mM NaH₂PO₄, 64 mM K₂HPO₄, 1 µM FeSO₄, 2 mM MgSO_4, and 0.1 mM CaCl₂). For methanesulfonate as a sulfur source, sulfur-negative M9 was prepared (2 mM MgCl₂ instead of MgSO₄). Isolate negative and carbon positive controls were performed for each media. Plates were incubated at 37 °C, shaking at 150 RPM and optical density read every 24 hours at OD₆₀₀ absorbance using an infinite 200Pro and i-control v2.0.10.0 (Tecan). Each experiment was performed in biological triplicate. Each well was blanked with a no-isolate control, then the mean used to determine growth, with a cutoff of 0.05 (limit of detection).

Coaxing experiments were performed as above, over 3 days. Isolates were added to media containing the tested substrate plus D-glycerol 10 mM on subculture day 1, 5 mM on subculture day 2, and 0 mM on subculture day 3. Prior to sub-culturing, cells were centrifugally washed as above and 5 µL transferred to 200 µL total media. Only subcultures on day 3 were used for analysis (0 mM D-glycerol).

Co-culture experiments

We selected three pairs of SLs and three corresponding pairs of substrates (four total substrates) for which one was predicted as a SL-specific core trait in SL A and the other predicted as a SL-specific core trait in SL B. For each SL, three isolates were selected from our in-house collections [4,26,70] and treated as biological replicates (S3 Data).

Isolate pairs were grown in single and co-culture using 0.4 µm pore Transwells (Sigma) and 24-well plates (Costar, Corning). Co-culture results were compared to single cultures and log₂-fold-change OD₆₀₀ values calculated.

To rule out strain-strain antagonism prior to co-culturing, a primary inoculum was streaked onto LB agar plates and grown for 8 hours at 37 °C. Then, primary inoculums of additional isolates were streaked perpendicularly. Plates were incubated for 16 hours at 37 °C and no growth inhibition of the perpendicular streaks was observed.

Co-culture results were compared to single cultures using OD₆₀₀ of pairs of isolates (S3 Data) that were physically separated by 0.4 µm pores using 24-well Transwells (Sigma). All cultures were grown at 37 °C, 100 RPM. Substrates were used in 1X M9 media at the following concentrations: 20 mM for galactitol, L-sorbose, L-tartrate, and 5 mM for L-hydroxyproline.

M9 media plus the tested substrate was added to the Transwell (100 µL) and the main well (800 µL). Isolates were grown overnight in 1X M9 containing 10 mM D-glycerol, then centrifugally washed as above. Cultures were diluted 1:100 in 0.9% NaCl, and 5 µL added to wells. The prototroph was added to the Transwell while auxotroph added to the main well. Cultures were grown and 75 µL was taken from wells at time point for OD₆₀₀ measurements. For isolate-substrate combinations where we observed slow growth rates, 24, 48, and 72-hour time points were used, otherwise time points at 0, 24, and 36 hours were used. OD₆₀₀ values were blanked using no-isolate media controls. For log₂ fold-change analysis, values of 0 were transformed to 1x10⁻⁵, but kept raw for other statistical analyses. After co-culture, isolates were then plated onto LB agar, confirmed uniform, distinct colony morphology, then re-grown under the same nutrient conditions in single culture.

Co-culture simulations

In silico co-culture experiments were performed using MICOM v0.32.2 [71] with the relevant K. pneumoniae metabolic models. Community models were built using the build function, then communities were grown in growth media matching co-culture experiments (M9 minimal media with various carbon sources) with tradeoff set to 0.5. Metabolite and reaction flux were then analyzed to determine metabolite sharing from prototrophs to auxotrophs.

