Extensive metabolite footprint of Eggerthella lenta in chemically defined media

To identify key nutrients and metabolic pathways required for growth of E. lenta, we first developed a custom chemically defined media formulation, referred to as Eggerthella Defined Media 1 (EDM1). We designed the initial EDM1 formulation by making several modifications to a recipe previously reported to support growth of many human gut bacterial isolates but not E. lenta [1]. We increased the quantity of L-arginine, removed sugars, and ensured the availability of all amino acids and vitamins/cofactors with fragmented or missing biosynthetic pathways in the E. lenta DSM 2243 genome (Virtual Metabolic Human database annotations [23]; Materials and methods, S1 Table]. The resulting media is composed of compounds typically present in the mammalian gut from microbial, host, and/or dietary sources. It supported robust E. lenta growth at a level comparable with standard culture conditions (Brain Heart Infusion media supplemented with 1% arginine; S1A and S1B Fig).

Using this platform, we sought to identify primary metabolites used and produced by E. lenta and the underlying core metabolic pathways active in the EDM1 condition. We used untargeted metabolomics to analyze culture supernatants of the type strain E. lenta DSM 2243 across 6 time points over its 50-hour growth curve in EDM1 batch culture (Fig 1A). We assigned putative identifications to metabolomics features by comparison with local library standards and public databases, with associated confidence levels for each identification based on a modified version of the Metabolomics Standards Initiative (MSI) framework (Materials and methods). After dereplication of features from positive and negative ionization modes, 4,095 features were detected, of which 636 (15.6%) were not detected in sterile control media (sample mean intensity >3× blank sample mean; Fig 1B). This large number of features likely represents a complex mixture of true known and unknown metabolites along with fragmentation products and other chemical artifacts, as is typically observed in untargeted metabolomics experiments [24]. A total of 612 features (14.9% of features overall) were significantly different in abundance between sterile controls and supernatants at the final time point (false discovery rate (FDR)-adjusted p < 0.1; Fig 1C), of which the majority (444, 72.5%) were increased in E. lenta cultures. Notably, the number of differentially abundant features at the final time point, both in total and among those assigned an identification, is substantially higher than previously reported metabolomic profiles of this species in ISP-2 and Mega media [25,26] (S1C Fig). This increased sensitivity was expected given our use of both chemically defined culture media and untargeted metabolomics.

Download:

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

Fig 1. Production and depletion of diverse metabolites by Eggerthella lenta DSM 2243 in chemically defined media.

(A) Number of metabolite features detected by tandem LC-MS in culture samples at each time point. Features are considered present if their average peak height in supernatant is greater than 3× the average peak height in blank samples. Using both positive and negative ionization modes, an increasing number of features not found in controls appear in culture supernatants over time. (B) Number of differentially abundant metabolite features compared with sterile control media at each time point, based on FDR-adjusted t tests of log-transformed peak heights. (C) Volcano plot of differentially abundant metabolite features at the final time point (50 hours) compared with sterile controls. P values shown on the y axis are based on Welch’s t tests comparing values at the final time point vs. sterile controls (Benjamini–Hochberg adjusted). (D) Heatmap of individual metabolite trajectories in cultures of E. lenta DSM 2243 grown in EDM1 batch culture. Features shown are those whose abundance was significantly different from controls (FDR-adjusted p < 0.1 and absolute log₂ fold change >0.75) at the final time point. Identified metabolites are labeled; the number in parentheses indicates the MSI confidence level for that identification (with 1 as highest confidence, see Materials and methods). Values shown are average log-transformed peak heights, scaled for each feature. The gray heatmap at the top indicates the average batch culture density at each time point of E. lenta DSM 2243 in EDM1 (normalized OD600). See also S1 and S2 Figs and S1 and S2 Tables. EDM1, Eggerthella Defined Media 1; FDR, false discovery rate; LC-MS, liquid chromatography–mass spectroscopy; MSI, Metabolomics Standards Initiative.

https://doi.org/10.1371/journal.pbio.3002125.g001

Metabolites of diverse chemical classes are modified by E. lenta (Fig 1C and 1D). Compounds produced by E. lenta tended to be amino acid and nucleic acid metabolites. As expected, these included ornithine and citrulline, suggesting activity from the arginine deiminase pathway, which is highly expressed by E. lenta in the presence of arginine [16]. However, other arginine-related metabolites were also produced at lower levels, including N,N-dimethylarginine, N₅-(1-iminoethyl)-ornithine, and homocitrulline, suggesting that arginine may also be metabolized via other pathways. Several other metabolites produced at lower levels appeared to be products of metabolism of other amino acids in the media, including 4-methyl-2-hydroxy-pentanoic acid (from leucine), indole-3-acetate and indole-3-lactic acid (from tryptophan), and 3-phenyllactic acid (from phenylalanine), consistent with one previous report of production of indole-containing compounds and phenyl acids by E. lenta [27]. Other metabolites produced in supernatants included the amino acids alanine, glutamate, glutamine, histidine, and lysine, as well as several intermediates in biosynthesis of both purines and pyrimidines (inosine, orotic acid, hypoxanthine, uridine, and thymidine). Overall, the set of metabolites produced by E. lenta supports its previously reported dependence on arginine catabolism but is highly multifaceted.

Of the 54 compounds in our EDM1 recipe, 22 were detected by untargeted metabolomics but just 3 were depleted significantly in E. lenta cultures (Figs 1D and S1D): arginine, riboflavin, and EDTA (which is likely reduced due to complexing with metal ions rather than from direct uptake or metabolism). This result suggested that most compounds were included in excess, leading us to reduce the concentration of several nondepleted amino acids for subsequent experiments (S1 Table). Interestingly, 5 of the identified metabolite features significantly depleted by E. lenta were not explicitly included in our defined media formulation, including guanine and 4 arginine dipeptides (Fig 1D). Since these compounds were found at low intensities and were annotated with high confidence, we inferred that they may be trace contaminants from commercial preparations of one or more EDM1 media components, which was partially supported by a preliminary follow-up analysis of fresh media (Materials and methods, S1E Fig). Regardless, the rapid depletion of these compounds indicates that their presence may influence growth and metabolic activity and reinforces the value of untargeted metabolomic profiling.

We examined the dynamics of metabolite production and depletion over the 50-hour growth of E. lenta in batch culture. Hierarchical clustering of metabolite trajectories indicated that among both produced and depleted features, some metabolites are produced/depleted rapidly early in growth, while others shift more dramatically later as the culture approaches stationary phase (Figs 1D and S1F). This observation suggests that 2 or more distinct growth phases may be occurring as resources are consumed from the media. Among identified metabolites, the trace guanine and arginine dipeptides are first depleted from the culture in early time points, while citrulline, inosine, and indole-3-lactic acid are produced at relatively higher rates (Fig 1D). In the later phase, arginine is depleted more rapidly, while alanine, 4,6-dihydroxypyrimidine, and various N-acetylated amino acid metabolites are produced.

To gain a better understanding of the contributions of individual nutrients to E. lenta growth, we systematically tested the effect of their removal from the media on growth of E. lenta DSM 2243 (Materials and methods, S2 Table). We collected growth curve data from EDM1 with and without each component and fit logistic growth models to the results, finding that 21 out of 41 compounds tested had a significant effect on the area under the growth curve (Wilcoxon rank-sum test, FDR-adjusted p < 0.2; S2A Fig). The only compounds whose individual removal fully prevented growth of E. lenta were arginine, tryptophan, riboflavin, biotin, and magnesium (although it is plausible that other compounds are required in trace amounts and were not fully removed by our preparation methods, particularly minerals such as iron). In general, removing amino acids most commonly tended to reduce carrying capacity, consistent with a role as carbon and/or energy sources, while removing vitamins had more varied effects on the growth curve (S2B Fig).

Acetate and arginine are key carbon and energy sources for E. lenta

Surprisingly, we found that sodium acetate contributed substantially to E. lenta growth in EDM1 (S2A Fig), even though it was included at a relatively low concentration (1 mM, compared to 57 mM arginine in EDM1). Since acetate is an abundant and variable metabolic byproduct of diverse human gut microbes [28], dependence on acetate could shape the ecological interactions of E. lenta in the human gut microbiota. Although our untargeted liquid chromatography–mass spectroscopy (LC-MS) workflow was not able to quantify acetate, we had observed accumulation of several N-acetylated compounds in supernatant (Fig 1D), suggesting that the amount of acetate incorporated into core metabolic pathways may be relatively small. However, acetate provided a dose-dependent increase in carrying capacity for E. lenta up to a concentration of at least 10 mM in EDM1 (Fig 2A). We therefore used a targeted derivatization and LC-MS/MS method to quantify acetate levels in supernatants from 3 strains of E. lenta (DSM 2243, AB8n2, and Valencia) grown in EDM1 with different acetate concentrations (0, 1, or 10 mM). Acetate was depleted to approximately the limit of quantification in cultures from the 1 mM acetate group, but not the 10 mM acetate group, confirming that a relatively small quantity is required for the observed level of E. lenta growth (S3A Fig). We tested the effect of replacing acetate with equimolar amounts of 10 other small carbon compounds, finding that no tested alternative compound provided a comparable benefit (Fig 2B). Based on these results, we chose to further investigate E. lenta’s acetate utilization pathways.

Download:

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

Fig 2. E. lenta uses acetate for nucleotide and peptidoglycan biosynthesis.

