Abstract
Despite evidence linking the human microbiome to health and disease, the mechanistic details of how the microbiota affects human physiology remain largely unknown. Metabolites encoded by bacteria are expected to play an integral role in the microbiota’s effect on its human host. Assigning function to these metabolites is therefore critical to determining the molecular underpinnings of the host-microbe relationship and ultimately developing microbiota inspired therapies. Here we use large-scale functional screening of small molecules produced by individual members of a simplified human microbiota to identify bacterial metabolites that agonize G-protein coupled receptors (GPCR). This analysis revealed a complex network of metabolite host receptor interactions and guided our identification of multiple microbiota derived agonists of GPCRs associated with diverse biological functions within the nervous and immune systems, among others. Collectively, the metabolite-receptor pairs we uncovered indicate that diverse aspects of human health are potentially modulated by structurally simple metabolites arising from primary bacterial metabolism.
Statement of Significance Bacteria residing within the human body have been shown to influence human health. It is likely that physiological responses to the human microbiota are mediated by the collection of small molecules encoded within these bacteria. In this study we use direct functional screening of small molecules produced by individual members of a simplified human microbiota to identify new G protein coupled receptor-metabolite interactions that seek to explain the molecular underpinnings of the microbiota’s influence on its human host.
Human bodies are home to diverse and ever-changing collections of bacteria. The ability of the microbiota to influence human health has been explored extensively. [1] The most common methods for studying host-microbe interactions have featured “omics” based-analyses that have examined genomic, transcriptomic, proteomic or metabolic differences between patient cohorts. [2–5] Although these informatics-based methods have served as powerful tools for uncovering correlations between changes in the microbiota and health and disease, they are somewhat limited in their ability to reveal the mechanistic details of how the microbiota might alter mammalian physiology.[6] Much of the influence the microbiota has on its human host is likely encoded in the collection of small molecules it produces or modulates.[7] However, the number of well-defined interactions between metabolites produced by human associated bacteria and discrete human receptors is dwarfed by the number of reports attributing biological phenotypes to the microbiome, highlighting the need for a more systematic characterization of microbiota encoded bioactive metabolites.
In the case of synthetic small molecules that have proved useful for therapeutically modulating human physiology (i.e., U.S. Food and Drug Administration (FDA) approved drugs) the majority (60-70%) function through just three classes of receptors: G-protein coupled receptors (GPCR), ion channels, or nuclear hormone receptors.[8] Many of these same proteins bind endogenous signaling molecules that regulate a wide range of physiological responses.[9] Based on the fact that these receptors play such an important role in how eukaryotic cells have evolved to translate external chemicals into biologic responses, it is likely the microbiota affects host physiology by modulating these same receptors with secreted metabolites.
Although healthy humans are colonized by hundreds, if not thousands of different bacterial species, the metabolic diversity they generate is likely limited by a high level of biosynthetic redundancy between bacterial species.[10] Due at least in part to this metabolic redundancy, it has been possible to use simplified human microbiomes (SIHUMIs) to model health and disease in murine models.[11, 12] In lieu of exploring random individual commensal species, we sought to conduct a more in-depth investigation of GPCR active microbiota encoded metabolites using bacteria from a model SIHUMI that contained a taxonomically diverse collection of commensal, health promoting and pathogenic bacteria. This consortium, which is composed of seven bacteria, assembled as a tool for studying gastrointestinal inflammation in the context of a healthy bacterial flora fulfills these general criteria and was therefore selected for use in this study.[13] Bacteria present in the SIHUMI consortium represent the major taxa found in the human microbiome and include beneficial bacteria (Lactobacillus plantarum, Bifidobacterium longum, and Faecalibacterium prauznitzii), non-pathogenic bacteria associated with disease (Bacteroides vulgatus and Ruminococcus gnavus), as well as clinically relevant pathogens (Escherichia coli LF-82 and Enterococcus faecalis).
We screened the metabolites produced by individually grown members of this SIHUMI consortium for agonism against 241 GPCRs. The resulting interaction map provides evidence, at the molecular level, for the existence of a complex network of microbial-host interactions, many of which involve receptors that have been modulated therapeutically with synthetic small molecules. Our characterization of interactions predicted by this analysis led to the discovery of both previously unrecognized as well as known microbiota encoded GPCR agonists. The structures of the active molecules we identified support the growing notion that simple bacterial metabolites arising from primary metabolic processes are likely to broadly impact human physiology.