Preparation of samples for metabolomics analysis

We used an untargeted Liquid Chromatography-Mass Spectrometry (LC-MS) approach to explore the metabolites present in the culture supernatant of putative cross-feeding prototrophs. We selected one representative prototroph for each SL-growth substrate combination: AJ155, INF359, INF120 (galactitol), INF171 (L-hydroxyproline), INF225 (L-sorbose), and INF354 (L-tartrate) and compared to no-isolate controls for each substrate. Briefly, overnight cultures in M9 minimal media + 20 mM glucose were washed twice in 0.9% NaCl, centrifuged at 8,000 x g for 10 min. Cultures were then standardized to McFarland Standard 0.45–0.55 in 0.9% NaCl, then inoculated into 5 mL M9 minimal media + substrates (as previously described in co-culture experiments). Performed with four biological replicates except negative controls which were performed in triplicate. Independent cultures were grown aerobically at 37 °C, shaking at 200 RPM for 18 hours. Cells were pelleted at 8,000 x g for 10 min, then supernatant filtered through a Acrodisc Supor PES Membrane 0.2 μm syringe filter (Pall). Supernatant was transferred to a new vial and frozen at −80 °C prior to transfer to the Monash Proteomics and Metabolomics Platform for LC-MS analysis, where 20 µL of cell culture supernatant was treated with 180 µL of ice-cold methanol containing a mixture standard isotope labeled amino acids acting as internal standards. The sample was then mixed at 4 °C for 10 min on a thermomixer before being subjected to centrifugation at 21,000 x g for 10 min at 4 °C. 900 µL of the supernatant was then transferred to an LC-MS vial for analysis.

LC-MS analysis of metabolites

We used both hydrophilic interaction liquid chromatography (HILIC) coupled to high-resolution mass spectrometry (HRMS). Samples were analyzed on a Vanquish Horizon coupled to Exploris 480 orbitrap mass spectrometer (Thermo Scientific). Chromatography was performed using an iHILIC-(P) Classic, (PEEK, 150 x 4.6 mm, 5µm, 200Å; HILICON, Sweden) with a gradient elution of 20 mM ammonium carbonate (A) and acetonitrile (B) (linear gradient time-%B as follows: 0 min-80%, 15 min-50%, 18 min-5%, 21 min-5%, 24 min-80%, 32 min-80%) at a flow rate of 500 µL/min. The mass spectrometer was operated in polarity switching modes with a resolution at 120,000. The electrospray conditions were: ionization voltage was 3.5 kV in positive and −2.5 kV in negative mode, ion transfer tube temperature = 325 °C; sheath gas = 60; Aux gas = 15; sweep gas = 2; vaporizer temp = 350 °C.

For accurate metabolite identification, a standard library of ~400 metabolites was analyzed before sample testing and accurate retention time for each standard was recorded. This standard library also forms the basis of a retention time prediction model used to provide putative identification of metabolites not contained within the standard library [72]. Acquired LC-MS/MS data was processed in an untargeted fashion using opensource software IDEOM [73], which initially used msConvert (ProteoWizard) [74] to convert raw LC-MS files to mzXML format and XCMS [75] to pick peaks to convert to peakML files. Mzmatch was subsequently used for sample alignment and filtering [76]. Peak intensities were compared for each prototroph versus the matched no-isolate control using a Wilcox Rank Sum test with Benjamini-Hochberg multiple testing correction. Missing and zero peak intensity data were imputed as 20% of the minimum peak intensity for each prototroph-control pair.

Data visualization and code availability

Statistical analysis and graphical visualization were performed using R v4.0.3 [77], RStudio v1.3.1093 [78], with the following software packages: tidyverse v1.3.1 [79], viridis v0.5.1 [80], RColorBrewer v1.1-2 [81], ggpubr v0.4.0 [82] ggpmisc v0.4.4 [83], aplot v0.1.6 [84], colorspace v2.0-2 [85], ggtree v2.4.1 [86], vegan v2.5-7 [63], rstatix v0.7.0 [87], and ggnewscale v0.4.5 [88].

All code used to generate metabolic models and simulate growth phenotypes, including media definitions, is available at https://github.com/kelwyres/Bactabolize. The KpSC pan-reference metabolic model is available at https://github.com/kelwyres/KpSC-pan-metabolic-model. All code used to generate figures, simulations, and perform statistical analysis can be found at (https://doi.org/10.6084/m9.figshare.24503737), alongside all metabolic models used in analyses.

