Human gut microbiota stimulate defined innate immune responses that vary from phylum to strain

A Screening Platform for Measuring Innate Immune Responses to Human Gut Derived Bacteria

To gain insight into the immunogenicity of gut microbiota strains, we assembled a bacterial culture library consisting of 277 strains isolated from 8 healthy individuals, 6 individuals with Crohn’s disease (CD) and 3 individuals with ulcerative colitis (UC) (Figure 1A, Supplementary Table 3). We constructed the library to capture bacterial species that were representative of the four core phyla in the gut microbiota (Figure 1A), including multiple strains of the same species to measure variation at the closest taxonomic rank. We co-cultured BMDC and BMDM with either the filtered conditioned media of each bacterial strain grown for 48 hours or the live bacterial culture washed of metabolites, and we measured secreted TNFα, IL-6, IL-10, IL-1β, and IL-23 in the co-culture after 24 hours using a high-throughput ELISA workflow (Figure 1B). We selected these five analytes because of their genetic, biologic and/or therapeutic association with mucosal immunity and immune mediated diseases of the gut (Feagan et al., 2016; Hue et al., 2006; Kühn et al., 1993; Ligumsky et al., 1990; Mudter and Neurath, 2007; Targan et al., 1997; Yen et al., 2006). Our high-throughput innate screening platform was reproducible with an average R² of 0.75 across biological replicates for the five analytes considered (Pearson correlation; Figure S1A). The platform also captured a broad dynamic range of dose-dependent responses to purified lipopolysaccharide (LPS, from Escherichia coli O55:B5), a potent TLR4 agonist (Figure S1B).

To understand innate immune responses to gut derived microbes in the context of different myeloid cells, we used a 54-member subset of the 277-strain library to compare responses to live bacterial cultures (washed with PBS and resuspended in RPMI supplemented with cysteine) between BMDM and BMDC (Supplementary Table 2). We also measured responses to the conditioned media of each bacterial strain collected during stationary phase. TNFα secretion by BMDM and BMDC in response to both live cultures and conditioned media were significantly correlated (live cultures R² = 0.28, p = 9.30×10⁻⁶, conditioned media R² = 0.69, p < 1×10⁻¹⁰, Pearson correlation; Figures S1C, S1D), with a proportionally higher signal observed in BMDC. TNFα secretion in response to live bacterial cultures did not correlate with responses to conditioned media (Figures S1E, S1F). Importantly, the assay was robust to variation in bacterial growth density across strains in the library, as TNFα secretion only weakly correlated with the optical density (600_nm) of the bacterial culture (R² = 0.02, p = 0.02, Pearson correlation; Figure S1G).

To address concerns that bacterial viability might influence our measurements, we compared the immunostimulatory activity of heat-killed strains versus live bacterial cultures. We observed a strong correlation between responses to live and heat-killed bacteria (TNFα R² = 0.22, p = 0.0002, IL-6 R² = 0.44, p = 3.7×10⁻⁸, Pearson correlation; Figure S1H) with no statistically significant differences in the immunostimulatory capacity of live versus heat-killed strains across phylum (TNFα p =0.48, IL-6 p = 0.28, ANOVA), demonstrating the viability of bacterial strains has little, if any, influence on the patterns of cytokine induction described. Based on these collective initial metrics, we focused our analysis on BMDC responses towards live cultures of gut derived bacteria since this model produced a larger dynamic range across all analytes considered, was more sensitive and more reproducible, and elicited a robust IL-23 response (Supplementary Table 2).

