At all phylogenetic levels, from phylum to genus, we identified four groups of body habitats that maintain a distinct pattern of numerically dominant bacterial taxa as profiled using the 16S rRNA gene (Figure 1a), as classified by the Ribosome Database Project (RDP) 29 . While only two phyla, the Firmicutes and Bacteroidetes, dominated the communities of all ten sites, their proportions and that of nearly all taxa in the sampled body habitats form groups as follows: Group 1, buccal mucosa, keratinized gingiva, and hard palate; Group 2, saliva, tongue, tonsils, and throat (back wall of oropharynx); Group 3, sub- and supra-gingival plaque; and Group 4, stool. The microbiota of Group 1 consisted mostly of Firmicutes followed in decreasing order of relative abundance by Proteobacteria, Bacteroidetes and either Actinobacteria or Fusobacteria (Figure 1a; Additional file 1). In comparison, Group 2 had a decreased relative abundance of Firmicutes and increased levels of four phyla: Bacteroidetes, Fusobacteria, Actinobacteria and TM7. Group 3, which consisted of both tooth plaque sites, had a further decrease in Firmicutes compared to Groups 1 and 2, with a marked increase in the relative abundance of Actinobacteria. Finally, the stool microbiota (Group 4) consisted of mostly Bacteroidetes (over 60%) followed by Firmicutes, with very low relative abundances of Proteobacteria and Actinobacteria, and less than 0.01% of Fusobacteria (Additional file 1).

These dramatic differences occurred consistently throughout the cohort, with closely juxtaposed body sites (for example, tongue dorsum (Group 2) and hard palate (Group 1)) possessing strikingly different microbial community structure even when considering the phylum level alone and independently of the structure of the tissue (Additional file 2). This supports strong local selective pressure on community membership even in the absence of disease, and these differences reach to the genus level (Figure 1a). In terms of genera, Group 1 was characterized by a very high relative abundance of Streptococcus, while Group 4 contained predominantly Bacteroides. In contrast, Groups 2 and 3, rather than having a single genus present at such high relative abundance, were characterized by a more even distribution of the most abundant genera. Streptococcus, Veillonella, Prevotella, Neisseria, Fusobacterium, Actinomyces and Leptotrichia were each present over 2% on average in Group 2. These seven genera plus Corynebacterium, Capnocytophaga, Rothia and Porphyromonas accounted for genera present at more than 2% in Group 3 (Figure 1a; Additional file 1).

Examining clade abundances at all taxonomic levels, we used the LEfSe (LDA Effect Size) system for biomarker discovery 30 to determine statistically significant biomarkers among these four groups within the digestive tract. These included both high and low abundance clades that significantly and consistently varied in abundance among and within body habitats, for example, in the three oral groups (Figure 1b; Additional file 3). For example, both the phylum Actinobacteria and individual taxa within the Actinomycetales were consistently more abundant on the tooth surfaces in Group 3 (Figure 1b; Additional file 3). When comparing Group 1 against the other three groups (a slightly more stringent setting than comparing all groups against each other as in Figure 1b and Additional file 1) two genera from the Firmicutes were identified as genus-level biomarkers: Streptococcus, from the Streptococcaceae (mean 47 ± 18% abundance in Group 1), and Gemella, from the Staphylococcaceae (mean 5.2 ± 5.1% abundance in Group 1) (Additional file 1). Although the Firmicutes phylum as a whole was most differentially abundant in Group 1, more specific taxa within the Firmicutes were detected as biomarkers for Groups 2 and 4 (Figure 1b; Additional file 3). For example, in Group 2, biomarkers, when compared to the other three groups, included Oribacterium and Catonella, members of the Lachnospiraceae, and Veillonella, a member of the Veillonellaceae (all Clostridia). The abundances of Veillonella and Prevotella overall were comparable in Group 2 (10.2 ± 5.4% versus 11.6 ± 7.3%, respectively), and both were identified as differentially abundant in this group. The other genus-level biomarkers for Group 2 detected at >2% were Porphyromonas (3.8 ± 4.2%) and Neisseria (6.6 ± 7.6%) (Additional file 1). Several genus-level biomarkers for Group 4 (stool) were also Firmicutes, mostly from the families Lachnospiraceae and Ruminococcaceae (Figure 1b; Additional file 3). These results support the overall consistency of the different microbial populations characterizing each of the four groups, and they also emphasize the need to take multiple levels of phylogenetic specificity into account when performing any analysis of the microbiome. Phylum relative abundances differentiated very distinct body habitats. As additionaly discussed below, these differences were reflected at the genus level within each body site in the healthy adult human.