(A) Growth of E. lenta DSM 2243 in EDM1 media with varying concentrations of sodium acetate. (B) Growth of E. lenta DSM 2243 in EDM1 media in which 1 mM sodium acetate is replaced with other small carbon compounds. (C) Trajectories of identified metabolite features responsive to acetate concentration in E. lenta EDM1 batch cultures. Values are scaled average log-transformed peak heights from untargeted metabolomics profiling of supernatants. Labels show metabolite identity and MSI confidence level in parentheses. Metabolites shown are those that were assigned an identity and that had significantly different trajectories in the 0 mM vs. 1 mM acetate group based on spline regression comparison with the R package santaR (FDR-adjusted p < 0.25). (D) SIRM profiling of E. lenta DSM 2243 in EDM1 media with ¹³C₂ labeled acetate. The number of compounds with labeled isotopologues detected at a peak area >10⁵ is shown for each sample group and time point, indicating incorporation of acetate into varied metabolites by E. lenta. (E) MIDs of intracellular metabolites. Each barplot shows the average isotopologue distribution in 1 mM and 10 mM acetate cultures. Compounds shown are those with an average labeled MID > 0.15 and a total peak area from labeled isotopologues of at least 10⁴ in at least 1 E. lenta DSM 2243 labeled acetate condition. (F) Hypothesized pathways for incorporation of acetate into E. lenta central carbon metabolism and into biosynthetic pathways to produce labeled metabolites. Circles indicate the number of carbon atoms in selected compounds and are colored green to indicate incorporation of ¹³C isotopes from external acetate. Compound names in bold were detected with the observed labeling patterns in either intracellular metabolite extracts or culture supernatants. Corresponding enzymes are annotated in the E. lenta DSM 2243 genome for all reactions shown and labeled in gray with the NCBI locus tag number. For pathways shown at a summary level (gluconeogenesis, pentose phosphate pathway, purine and pyrimidine biosynthesis, peptidoglycan biosynthesis), only the first enzyme in the pathway is labeled on the plot. See also S3–S5 Figs and S1 Data. EDM1, Eggerthella Defined Media 1; FDR, false discovery rate; MID, mass isotopologue distribution; MSI, Metabolomics Standards Initiative; SIRM, stable isotope-resolved metabolomics.

https://doi.org/10.1371/journal.pbio.3002125.g002

First, we used our untargeted LC-MS metabolomics workflow to compare metabolites in supernatant over time from the same 3 E. lenta strains grown in EDM1 with different acetate concentrations (E. lenta DSM 2243 shown in Fig 2C; AB8n2 and Valencia in S3B and S3C Fig). Using smoothing spline models, we found that many produced or depleted compounds had significantly different abundance trajectories across the growth phase (FDR-adjusted p < 0.25) depending on the presence of acetate. These included pyrimidine metabolites, N-acetylated amino acids, amino acid metabolites including indole-3-lactic acid and 2-hydroxyglutaric acid, and 423 unidentified metabolite features (Fig 2C). Of the 612 features produced by E. lenta, 53.4% had significantly different trajectories in the no acetate condition. Most differentially abundant compounds were associated with cell density and produced by E. lenta at higher levels when grown with higher acetate concentrations, reinforcing the general loss of biomass production in the absence of acetate.

To identify the specific pathways by which acetate is metabolized by E. lenta, we next profiled metabolites in the supernatant across time during growth of the same 3 strain isolates of E. lenta with ¹³C₂ acetate provided as a stable isotope-labeled substrate (DSM 2243 in Fig 2D–2F; 2 additional strains in S4 Fig). We detected the incorporation of ¹³C labeled atoms in 52 features in E. lenta supernatants from any strain at the final time point, of which 24 were previously identified as responsive to acetate concentrations (Materials and methods, Figs 2D and S4). Acetate was incorporated into diverse products across metabolite classes but was found at the highest enrichment levels in nucleotide and carbohydrate metabolites (S1 Data).

Because many core metabolites are not produced in excess or secreted during growth, we also analyzed intracellular metabolites from extracts collected at a single time point in the late-exponential growth phase. Labeled intracellular compounds included glutamate, glutamine, sugars, nucleotide metabolites, and UDP-N-acetyl-glucosamine, a primary component of peptidoglycan (Fig 2E), as well as 7 labeled compounds of unknown identity. The signal from carbohydrate-related compounds including glucose-6-phosphate and UDP-N-acetyl-glucosamine was almost exclusively from labeled isotopologues (97.5% in 1 mM acetate and 100% in 10 mM acetate), indicating that synthesis of these compounds using acetate may be more efficient than any alternative non-acetate–dependent pathways available to E. lenta in the EDM1 condition.

Acetate-derived extracellular and intracellular metabolites were consistent across the 2 additional strains of E. lenta. While the overall rate of acetate incorporation differed between the 3 strains, the set of extracellular and intracellular labeled compounds was fully consistent. Isotopic enrichment for 2 additional extracellular metabolites (malonic acid and 3-hydroxy-myristic acid) was identified in both of these strains as well as 4 additional intracellular metabolites in one or both strains (all of unknown identity), confirming that acetate is incorporated by E. lenta into varied biosynthetic pathways (S4 Fig).

Based on these results and metabolic gene annotations of the E. lenta DSM 2243 genome, we hypothesized that E. lenta converts acetate to acetyl-CoA via acetate kinase (ELEN_RS08645) and phosphate acetyltransferase (ELEN_RS08640). Acetyl-CoA could then be used as a carbon source via 2 routes: conversion to glutamate by a partial citric acid cycle, and synthesis of pyruvate by the enzyme pyruvate-ferredoxin oxidoreductase (PFOR; ELEN_RS10770) (Fig 2F). This hypothesis is consistent with the organization of the E. lenta DSM 2243 genome, as 2 of the 3 enzymes required for conversion of acetyl-CoA to glutamate are colocated (aconitate hydratase and isocitrate dehydrogenase, ELEN_RS11710, ELEN_RS11715). Genes for another partial component of the citric acid cycle—fumarate hydratase and malate dehydrogenase—are colocated in another region of the genome (ELEN_RS056; [70–90]), suggesting that they may act in a separate functional role. Taken together, these data suggest that E. lenta uses acetate as a key carbon source for synthesis of biomass components, in tandem with ATP generation from arginine catabolism, anaerobic respiration, and/or other unknown pathways.

However, we inferred that acetate is likely not the sole carbon source used by E. lenta in EDM1, given the relatively low concentration required for growth promotion and the abundance of unlabeled isotopologues detected for many produced compounds (S1 Data). We wondered whether arginine or ornithine may also be substrates for synthesis of biomass components, or if arginine is exclusively catabolized to ornithine for ATP production, as suggested by one previous study in rich media [29]. We first confirmed that citrulline, but not ornithine, can replace arginine with nearly equivalent growth in EDM1, replicating a previous result in rich media ([16]; S5A Fig). We then analyzed intracellular and extracellular metabolites from E. lenta DSM 2243 growing in EDM1, this time with ¹³C₆ L-arginine as a stable isotope-labeled substrate. We found by far the largest composition of ¹³C enriched isotopologues in ornithine, citrulline, and other closely related compounds (S5B–S5E Fig), indicating that arginine is predominately processed by the arginine deiminase pathway. However, we observed M+1 enrichment (i.e., incorporation of a single ¹³C carbon atom from arginine) in produced glutamine, orotic acid, and pyrimidines, among others (S5C and S5D Fig), suggesting biosynthesis from the carbamoyl phosphate intermediate. Labeled M+5 isotopologues of proline and prolinamide also appeared at low levels at later time points, likely indicating a slower flux producing these compounds from accumulated ornithine (S5D and S5E Fig). Yet in total, only 29/324 features were detected with ¹³C enrichment for 5 or more carbon atoms in intracellular extracts, and most appeared closely related to arginine, citrulline, and ornithine (S1 Data). These results confirm that arginine is primarily an energy source and not a major biosynthetic precursor for E. lenta (S5F Fig).

A genome-scale metabolic model of the E. lenta type strain recapitulates growth, metabolite, and gene expression phenotypes

COnstraint-based Reconstruction and Analysis (COBRA) is a set of computational tools that has been applied to interpret -omics data and optimize metabolic activities for various microbes of importance in basic science, metabolic engineering, and medicine [30–32]. It has been proposed as a promising strategy to predict phenotypes and design modifications to complex host-associated microbial communities by synthesizing information about the physiology of individual members and the available nutrients into a rational framework [8,33,34]. However, the value of such a framework is dependent on its ability to accurately describe the contributions of metabolically active community members. The reconstructions currently available for many anaerobic microbes have only been curated to a limited degree and remain minimally validated. Therefore, we used our in vitro platform to curate and analyze a genome-scale metabolic network model of E. lenta DSM 2243 growth in EDM1 and assessed the degree to which this model can explain E. lenta metabolic phenotypes across conditions.

We obtained a genome-scale metabolic reconstruction from the AGORA database version 2.0.0 [35], which we term iEL2243_2. Initial testing indicated that the model was incapable of biomass production in EDM1 media, so we performed additional curation of model reactions and transporters (S3 Table). We curated the reconstruction based on genome annotations from multiple sources [36–38] and added transporters for strongly depleted and produced compounds that were identified with high confidence in our metabolomics data. Throughout this process, we compared model results with experimentally observed growth in chemically defined media conditions, using these results to inform the curation process and add missing reactions where supported by experimental data. We simulated metabolic fluxes in different conditions by converting defined media concentrations into estimated maximum nutrient uptake rates for each compound, assuming a fixed and evenly mixed concentration of E. lenta cells (Materials and methods). Nutrient constraints for other metabolites detected in EDM1 media by untargeted metabolomics were applied qualitatively following previous work (Materials and methods; [39]). While these models are typically validated by comparison with gene essentiality data [40], the tools to generate such data are not yet available for E. lenta. We instead evaluated whether the model was consistent with observed metabolite utilization and production and with gene expression during exponential growth in EDM1, and whether predicted essential genes were conserved across strain genomes. We did not use metabolite production or transcriptomic data to refine the model so that we could instead use these datasets to independently evaluate the effectiveness of the growth-based curation.