Innate Cytokine Reponses to Gut Microbial Strains Are Taxonomically Enriched and Not Explained by Single TLR Agonists

BMDC responses towards live cultures of gut derived bacteria led to a broad range of cytokine profiles that significantly varied by several orders of magnitude across strains (TNFα ANOVA p < 1×10⁻¹⁰; Figure 1C) by several orders of magnitude. To place these responses in the context of known stimuli, we considered how innate cytokine responses to gut derived bacteria compared to responses towards 8 common purified TLR ligands. We calculated dose-response relationships with each of these individual TLR agonists, and selected concentrations which elicited the highest response or at which the signal saturated (Figure S1I). Stimulating BMDC with these common TLR ligands did not capture the range of responses observed upon stimulation with the commensal bacterial library (Figure 1C). We also compared gut-derived bacterial responses to three well-characterized enteric pathogens (Listeria monocytogenes, Salmonella enterica and Shigella flexneri). The cytokine responses to these pathogens largely fell in the upper part of the cytokine distributions (Figure 1C), suggesting that innate responses towards the majority of commensal isolates are on average lower than responses towards common enteric pathogens. However, all cytokine responses to these common pathogens were contained within the range observed across the 277-strain library (Figure 1C). Together, these results indicate that innate responses to gut derived bacterial strains can be as strong as responses towards pathogenic bacteria, span a tremendous dynamic range, and cannot be reproduced by individual TLR ligands. Given long-standing efforts to identify commensal strains that contribute to inflammatory responses in inflammatory bowel diseases (IBD) (Onderdonk et al., 1987; Rath et al., 1999, 1996; Seo et al., 2015), we tested if commensal strains isolated from patients with UC and CD elicited specific cytokine signatures. We found no enrichment for greater cytokine responses in IBD-associated bacterial strains compared to strains derived from healthy donors (Figure S1J) and there were no significant differences in the overall mean level of TNFα, IL-6, IL-10, IL-1β and IL-23 produced in response to healthy donor strains versus strains derived from UC or CD donors (Figure S1K).

To examine if certain bacterial taxa associate with the induction of particular cytokine signatures, we applied common clustering and dimensional reduction algorithms. K-means clustering of cytokine responses (k = 4; Figure S1L) segregated bacteria into four groups that broadly associated with phyla (Figure 1D). Firmicutes were enriched in a cluster encompassing the lowest cytokine responses across all analytes (Cluster 1 Firmicutes p = 1.51×10⁻⁴, Fisher’s exact test; Figure 1E), while Bacteroidetes were enriched in a cluster comprising a milder cytokine phenotype (Cluster 2 Bacteroidetes p = 8.92×10⁻⁶, Fisher’s exact test; Figure 1E). Proteobacteria were enriched in a cluster with the highest IL-1β and IL-23 induction (Cluster 4 p = 1.24 ×10⁻⁶, Fisher’s exact test; Figures 1E), and promoted greater secretion of IL-1β and IL-23 compared to the other phyla in the library (p < 0.01, ANOVA with Tukey correction; Figure S1M). Principal component analysis discriminated between a colinear TNFα/IL-10 axis and an orthogonal IL-23/IL-1β Proteobacteria driven response (Figure S1N).

We next looked for genera within the library that were enriched in the top extremes of the cytokine responses (using the 75^th quantile as a cutoff; Figure 1F). Focusing on the four genera with the highest number of strains per genus in the library (Bacteroides n = 67 strains, Bifidobacterium n = 37 strains, Enterococcus n = 23 strains and Escherichia n = 15 strains), we observed that Enterococcus members were enriched for IL-10 induction (Figure 1F) and induced on average more IL-10 than Bacteroides, Bifidobacterium and Escherichia (p < 10⁻⁵, ANOVA with Tukey post hoc test; Figure 1G). Members of Escherichia were enriched for strains in the upper quartile of IL-6, IL-23, and IL-1β induction (Figure 1F), and induced more IL-23 than Bacteroides, Bifidobacterium and Enterococcus (p < 10⁻⁵, ANOVA with Tukey post hoc test; Figure 1G), which is consistent with studies suggesting these bacteria can elicit Th17-biased immunity in vivo (Atarashi et al., 2015; Britton et al., 2020; Viladomiu et al., 2017). Bacteroides were enriched within strains that elicited high levels of IL-6 (Figure 1F), and IL-6 production in response to Bacteroides was higher than towards Bifidobacterium and Enterococcus (p < 0.05, ANOVA with Tukey post hoc test; Figure 1G). Together, these results demonstrate that innate immune responses towards gut microbial strains vary across taxa, with broad enrichments in cytokine response at the phylum level and more refined enrichments at the genus level.