Metabolic traits are bimodally distributed across the population

The study dataset included 7,835 high-quality KpSC genomes covering all seven taxa in the complex (Fig 1A) and 1,931 STs (S1 Data). Pan-genome analysis [59] identified 61,595 gene clusters, 11,184 (18.2%) of which were matched to a total of 2,601 unique KEGG Orthologs [61] with associated substrate-reaction data. This represented a lower bound estimate of the proportion of ‘metabolic genes’ in the KpSC pan-genome, as 38.9% of gene clusters could not be assigned functional annotations (hypothetical proteins). Metabolism-associated KEGG Orthologs (hereafter ‘metabolic orthologs’) were bimodally distributed (Fig 1B and S4 Data), with 1,499 orthologs (57.4%) present in ≥95% of genomes and 899 rare orthologs present in <15% of genomes. This bimodal distribution is consistent with the general distribution of accessory genes in the KpSC [4], and other bacterial species with similarly large effective population sizes [64,89].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Population-level metabolic diversity in the K. pneumoniae species complex.

A: Neighbor-joining tree of the 7,835 KpSC genomes used in this study. Tip colors indicate species as labeled. An interactive version found at https://microreact.org/project/pcTMwQZAGCsqqhbEZzaj11-a-metabolic-atlas-of-the-klebsiella-pneumoniae-species-complex. B: Stacked bars show pan-genome content by metabolic gene (KEGG) category as indicated in legend (left). Population distribution of KEGG orthologs associated with metabolic genes (‘metabolic orthologs’, right). The light gray background shows rare orthologs (>0–15%), the medium gray shows intermediate orthologs (>15–<95%) while the dark gray shows core orthologs (≥95%). C: Stacked bars show population frequency categories for predicted substrate usage, coloured by substrate type as indicated in legend. The data and phylogeny underlying this Figure can be found at Figshare (https://doi.org/10.6084/m9.figshare.24503737) and S4 Data and S5 Data.

https://doi.org/10.1371/journal.pbio.3003559.g001

To analyze the phenotypic diversity across the KpSC, a strain-specific metabolic model was generated for each genome using Bactabolize [65], which was previously shown to produce models with overall 93.2% and 78.1% predictive accuracies in aerobic and anaerobic conditions, respectively (data for 37 diverse KpSC each tested for growth using 124 distinct carbon substrates via the Omnilog microarray; range 5.4%–100.0%, median 100.0% accuracy per substrate in aerobic conditions; range 0.0%–100.0%, median 88.9% accuracy per substrate in anaerobic conditions [66]). Growth phenotypes were predicted for 345 carbon, 192 nitrogen, 68 phosphorus, and 34 sulfur substrates as sole sources in minimal media under aerobic and anaerobic conditions (a total of 1,278 conditions). Individual strain-specific models comprised a mean of 1677.0 genes ± 17.3 (standard deviation, SD) and 3330.1 reactions ± 13.5 (SD). Individual isolates were predicted to grow in a mean of 678.2 ± 17.9 (SD) conditions. Seventeen isolates were identified as outliers with fewer positive growth phenotypes in anaerobic conditions, likely due to sequencing coverage issues rather than genuine biological deficiencies [90] (S5 Data, S1A Fig, and S1 Text).

The distribution of positive growth predictions mirrored that for metabolic gene content. At the population-level, 643 conditions (47.6%) were predicted to support growth for ≥95% isolates (‘core’, Fig 1C and S5 Data) and were comprised of 349 aerobic and 294 anaerobic conditions, which systematically confirmed the facultative anaerobe status of the KpSC. These 643 core conditions consisted of 207 total substrates (note that some substrates can be tested as sole sources of multiple elements, e.g., carbon and nitrogen), with at least 204 of these being known human metabolites, 134 diet-related compounds, and 203 gut microbiome metabolites (S5 Data). In contrast, 509 modeled growth conditions (43.6%) did not support growth of any isolates (224 aerobic and 285 anaerobic conditions). Carbon substrate usage displayed the greatest variability, with 41 aerobic and 37 anerobic conditions predicted to support growth in <95% isolates (Figs 1C and S1B).