This process resulted in a model with 1,244 reactions linked to 727 gene annotations and 1,218 metabolites (Fig 3A). The largest number of reactions were in the subsystems of fatty acid synthesis, extracellular transport, and glycerophospholipid biosynthesis (Fig 3B). Flux balance analysis of the final model estimated the maximum growth rate of E. lenta DSM 2243 in EDM1 to be 0.96 h⁻¹, higher than experimental values (median 0.32 h⁻¹; S2B Fig). The existence of a difference between these values is not surprising given that organisms do not necessarily grow at their theoretical maximum growth rate, and growth constraints may exist that are not encoded in the metabolic network model [40]. However, the relatively large discrepancy indicates that additional modifications to the biomass equation may further improve the model. In particular, we found that the maximum growth rate was sensitive to the biomass requirements for flavin compounds and for ATP (S6 Fig).

Download:

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

Fig 3. A curated genome-scale metabolic model of E. lenta DSM 2243 partly explains growth phenotypes across conditions.

(A) Summary of the curated reconstruction of E. lenta DSM 2243 indicating the number of genes, reactions, and metabolites in the original and curated models, and the share of those required to be active for growth in EDM1 based on pFBA. (B) Summary of the total number of reactions by subsystem, and the share of each subsystem predicted to be active in EDM1 (only the top 20 subsystems are shown). (C) Acetate and L-arginine uptake dependencies inferred by the model. In the final curated model (red lines), the maximum growth rate decreases with decreasing availability of both L-arginine and acetate, qualitatively consistent with experimental data. A previous model incorporating a carbon monoxide dehydrogenase reaction based on [20] (blue lines) failed to recapitulate the expected dependencies. (D) Confusion matrix summarizing a comparison of growth/no growth between the iEL2243_2 model vs. experimental observations for leave-one-out media conditions. The labeled p-value is based on a Fisher exact test. (E) Full set of quantitative comparisons underlying panel (D). Each column shows the FBA-inferred maximum growth rate in the EDM1 condition with a media component removed, paired with the experimentally observed area under the empirical growth curve for that condition. A gray tile indicates zero growth. (F) Comparison of shifts in metabolomics data with uptake and secretion rate ranges inferred for the same compounds by FVA. Metabolites that can only be imported according to FVA were decreased in metabolomics data, while those with potential for being produced were indeed produced. (G) Comparison of absolute fluxes inferred by pFBA with gene expression of linked enzymes of E. lenta DSM 2243 during exponential growth in EDM1. Within flux quantiles (on the x-axis), genes are expressed at a wide range of levels, but genes linked to reactions with the highest fluxes are generally highly expressed. See also S6 and S7 Figs and S3 and S4 Tables. EDM1, Eggerthella Defined Media 1; FBA, flux balance analysis; FVA, flux variability analysis; pFBA, parsimonious flux balance analysis.

https://doi.org/10.1371/journal.pbio.3002125.g003

The initial model with nonzero growth in EDM1 did not recapitulate the experimentally observed dependencies on either arginine or acetate (Fig 3C). We noticed that this lack of dependency was linked to the inclusion of Wood–Ljungdahl acetogenesis reactions in the model, previously suggested to be present in E. lenta [20,22]. The presence of these reactions allowed the model to draw on an effectively unlimited source of acetyl-CoA from CO₂ and H₂. Regardless of whether the previous annotation of this pathway (which has not been biochemically validated) is correct, reductive acetogenesis may not be thermodynamically favorable during in vitro growth in our anaerobic chamber, where the H₂ concentration is ≤5% [41]. Blocking model flux through the carbon monoxide dehydrogenase reaction of this pathway increased growth dependency on uptake of both arginine and acetate, reflecting our experimental observations (Fig 3C). The model also found no growth benefit from pyruvate, citrate, and other fatty acids based on a lack of annotated transporters for these compounds, consistent with experimental results.

In another key curation step, required to enable biomass production by the model in EDM1, we noticed that E. lenta lacks an annotated gene for the enzyme enoyl-acyl protein carrier reductase, which performs the elongation in the typical type 2 fatty acid synthesis pathway used in bacteria. Because fatty acid biosynthesis is essential and previous studies have noted a high level of diversity in this essential step among bacterial genomes [42], we preserved this step in the model without any current gene annotation. This gap may indicate a novel enzyme family performing this conversion (S3 Table).

We applied the iEL2243_2 model to predict growth phenotypes across our leave-one-out chemically defined media conditions, finding that these were generally consistent with some remaining notable exceptions (Fig 3D; overall Matthews correlation of 0.35, Fisher exact test odds ratio = 9.1, p = 0.06). Amino acid dependencies matched well between the model and experimental data, with the exception of cysteine, which likely provides a benefit as a reducing agent that is not accounted for by the model [43]. Vitamin dependencies were also generally consistent, with the notable exception of folate, which had no effect on growth despite the lack of several genes for reactions in the canonical folate biosynthesis pathway and the absence of a known dihydrofolate reductase enzyme [44]. The phenomenon of presumed-essential but absent folate genes in bacterial genomes has been recognized previously [44–46], suggesting the possible existence of undiscovered alternative enzymes. Notably, growth was negatively affected by the removal of the folate precursor p-aminobenzoate (S2A Fig). Most of the remaining discrepancies between the model and the growth data are in conditions in which metal ions were removed, which were expected to be required by the model (Cu²⁺, Ca²⁺) but were not essential based on our experiments (Fig 3E). However, these likely reflect difficulties in fully removing trace minerals in our experiment rather than errors in the model reconstruction.

While we curated the model based on growth data, we did not incorporate our metabolomics data except to add transporters for highly differentially abundant metabolites. Even so, we found that there was a high correspondence between observed metabolite shifts and the possible uptake and secretion fluxes inferred by flux variability analysis (FVA) of the model. FVA identifies the range of fluxes for each reaction that are compatible with near-maximum growth. All 37 identified metabolites present in both the model and our metabolomics data displayed experimental shifts in abundance qualitatively compatible with inferred flux ranges (Fig 3F), providing additional support for model quality.

We further compared the iEL2243_2 inferred flux profile with RNA-Seq data from E. lenta growing in this condition, which was not used for model curation (Materials and methods). Genes linked to active reactions were 3.4 times more likely to be in the top half of metabolic genes by expression level than those linked to inactive reactions (71.9% of active versus 42.9% inactive, p < 2.2 × 10⁻¹⁶). Expression level and absolute flux magnitude were correlated across all genes linked to metabolic reactions (Spearman rho = 0.34, p < 2.2 × 10⁻¹⁶; Fig 3G). While we would not expect a perfect correlation between expression and metabolic flux, correspondence between the two provides support that our model has correctly identified pathways with high activity.

Having established consistency with experimental data, we next examined overall reaction fluxes and key pathways in the final model. We found that fewer than half of reactions were predicted to be active in EDM1 by parsimonious flux balance analysis (pFBA; Fig 3A). In the pFBA solution, acetate is incorporated into a partial reductive citric acid cycle via pyruvate formate oxidoreductase (PFOR), which then feeds lipid and carbohydrate biosynthesis pathways, consistent with our SIRM results and with our RNA-Seq data, where PFOR was one of the most highly expressed genes. The vast majority (99.6%) of arginine uptake flux was directed to ATP generation, and 58.4% of ATP generation was sourced from the arginine deiminase pathway (which contains the first, third, fourth, and fifth most highly expressed protein-coding genes in our RNA-Seq data; S4 Table). The remainder of ATP generation in the pFBA solution was attributed to anaerobic respiration via an ATP synthase reaction, although the specific electron transport chain substrates were not clear. However, consistent with this hypothesis, genes linked to respiration were expressed at moderate levels, including ATP synthase subunits and an Rnf electron transport complex, and E. lenta is known to have a large number of poorly characterized enzymes potentially involved in electron transfer [21,47]. The model also identified the regeneration of NADP+ via transaminase reactions (using mainly pyruvate and/or branched chain amino acids) and glutamate dehydrogenase as a key high-flux pathway.

Finally, we applied the model to predict the effects of knocking out individual reactions on growth of E. lenta. Around 15.3% of all reactions in iEL2243_2 were predicted to be essential in any condition and 19.4% to be essential in EDM1. These reactions tended to be involved in lipid metabolism, cell wall biosynthesis, and transport of essential metabolites (S7A Fig). Genes linked to reactions whose removal reduced growth to <70% of wild-type levels were found in a greater number of E. lenta strain genomes than other genes (Wilcoxon rank-sum test, p = 0.001; S7B Fig) and were more likely to be part of the core genome (found in all strains; Fisher exact test odds ratio = 1.74, p = 0.0002). Overall, while significant manual curation was required for the model to recapitulate realistic growth in EDM1, our updated model is able to predict and interpret many aspects of E. lenta growth and metabolic activity across conditions.

The strain-variable E. lenta metabolome is enriched for nucleotides and cell wall metabolites and can be linked to genome variation

Our initial efforts to characterize E. lenta core metabolism focused mainly on the type strain. However, E. lenta has an open pan-genome and established variability in secondary and xenobiotic metabolism [26]. We therefore evaluated the extent to which the metabolic profile of this species is conserved across a larger number of strain isolates. We used untargeted metabolomics to profile stationary phase supernatants of 30 strains grown in EDM1 (S5 Table, S8A and S8B Fig) and used linear models to identify features with significant strain-associated differences in abundance. Over half of the features produced by the UCSF DSM 2243 type strain (52.8%) were variable across strains of E. lenta (Fig 4A), and 1,097 features produced by at least 2 other strains were not produced by the type strain. Divergence in metabolite profiles between strains was not associated with phylogenetic divergence based on an alignment of core genes (Procrustes analysis, p = 0.31; Fig 4B), consistent with previous findings from untargeted metabolomics profiling of these strains in rich media with a different metabolomics platform [26]. Overall metabolite profiles were moderately associated with presence/absence patterns of variable gene families between strains (p = 0.04; Fig 4B), indicating that the presence or absence of biosynthetic genes and pathways only partly explains variation in the metabolome and that other factors like gene regulation and enzymatic activity may also play a substantial role. Interestingly, both metabolite and gene presence/absence profiles could be partly explained by the geographic region of origin of each strain (PERMANOVA; metabolites R² = 0.22, p = 0.04; genes R² = 0.36, p = 0.02; S8B Fig), suggesting a possible influence of geography on the distribution of accessory metabolism. Other metadata variables, including host health status and isolation body site, were not significantly associated with metabolite profiles.