Combinations of TLR Ligands Reflect Cytokine Responses Towards Gut-Derived Bacterial Stains

Although macrophages and dendritic cells express TLRs (Reis E Sousa, 2004; Visintin et al., 2001), it is unclear how the cell responds to the complex mixture of PAMPs on the surface of bacteria. To more accurately model a scenario in which a cell responds to a bacterial organism, we assembled from a set of eight TLR agonists all 255 possible combinations of 8 choose K (8 choose 1, 8 choose 2, etc…; Figure 2A). Responses following stimulation by these 255 different TLR ligand combinations increased on average with the value of K (TNFα ρ = 0.55, p < 1×10⁻¹⁰, IL-6 ρ = 0.52, p < 1×10⁻¹⁰, IL-10 ρ = 0.51, p < 1×10⁻¹⁰, IL-1β ρ = 0.34, p = 2.48×10⁻⁸, Spearman correlation; Figures 2A, 2B). We found that combinations of three TLR agonists (K = 3) elicited the largest range of responses (Figure S2A), and that as K increased the variance in responses decreased (Figure S2B). Overall, the response to these 255 different TLR ligand combinations could explain the vast majority of the dynamic range of BMDC responses to the 277 bacterial strains across all analytes (Figures 2C, S2C), providing an opportunity to use these synthetic TLR combinations to gain insights into how bacteria elicit particular cytokine responses.

Pairwise Synergies Between TLR Ligands Boost Cytokine Responses

We first assessed the individual impact of each TLR ligand (i.e., the main effects) on cytokine induction in the context BMDC responses to all 255 synthetic TLR ligand combinations. Unsurprisingly, LPS had the largest effect (TNFα p < 0.0001, Student’s t-test), followed by cytosine guanosine dinucleotide (CpG) (TNFα p < 0.0001, Student’s t-test) and Pam3CSK4 (TNFα p < 0.0001, Wilcoxon test; Figures 3A, S3A). To uncover synergy between TLR agonists and cytokine response, we used stepwise linear regression and elastic net regression to model pairwise interactions in our dataset. Both stepwise and elastic net regressions generated similar coefficient estimates (R² = 0.86, p = 1×10⁻¹⁰, Pearson correlation; Figure S3B). We found synergy between LPS and flagellin (Supplementary Table 4) in the production of TNFα (p = 0.019, Student’s t-test with BH correction), consistent with studies suggesting that TLR4 and TLR5 physically associate to trigger inflammatory responses (Hussain et al., 2020). LPS and CpG also acted synergistically to enhance the production of TNFα, IL-6, IL-10 and IL-23 (TNFα, IL-6, IL-10, IL-23 p < 0.0001, Wilcoxon test with BH correction, Figure 3B, Supplementary Table 4), which corroborates previous work demonstrating that TLR4 and TLR9 non-additively amplify cytokine signaling (De Nardo et al., 2009; Napolitani et al., 2005). Notably, LPS and CpG together reduced the level of IL-1β compared to LPS alone (p < 0.001, Wilcoxon test with BH correction; Figure 3B), perhaps via an IL-10 negative feedback mechanism (Abrams et al., 1991).