The four observed groups differed significantly not only based on their specific microbial compositions, but also by several ecological summary statistics. Most strikingly, after comparing every pair of samples using the Bray-Curtis measure of beta diversity 31 , within-group distance was very significantly lower (greater similarity) than between-group distance (lower similarity) for all four groups (Additional file 4; Table 1; all P < 10^-20). The coarse level of species richness measurement offered by phylotype data did not distinguish strongly among any body habitats, but evenness and the resulting within-community alpha diversity ranged widely among groups as measured by the inverse Simpson index 32 (Additional file 5). For example, the Group 1 body sites together averaged below a relative diversity of 5.3, Group 2 ranged from 7.3 ± 3.0 (tonsils) to 10.6 ± 3.1 (saliva), the plaques in Group 3 had average diversities of 9.6 ± 3.1 and 9.8 ± 3.0, and Group 4 (stool) declined to a mean of 4.6 ± 2.9. The lower diversities in Group 1 are largely an effect of Streptococcus abundance, and likewise the gut microbiota's diversity is lowered by the prevalence of the Bacteroides in these data (both detailed above and below). These differences are highly statistically significant (for example, Group 1 versus 2 P < 1e-50 by t-test) and provide evidence in support of the four-group distinction at the levels of both individual bacterial clade and overall ecological structure.

Bacterial phyla originally thought to be exclusively environmental have recently been observed to possess human host-associated membership 33 34 35 36 . This phenomenon was widely observed within this normal population. The phylum TM7 was highly prevalent, detected in at least one sampling site of the upper digestive tract of 85% of subjects and in the stool of 13.6% of the subjects (Additional file 6). The phyla SR1 and Synergistetes were present in at least one upper digestive tract site of 65.3% and 58.5% of the subjects and in the stool of 1.4% and 8.8% of the subjects, respectively. The phylum Verrucomicrobia, represented mainly by the genus Akkermansia 35 , and the phylum Lentisphaerae, represented by the genus Victivallis 34 , were present in the lower digestive tract of 41.5% and 15.0% of the subjects and in the upper digestive tract of 13.6% and 1.4% of the subjects. TM7 bacteria accounted for a mean of 3.1 ± 5.7% of the saliva population and 0.6 ± 1.2% of the bacteria found in plaque communities (Figures 1a and 2; Additional file 1). The SR1 phylum was also most abundant in saliva (mean 0.4 ± 1.2%), and both TM7 and SR1 phyla were found at trace amounts in stool. While these phyla were varyingly prevalent (Figure 2), they occurred near-uniformly at low but significantly non-zero abundances, which highlights their lack of detection in smaller studies without deep high-throughput sequencing.