Results
Culturing bacteria and GPCR screening
Bacteria from the SIHUMI consortium were individually fermented under anaerobic conditions in separate large scale (20L) culture vessels (Figure 1). After 10 days of static fermentation at 37 °C, hydrophobic resin was added directly to each culture. The resulting suspension was mixed to allow organic metabolites present in the fermentation broth to bind to the absorbent resin. Metabolite loaded resin was then collected by filtration, washed, and the bound metabolites were eluted with acetone. Each resulting crude metabolite extract was partitioned into 9 metabolite-rich fractions using reversed phased flash chromatography. A small aliquot of each fraction was arrayed for use in high-throughput GPCR screening. The remaining material was saved for follow up assays and for use in molecule isolation and structure elucidation studies. Although this pre-fractionation process increases the number of samples to be screened it simplifies the complexity of the crude culture broth extracts, which improves the signal in the primary screen thereby increasing the diversity of interactions that are identified and facilitating the downstream isolation of bioactive compounds. In addition to the bacterial fermentations, media not inoculated with bacteria was processed under identical conditions to control for possible bioactivity of small molecules derived directly from the media. The resulting library of bacterial metabolites was then screened with a cell-based assay for fractions that could agonize members of a panel of 241 GPCRs (Table S1). Specifically, a collection of recombinant cell lines engineered to measure β-arrestin recruitment by individual GPCR targets (β-arrestin recruitment assay) was used. For GPCRs with well characterized endogenous ligands, a maximum value for β-arrestin recruitment (100%) was set by exposing the recombinant cell line to a known agonist (Table S1). In the case of orphan receptors (i.e., receptors without well characterized endogenous ligands), normalization of β-arrestin recruitment was performed by assigning the vehicle control to 0% activity. Hits were classified as such if a fraction induced a GPCR response to >30% of the control ligand (>50% for orphan GPCRs) and the comparable media control fraction showed <30% activity against the same GPCR (<50% for orphan GPCRs).
The bacterial fraction library induced β-arrestin recruitment above our hit threshold levels for 67 of the 241 individual GPCR reporter cell lines we tested (Figure 2A-B, Table S2-S3). Of these 67 GPCRs, 54 did not show a strong background signal from the corresponding media control fraction, suggesting they were responding to bacterially encoded metabolites. Manual review of these 54 hits led us to de-prioritize 15 of these GPCR-fraction pairs due to high background of either the receptor or fraction (Table S4). The remaining 39 GPCRs were re-assayed in replicate; 22 of these GPCRs showed reproducible β-arrestin recruitment in response to 1 or more bacterial fractions (Figure 2C). Based on data from the Human Protein Atlas most of the receptors that reproducibly responded to bacterial metabolites are expressed at body sites regularly exposed to the microbiota (Figure 2C).[14] Notably, a large number of the receptors that were reproducibly agonized by microbiota encoded metabolites are also targeted by FDA approved drugs, indicating that receptors with proven physiological relevance are potentially modulated by bacterial ligands (Figure 2C). To identify specific GPCR-active metabolites, we used bioassay-guided isolation to purify metabolites from the large-scale culture broth fractions and de novo structure elucidation methods to determine their structures.
Known and novel ligands for hydroxycarboxylic acid receptors
A number of receptors agonized in our screen are known to respond to bacterial ligands. As an initial validation exercise we characterized activities expected to arise from well-known bacterial GPCR agonists. The hydroxycarboxylic acid receptors, GPR81, GPR109A and GPR109B, are agonized by both human and bacterial ligands.[15] Bioassay guided fractionation of GPR109A active fractions from cultures of both L. plantarum and R. gnavus yielded nicotinic acid (Vitamin B3) as the active metabolite (Figure 3A, Figure S1). Nicotinic acid, an essential nutrient acquired either through diet or gut bacteria, is the most extensively studied non-endogenous ligand for this receptor. Its ability to regulate lipid metabolism in hyperlipidemic patients is well established in the clinic.[16] The identification of this well characterized and in vivo validated ligand receptor pair suggests that data generated in our screen has the potential to uncover new biologically relevant metabolite GPCR interactions.
Fractions derived from cultures of both E. coli LF82 and L. plantarum agonized a second hydroxycarboxylic acid receptor, GPR109B. Bioassay guided fractionation did not identify the endogenous ligand produced in humans, 3-hydroxyoctanoic acid, but instead yielded phenylpropanoic acid as the active metabolite (Figure 3B, Figure S1). This previously unknown GPR109B agonist, elicited a similar GPCR response than did 3-hydroxyoctanoic acid (Figure 3B). While the EC50 values for the known and novel ligands (304 μM and 208 μM, respectively) are higher than is often seen for endogenous GPCR ligands (Table S1), no more potent GPR109B agonists have been identified, outside of those derived synthetically.[17] Whether this is an inherent attribute of the receptor or represents a failure to identify the natural human ligand for this receptor remains to be seen.