Download:

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

Fig 4. Extensive within-species variation in E. lenta metabolites can be linked to variable gene families.

(A) Volcano plot of metabolite features detected in stationary phase supernatants of E. lenta DSM 2243 (UCSF lab strain) vs. sterile controls. P values are based on Benjamini–Hochberg corrected Welch’s t tests. Features are colored based on whether their classification as significantly produced or depleted (increased or decreased based on linear regression models of peak area by strain, with FDR-adjusted p-value < 0.1 and log2FC > 0.5) is consistent across 28 other E. lenta isolates and 1 isolate of Eggerthella sinensis profiled in the same experiment. (B) Procrustes analysis of overall metabolite profiles compared with genome features. The upper plot shows a rotated Procrustes superimposition of average metabolite profiles for each isolate (red points) and the phylogenetic distance between them based on an alignment of core genes (blue points). The lower plot shows a superimposition of metabolite profiles and profiles based on the presence/absence of variable gene clusters (purple points). (C) The left-hand plot shows the distribution of strain-variable features in various ClassyFire chemical superclasses, based on Feature-based Molecular Networking with GNPS. The number in parentheses for each class indicates the total number of features with that assignment. The right-hand plot shows the number of strains producing a given feature within each superclass, among features produced by any Eggerthella isolate. Each point represents a single feature, and its position on the x-axis indicates the number of strains for which that feature was significantly increased (FDR-adjusted p-value < 0.1 and log₂ fold change > 0.5) in supernatants compared with controls. (D) Feature abundances of example metabolites across strains. The first 2 panels show 2 strain-variable unidentified features associated with the presence of specific strain-variable gene families—putatively identified as the 2 dominant naturally occurring isotopes of an [M+Cl−] adduct of the teichoic acid component ribitol. The points indicate the log-transformed abundances of these features for each strain. The dotted line in each panel indicates the average level of that feature in sterile controls. Points in dark blue represent strains whose genomes contain genes for a ribitol-5-phosphate cytidylyltransferase (tarJ) and ribulose-5-phosphate reductase (tarI) not found in other genomes. The third and fourth panels show a highly conserved identified metabolite (ornithine) compared with a strain-variable identified metabolite (pantothenic acid). Points in white in the pantothenic acid panel indicate strains whose genome lacks the final step in the biosynthetic pathway for this metabolite. Points are shown as mean and standard error across 3 replicates. The order of strains on the y-axis matches their phylogeny, shown in S8A Fig. (E) Strains lacking a gene annotated as pantothenate synthetase deplete pantothenic acid completely from media (previous panel) and have a substantial growth defect when grown in the absence of pantothenic acid (left panel). Closely related strains that do possess this gene are unaffected by the removal of pantothenic acid (right panel). Carrying capacity is estimated based on a logistic growth model fit by the R package growthcurver. See also S8 and S9 Figs and S5–S7 Tables.

https://doi.org/10.1371/journal.pbio.3002125.g004

While strain-variable metabolites were quite diverse, they were enriched for certain chemical classes. Approximately 92.0% of strain-variable metabolites had no identity information, a similar ratio to the total number of metabolite features (91.8% of features in the whole dataset). Among other features, organic acids (which included many amino acid metabolites) were the least likely to be strain-variably produced. In contrast, organic oxygen compounds (which included several features identified as sugars) and nucleotide metabolites were more likely to be strain-variably produced, and organic heterocyclic compounds and benzenoids were enriched for strain-variable depletion (Fig 4C). The share of strains producing any individual feature varied widely (Fig 4C), although the largest number of features (76.8%) were produced by either only a few (<4) strains or nearly all (>27) strains (S8C Fig).

Given the large share of unidentified metabolites in our dataset, we evaluated whether linking strain-variable metabolites with strain-variable genes could inform metabolite annotations. We performed an association analysis between metabolite feature abundances and the presence of specific accessory gene families, applying a method developed for previous analysis of this E. lenta strain collection [26]. A full 39.0% of metabolite features were significantly associated with the presence of one or more variable gene families (FDR-adjusted p < 10⁻⁴). Using stricter filtering criteria for significance, effect size, and separability, 84 metabolite features (1.3%), of which 80 had no annotation, were linked with the presence of variable genes (S6 Table, Materials and methods). Gene families linked to these features were enriched for KEGG annotations in sulfur metabolism (q = 0.00017), ABC transporters (q = 0.02), porphyrin metabolism (q = 0.03), and biosynthesis of nucleotide sugars (q = 0.049), consistent with the profile of identified variable metabolites.

As a case study, we further examined 2 of the top hits from this analysis, 2 closely related but unidentified metabolite features highly associated with the presence of 2 adjacent gene families (Fig 4D). These gene families were annotated by Prokka [48] as ribulose-5-phosphate reductase 1 (tarI) and a ribitol-5-phosphate cytidylyltransferase (tarJ), which are essential enzymes in the biosynthesis of CDP-ribitol teichoic acid. Teichoic acids are an abundant component of the cell wall of gram-positive bacteria that can take multiple forms and can be synthesized with either CDP-glycerol or CDP-ribitol subunits [49–51]. Interestingly, the m/z values and retention times of the linked features were consistent with an annotation as the 2 dominant [M+Cl]− naturally occurring isotope adducts of a 5-carbon sugar alcohol—i.e., potentially ribitol, xylitol, or a related compound. The annotation of these features as a sugar alcohol was further supported by comparison with a ribitol standard spiked into sterile EDM1 (S8D Fig).

Further examination of the tar/tag biosynthetic gene cluster in which these genes are located revealed extensive strain diversity, with 10 different gene arrangements across the 30 isolates (S8E Fig), suggesting recent positive selection possibly as a form of phage defense [52,53] or host immune interaction [54]. Most genomes have one or more genes with homology to E. coli arnC genes in this region, indicating that the products may be lipoteichoic acids anchored to the cell membrane rather than wall teichoic acids [49]. Among E. lenta genomes without tarI and tarJ, all except the type strain have a tagD gene in the same region instead, which catalyzes the synthesis of CDP-glycerol subunits instead of CDP-ribitol (S8E Fig) and would be consistent with the absence of extracellular ribitol in those strains. Two other metabolite features were associated with the presence of other members of this gene cluster, possibly indicative of other strain-variable cell wall components (S6 Table). This example illustrates that comparative multiomics can be a powerful strategy to identify and begin to decipher the functional consequences of strain variation, even when metabolite identities are not confirmed.

In addition to the unbiased association analysis above, we also assessed whether strain variation in metabolites of known identity could be predicted based on relevant gene annotations. We created genome-scale metabolic reconstructions of a subset of strains included in this experiment (n = 24, using the DEMETER pipeline), curated them using a limited version of the process applied to the type strain (Materials and methods), and again predicted growth and reaction fluxes in EDM1 and in leave-one-out media conditions using flux balance analysis. Across the metabolic networks of E. lenta strains, most reactions were conserved, including arginine metabolism and central carbon metabolism (S7 Table). Tryptophan and riboflavin auxotrophies were also predicted to be conserved across strains. Variable reactions tended to be in the subsystems of transport, fatty acid biosynthesis, cell wall biosynthesis, and nucleotide interconversion (S9A Fig). Consistent with the central role of arginine metabolism, ornithine and citrulline levels in our metabolomics dataset were very consistent across strains. Ornithine was among the least variable metabolite features (Fig 4D), and one of the most correlated with biomass (as estimated by optical density, Spearman rho = 0.36, FDR-adjusted p = 0.1).

While the predicted effects of most compounds on growth were similar or identical across strains (S9B Fig), we noticed a clear difference in pantothenic acid dependence, as a subset of strains were predicted to be unable to grow in its absence. These strains lack the final enzyme in the biosynthesis pathway for pantothenic acid, which is itself a precursor of coenzyme A. Pantothenic acid was depleted to varying degrees in our metabolomics data, reaching the lowest levels in strains that lack pantothenic acid synthase (Fig 4D). Notably, M+2 isotopologues of pantothenic acid were also detected in supernatants from the acetate SIRM experiment, corroborating that at least 3 E. lenta strains synthesize this vitamin de novo (S1 Data and Figs 2F and S4B). We tested growth of pantothenate synthase-lacking strains in comparison with a subset of genetically similar strains in EDM1 with or without pantothenic acid, confirming that strains without this gene family had a greatly reduced carrying capacity in the absence of pantothenic acid (Fig 4E) and highlighting the ability of curated genome-scale models to predict phenotypic differences. Overall, our analysis of strain variation in metabolite profiles is consistent with a model in which E. lenta’s distinctive central carbon and energy metabolism is a core species trait, while more peripheral biosynthetic pathways including synthesis of cofactors and cell surface components can vary freely to adapt to specific microenvironments [55].

Comparison of E. lenta’s metabolic profile in vitro and in vivo identifies shared signatures and usage of a novel nutrient

Having characterized the metabolic profile of the E. lenta species in a simplified in vitro environment, we next asked how these findings compare with its metabolic activity in a host, and whether our in vitro platform could help identify metabolic processes performed by E. lenta within the gastrointestinal tract. We monocolonized germ-free (GF) mice with one of 3 strains of E. lenta by oral gavage, collected serum and intestinal contents after 2 weeks of colonization, and profiled metabolites using the same LC-MS/MS untargeted metabolomics workflow as above. We identified features that were significantly differentially abundant in E. lenta-colonized mice versus their GF counterparts using linear mixed models and compared these features with our in vitro metabolomics datasets, identifying metabolites consistently shifted by the presence of E. lenta across environments. After data processing, quality filtering, and dereplication, we obtained a dataset of 19,714 metabolite features from intestinal samples. Of these, 9.5% were significantly differentially abundant (FDR-adjusted p < 0.1) in response to colonization with at least 1 strain in at least 1 segment of the intestinal tract, indicating a substantial metabolic impact of E. lenta on the intestinal environment (S10A Fig). Interestingly, only 0.95% of features were significantly shifted in the ileum, compared with 6.3% in colon and 4.8% in the cecum. In a previous monocolonization experiment, E. lenta DSM 2243 colonized the ileum at roughly an order of magnitude lower abundance than the cecum and colon of GF mice [26]. Additionally, only 13 features (0.26%) were differentially abundant in serum in response to any of the 3 strains. Overall separability of metabolite profiles between GF and colonized was also highest in the cecum and colon (Fig 5A). These results indicate that E. lenta’s strongest metabolic effects are restricted to the lower intestinal tract.