Metabolic variation is structured within and between species

The set of core growth conditions for each KpSC taxon was greater than the species complex as a whole (11–39 additional conditions each). Each taxon was associated with a distinct core growth profile mirrored by distinct core metabolic ortholog profiles as expected [70,91] (S2–S4A Figs, S4 Data, S5 Data, and S1 Text), and there was evidence of specialization at the level of macromolecular functions as determined by COG analysis (S5 Fig, S1 Data, and S1 Text). These predicted growth capabilities were generally consistent with biochemical tests described in formal species definitions [92–95]. However, there were a small number of notable discrepancies that indicate that species identification protocols should be revisited (see S1 Text and S6 Data for details).

While individual KpSC taxa are clearly defined by distinct metabolic profiles, there was also clear evidence of intra-taxon variation (S2 and S3 Figs). Four taxa were represented by sufficient genomes for comparisons (n > 200 each); K. pneumoniae (hereafter Kp, n = 6,652), Klebsiella quasipneumoniae subsp. quasipneumoniae (Kqq, n = 201), K. quasipneumoniae subsp. similipneumoniae (Kqs, n = 285), and Klebsiella variicola subsp. variicola (Kvv, n = 672). In Kp, 938 metabolic orthologs were variably present in <95% genomes (544, 586, and 738 orthologs variable among Kqq, Kqs, and Kvv, respectively) and 79 conditions were predicted to support variable growth (81, 78, and 67 for Kqq, Kqs, and Kvv). We hypothesized that this variation is not randomly distributed in the population, and our analysis of all-vs-all genome pairs within taxa showed statistically significant positive relationships between pairwise Average Nucleotide Identity (ANI, generally indicative of ancestral relatedness) and pairwise Jaccard similarity across metabolic genes (general linear hypothesis test, p < 2 × 10⁻¹⁶ for each of the four well-sampled taxa, s 2A). Comparison of the regression gradients indicated significant differences between taxa; however, the R² values, which indicate how well the models fit the data, were low (range 0.04–0.48), suggesting a large amount of variation in metabolic gene Jaccard similarities was not explained by the relationship with ANI.

Next, we used the recently established core-genome Life Identifier Number (LIN) code scheme to assign genomes to SLs [54]. These SLs represent deep-branching phylogenetic clades, separated by hundreds of years of evolutionary divergence (with notable exceptions discussed below). Of the 7,835 genomes, 7,833 were assigned to 1,418 distinct SLs and two were unassigned due to insufficient alleles detected.

Pairwise Jaccard similarities for metabolic ortholog profiles were significantly higher within SLs (median 0.956, IQR 0.947–0.965) than between SLs within the same species (median 0.923, IQR 0.914–0.932), p < 0.0001 (Kruskal–Wallis test). Mirroring this, pairwise Jaccard similarities for predicted substrate usage profiles were significantly higher for pairs within SLs (median 0.990, IQR 0.983–0.997) than between SLs (median 0.97, IQR 0.96–0.98) for both aerobic and anaerobic substrate usage (p < 0.0001 for all comparisons, Kruskal–Wallis test with Holm correction, followed by Dunn’s post-hoc, Fig 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Metabolic heterogeneity within taxa is nonrandomly distributed.

A: Pairwise Average Nucleotide Identity (ANI) vs. pairwise Jaccard similarity of metabolic orthologs within each of the well-sampled taxa (n > 200 genomes). B: Distributions of pairwise (all vs. all) Jaccard Similarity of simulated substrate usage for pairs of genomes within vs. between SLs in the same species. Significance asterisks excluded as each group was significantly different from every other group (p < 2 × 10⁻¹⁶), calculated using a nonparametric Kruskal–Wallis test with Holm correction, followed by Dunn’s post-hoc test. The data underlying this Figure can be found at Figshare (https://doi.org/10.6084/m9.figshare.24503737).