Phenylpropanoic acid is not produced natively by humans. Its presence in human fecal and sera samples has been attributed to either de novo biosynthesis by bacteria or microbial transformation of dietary compounds, most notably by species of Clostridium.[18–21] While we believe phenylpropanoic acid is the first microbiota derived agonist to be identified for GPR109B, in a screen of synthetic molecules, several aromatic D-amino acids were found to be GPR109B agonists that can trigger chemoattraction signaling pathways in leukocytes.[22, 23] In a quantitative analysis of human fecal water, phenylpropanoic acid was reported in healthy patients at an average concentration of 77.30 μg/mL (513 μM).[18] At this concentration, production of phenylpropanoic acid by gut bacteria would be high enough to expect agonism of GPR109B. An analysis of the concentration of phenylpropanoic acid in a larger number of patient samples will be required to determine whether phenylpropanoic acid dependent agonism of GPR109B is likely to be a common phenomenon in the gut.
Known and novel ligands for neurotransmitter receptors
Our GPCR interaction map revealed numerous bacterial fractions that strongly agonized neurotransmitter receptors, a key component of the gut brain axis (Figure 1C).[24] Bacterial produced aromatic amines, most notably tryptamine, have recently been reported as agonists of neurotransmitter receptors, particularly serotonergic GPCRs (5-hydroxytryptamine receptors, HTRs).[25] A majority of bacteria in this SIHUMI produced fractions that agonized HTRs (Figure 3C). Isolation of the active metabolite yielded tryptamine, which was produced in varying quantities by members of this SIHUMI (Figure S3). These results agree with various reports that HTRs are responsive to a wide array of bacteria due to the generality of tryptamine production across species.[26]
In fractions from multiple bacterial species we observed agonism of the D2-type dopamine receptors (DRDs), DRD2 and DRD3 (Figure 3D). Bioassay-guided isolation led to the aromatic amine tyramine as the major metabolite responsible for DRD agonism in these fractions. Tyramine arises from decarboxylation of tyrosine and differs from dopamine only by the absence of a second hydroxyl on the aromatic ring. It is reported to accumulate to μM levels in the gastrointestinal tract, a phenomenon which has been attributed to production by human microbiota.[27] While no biological significance has been assigned to the microbiota dependent accumulation of tyramine in animal models, it is sufficiently potent that its observed concentration in the gastrointestinal tract is high enough to agonize D2 subtype DRDs.
In contrast to the broad activation seen for DRDs and HTRs across extracts from all of the bacteria in this consortium, a specific response to fractions from E. coli LF82 was detected for a member of the histamine receptor (HRH) family, HRH4. Our inability to retain HRH4-activity when using hydrophobic chromatography during the bioassay guided purification process suggested that the active molecule was highly polar. We did not, however, expect that the activity was due to bacterially produced histamine, as the active fraction did not agonize other HRH family receptors and we could not detect histamine by LC-MS or NMR. We ultimately found the polyamine cadaverine to be the metabolite responsible for HRH4 agonism (Figure 4A). The activity of cadaverine was confirmed using a commercial standard (EC50 1.1 μM) (Figure 4B). In addition to cadaverine, bacteria commonly produce a number of other simple polyamines including agmatine, spermidine and putrescine.[28] To explore the promiscuity of HRH4 agonism by polyamines, we tested synthetic standards of these metabolites for the ability to induce β-arrestin recruitment by each member of the HRH receptor family. Agmatine and putrescine showed limited activity against HRH4 (Figure 4C), while spermidine did not show activity against any receptor in the family. The inability for humans to biosynthesize cadaverine suggests that the influence of polyamines on histamine signaling pathways is likely specific to bacterial metabolism.