Clades populated with known bacterial oral pathogens were well represented in this reference adult cohort, typically with moderate to high population penetrance but low relative abundance in each individual. Among the Spirochetes, Treponema species are associated with periodontal and endodontic diseases 37 38 and were present in at least one of the upper digestive tract sites of 96% of this disease-free population (and in all the nine oral sites of 6.8%). Treponema had a variable relative abundance among the oral body habitats, with highest representation in the subgingival biofilm (mean 2.2 ± 4.1%) and non-zero abundances in several other sites, for example, palatine tonsils (Figure 3; Additional file 1). In contrast, a minority of stool samples (3.4%) contained trace levels of Treponema. The previously published rarity and specificity of Brachyspira to the gut was confirmed by its detectable presence in only one stool sample (226 stool samples in total; Additional file 7) and absence from all the upper digestive tract sites (1,879 samples; Additional file 7). Other periodontal pathogens were lower in abundance. Aggregatibacter were found mostly along the tooth surfaces (Group 3; mean 0.4 ± 0.7/0.8% from supra- and sub-gingival biofilms), and Megasphaera were found mostly in Group 2 (from mean 0.4 ± 0.6% in the tonsils and tongue dorsum to 0.8 ± 0.9% in saliva). Bifidobacteriaceae, implicated in the formation of caries 39 40 , were very rare at all oral sites (means <0.03%), but possessed high prevalence (40.8%). In the stool, the genus Bifidobacterium was most represented with a low mean relative abundance of 0.08 ± 0.3%. The low abundance of Bifidobacteriaceae in the oral cavity may be a reflection of the lack of carious lesions in this healthy subject population. Porphyromonas , which includes Porphyromonas gingivalis (one of the most studied oral pathogens) and non-pathogeneic strains, was present in the upper digestive tract of all the subjects (mean 3.0 ± 3.8%, 3.8 ± 4.2%, and 3.0 ± 3.5% in the three oral groups, respectively) and in 25% of the lower digestive tract samples, though in very low abundance in the stool (Additional file 1). Tannerella, thought to incur similar host phenotypes, was present in the upper digestive tract of 97.3% and in the stool of 3.4% of the subjects. Both genera, Porphyromonas and Tannerella, were almost uniquely distributed in average abundance among individual body sites within the oral cavity, whereas the other relevant genera in the family Porphyromonadaceae (Parabacteroides , Barnesiella , Odoribacter, and Butyricimonas) predominantly colonize the stool (Figure 4).

Genera that include important human pathogens colonizing the throat/tonsils - Streptococcus pneumoniae, Streptococcus pyogenes, Neisseria meningitidis, and Haemophilus influenzae - were all well represented in the microbiota of the upper digestive tract sites (Figure 1a; Additional file 6). The known difficulty of performing species-level identification from 16S rRNA pyrosequencing experiments 41 precluded the determination of prevalence for these specific pathogens in this cohort. The genus Moraxella, which includes the common sinus pathogen Moraxella catarrhalis, was detected in the upper digestive tract microbiota at low relative abundance, reaching a mean >0.5% only in the throat (Additional file 1). Interestingly, the high standard deviation (4.7%) of the relative abundances of Moraxella in the throat suggested variable colonization within this population.

In the lower intestinal tract, genera containing known pathogens were typically low in both prevalence and relative abundance. Helicobacter, implicated in gastrointestinal diseases, appeared in only 1.4% of stool samples in trace amounts while studies of Helicobacter pylori stool antigen prevalence in healthy European adults ranged up to 33% 42 . Enterobacteriaceae abundances were uneven among individuals in the gut and within each individual among body sites, with the most abundant genus being the Escherichia/Shigella complex (mean 0.1 ± 0.67%), which was detected in 33% of stool samples. Finally, Faecalibacterium, a genus of considerable interest due to its observed decrease in abundance in active Crohn's disease 43 44 45 46 , was highly represented in the stool (98% of subjects and mean 4.6 ± 5.2%) but present only at trace levels in the oral cavity (always below 0.05%), suggesting that it may be adapted to a very specific niche within the gut.

Within the oral cavity, the Group 3 sub- and supra-gingival plaque bacterial communities were most distinct and differed strongly from the other body sites, but further differences characterized each of these two sites individually. The tooth surface adjacent to the soft gingival tissues specifically comprises two distinct ecological niches, supragingival, and subgingival (Additional file 8). The supragingival region sits above the gingival margin, exposed to the oral cavity, bathed in saliva and exposed to ingested substances; the subgingival region is bathed in a serum transudate that flows from the base of the crevice outward to the oral cavity. A key known physiological difference between these two regions is the lower redox potential found subgingivally 47 . Correspondingly, we observed differences in the non-diseased plaque biofilm communities from these two regions distinguished by proportional shifts consistent with these physiological distinctions (Figure 5a; Additional file 1). Shifts at the phylum level were driven by subgingival increases in the obligate anaerobic genera Fusobacterium, Prevotella, and Treponema, and by lesser relative abundances of Dialister, Eubacterium, Selenomonas, and Parvimonas. In contrast, groups significantly increased in the supragingival plaque consisted predominantly of facultative anaerobic genera, including Streptococcus, Capnocytophaga, Neisseria, Haemophilus, Leptotrichia, Actinomyces, Rothia, Corynebacterium, and Kingella (Figure 1a; Additional file 1). This suggests that along these tooth surfaces, where direct bacterial interaction with host cells is diminished, oxygen availability - an environmental factor - may be a major driver of community composition.