To uncover other nonadditive effects between TLR ligands, we segregated our data into conditions where LPS was present or absent. There was a larger dynamic range in response to ligand combinations which did not include LPS (Figure 3C), suggesting that 1) the presence of LPS compresses the dynamic range by increasing baseline signal and 2) innate immune responses can be exquisitely scaled depending on which ligands are present. Pam3CSK4 had a large individual effect when LPS was excluded (Figures 3D, S3C), implicating TLR2 as an important signaling potentiator in our model system. Also considering combinations without LPS, we found significant synergy between Pam3CSK4 and poly I:C (a TLR3 ligand) in the induction of TNFα, IL-6 and IL-10 (TNFα p < 0.05, IL-6 p < 0.001, IL-10 p < 0.01, Wilcoxon test with BH correction; Figure 3E). TLR2 and TLR9 crosstalk is thought to contribute to host defense against pathogenic bacteria and parasites (Bafica et al., 2006, 2005), and we found that double-stranded DNA from E. coli (a TLR9 ligand) and Pam3CSK4 together enhanced TNFα, IL-6, and IL-1β (TNFα p < 0.001, IL-6 p < 0.01, IL-1β p < 0.001, Wilcoxon test with BH correction; Figure 3F). Notably, CpG (also a TLR9 ligand) and Pam3CSK4 did not synergize, suggesting that ligand specific factors may be important in regulating the scale of the response (Iwasaki and Medzhitov, 2010; Long et al., 2009). Together, these data suggest that pairwise synergies between TLRs occur in the context of complex combinations of TLR agonists.

To gain insight into the types of TLR ligand combinations that correspond to particular groups of gut derived bacteria, we used supervised machine learning to classify the response of each TLR ligand combination using responses to the 277-member bacterial strain library as a training dataset. Specifically, we assigned each TLR ligand combination a membership to one of the four clusters that broadly associated with each phylum (Figure 1C) using a k-nearest neighbor algorithm (Figure 3G). We found that ligand combinations containing LPS were predominantly assigned into clusters 2, 3 and 4, which associated with responses to Bacteroidetes and Proteobacteria (Figures 3G, 3H). The complexity (K) of the TLR ligand combination also factored into assignments, as combinations of K = 2 and K = 3 were enriched in the Firmicutes associated cluster (i.e., cluster 1), while combinations of K = 4 and K = 5 were enriched in the Bacteroidetes associated cluster (i.e., cluster 2; Figure 3H). These analyses suggest that the absence of LPS with low-complexity TLR ligand combinations best replicate a response to Firmicutes, while the presence of LPS plus intermediate-complexity TLR ligand combinations broadly explain responses to Bacteroides. More generally, TLR ligand combinations can be engineered to replicate responses to specific bacterial taxa, suggesting that defined TLR stimulation can model aspects of innate immune responses to bacteria from the human gut microbiota.

Sensing of Gut Derived Bacteria by TLR2 and TLR4 is Taxonomically Defined

To determine the immune receptors necessary for the innate immunogenicity of gut derived bacteria, we subsampled the 277-member commensal library into a smaller 54-member library containing representatives from all four main gut phyla and all four functional clusters (Figure 1), and screened responses using BMDC derived from TLR knockout (KO) animals. We focused on the receptors TLR2 and TLR4 because the ligands for these receptors (Pam3SCK4 and LPS) exerted strong signal potentiation in the combinatorial TLR ligand experiments (Figure 3). First, we confirmed that wild-type BMDC responses to the subsampled 54-member library correlated with responses to the original 277-member library (Figure S4A). We found the production of cytokines in response to gut derived bacteria was completely abrogated in BMDC deficient in MyD88, the signaling adaptor protein used by many cell surface TLRs, including TLR2 and TLR4 (p < 0.0001, paired Student’s t-test; Figure S4B; Hou et al., 2008). We then explored innate immune responses to gut microbes using BMDC deficient in TLR2 and TLR4 (TLR2^KO or TLR4^KO; Wooten et al., 2002). To validate our KO system, we measured the dose-response to LPS and confirmed that cytokine levels diminished as expected (Figure S4C). We then examined responses to gut derived bacteria using TLR2^KO and TLR4^KO BMDC and compared these to responses by wildtype BMDC. Using hierarchical clustering, we grouped the fold-change per bacterial strain over all cytokines and over all KOs, and found that strains segregated by phyla (Figure 4A). Actinobacteria and Firmicutes were largely resistant to the loss of TLR2 or TLR4 (Figures 4B, S4D), while responses to Proteobacteria depended on TLR4 but not TLR2 (p < 0.001, Students t-test; Figures 4B, S4D). In contrast, IL-6 and IL-10 secretion in response to Bacteroidetes decreased with the loss of either TLR2 or TLR4 (IL-6 TLR2 p < 0.001, IL-6 TLR4 p < 0.001, ANOVA with Tukey’s correction; Figures 4B, S4D). Of note, we find in rare instances higher signal in response to Proteobacteria by TLR2^KO BMDC compared to wildtype BMDC (TNFα p < 0.01, Figure 4B), which may arise from compensatory signaling through other TLR receptors such as TLR4.