Cadaverine is produced by bacteria through the decarboxylation of lysine (Figure 4A), while agmatine and putrescine are derived from arginine. In a number of bacteria, including many associated with the microbiota (Table S5), cadaverine is encoded by both the constitutive ldc gene cluster as well as the cad gene cluster, which is induced at low pH (pH <6.8).[29] High level production of cadaverine by the CadA lysine decarboxylase plays a role in protecting against acid stress.[30]. As the pH of the digestive system varies longitudinally and features multiple acidic sections (e.g., cecum pH ~5.7), high level production of polyamines by cad gene cluster containing bacteria is likely to occur at numerous sites in the GI tract. The biological relevance of gastrointestinal production of polyamines remains unclear, however host responses to polyamines have been reported in various contexts.[31, 32] Interestingly, although histamine receptor subtypes differ in their associated functions and their distribution throughout the human body, HRH4 is expressed in the gastrointestinal tract and altered expression levels have been linked to inflammatory responses that are related to inflammatory bowel diseases and cancer.[33]
A growing number of studies have uncovered connections between gut microbiota and the nervous system.[34, 35] Our exploration of microbiota encoded neurotransmitter receptor agonists expands the mechanistic evidence for simple biogenic amines serving as potentially widespread modulators of the gut-brain axis.[14] These data imply that microbiota-dependent dopaminergic, serotonergic and histaminergic responses likely represent general signaling events in the gastrointestinal tract with varying activation profiles depending on the specific collection of bacteria present in an individual’s microbiome.
Structurally distinct lipids agonize diverse GPCRs
Lipids, which represent diverse GPCR active ligands [36, 37], predominantly elute very late in our fractionation protocol (Figure 5A). Based on the receptor interaction map we could initially classify GPCRs as lipid responsive if they were agonized by the late lipid-enriched fractions of the extract library. A subset of receptors, including GPR120, CNR2, GPR171, GPR132, responded broadly to the lipid fraction from most of the consortium, whereas other responses were specific to particular species (BAI1, NMU1R, UTR2). HPLC-charged aerosol detection analysis of the lipid fractions indicated they contained not only mixtures of simple saturated fatty acids but also other more complex lipid species (Figure 5B). Marrying unique receptor activity profiles with unique lipid signals guided us to previously unrecognized bacteria encoded GPCR agonists.
The brain angiogenesis factor 1 (BAI1) receptor was agonized by lipid fractions from the Gram-negative bacteria in the consortium: E. coli and B. vulgatus. The E. coli LF82 lipid fraction showed the most potent agonism of BAI1 and therefore it was selected for further analysis. Bioassay guided fractionation identified the BAI1 agonist as the cyclopropyl-containing lipid 9,10-methylenehexadecanoic acid (EC50 11 μM). Synthetic 9,10-methylenehexadecanoic acid, but no saturated lipids we tested agonized BAI1, confirming the specificity of the receptor reflected in the initial GPCR activity map (Figure 5C). The enzyme cyclopropane-fatty-acyl-phospholipid synthase (Cfa) uses the one carbon donor S-adenosyl-L-methionine to generate cyclopropyl lipids from unsaturated fatty acids (Figure 5G). Cyclopropane-containing fatty acids are important membrane components in Gram-negative as well as mycolic acid bacteria.[38] Macrophages use BAI1 as a pattern recognition receptor to sense Gram-negative bacteria and to induce selective phagocytosis and antimicrobial responses; 9,10-methylenehexadecanoic acid may represent a previously unrecognized recognition motif for innate immune responses.[39–42]
Two peptide receptors NMU1R (neuromedin receptor 1), which mediates satiety and peristalsis in the gut [43, 44] and the vasoconstriction inducing urotensin 2 receptor (UTR2) responded specifically to lipid fractions generated from B. vulgatus. Isolation of the active metabolite yielded the anteiso-methyl branched-chain fatty acid 12-methyltetradecanoic acid (aiC15:0) (Figure 5D). Anteiso-fatty acids (ai) contain an alkyl branch at the ante-penultimate carbon in contrast to iso-fatty acids (i) which branch at the penultimate carbon. Both synthetic and natural aiC15:0, but no simple fatty acids we tested, agonized NMU1R (EC50 125 μM) and UTR2 (EC50 191 μM). Lipid sensitivity of NMU1R and UTR2 appears specific to aiC15:0, as fatty acids with even slightly modified branching patterns (iC15:0) or carbon chain length (aiC17:0) displayed minimal agonist activity (Figure 5E-F). Methyl branched fatty acids arise from the use of a branched primer in place of acetyl CoA in normal fatty acid biosynthesis. In the case of anteiso-methyl-branched fatty acids, 2-methyl-butyryl-CoA, which is derived from isoleucine is used to prime fatty acid biosynthesis (Figure 5G). The selectivity for branched primers lies with the β-ketoacyl acyl carrier protein synthase (KAS III or FABH) that carries out the first condensation in fatty acid biosynthesis. Anteiso-methyl fatty acids are predominantly produced by Gram-positive FABH enzymes.[45, 46] Roughly 10% of bacteria have lipid pools enriched in branched chain fatty acids.[45] B. vulgatus, is among those bacteria enriched in branched chain fatty acids and maintains aiC15:0 as ~30% of its total fatty acid repertoire.[47]
Bacteria are known to produce diverse and oftentimes taxa specific, collections of lipids. The examples described here from examining even this minimized model microbiome suggest the potential for markedly different receptor activation profiles and hence biological consequences depending on the specific lipid signature encoded by an individual’s microbiome. For BAI1, NMU1R and UTR2 our data suggests that they differentially respond to lipids produced by largely Gram-positive or Gram-negative bacteria indicating that their activities will fluctuate with changes in the gross taxonomic composition of a microbiome.