Download:

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

Fig 5. Comparison between E. lenta’s metabolic footprint in vivo and in vitro reveals shared metabolite signatures.

(A) Principal component analysis of untargeted metabolomics profiles of intestinal contents and of serum. (B) Comparison of the effect of E. lenta on metabolite features detected in both EDM1 cultures and monocolonized mice. Each point represents a metabolite feature detected in both datasets. The x axis indicates the log₂ fold change of each feature in supernatants compared with sterile controls, compared with the estimated log₂ fold change of that feature in monocolonized mice compared with GF mice, based on linear mixed effects models after adjusting for covariates. Points are colored green if the feature is significantly differentially abundant in gnotobiotic mice and is shifted in the same direction by the corresponding strain in the stationary phase in vitro experiment. See also S10 Fig and S2 Data. EDM1, Eggerthella Defined Media 1; GF, germ-free.

https://doi.org/10.1371/journal.pbio.3002125.g005

We assessed the extent to which metabolite features produced by E. lenta in cell culture are detectably shifted by the presence of E. lenta in mice. To do so, we integrated our processed metabolomics datasets by linking metabolite features across datasets with highly similar m/z, RT, and MS2 spectra (see Materials and methods). Based on this analysis, 37.2% of identified metabolite features in intestinal contents and 12.2% of features overall were also detected in vitro (S10B Fig). We compared the estimated log₂ fold change of each linked feature in vitro with the corresponding shifts in vivo (full set in S2 Data; Fig 5B shows the comparison with the strain collection dataset in Fig 4; S10C Fig shows a comparison with the dataset in Fig 1). A total of 202 features significantly increased by the presence of E. lenta DSM 2243 in cecal contents were also increased in one of our EDM1 in vitro datasets, providing support that they are directly produced by E. lenta in vivo. These features represented 78.9% of the set that could be linked across datasets and 20.9% of the full set of E. lenta DSM 2243-increased features in cecal contents. Only 18 metabolites depleted in cecal contents were similarly depleted in vitro, but only 3 of the other 405 depleted features were detected in vitro at all, indicating that E. lenta likely uses a much richer set of nutrients in vivo than those available in EDM1. Overlapping produced and depleted metabolites were found in greater abundance in the cecum and colon than the ileum and serum (Fig 5B), again suggesting a greater metabolic footprint of E. lenta in the lower gastrointestinal tract relative to other sites.

Ornithine was among the most increased features across sampling sites and strains, consistent with our in vitro data (Figs 6A and S11A). Other features that were consistently increased across strains included 5-methyluridine, citrulline, glutamine, and lysine (S2 Data). Interestingly, arginine was only significantly reduced by colonization with one of the 3 E. lenta strains in this experiment (S11B Fig). However, most other proteinogenic amino acids were increased in abundance in intestinal contents in colonized mice compared with GF (S11C Fig), likely due to differences in host activity, so the absence of an increase in arginine may be consistent with arginine usage by E. lenta. Arginine levels were also more variable within strain colonization groups compared with GF (S11B Fig), even though levels of the primary products ornithine and citrulline were relatively consistent. This observation may reflect heterogeneity in host amino acid absorption and metabolism in response to E. lenta colonization.

Given these results, we evaluated what other substrates may be used as carbon or energy sources by E. lenta in vivo. The metabolites most strongly depleted by the presence of E. lenta DSM 2243 in the intestinal tract included several fatty acids conjugated with carnitine as well as multiple other nitrogen-containing metabolites: saccharopine and agmatine (Fig 6A). We chose to investigate agmatine usage further for several reasons: its chemical similarity to arginine, the presence of known agmatine utilization genes in the E. lenta type strain genome, evidence of a consistent decrease across all 3 strain colonization groups (S11A Fig), and its multiple roles as a microbial metabolite and a host metabolite involved in regulation of cell division and neural signaling [56]. The E. lenta DSM 2243 genome contains 2 complete and 2 partial operons encoding genes for the agmatine deiminase pathway. This pathway operates analogously to the arginine deiminase pathway, with ATP production via carbamate kinase as the final step (Fig 6B). Despite this similarity, the agmatine deiminase enzyme family is highly structurally distinct from arginine deiminase [57]. Presence of this pathway is conserved across strains, as other E. lenta genomes contain anywhere between 1 and 4 copies of the key genes for agmatine deiminase and putrescine carbamoyltransferase (KEGG; S8 Table). Additionally, a transcriptional regulator found adjacent to this operon in some strains was previously associated with E. lenta competitive fitness in vivo [26].

Download:

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

Fig 6. Agmatine can replace arginine as an energy source for E. lenta.

(A) Identified metabolite features with the highest estimated effects in E. lenta DSM 2243-colonized mice compared with GF. Each point indicates the effect size of that feature in a particular sample site (denoted by shape). (B) Model of the agmatine deiminase ATP-generating pathway [57]. Three copies of an operon containing genes for all 3 of the labeled enzymes are annotated in the E. lenta DSM 2243 genome. (C) Growth of E. lenta DSM 2243 in EDM1 where arginine has been fully or partially replaced with agmatine sulfate. Curves show mean ± standard error for 4 replicates. (D) Induction of the agmatine deiminase pathway in E. lenta DSM 2243 cultures in response to the addition of agmatine. The volcano plot shows the log₂ fold change and FDR-adjusted p-values of agmatine-treated cultures compared to vehicle (as estimated by negative binomial differential abundance models with DESeq2). See also S11 Fig and S8 and S9 Tables. EDM1, Eggerthella Defined Media 1; FDR, false discovery rate; GF, germ-free.

https://doi.org/10.1371/journal.pbio.3002125.g006

Based on these observations, we predicted that E. lenta may be able to grow in the absence of arginine if it is supplied with agmatine as an alternative energy source. A flux balance analysis simulation of E. lenta in agmatine-based EDM1 predicted a somewhat reduced maximum growth rate (0.54 versus 0.96 h⁻¹) in this condition, with arginine synthesized for protein via its annotated biosynthetic pathway from glutamate. Indeed, we found that replacing arginine with agmatine introduced a growth lag but resulted in a slightly higher final carrying capacity than the equivalent amount of arginine (Fig 6C). We additionally investigated agmatine-responsive genes using RNA-Seq. We grew E. lenta DSM 2243 in a formulation of EDM1 with 70% of the standard levels of arginine and acetate, treated cultures with either concentrated agmatine solution or water, and extracted RNA for sequencing. The genes most strongly induced by treatment with agmatine were 2 copies of putrescine carbamoyltransferase, 1 copy of agmatine deiminase, and a transporter in the same operon (Fig 6D and S9 Table). Genes in the second complete agmatine deiminase operon (ELEN_RS110; [5–15]) were not differentially expressed, suggesting that the annotation of this second operon may be incorrect and/or may be involved in metabolism of a related compound. Interestingly, the most strongly down-regulated genes were 2 transport-related genes adjacent to the energy-conserving hydrogenase (Ech) complex (ELEN_RS078; [45–50]), one of which has structural homology to the arginine-ornithine antiporter found in the arginine deiminase operon (ELEN_RS09745, 27.6% identity). These results indicate that E. lenta can generate ATP from agmatine as a distinct alternative to arginine both in vitro and in vivo and has extensive genetic machinery to efficiently and specifically use each of these compounds.

Experimental model and subject details

Bacterial strains

Strain isolates analyzed in this work are described in [26]. All experiments were performed in an anaerobic chamber with 2% to 5% hydrogen gas, 20% carbon dioxide, and the balance nitrogen, with growth in a 37°C incubator. Standard BHI media (VWR 90003–040) supplemented with 1% L-arginine (referred to below as BHI+) was used for culturing outside of defined media experiments.

Defined media formulations and preparation

Standard composition of the EDM1 media and related formulations are provided in S1 Table. As specified in S1 Table, some experiments were performed using the initial formulation of the media, and others using a simplified form based on the results of leave-one-out growth experiments. For most components, 30 to 1,000× stock solutions were prepared following [76]. Stock solutions were sterilized with a 0.22-μm syringe filter and stored at −20°C. Amino acids and uracil were typically added together directly from powder into a combined 2× stock solution, which was then filter sterilized with a 0.22-μm vacuum filter, except when preparing individual leave-one-out amino acid growth experiments. Most versions used ATCC Trace Mineral and Vitamin Mix Supplements (MD-TMS and MD-VS), except for experiments to test leaving out individual components of these mixes. Media formulations were allowed to equilibrate in an anaerobic chamber (Coy) for at least 24 hours prior to use. To investigate the sources of trace background compounds observed in untargeted metabolomics profiling of defined media, we also prepared fresh solutions of arginine, uracil, and the combined amino acid stock described above, collected them directly with the sample collection, preservation, and extraction methods described below, and analyzed them with the same metabolomics workflow.

Bacterial culture and growth assays

For growth and metabolomics experiments, glycerol stocks were first streaked on BHI+ agar plates and incubated at 37°C for 2 to 3 days. Individual colonies were inoculated into 3 to 4 mL liquid BHI+ and incubated at 37°C for 40 to 48 hours, or until approximately early stationary phase. Culture optical density (600 nm wavelength absorbance, OD600) was measured using a Hach DR1900 spectrophotometer. A volume of 1 mL samples of BHI starter cultures were then centrifuged at 1,568 rcf for 4 minutes in a microcentrifuge (Thermo Scientific mySpin 12) in the anaerobic chamber and resuspended in 1 mL sterile phosphate-buffered saline (PBS). For leave-one-out experiments, the resuspended cells were washed by centrifuging and resuspending in PBS again. The resulting suspension was vortexed and diluted to an approximate OD600 of 0.1 and used as inoculum into defined experimental conditions.