To gain further insight into the engagement of specific TLRs by gut derived bacteria, we assembled conditioned media from members of the full 277-member bacterial strain library, and assayed TLR2 and TLR4 stimulation using HEK-Blue reporter cell lines expressing murine TLR2 or TLR4. We found that Actinobacteria and Bacteroidetes broadly stimulated TLR2 (Figure 4C), while Proteobacteria selectively stimulated TLR4 (p < 0.0001, ANOVA with Tukey’s correction; Figure 4C). Remarkably, few bacteria in the collection stimulated both TLR2 and TLR4 (R² = 0.01, p = 0.13, Pearson correlation; Figure 4D).

We next compared responses to the genera most represented in the bacterial library using TLR2^KO and TLR4^KO BMDC. Clostridium and Enterococcus did not depend on either TLR2 or TLR4 (Figures 4C, S4E). Interestingly, Bifidobacterium stimulated TLR2 (Figure S4F), but IL-6 secretion in response to Bifidobacterium was not significantly affected by the absence of TLR2 (Figure S4G), suggesting redundancy in the sensing of this genus. In contrast, IL-6 and IL-10 responses to Bacteroides were strongly TLR2 dependent (IL-6 p < 0.001, IL-10 p < 0.01, ANOVA with Tukey correction; Figure S4G) and IL-6 responses to Bacteroides were more dependent on TLR2 than on TLR4 (IL-6 p < 0.05, ANOVA with Tukey correction; Figure S4G). Meanwhile, responses to Escherichia were lost in TLR4-deficient cells (IL-6 p < 0.01, TNFα p < .05, IL-10 p < 0.01, ANOVA with Tukey correction, Figure S4G), but were unaffected by TLR2 deficiency (Figure 4C). Taken all together, these data suggest there are phylum and genus-level distinctions in innate immune sensing of commensal microbes by TLR2 and TLR4.

We also considered whether innate immune recognition of bacteria was constrained by the expression of TLRs intrinsic to the C57Bl/6J mouse model. To address this question, we compared responses to bacterial strains using BMDC derived from the BALB/c background to responses from BMDC derived from the C57Bl/6J background. The differences in TLR expression between these two inbred mouse strains are well-characterized: cells from BALB/c mice express higher levels of TLR2, TLR4 and TLR5 and lower levels of TLR9 than cells from C57Bl/6 mice (Liu et al., 2002). Interestingly, we found a strong correlation between the responses of BMDC derived from C57Bl/6J and BALB/c mice (TNFα R² = 0.77, p < 1×10⁻¹⁰, IL-6 R² = 0.88, p < 1×10⁻¹⁰, Figure 4E), which suggests that the expression of TLRs is not a major driver of the specific innate immune responses to our bacterial strains.

Differences between gut derived bacterial strains drive variation in innate immunogenicity

Having identified broad taxa that associate with specific innate cytokine responses, we next examined more granular enrichments in innate immune potentiation. We focused our analyses on the gram-negative genera Bacteroides and Escherichia given their near ubiquity in the human gut microbiome, our data suggesting a differential reliance on TLR2 and TLR4 signaling, and previous studies showing non-uniform effects of B. ovatus and E. coli strains on the mucosal adaptive immune system (Atarashi et al., 2015; Britton et al., 2020; Viladomiu et al., 2017; Yang et al., 2020). We found that Bacteroides species exhibited a range of TLR2 stimulation (p = 0.007, ANOVA), with B. ovatus among those which most highly stimulated TLR2 (Figure 5A). Using a collection of 10 distinct B. ovatus strains, we found that IL-6 secretion in response to B. ovatus strains was indeed TLR2 dependent (p < 0.001, paired Student’s t-test with BH correction), and not dependent on TLR4 (Figure 5B). The loss of TLR2 also diminished TNFα secretion in response to B. ovatus strains more than the loss of TLR4 (p < 0.0001, paired Student’s t-test with BH correction; Figure 5B). We also measured cytokine induction by 10 unique species of Bacteroides. We observed significant variation in TNFα, IL-6, and IL-10 induced by different species of Bacteroides (TNFα p = 3.39×10⁻⁶, IL-6 p = 0.004, IL-10 p = 0.016, ANOVA; Figures 5C, S5A), which all together highlight species-level differences in innate immune recognition and potentiation.