Analysis of mice colonized with the seven strain SIHUMI consortium
In parallel with our in vitro screen studies, we used high-resolution mass spectrometry-based metabolomics to compare germ free and SIHUMI consortium colonized mice. For this analysis, cultures of individually grown bacteria from the consortium were combined and the mixed sample was gavaged into germ-free C57BL/6 mice. PCR based species analysis of DNA extracted from the stool of animals three days post inoculation confirmed their colonization by the consortium. Ten days post colonization the lumen material (cecal stool) was collected from the germ-free controls as well as the SIHUMI colonized animals. Using targeted high-resolution mass spectrometry, we looked for differences in metabolite accumulation in these samples (Figure 6).
Targeted MS analysis of cecum extracts revealed that all but one of the GPCR-active metabolites we identified was enriched in these mice compared to their abiotic counterparts (Figure 6, Table S7), suggesting a largely parallel biosynthesis in laboratory grown mono-cultures and the consortium in vivo. The lone exception was phenylpropanoic acid. Our failure to identify phenylpropanoic acid in this specific model study does not preclude its potential biologic relevance in other settings, especially in light of the fact that it has been seen at μM levels in human patient samples.[18] The lack of in vivo production is likely due to low production of this metabolite by the specific strain used in this consortium, which is supported by the low GPR109B activity we observed in our original fraction screen.
Discussion
Phenylpropanoic acid, cadaverine, 9-10-methylenehexadecanoic acid, and 12-methyltetradecanoic acid add to a growing list of structurally simple molecules that are capable of modulating human signaling pathways that underlie diverse clinically relevant areas of physiology including immune recognition, neurotransmission, and inflammation.[35, 48] The biosynthetic simplicity of these metabolites combined with their abundant starting materials likely drives their high titers in the gut and potential broad biological relevance. Expanding functional screening to include not only more bacteria but also additional culture conditions and receptor families will undoubtedly provide additional insight into the biochemical mechanisms and small molecules underlying human-microbiome interactions. For example, a study that was published while the work reported here was under review examined a different collection of bacteria and identified different GPCR active metabolites.[49] Advancements in laboratory culturing techniques now allow for a majority of gut bacteria to be cultured from fecal samples.[50, 51] Systematic functional screening of metabolites produced by this growing collection of bacteria is likely to be a rewarding avenue for developing mechanistic hypotheses that can be tested in specific animal models.
Materials and Methods
Full description of the experimental procedures used can be found in Supplemental Material.
Author Contributions
D.A.C., L.J.C. and S.F.B. designed experiments. D.A.C., J.A.K. and P.M.L. performed analytical chemistry. S.M.H. and L.J.C. performed murine experiments. A.J.P., A.R. and J.R.C. designed, performed, and analyzed mass-spectrometry experiments. D.A.C. and S.F.B. wrote manuscript.
Competing Interests
S.F.B. is the founder of LODO Therapeutics.
Additional data table S1 (separate file)
Table S1 List of GPCRs tested with control compounds indicated
Table S2 Raw screening data for GPCR agonist screen with tested positive controls
Table S3 Raw screening data for Orphan GPCR agonist screen
Table S4 List of hit filtering by receptor type
Table S5 a. Raw cadA BLAST analysis of HMRGD. B. cadA PFAM analysis of HMRGD
Table S7 Raw targeted HRMS data
Acknowledgements
All bacterial strains were generously provided by Daniel Mucida. High-resolution mass spectrometry of purified compounds was performed by Rockefeller University Proteomics Core. We are grateful to C. Fermin, E. Vazquez, and G. Escano in the Precision Immunology Institute at the Icahn School of Medicine at Mount Sinai (PrIISM) Gnotobiotic facility and Microbiome Translational Center for their help with gnotobiotic experiments. Funding was provided by the Bill and Melinda Gates Foundation (OPP1168674) and the National Institutes of Health (5R01AT009562–02).
Footnotes
Supplemental files updated