https://doi.org/10.1371/journal.pbio.3003559.g002

Finally, we sought to explore structuring among accessory metabolic traits, using a co-occurrence analysis. Twelve pairs of co-occurring substrate-usages were identified and 11 of these pairs were mechanistically-linked, e.g., the same substrate tested as different element sources (n = 5 pairs) or substrates for which the catabolic pathways are overlapping (n = 4 pairs, see S1 Text). Just one co-occurring phenotype pair (formamide and Fe(III)dicitrate as sources of carbon) was not obviously mechanistically-linked, indicating that they may be co-selected or encoded by genes that are physically linked in the genome or on plasmids. Accordingly, we have identified the genes enabling use of formamide and Fe(III)dicitrate as sources of carbon co-located on plasmids in the completed genomes of multiple unrelated strains, e.g., INF008 (SL221), INF044 (SL258), and INF119 (SL17).

Major SLs are defined by unique metabolic profiles

To explore the association between metabolism and population structure at higher-resolution, we calculated the relative prevalence of predicted growth phenotypes within 48 Kp SLs that were each represented by genomes from broad geographies and time-spans (≥15 genomes, 3–12 geographic regions, 5–86 years). These included SLs corresponding to globally-distributed multi-drug resistant (SL14, SL15, SL17, SL29, SL37, SL101, SL147, SL258, and SL307) and hypervirulent clones (SL23, SL25, SL65, and SL86) [3]. The data showed that each SL is associated with a distinct, core metabolic fingerprint, indicating metabolic specialization (Figs 3, S4B, S6, and S5 Data), with 253–316 core aerobic and 198–223 core anaerobic conditions each. Note that in some cases the SL-specific core was smaller than the Kp species core reported above using the ≥95% genome threshold (363 aerobic, 307 anaerobic conditions), because most SLs represent far less than 5% of the total Kp genome collection. This contradiction highlights the influence of sample size and diversity on core definitions, which should be considered as indicative of the general trends in the population. We tested for the impact of sampling biases on our SL comparisons. The data show no correlation between the number of SL core traits and either of sample size or temporal range (S7A and S7B Fig, Pearson’s product moment correlation, r = 0.20 and 0.04, p = 0.1751 and 0.7774, respectively). There was a moderate positive correlation between the number of SL core traits and number of geographic regions represented (S7C Fig, r = 0.49, p = 0.0004). However, the latter is counter intuitive and seems to be driven by a small number of SLs that have a low number of core traits, rather than a genuine trend (S7C Fig). We are therefore confident that the trends we describe are not the result of sampling biases. Nonetheless, we acknowledge that the precise sets of SL-specific core traits listed here should be considered an approximation, because the addition of further diverse genomes for any given SL could result in the removal of a minority of traits from its core list.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distinct metabolic fingerprints of SLs within species.

Heatmap showing frequency of variable aerobic substrate usage across 48 global K. pneumoniae sub-lineages (SL). Rows are ordered by phylogeny. SL labels are coloured to indicate the globally-distributed clones described in [3]: Blue shows hypervirulent while red shows multi-drug resistant. Substrate shown along X-axis. The element source of each substrate is semi-colon separated and abbreviated for brevity: C = Carbon, N = Nitrogen, and S = Sulfur. O2 indicates aerobic conditions. Frequency of substrate usage indicated by shading as shown in legend. Where available, substrate-specific prediction accuracies are indicated in coloured circles below the X-axis. Number of isolates per each SL shown in bars. Anaerobic growth predictions shown in S6 Fig. The data underlying this Figure can be found in S5 Data.