The pharynx is the site of carriage of a number of important bacterial pathogens that impact both healthy and immunocompromised individuals. LEfSe analysis of all samples did not identify any genus-level organisms characteristic of the microbiome of the throat and/or tonsils consistently present above our limit of detection. For example, when throat and tonsil samples were compared to the mouth sites, the genera Butyrivibrio and Mogibacterium (both from the phylum Firmicutes) were identified as differentially abundant, but both were present at only low levels (mean 0.057 ± 0.09%, and 0.188 ± 0.316%, respectively, corresponding to only approximately 1 to 5 sequences/sample; Additional file 1). The palatine tonsils, located in the oropharynx, are unique among the sites sampled in this study as the only lymphoid tissue. However, the genus-level tonsil-specific biomarker when compared to the mouth, Peptococcus, was again present at very low relative abundance (mean 0.049 ± 0.079%; Additional file 1). This lack of throat- or tonsil-specific biomarkers among bacterial taxa with a relative abundance >1% likely reflects the similarity of the microbiome of these two oropharyngeal sites with those of the tongue dorsum and saliva (Group 2 in Figure 1) despite their differences in tissue type (Additional file 2). This observation is supported by the comparison of the complete Group 2 with all other groups, which revealed distinct and abundant biomarkers as discussed above (Figure 1b; Additional file 3). Microbiota composition and the path of swallowed saliva suggest a potential role of saliva as one of the host factors influencing microbiota of Group 2.

After analyzing the microbiota of body habitats within the digestive tract, we next asked if there were bacteria whose differential abundance characterized the digestive tract as a whole. The non-mucosal sites sampled in the HMP included anterior nares, both post-auricular creases (crease behind the ear), and both antecubital fossae (inner elbow crease). Propionibacterium , Staphylococcus , and Pseudomonas were identified as biomarkers for the non-mucosal sites, based on a LEfSe analysis of all ten digestive tract sites versus the non-mucosal sites (Figure 5b). However, no genus-level biomarkers were identified as uniquely present throughout the digestive tract microbiota. The unclassified Veillonellaceae and Porphyromonadaceae (Figure 5b) are unlikely to be true biomarkers due to their low representation. Further analysis was impaired by the lack of reference sequences for them within RDP. Members of Veillonellaceae and Porphyromonadaceae families were much less abundant at non-mucosal sites, and were essentially absent from the HMP vaginal samples, suggesting that their adaptation is to the digestive tract mucosa rather than mucosal surfaces in general .

Bacterial genera membership overlap in the same subject between oral and stool samples was limited when considered if present in at least 45% of the subjects. It included Bacteroides, Faecalibacterium, Parabacteroides, Eubacterium, Alistipes, Dialister, Streptococcus, Prevotella, Roseburia, Coprococcus, Veillonella, Oscilibacter, and yet-to-be-classified genera from a subset of families (Additional file 6). Interestingly, the presence of genera in a large portion of subjects was not related to a stable relative abundance in the microbial communities, as Bacteroides and Dialaster were among the four most variable genera among subjects. In contrast, Prevotella, Veillonella and Streptococcus were the genera with the most consistent presence in the subject population (Figure 3). The importance of Lachnospiraceae, Veillonellaceae, and Porphyromonadaceae families in the healthy digestive tract microbiome was indicated by their relative abundance among all body habitats and among subjects (Figures 1 and 4; Additional files 1 and 6). Bacteria of the Lachnospiraceae and Veillonellaceae families specifically were present in all subjects' oral cavities and stools (Additional file 6). Porphyromonadaceae were present in the oral cavity of all subjects and the stool of 95.9% of subjects (Additional file 6), although their relative abundance of member genera varied by body habitat (Figure 4). Porphyromonas was present primarily in the oral sites, while Parabacteroides, Barnesiella, Odoribacter and Butyricinomonas were predominant in the stool (Figure 4). The significance of this variation in genus distribution was confirmed by LEfSe conmparisons of the upper (oral) and lower (stool) digestive tract sites (Additional file 3). In contrast, Tannerella (Figure 4) was present in most oral sites, but due to a lower relative abundance specifically in the keratinized gingiva, it was not found to be statistically signicant between the oral and gut sites. The pattern of variable genus distribution between the upper and lower parts of the digestive tract holds for the Lachnospiraceae and Veillonellaceae as well (Figure 4), again suggesting a pattern of niche specialization among human body habitats extending from the bacterial family level down to specific genera.