Growth assays were performed in standard 96-well microplates (Corning) at 37°C with a microplate reader (Biotek Eon or PowerWave). Around 180 μL of defined media were pipetted into each well, followed by 20 μL of inoculum. All experiments included at least 3 sterile control wells for each condition, into which 20 μL of sterile PBS was pipetted to establish consistent background OD600 measurements. Replicate wells were distributed pseudorandomly across the plate to control for plate layout effects, and inoculated wells were always paired with an adjacent control well of the same condition. For each condition, 3 to 6 replicates were included. Plates were sealed with a transparent Breathe-Easy sealing gas exchange membrane (RPI). Every 30 minutes, plates were shaken at medium speed for 40 seconds, after which OD600 readings were performed.

After large metabolomics and RNA-Seq experiments (see below), culture purity was checked by plating and 16S rRNA gene Sanger sequencing, using standard primers (8F AGAGTTTGATCCTGGCTCAG and 1542R AAGGAGGTGATCCAGCCGCA). In the time course metabolomics experiment shown in Fig 1, cross-contamination of sterile control tubes occurred after the fourth time point. Growth in both culture and control tubes was confirmed as pure E. lenta by plating and 16S rRNA sequencing.

Sample collection for metabolomics

Time course experiments were conducted in tubes in the anaerobic chamber in a 37°C incubator. For all metabolomics experiments, 3 independent culture replicates were included for each condition, with an equal number of uninoculated control tubes. Starter cultures and inocula were prepared as described above for growth assays. Approximately 5 mL of defined media was added to VWR glass culture tubes (53283–800) with screw caps. The PBS-washed inoculum was added to culture tubes to obtain an approximate starting OD600 of 0.001. A preliminary growth assay was conducted to define time points spanning the exponential growth phase in the tested conditions. At each time point, OD600 measurements of all inoculated tubes were first measured using a Hach DR1900 spectrophotometer, with a paired control tube to normalize for the background. About 100 μL from each tube were then transferred into a 96-well microplate, which was sealed and removed from the anaerobic chamber. Plates were centrifuged at 1,928 rcf at 4°C for 8 minutes, after which supernatants were collected into fresh polypropylene tubes or plates, sealed, and flash-frozen in liquid nitrogen.

Two time course experiments were carried out with stable isotope-labeled substrates. Experimental groups included conditions in which sodium acetate in the defined media was replaced with ¹³C₂ labeled sodium acetate (Sigma-Aldrich 282014), along with a matched experimental group with the same concentration of unlabeled substrate. The same procedure was followed for the arginine labeling experiment, using ¹³C₆ labeled L-arginine HCl (Sigma-Aldrich 643440).

For the comparative strain metabolomics experiment, 96-well polypropylene deep well plates were prepared with 800 μL of fresh media in each well. Starter cultures and inocula for 29 isolates of Eggerthella lenta and 1 isolate of Eggerthella sinensis [26] (S5 Table) were prepared as described above for growth assays, except without final dilution, and 80 μL was used to inoculate wells, leaving a blank well in between every culture well to prevent cross-contamination. After 72 hours, OD600 measurements were taken, plates were centrifuged, and supernatants were collected as described above.

Targeted quantification of acetate

A subset of unlabeled supernatant samples from the acetate labeling time course were shipped to Stanford University on dry ice for targeted quantification of acetate.

Samples (20 μL) were first mixed with an internal standard solution (30 μL, 1 mM phenylpropionate-d9) in a V-bottomed, poly(propylene), 96-well plate, and extracted by mixing with 3 sample volumes of extraction solution (75% acetonitrile, 25% methanol). The plate was covered with a lid and centrifuged at 5,000 rcf for 15 minutes at 4°C. Supernatant was collected for derivatization before subjecting to LC–MS analysis.

Samples were processed using a derivatization method targeting compounds containing a free carboxylic acid. Extracted samples were mixed with 3-nitrophenylhydrazine (NPH; 200 mM in 50% acetonitrile) and N-(3-dimethylaminopropyl)-N′-ethylcarbodiimide (120 mM in 6% pyridine) at a 2:1:1 ratio. The plate was sealed with a plastic sealing mat (Thermo Fisher Scientific, #AB-0566) and incubated at 40°C, 600 rpm in a thermomixer for 60 minutes to derivatize the carboxylate-containing compounds. The reaction mixture was quenched with 0.02% formic acid in 10% acetonitrile:water before LC–MS.

Samples were injected via refrigerated autosampler into mobile phase and chromatographically separated by an Agilent 1290 Infinity II UPLC and detected using an Agilent 6545XT Q-TOF (quadrupole time of flight) mass spectrometer equipped with a dual jet stream electrospray ionization source, operating under extended dynamic range (1,700 m/z). Chromatographic separation was performed using an ACQUITY Bridged Ethylene Hybrid (BEH) C18 column 2.1 × 100 mm, 1.7-μm particle size, (Waters, Milford, MA), using chromatographic conditions published elsewhere [6]. MS1 spectra were collected in centroid mode, and peak assignments in samples were made based on comparisons of retention times and accurate masses from authentic standards using MassHunter Quantitative Analysis v10.0 software from Agilent Technologies. Acetate was quantified from calibration curves constructed with acetate-d4 as a standard using isotope-dilution MS with phenylpropionate-d9 as the internal standard. Calibration curves were performed in a modified base form of EDM1 lacking amino acids and other carboxylic acids. A background level of 1.05 mM of acetate was subtracted to obtain the final quantities.

A plate layout error for supernatant samples from time points 4 to 7 in this experiment was noted based on the resulting acetate concentrations and corrected across datasets.

Untargeted metabolomics

Bacterial culture supernatant and sterile media, used in culture, were thawed on wet ice. Once thawed, samples were homogenized by inversion 5 times. Extracellular culture supernatant samples were prepared as follows: 20 μL of culture supernatant were extracted using 80 μL of a chilled extraction solvent at −20°C (1:1 acetonitrile:methanol, 5% water containing stable isotope-labeled internal standards). Samples were homogenized via pipette action, incubated for 1 hour at −20°C, centrifuged at 4°C at 6,000 rcf for 5 minutes. The supernatant was transferred to a new plate and immediately sealed and kept at 4°C prior to prompt analysis via LC-MS/MS.

Intestinal samples (colon, cecum, and ileum) were prepared individually using a single protocol as follows. Samples were kept frozen on dry ice and massed to at least 10 mg. Approximately 4 mL of −20°C extraction solvent (2:2:1 methanol:acetonitrile:water + stable isotope labeled internal standards) were added per milligram of intestinal sample. Six to eight 1 mm zirconia silica beads were added to each sample followed by prompt bead beating (15 Hz, for 10 minutes). Following a 1-hour incubation in the −20°C freezer, samples were centrifuged at 4°C at 18,407 rcf for 5 minutes. Supernatant was collected and stored at −20°C prior to centrifugal plate filtration (0.2 μm polyvinylidene difluoride (PVDF), Agilent Technologies, Santa Clara CA) at 4°C at 4,122 rcf for 3 minutes. Collection plate was sealed and maintained at 4°C prior to prompt analysis.

Serum samples were first thawed on wet ice. Around 20 μL of serum was extracted with 4 volumes of methanol, containing stable isotope-labeled internal standards. Samples were homogenized by vortexing for 20 seconds and placed in a −20°C for 1 hour to maximize protein precipitation. After freezer incubation, samples were centrifuged at 4°C at 18,407 rcf for 5 minutes. Supernatant was removed and dried under vacuum via centrivap (Labconco). Dried samples were then resuspended in 30 μL of 80% acetonitrile in water containing exogenous standard CUDA at 60 ng/mL. Samples were maintained at 4°C prior to prompt analysis.

Within each analysis batch, a small amount of each sample was removed and combined to create multiple technical replicate “pools,” which were analyzed intermittently throughout the analysis. These pools were used as external standards to ensure instrument stability across the batch. Additionally, method blanks were created using LC–MS-grade water in place of supernatant, sterile media, serum, or intestinal contents. These blanks were used to ensure that reported metabolites were not inadvertently added during sample preparation.

Samples, sterile media, pools, and blanks were promptly added to a Thermo Vanquish Autosampler at 4°C in a Vanquish UHPLC (Thermo Fisher Scientific, Waltham, MA). Chromatographic separation was performed using an ACQUITY Bridged Ethylene Hybrid (BEH) Amide column 2.1 × 150 mm, 1.7-μm particle size, (Waters, Milford, MA), using chromatographic conditions published elsewhere (HILIC method described in the Supplementary Methods of [25]). Samples were analyzed on a Thermo Q-Exactive HF orbitrap mass spectrometer operated utilizing data-dependent acquisition of MS2. Data were acquired independently in positive and negative modes via subsequent injections.

Stable isotope-resolved metabolomics (SIRM)

Intracellular extract samples were prepared with the following procedure, which was optimized for lysis of thick gram-positive cell walls: 600 μL of culture was transferred to an Eppendorf tube in anaerobic conditions and subsequently centrifuged at 10,000 rcf for 3 minutes at 4°C, after which the supernatant was removed and the samples were immediately flash frozen to quench metabolites. To each pellet, 300 μL of cold methanol was then added, followed by sonication on ice for 5 minutes and then shaking at 4°C for 4 to 12 hours. Samples were then centrifuged at 4°C at 15,000 rcf for 8 minutes, after which 120 μL of supernatant was transferred to fresh tubes and stored at −80°C until analysis. Prior to analysis, intracellular samples were dried at room temperature via Centrivap Benchtop Concentrator (Labconco). Samples were resuspended in 60 μL of a chilled solution of 1:1 methanol and acetonitrile, with 24% water at −20°C containing the internal standards CUDA and VAL-TYR-VAL each at 60 ng/mL. Samples were centrifuged at 4°C, 4,122 rcf for 5 minutes, and the supernatant transferred to a vial and immediately capped for LC-MS analysis.