We next examined differences between strains of the same species. Comparing cytokine induction by 10 strains of B. ovatus, we found the induction of TNFα, IL-6, and IL-10 varied significantly (TNFα p = 4.79×10⁻⁶, IL-6 p = 1.98×10⁻¹⁰, IL-10 p = 8.33×10⁻⁵, ANOVA; Figures 5D, S5B). In addition, the induction of TNFα, IL-23, IL-6, IL-10 and IL-1β varied significantly across 14 distinct strains of E. coli (TNFα, IL-23, IL-6, and IL-10 p < 1×10⁻¹⁰, IL-1β p = 2.33×10⁻¹⁰, ANOVA; Figures 5E, S5C), which demonstrates how different strains within the same species of bacteria can drive varied innate immune responses.

We also considered whether differences between gut derived bacterial strains could associate with in vivo immune phenotypes. B. ovatus is a known inducer of fecal IgA, and for reasons unknown, some strains of B. ovatus are capable of inducing higher IgA than other B. ovatus strains (Yang et al., 2020). We found that strain-level in vitro TNFα responses to B. ovatus correlated with in vivo fecal IgA levels in gnotobiotic mice monocolonized with the corresponding B. ovatus strain (R² = 0.43, p = 0.047, Pearson correlation; Figure 5F), suggesting that differences in the immunogenicity across strains measured in vitro can associate with immune phenotypes in vivo. Of note, we did not find a significant correlation between IL-6 in response to B. ovatus and fecal IgA levels (Figure S5D) and the association between TNFa and IgA is likely not causative but representative of the differential interactions that occur between distinct bacterial strains and the host immune system.

One of the most striking examples of inter-strain variation that we identify is the varied induction of IL-23 across strains of E. coli (Figure 5E). Some strains of E. coli are known to induce Th17 cells in vivo (Britton et al., 2020; Viladomiu et al., 2017) and we wondered if genomic sequence similarity between strains of E. coli could explain differences in the capacity to elicit IL-23 in vitro. We found no correlation between whole genome similarity between pairs of E. coli strains and their differential ability to induce IL-23 (R² = 0.012, p = 0.23, Pearson correlation; Figure S5E). However, when we looked at the genomes within sub-species of E. coli with >65% similarity (Figure S5E), we found that the pairwise genomic similarity over the whole genome negatively correlated with the difference in IL-23 induction (right R² = 0.25, p = 0.007, Pearson correlation; Figure 5G), suggesting that evolution may explain variation in IL-23 induction among closely related strains but does not explain variation between strains that share less sequence homology. To increase the resolution of our approach, we examined the relationship between defined genetic loci on the E. coli chromosome and the capacity to elicit IL-23 from myeloid cells (Figures S5F and S5G, Extended Data). Using comparative genomic approaches, we found that variation at three specific genomic loci could explain the differences in the capacity for E. coli strains to induce IL-23 (FunGC 43 ρ = 0.37, p = 2.9×10⁻⁸; FunGC 52 ρ = 0.28, p < 1×10⁻¹⁰; FunGC 56 ρ = 0.43, p < 1×10⁻¹⁰; Spearman correlation; Figures S5G and S5H, Supplementary Table 5). One of these significant genomic locations is associated with membrane transport of bacterial peptidoglycans (FunGC 56 ρ = 0.43, p < 1×10⁻¹⁰; Spearman correlation Figure S5H).