In addition to relative abundances of bacterial organisms based on 16S rRNA genes, we examined the abundances of microbial metabolic pathways as profiled from metagenomic shotgun sequencing of a subset of the available body habitats 48 . These data from the HMP included one body site within each of the four digestive tract groups: the buccal mucosa (Group 1), the tongue dorsum (Group 2), the supragingival plaque (Group 3) and the stool (Group 4). The data analyzed below include the relative abundances of individual enzyme families (Kyoto Encyclopedia of Genes and Genomes (KEGG) Orthologous groups (KOs) 49 ) and of complete metabolic modules (KMods) (Figure 6; Additional file 9).

Bacterial cells use a wide variety of aerobic or anaerobic degradation pathways as energy sources, and this was most evident in the differences in relative abundance of specific sugar transporters when comparing the oral sites to the gut. PTS transporters for small sugars were most abundant in the oral cavity and were represented for monosaccharides by mannose (M00276) and fructose (M00304) transporters, as well as the transporter of galactosamine (M00287), derived from the breakdown of sugar-decorated glycoproteins. The supragingival plaque microbiome was enriched for threhalose (M00270, M00204), alpha-glucosides (M00201, M00200), and cellobiose (M00206) transport; in contrast, the stool microbiome was enriched for the transport of lactose/arabinose (M00199) and oligogalacturonide (M00202), and for the degradation of the larger dermatan (M00076), chondroitin (M00077) and heparin (M00078) sulfate polysaccharides. Surprisingly, while anaerobiosis-related pathways were expected throughout the digestive tract, putrescine transporters in particular were most represented in the oral cavity (M00193, M00300). This is of potential interest as concurrent production and import of putrescine is a delicate balance, and excess putrescine is one source of halitosis 50 .

Consistent with what is known about the function of the colonic gut bacteria, we observed several prominent enzymes and metabolic pathways most abundantly in the stool metagenome. For instance, β-glucosidase (K05349) was specifically abundant in the gut microbiota and not at oral sites; this enzyme is critical in the pathway of cellulose breakdown to β-D-glucose. Concomitantly, given that the Embden-Meyerhoff pathway is also known to be the major route for glucose metabolism to pyruvate in the colon, the highly associated glycolysis pathway module (M00001) was also significantly enriched in stool 51 . This finding is further in agreement with the 16S rRNA gene sequencing data, which included prevalent Ruminococcus in stool that are important colonizers of plant-derived material in the gut and possess cellulolytic activity 52 . The stool bacteria were also uniquely associated with pathways for ammonia (M00028, urea cycle, and M00029, ornithine biosynthesis) and methane (M00356 and M00357, both methanogenesis) production; the prominence of these enzymes is consonant with the colonic microbiome as a significant source of ammonia production. In fact, targeting the colonic microbiome with antibiotics has been shown to be a successful therapy in acquired diseases of hyperammonemia such as encephalopathy complicating hepatic insufficiency 53 . Relatedly, compared to upper digestive tract sites, there was very high abundance of a specific multiple antibiotic resistance protein (K05595) and association with the pyruvate:ferredoxin oxidoreductase pathway, which, due to its role in converson of metronidazole to its active nitroso form, can also determine sensitivity to this antibiotic. These potential pathogenically linked behaviors are of course in addition to the expected colonic bacterial activities detected for producing energy from undigested cellulose, nitrogen-containing compounds, and vitamins and cofactors important in support of basic metabolic pathways.