Extracellular supernatant extraction for SIRM was performed as described above (Untargeted metabolomics section) with one modification. In SIRM samples, deuterated internal standards were replaced with CUDA and Val-Tyr-Val to enable untargeted enrichment analysis. LC-MS/MS analysis conditions for SIRM were identical to those used for standard untargeted metabolomics.

Untargeted metabolomics data processing

Untargeted metabolomics datasets were processed using MS-DIAL version 4.60 [77]. Metabolite features with intensity not greater than 3-fold elevated in samples compared to mean blank intensity were removed. Annotations were assigned using both local [25] and global (Mass Bank of North America) tandem mass spectral libraries. Annotation confidence scores were assigned based on MSI best practices [78,79]. Briefly, MSI level 1 denotes library matches of accurate mass (m/z), retention time (RT), and tandem mass spectra (MS2). MSI level 2 follows the same rules as MSI 1 but allows for partial matching of MS2 spectra—as is prone to occur when experimental spectra are convoluted. MSI level 3 denotes a high scoring and visually confirmed match of MS2 spectra. In a modification from the original definition of the MSI, we assign MSI level 4 when exact stereospecificity cannot be determined by MS2 and chromatographic separation. MSI level 4 is often assigned to sugars, lipids, and polyphenols. MSI levels 1 and 2 could only be assigned to metabolites in our local library, for which authentic standards have been analyzed in the same chromatographic conditions as the samples being annotated. Library matches were within 0.0025 m/z and 0.25 minutes retention time for all annotations assigned MSI levels 1 and 2 (comparisons are shown for the time course dataset in S12 Fig). Postprocessing was performed using MS-FLO [80] for removal of erroneous features.

Processed datasets were further analyzed using Feature-based Molecular Networking and MolNetEnhancer in the GNPS web platform version 22 [81–84], which assigned ClassyFire chemical classes to features based on molecular networking, independently of whether they were assigned a library identity.

To merge positive and negative ionization mode datasets from the same samples, duplicate features across datasets were identified as those with an expected mass difference of less than 0.02, a retention time difference of less than 0.1, and a Pearson correlation across samples of at least 0.7. If one or both members of a pair of duplicate features were assigned an identification, the feature with lower (more confident) MSI score was retained in the merged dataset. Otherwise, the positive mode feature was retained. The other feature was removed for downstream analysis.

Prior to statistical analysis, initial untargeted metabolomics feature tables were filtered to remove features with a high coefficient of variation across replicate samples (>50%) and to remove potential technical outlier samples where the total signal from all features differed from the assay median by >50%. Log₁₀-transformed intensities were used for most statistical analysis, with the exceptions of SIRM datasets and the comparative strains dataset (for which values were approximately normally distributed without transformation). A pseudocount equal to 0.25 times the minimum nonzero value was added to the peak intensities for each feature before log transformation. Heatmaps of metabolite abundances were generated using the ComplexHeatmap package v2.13.1 [85].

SIRM data processing

Intra- and extracellular untargeted data generated from SIRM experiments was analyzed separately using Compound Discoverer version 3.3 (Thermo Scientific, Bremen, Germany). Samples treated with labeled compounds were always paired with matched samples treated with unlabeled compounds in order to correct for naturally occurring isotope abundances. Unlabeled samples were used for compound detection and formula assignment via isotope pattern-based prediction, spectral library matches, or mass lists matches. The isotope patterns and formulas from the sample files then served as a reference for the detection of potential isotopologues per compound in the labeled sample type.

Specifically, the workflow consisted of the following nodes in Compound Discoverer: Input Files→ Select Spectra→ Align Retention Times (ChromAlign)→ Detect Compounds (Legacy) → Group Compounds→ Predict Compositions→ Search Mass Lists→ Search mzCloud→ Mark Background Compounds→ Assign Compound Annotations→ Analyze Labeled Compounds→ Descriptive Statistics→ Differential analysis.

The default settings from the “Stable Isotope Labeling w Metabolika Pathways and ID using Online Databases” workflow were used, with the following modifications:

1. Detect Compounds (Legacy): General–Min.Peak Intensity: 10000; Ions: [M+H]+1 or [M-H]-1 for positive and negative mode experiments respectively.
2. Group Compounds: Peak Rating Filter–Peak Rating Threshold: 4; Number of Files: 3.
3. Search Mass Lists: Search Settings–Mass Lists: Combined Hilic Mass mzRT library; Use Retention Time: True; RT Tolerance: 0.3 min; Mass Tolerance: 5 ppm.
4. Search mzCloud: DDA Search–Match Factor Threshold: 85
5. Mark Background Compounds: General–Max. Sample/Blank: 3
6. Assign Compound Annotations: Data Sources–Data Source #1: mzCloud Search; Data Source #2: MassList Search; Data Source #3: Predicted Compositions.
7. Analyze Labeled Compounds: Pattern Analysis–Intensity Threshold [%]: 2

Positive and negative polarity files were analyzed initially as separate studies with the following study definitions: Study factors including strain, replicate, substrate concentration, sample type, and time point were assigned. Sample types were assigned as either sample (unlabeled), labeled, or blank. These study factors interfaced with several nodes to reduce undesirable features and maximize reporting of quality high-intensity peaks with potential for accurate measurement of ¹³C incorporation.

Results were filtered for nonblank formula assignment and absence in background samples. The MSI levels and labeling status for persisting entries were manually inspected for each compound and annotated onboard via custom tags. MSI levels were assigned based on the criteria previously described to match MS-Dial output. The mass isotopologue distributions were plotted to ensure reproducibility between replicates of various time points and detect anomalous labeling trends. The absence of reported enrichment in control samples processed as labeled samples was verified. A minimum threshold of 3% combined enrichment across all isotopologues other than M+0 was applied. This threshold was necessary for less abundant peaks where the ¹³C natural isotopic abundance correction introduces uncertainty in the M+1 and M+2 isotopologues.

A specification of the full Compound Discoverer workflow is available at https://github.com/turnbaughlab/2022_Noecker_ElentaMetabolism.

RNA-Seq

E. lenta DSM 2243 glycerol stocks were plated on BHI+ and incubated anaerobically at 37°C for 3 days. A single colony was then inoculated into 5 mL of BHI+ liquid culture and incubated at 37°C for 48 hours. A volume of 1 mL of the resulting culture was then centrifuged, washed once, and resuspended in equal volume PBS; all in anaerobic conditions. To obtain a starting OD600 of 0.01, 220 μL of this inoculum were transferred into culture flasks containing 20 mL of EDM1 (70% carbon source reduced version; see S1 Table). After 20 hours of growth (early or mid-exponential phase), these cultures were treated with an additional 1.9 mL of sterile water or filter-sterilized solution containing either L-arginine (to reach a final concentration of 86 mM), sodium acetate (final concentration of 14.5 mM), or agmatine sulfate (final concentration of 30 mM). After 18 more hours (late exponential phase), 7.5 mL of each culture was collected into 15 mL conical tubes containing 5 mL of Qiagen Bacterial RNA-Protect (#76506). Cultures were centrifuged at 2,800 rcf at 4°C for 10 minutes, after which the supernatant was carefully removed. Pellets were extracted directly using the Qiagen RNeasy Mini kit (#74104) with modifications for difficult-to-lyse gram-positive bacteria. Samples were maintained on ice in between all disruption and centrifugation steps throughout the protocol. Briefly, 200 μL of TE buffer containing lysozyme (15 mg/mL, #L4919) and 20 μL of Qiagen Proteinase K (#19131) were added to each pellet, vortexed gently, and incubated at room temperature for 10 minutes with shaking on an Eppendorf ThermoMixer at 900 rpm. Around 700 μL of Buffer RLT was then added to each tube and vortexed, after which the full contents were transferred to MP Biomedical Lysing Matrix E tubes (#116914500) and disrupted mechanically in a BioSpec Mini-Beadbeater-96 for 50 seconds. After disruption, tubes were centrifuged for 3 minutes at 15,000 rcf, and 850 μL of supernatant was transferred to fresh tubes. To each sample, 590 μL of 80% ethanol was added and mixed by pipetting, after which lysates were transferred to Qiagen RNeasy spin columns and washed, following the RNeasy Mini kit QuickStart protocol including a single on-column DNase digestion (DNase #79254). After purification, RNA was eluted twice into 30 μL of nuclease-free water. RNA integrity was checked using an Agilent TapeStation 4150 and stored at −80°C.

RNA library preparation and sequencing was performed by the Microbial Genome Sequencing Center/SeqCenter. Samples were DNase treated with Invitrogen DNase (RNAse free, #AM2222). Library preparation was performed using Illumina’s Stranded Total RNA Prep Ligation with Ribo-Zero Plus kit (#20040529) and 10 bp IDT for Illumina indices. Supplementary oligonucleotide probes specific to E. lenta rRNA and other highly expressed noncoding RNAs were incorporated during Ribo-Zero depletion (S10 Table). Sequencing was done on a NextSeq 2000 with 2 × 50 bp reads. Demultiplexing, quality control, and adapter trimming was performed with bcl-convert v3.9.3.

Reads were trimmed and quality filtered using fastp v0.20.0 [86] with the following parameters:—trim_poly_g—cut_front—cut_tail—cut_window_size 4—cut_mean_quality 20—length_required 15. The Hisat2 aligner v2.2.1 [87] was used to map reads to the E. lenta DSM 2243 reference genome, downloaded from NCBI RefSeq (GCF_000024265.1). Gene-level read counts were obtained using the corresponding NCBI annotations and the featureCounts function in the R package Rsubread v2.6.4 [88], with the minimum quality score set to 1.