https://doi.org/10.1371/journal.pbio.3003559.g003

No pair of SLs shared the same core growth profile, even pairs that are thought to share relatively recent common ancestry, such as SL14 and SL15 (common ancestor ~45 years ago [4]) and SL65 and SL25 (common ancestor yet to be dated) [96] (Figs 3 and S6). However, there was also clear evidence of diversity within SLs, with 3–102 variable growth phenotypes each, a pattern that was replicated by metabolic ortholog distributions (S3, S6 Figs, and S1 Text). After excluding growth conditions predicted to support growth among ≥95% (46/48) K. pneumoniae SLs (n = 615 conditions, 340 aerobic and 275 anerobic, corresponding to 202 total substrates) and those that did not support any growth among the 48 SLs (n = 525 conditions), the data support a broad classification of phenotypes as follows: i) 89 that were core in some SLs (range = 1–45 SLs) but variably present in others (range = 1–17) which we call ‘SL-specific core traits’; ii) 13 that were not core in any SL (accessory in 3–46 SLs, including the only pair of co-occurring phenotypes that was not mechanistically-linked) (Figs 3 and S6). The SL-specific core traits could be further divided into those that were core to the majority (≥66%) of SLs (n = 57 phenotypes that we call ‘majority SL-specific core traits’), those that were core to an intermediate number of SLs (≥25%–<66%, n = 13 phenotypes, corresponding to six substrates, that we call ‘common SL-specific core traits’) and those that were core in only very few SLs (<25%, n = 13 ′rare SL-specific core traits’, eight aerobic and five anerobic, corresponding to seven total substrates).

We have previously tested the predictive accuracy of 16 of 68 variable aerobic growth phenotypes, 14 of which were estimated ≥70%, including 12 estimated ≥90% (see Fig 3), while the accuracy for 11 of 16 tested anaerobic growth phenotypes was estimated ≥70%, including three estimated ≥90% (see S6 Fig). We propose that the majority SL-specific core traits and the accessory traits are unlikely to play a major role in defining SL-specific evolutionary trajectories or ecological interactions, and hence we focused our follow-up investigations here on the common and rare SL-specific core traits. Our prior work had confirmed predictive accuracies ≥97% for five of the six common SL-specific core substrates; L-sorbose, L-tartrate, galactitol, D-tagatose, and L-galactonate as sole sources of carbon in aerobic conditions [66] (tested using the Omnilog phenotyping microarray, for 37 diverse KpSC isolates). The sixth, fructoselysine, is not present in the Omnilog array and could not be tested here due to its high cost. One of the seven rare SL-specific core substrates, allantoin, was previously tested as a sole source of carbon in aerobic conditions, for which the predictive accuracy was 89.2% (37 isolates tested [66]). Here, we tested growth using allantoin as a sole source of nitrogen and found the model accuracy was much lower (15% for 13 isolates tested). We also tested the six other rare SL-specific core growth substrates (13 isolates tested for each), confirming high accuracy (85%) for L-hydroxyproline as a carbon source, 77% accuracy for each of butyrate and methanol as sources of carbon, 46% for xylitol as a source of carbon, 31% for methanesulfonate as source of sulphur, and 15% for acetoacetate as a source of carbon (S2 Data, see S1 Text for more details, Fig 3).

SLs co-operate through cross-feeding metabolites

We have shown substrate usage varies across the KpSC and some of this variation is SL-dependent. We hypothesized that such diversity promotes the maintenance of distinct SLs by limiting inter-lineage competition and/or promoting commensalism. We tested this hypothesis using in vitro co-culture experiments.

We selected four carbon substrates (galactitol, L-tartrate, L-sorbose, and L-hydroxyproline) for use in three independent substrate combinations and corresponding SL pairs for which predicted substrate-usages were mutually exclusive SL-specific core traits. One isolate from each SL was selected for co-culture aerobically in each of two conditions (minimal media plus each of the SL-specific core substrates), such that each experiment contained one isolate that was able to utilize the substrate as a sole carbon source (the prototroph) and one that was not (the auxotroph). In short, Strain A as an auxotroph and Strain B as a prototroph, after which we then swapped their roles via use of another substrate, so that A would be a prototroph and B would be an auxotroph. This was done within each pair to evaluate reciprocal cross-feeding between SLs under different conditions. Growth of the auxotroph was measured using OD₆₀₀ and compared to the same isolate growing in monoculture in the same media. Co-cultured isolates were physically separated by a Transwell membrane but had shared access to the growth medium (Fig 4A). A unique aspect of this experiment was the use of independent isolates from the same SLs as biological replicates rather than replicates of the same isolate. This allowed us to limit the impact of strain-specific growth biases while demonstrating conserved behavior across a greater number of independent isolates. SLs representing clinically-relevant clones were selected (multi-drug resistant SL37 versus hypervirulent SL29, hypervirulent SL86 versus hypervirulent SL23, and multi-drug resistant SL17 versus SL29).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Nutrient prototrophic lineages support growth of auxotrophic lineages.