A subset of the bacterial strains present in our library originated from stool that was later transferred to human recipients with recurrent Clostridioides difficile infection as fecal microbiota transplantation (FMT) therapy (Aggarwala et al., 2021). To determine if the innate immunogenicity of strains changes across human transmission and over time, we measured innate immune responses to strains isolated from the donor microbiota before transplant and compared these to responses to the identical strain isolated from the microbiota following transplantation into a recipient human host. We surveyed 11 strains from 7 different species and found no differences in the mean TNFα and mean IL-6 induced by strains isolated from the original donor compared to strains isolated from the recipient following FMT (TNFα p = 0.92, IL-6 p = 0.58, paired t-test; Figure 5H). One FMT recipient was sampled longitudinally and we found that strains re-isolated 11 months post-FMT from this donor elicited comparable TNFα and IL-6 induction compared to isolates from the original donor (FMT timepoint 2, TNFα p = 0.85, IL-6 p = 0.76, paired t-test; Figure 5H), which suggests that innate immunogenicity of human gut strains is stably maintained across human transmission and across time.

Microbiota-driven TLR2 stimulation drives colonic regulatory T cell induction in vivo

In vitro platforms provide the scale to identify patterns that can be validated and explored further in more complex models. We wondered whether the taxonomically-based patterns in innate immune recognition that we observe in our in vitro screens could explain immune responses in vivo. First, we confirmed that the cytokine profiles elicited by bacterial strains in vitro were broadly reflective of the cytokine profiles observed in vivo. Using data from our in vitro screen (Figure 1), we colonized mice with mixtures of strains that drove differential TNFα and IL-6 induction in vitro (TNFα p = 0.03, IL-6 p = 0.02, Student’s t-test, Figure S6A). We found significantly higher TNFα production within colonic cDC2 and CD11b⁺ CD103⁺ dendritic cells in mice colonized with the consortia containing ‘high inflammatory’ strains derived from Cluster 1 and Cluster 4 (p= 0.029, Wilcoxon test, Figures S6B and S6C). Collectively, these data demonstrate that strains with diametric immune phenotypes in vitro elicit broadly translatable phenotypes in vivo.

We next asked whether distinctions in innate immune sensing of commensal microbes by either TLR2 or TLR4 associates with specific immunity in vivo. Given that B. ovatus is sensed by TLR2 (Figures 5A and 6A) and that innate cytokine production in response to B. ovatus is dependent upon TLR2 signaling (Figure 5B), we reasoned B. ovatus may drive TLR2-mediated immunity in vivo. Blockade of TLR2 in vivo is sufficient to diminish the number of colonic IL-10 producing lymphocytes (Mishima et al., 2019). Furthermore, the induction of FoxP3⁺ Treg cells is diminished in TLR2^KO mice (Netea et al., 2004) and augmented by the application of TLR2 agonists (Round and Mazmanian, 2010). We therefore hypothesized that B. ovatus would be a potent inducer of intestinal FoxP3⁺ Treg cells. We first confirmed that strains of B. ovatus specifically stimulate TLR2 (compared to E. coli p = 1.22×10⁻⁴, Student’s t-test; Figure 6A). We then compared the levels of FoxP3⁺ Treg cells in the colonic lamina propria of mice monocolonized with strains of B. ovatus to mice monocolonized with strains of E. coli, which is sensed by TLR4 and not TLR2 (Figure 6A). We found that all B. ovatus strains potently induced FoxP3⁺ Treg cells in the colon, and that the levels of FoxP3⁺ Treg cells were significantly higher in mice monocolonized with B. ovatus than in mice monocolonized with E. coli (p = 1.67×10⁻³, Student’s t-test; Figures 6B and S6D). Furthermore, the levels of colonic RORγt⁺ Treg cells were also higher in mice monocolonized with B. ovatus compared to mice monocolonized with E. coli (p = 3.32×10⁻⁵, Student’s t-test; Figure 6B), which together suggests that species level differences in TLR2 and TLR4 recognition associate with alterations in colonic FoxP3⁺ Treg cell induction.