Construction, curation, and analysis of metabolic reconstructions

Genome-scale metabolic reconstructions were created from genome sequences of 25 E. lenta strains [26] using the DEMETER pipeline [89]. Briefly, DEMETER performs systematic refinement of a draft genome-scale reconstruction, in this case generated through KBase [90]. Based on manually gathered experimental data, gap-filling solutions that had been manually determined in a subset of reconstructions are propagated by DEMETER to newly reconstructed strains. Moreover, DEMETER ensures correct reconstruction structure through use of a curated reaction and metabolite database and removes futile cycles resulting in unrealistically high ATP production. A test suite ensures agreement with the input experimental data and verifies model features such as mass and charge balance and feasible ATP production. The E. lenta DSM 2243 reconstruction [35] underwent additional refinement of reactions and gene annotations against manually performed comparative genomics analyses.

Reconstructions were analyzed using the Cobra Toolbox version 3.0 [91] in Matlab version 2018b, with the IBM Cplex solver version 128. Defined media concentrations were mapped from compound names to BiGG metabolite IDs [92] and converted to cell uptake rates over the duration of E. lenta’s exponential growth phase in batch culture (S1 Table) using the concToCellRate function in the Cobra Toolbox and an approximate cell dry weight of 3.3 × 10⁻¹³ g, calculated based on colony forming units and dry biomass quantification from 2 aliquots of a late-exponential phase EDM1 culture. Additional compounds detected in sterile culture media with high confidence based on untargeted metabolomics were included in the simulation media with a fixed maximum uptake rate of 1 mM/gDW/h.

The collection of E. lenta strain reconstructions included 2 reconstructions of the type strain: the DSM 2243 reconstruction, which had undergone additional comparative genomics curation with PubSeed [93], and a slightly smaller and less refined reconstruction included in the AGORA2 collection [35] based on genome resequencing of the ATCC 25559 version of the type strain. Neither reconstruction initially displayed nonzero growth in EDM1 using flux balance analysis. In order to facilitate interpretation of flux balance analysis (FBA) results and avoid excess gap-filled reactions, we used the simpler E. lenta ATCC 25559 type strain reconstruction as the basis for subsequent curation and analysis. We transferred reactions present in the DSM 2243 reconstruction into this version if they were supported by genome annotations from other sources (Prokka [48], GapMind [37]) and/or by experimental growth or metabolomics data. We also performed several additional custom curations. Transporters were added in a strain-specific manner for metabolites identified with high confidence (MSI level 1) and detected as secreted or depleted with a log₂ fold change greater than 2 in the stationary phase strain collection metabolomics dataset (Fig 4). Several pathways were also modified based on growth assay results and/or pathway annotation software [36,37]. Curations were checked for viable growth in EDM1 using flux balance analysis. Reconstructions for the other 23 strains were only curated to ensure growth in EDM1 and to allow import/export based on metabolomics data, but not based on genome analysis with GapMind [37] or the results of leave-one-out growth experiments, since those were only performed using the type strain. A complete summary of all curation steps is found in S3 Table.

FBA, pFBA, and FVA were performed using the Cobra Toolbox functions optimizeCbModel, minimizeModelFlux, and fastFVA, respectively. Flux variability ranges are reported for 99% of the maximum growth rate.

Metabolite uptake and secretion ranges estimated by FVA were compared with the stationary phase strain collection metabolomics dataset (shown in Fig 4). To compare metabolite data with FVA estimates, identified metabolites were first mapped from InChIKey metabolite IDs to KEGG IDs using the CTS Convert utility [94] implemented in the R package webchem v1.1.2 [95]. KEGG IDs were then mapped to BiGG IDs using the BiGG database [92] and manually checked for consistency with compound IDs in the AGORA models. For purposes of this analysis, metabolites were considered produced if they had a log₂ fold change greater than 0.5 in supernatants from at least one of the 3 type strain isolates included in the experiment (E. lenta DSM 2243—UCSF, E. lenta ATCC 25559, and E. lenta DSM 2243—DSMZ), and depleted if the log₂ fold change was less than −0.5.

To compare gene expression values with model flux estimates, we first ran pFBA and FVA for the modified EDM1 condition used for RNA-Seq (with 70% of the standard levels of arginine and acetate). We obtained the set of genes linked to reactions in the iEL2243_2 reconstruction, using NCBI BLASTn to map genes between different sets of annotations. Genes linked to multiple reactions were counted multiple times for each reaction, and vice versa. Only genes linked to reactions in the original ATCC 25559 reconstruction were included.

Similarly, to compare reaction knockout predictions with strain variation, genes linked to reactions in the original ATCC 25559 reconstruction were mapped to annotations used in a previous pan-genome analysis of 31 nonclonal E. lenta genomes [26]. In this previous analysis, amino acid sequence families were clustered across genomes using ProteinOrtho [96] with cutoffs of 60% identity and 80% coverage. This analysis was then used to determine the number of strains in which each gene family in the ATCC 25559 reconstruction was present and to compare that distribution with the effects predicted by knockout analysis of the unconstrained model.

Strain comparative metabolomics analysis

Metabolites were classified as strain-variably produced/depleted if they were differentially increased/decreased (FDR-adjusted p < 0.1 and absolute log₂ fold change > 0.5) in supernatants from at least 1 isolate strain but fewer than 29 (of the 30 isolates included in this experiment).

The phylogenetic and comparative genomics analyses used in this study were previously reported, including a core gene phylogenetic tree (Phylophlan), gene family clustering across strains (ProteinOrtho) and Prokka and GhostKoala annotation of all genomes [26].

Procrustes analysis was performed using the R package vegan v2.6–2, with evaluation of significance using the protest function [97]. The E. sinensis isolate was excluded from Procrustes analysis to avoid skewing the distribution of phylogenetic distances.

The gene-metabolite association analysis was performed as described previously [26], with different cutoffs for prioritization. Briefly, all observed patterns of gene family presence-absence (based on clusters of 60% identity and 80% coverage) were enumerated across the collection of genomes. Log-transformed metabolite intensities were then tested for association with each presence-absence pattern using Welch’s t tests. Using an initial cutoff of an FDR-adjusted p-value of 10⁻⁴, 39.0% of metabolite features were significantly associated with a gene cluster by this method. To further restrict results to those features most likely to depend on the presence of a gene, we filtered gene-metabolite links using 2 additional separability criteria. First, the difference in median log₁₀ metabolite values between strain samples with and without the gene was required to be at least 0.4. Secondly, the 10th percentile log₁₀ metabolite value for strains with the gene was required to be at least 0.4 above the maximum value in controls, and the 90th percentile value for strains without the gene was required to be lower than that value. Finally, only the highest association for each metabolite feature was retained. This additional filtering resulted in the final table of 84 gene family-metabolite links. KEGG pathway enrichment analysis of the final gene set was performed using clusterProfiler v4.0.5 [98] with a p-value cutoff of 0.1.

Cross-dataset untargeted metabolomics analysis

As described above, untargeted metabolomics datasets from supernatant, mouse intestinal contents, and serum were collected using the same chromatography and mass spectrometry methods. Pairs of features were compared across these datasets and linked if they were within 0.007 m/z, 0.5 minutes retention time and had a cosine similarity of at least 0.205 between their MS2 spectra for positive ionization mode and 0.251 for negative ionization mode. Features for which MS2 spectra were not collected were linked to other features within 0.001 m/z and 0.2 retention time. Linked feature pairs were also required to be annotated as the same adduct. These m/z and retention time thresholds were chosen based on examination of the distributions of pairwise differences between features. The cosine similarity cutoffs were chosen as the 99.5th percentile of cosine similarity between a large sample of unrelated feature pairs: specifically, all pairwise comparisons of 2 sets of 200 randomly sampled features with retention times differing by at least 1 and m/z differing by at least 0.01. This procedure was repeated separately for positive and negative ionization mode datasets. Under these criteria, only approximately 0.5% of linked features assigned an identity were linked to features with a conflicting identity. Linked pairs of features were merged into shared metabolite IDs that were carried forward for cross-dataset analysis and comparison.

Quantification and statistical analysis

All statistical analyses were performed in R v4.1.1, with data visualizations generated using the ggplot2 package v3.3.6 [99]. Statistical tests, sample size, and standard error are reported in the figures and figure legends. Benjamini–Hochberg FDR correction was used to adjust for multiple comparisons in all cases.

Untargeted metabolomics statistical analysis

Differential trajectories across time series datasets were assessed using spline models implemented in the santaR package v1.2.3 [100]. Differential abundance analysis between supernatant samples and sterile controls at the final time point was performed using Welch’s t tests after checking assumptions of normality. In one case where cross-contamination of sterile control tubes occurred at later time points, features at those time points were compared with control samples from the last uncontaminated time point.

Differential abundance analysis for the comparative strains dataset was performed using linear models with each strain identity as a covariate. In the gnotobiotic mouse experiment, each mouse was considered an independent experimental unit; however, hierarchical random effects statistical models were used to account for shared covariance between mice from the same cage. Specifically, differential abundance analysis for the gnotobiotic mouse intestinal dataset was performed using linear mixed models with the R package lmerTest v3.1–3 [101], incorporating fixed effects for intestinal site, colonization group, and the interaction between them, and nested random effects for cage and animal. The difflsmeans function with Benjamini–Hochberg multiple hypothesis adjustment was used to evaluate the statistical significance of differences of each colonization group versus GF under this model.

Statistical analysis of growth curves

Growth curves were normalized based on average time-matched readings from blank control wells. Normalized values were used to fit logistic growth models for each well using the R package growthcurver v0.3.1 [102]. Low-quality model fits (sigma > 0.1) were removed prior to calculation of summarized parameter values.

Targeted metabolomics statistical analysis

Differential abundance analysis was performed using a linear model with terms for time point, strain, and their interaction. Differences from controls under the resulting model were estimated using Dunnett’s method as implemented in the package emmeans v1.7.5 [103].

RNA-Seq statistical analysis

Differential expression analysis was performed using negative binomial generalized linear models implemented in the DESeq2 package v1.32.0 [104].