Co-culture growth assays between prototrophs and auxotrophs under various substrate usage conditions. Time points captured were at 0, 24, 36, 48, and 72 hours depending upon the substrate pair (Materials and methods). A: Graphic demonstrating the experimental design measuring commensalism between an auxotrophic strain unable to utilize a carbon source and a prototrophic strain which can. The growth of the auxotroph was compared in single vs. co-culture conditions. B: Growth results (OD₆₀₀) over time for each auxotrophic strain in the bottom well (n = 3 strains per SL as biological triplicates) in single and co-culture with a prototroph for the substrate. C: Log₂-fold-change distribution of growth results (OD₆₀₀) at each time point under different culture conditions. Statistical significance comparing the OD₆₀₀ values of co-culture to single culture is shown above plot (Dunn’s post-hoc with Holm correction, *p < 0.05, **p < 0.01, ***p < 0.001). The data underlying this Figure can be found in S3 Data.

https://doi.org/10.1371/journal.pbio.3003559.g004

Our data showed that co-culturing auxotrophs with prototrophs in the presence of a SL-specific core substrate significantly increased the OD₆₀₀ of the auxotrophs (Fig 4B) with mean log₂-fold-change 3.78 ± 2.64 SD at 24 hours, p = 0.0128; 4.02 ± 2.34 SD at 36 hours, p = 0.0001; 4.94 ± 1.14 SD at 48 hours, p = 0.0039; and 4.35 ± 1.27 at 72 hours p = 0.0039; Kruskal–Wallis + Dunn’s post-hoc with Holm correction (Fig 4C). To confirm this phenomenon only occurred in the presence of a cross-feeding partner and eliminate the possibility of contamination or horizontal transfer of the genes conferring the prototroph phenotype, aliquots from each co-culture well were plated onto LB agar to confirm a single, distinct colony morphology. Colonies were then inoculated into M9 minimal media plus the auxotrophic substrate as a sole carbon source, confirming the original growth behavior of each isolate (S3 Data).

To further explore these pairwise strain interactions, we performed in silico co-culture simulations, which indicated that prototrophs were likely cross-feeding the auxotrophs by catabolising the sole carbon substrates, then exporting metabolites that were capable of supporting auxotroph growth (see S1 Text). Subsequent metabolomics analysis of representative prototroph culture supernatants supported the presence of many of the implicated metabolites (S6 Data, S8 Data, and S8 Fig): Between two and 13 metabolites predicted to be involved in cross-feeding were detected with significantly higher peak intensity in the prototroph culture supernatant compared to the no-isolate controls (p < 0.05 by two-tailed Wilcox Rank Sum Test with Benjamini-Hochberg multiple testing correction). Among these, between two and 11 metabolites were each detected with >1 log₂ fold intensity increase, indicating a substantial increase in the quantity of the metabolite present in the prototroph supernatant as compared to the matched control (S6 Data). Metabolites with substantial intensity changes included those involved in central carbon metabolism and/or that have been previously shown to support growth for all or almost all KpSC, e.g., alanine, aspartate, glycerate, glutamate, lactate, malate, and pyruvate [70] (labeled in S8 Fig).

We cannot rule out the possibility that the putative cross-feeding metabolites were released following cell lysis; however, we note that the associated peak intensity shifts were among the most prominent of all metabolites tested, including ubiquitous intracellular metabolites, e.g., other amino acids (see S8 Fig). Therefore, it is unlikely that cell lysis alone explains the metabolite quantities we detected. Furthermore, each of the substrates named above is known to be actively exported by the closely related organism E. coli [97–102], and a tBLASTn search showed that all six of the representative prototrophs carried orthologs of genes encoding the relevant substrate exporter proteins characterized from E. coli (alanine, glutamate, malate, lactate, and pyruvate) or Tetragenococcus halophila (aspartate). (We were unable to identify any known exporters of glycerate. Also see S1 Text.)