To determine if supplementation of a host harboring a low TLR2-inducing E. coli strain with a high TLR2-inducing strain could boost colonic Treg induction, we co-colonized mice with a strain of E. coli and a strain of B. ovatus. We found that colonization with both B. ovatus and E. coli increases the proportion of FoxP3⁺ and RORγt⁺ FoxP3⁺ cells in the colon to similar levels observed following colonization with a a high TLR2-inducing B. ovatus strain (FoxP3⁺ p =0.06 and RORγt⁺ FoxP3⁺ p =0.001, ANOVA with Tukey correction, Figure 6C). To specifically test the importance of TLR2 stimulation, we administered Pam3CSK4 (a selective synthetic TLR2 agonist) to animals colonized with a ‘low-Treg’ inducing strain of E. coli alone and found it was sufficient to boost the levels of colonic FoxP3⁺ Treg cells (E. coli + Pam3CSK4 versus E. coli p =0.02, ANOVA with Tukey correction, Figure 6C). Of note, administration of Pam3CSK4 to germ-free animals did not boost Treg induction, suggesting a microbial component in addition to direct TLR2 agonism may be required for Treg induction (Figure 6D). We found that blocking TLR2 signaling in vivo with a monoclonal antibody in animals colonized with B. ovatus significantly reduced the proportion of RORγt⁺ FoxP3⁺ Tregs induced in the colon (p 0.03, ANOVA with Tukey correction, Figure 6C). Together these data suggest that microbiota-driven TLR2 stimulation can drive colonic Treg induction.

To extend the role of TLR2 signaling in FoxP3⁺ Treg induction across more bacterial strains, we compared FoxP3⁺ Treg induction in mice monocolonized with 5 additional strains of Bacteroidetes spanning the TLR2 stimulation distribution (Figure 6D). We found these strains differentially induced the proportion of FoxP3⁺ Tregs and RORγt⁺ FoxP3⁺ Tregs in the colonic lamina propria (p = 0.0006 and p < 1×10⁻¹⁰ respectively, ANOVA, Figure 6E). We compared the induction of FoxP3⁺ Treg in vivo with in vitro TLR2-stimulating activity. No ‘TLR2 high’ species induced low levels of colonic FoxP3⁺ Tregs (Figure 6F). Of note, although B. fragilis was the lowest TLR2 stimulator in vitro, it induced a modest proportion of Tregs (Figure 6E), which is not surprising given that polysaccharide A on the surface of B. fragilis is known to induce colonic Tregs (Mazmanian et al., 2008). B. massiliensis also elicited high colonic FoxP3⁺ Treg induction despite being a ‘TLR2 low’ species. These data collectively suggest that the degree of TLR2 stimulation by species of Bacteroidetes is one factor controlling the induction of FoxP3⁺ Tregs but other mechanisms can play an important role.

RESOURCE AVAILABILITY

EXPERIMENTAL MODEL AND SUBJECT DETAILS

METHOD DETAILS

QUANTIFICATION AND STATISTICAL ANALYSIS

Analyses throughout this study were performed using RStudio v1.1.463. All p values less than 1×10⁻¹⁰ were rounded. P-values, the tests used and the depictions of error estimates are reported in the figure legends. Heatmaps were constructed using either ComplexHeatmap (Gu et al., 2016) or ggplot2. For plots depicting fold-change of knockout responses relative to wildtype, the bottom 3% of responses per analyte were thresholded for visualization purposes. K-nearest neighbor classification was performed using the R package class (k=15). Stepwise regression (using AIC minimization) was applied using R package MASS. Procedures for fitting elastic net regression (λ selected using 10-fold cross validation) were performed with the R package glmnet. The multiple linear regression for modelling responses to TLR ligand combinations including interaction terms is:

Where x₁ = TLR ligand 1, x₂ = TLR ligand 2 …, x_i = TLR ligand i, x_j = TLR ligand j, n = number of parameters in the model including the intercept, main effects and interaction terms.