Although HMP protocols were optimized for bacterial sequences, shotgun sequencing also provides an initial means of assessing the community structure of non-16S assayable microbes. As reported in Additional file 10, the fractions of Archaea (0.04% in the stool; below the detection threshold in the oral cavity) and small eukaryotes (0.34% in the buccal mucosal; <0.1% in the other body sites) detected here proved to be very small. Although this may be due in part to the HMP's specific DNA handling protocols 10 , this suggests that 16S rRNA gene-based community surveys provide an accurate overview of these digestive-tract associated microbial communities. Likewise, ribosomal and shotgun sequencing in the HMP cohort have been compared elsewhere and provide consistent quantitative estimates of genus-level abundances 54 without systematic phylum-level biases.

A subset of the HMP's microbiome samples was assayed with both shotgun metagenomic and 16S rRNA gene sequencing. This allowed us to assess co-variation of microbial abundances with inferred metabolic pathways. Strong correlations between the abundances of bacterial clades (from 16S rRNA data) and gene or pathway abundances (from metagenomic data) in some cases clearly highlighted genes carried by these organisms, and in others denoted less clear pangenomic elements or metabolic dependencies. An example of the former was the arabinofuranosyltransferase genes aftA and aftB (K13686 and K13687). These genes were only present in the tooth surface habitats and are known to be encoded by Corynebacterium, a biomarker of the plaques as discussed above. This was confirmed by the genes' strong association with Corynebacterium clades in these data (Spearman correlation 0.76, P-value <1e-15; Additional file 11). Archaea were not included in our analysis, as these were not detectable by 16S rRNA gene sequencing (due to lack of the conserved sequence in the primers used) and were poorly represented in the shotgun sequencing data.

The acquisition and export of metals for bacterial homeostasis and for pathogenicity is ubiquitous throughout the human microbiota, with iron being most generally necessary. Iron transporters were widely distributed among the microbiota of all four body sites, but again, the specific mechanisms of iron uptake and sequestration differed as needed for niche specialization. One use of the iron is its incorporation in porphyrin, and there was a wide distribution of cytochrome c heme lyase (K01764), which appeared to be ubiquitous and was not strongly associated with individual organisms. Conversely, uroporphyrinogen synthase (K01719) occurred at higher relative abundance in stool, inversely associated with members of the Clostridiales (Spearman correlation -0.79, P-value <1e-15; Additional file 11). This can be contrasted to protoporphyrinogen oxidase (K00231) in the oral cavity, which is potentially linked to the Prevotella enrichment (Spearman correlation 0.71, P-value <1e-15). Within the oral cavity specifically, coproporphyrinogen oxidase (K00228) and protoporphyrinogen oxidase (K00231) were both more abundant on the tongue and in supragingival plaque than on the buccal mucosa, expected to be linked to the increased relative abundance of Porphyromonas and Prevotella on those surfaces 55 (Figure 1; Additional file 11).

Metal export and utilization were likewise ubiquitous throughout the microbiota, but differed in the prevalence of specific mechanisms. Most genes encoding exporters needed for heme tolerance 56 , such as MtrCDE (K00579, K00580) and HrtAB (K09814, K09813), were present at low levels throughout the digestive tract, although MtrCDE was somewhat enriched in the more anaerobic habitats, stool and plaques. None were significantly associated with specific organisms in these data. The gene encoding hemerithryn (K07216) was detected at multiple body sites but was highly enriched in stool. This enzyme for iron utilization is most often found in members of the Methylococcaceae family 57 , but these were again not detectable in this study due to their absence from the RDP 16S rRNA database (see Materials and methods). Intriguingly, the hemerithryn (K07216) gene consistently associated with members of the Clostridiales when present in the gut (Spearman correlation 0.72, P-value <1e-15; Additional file 11). Finally, other metals, including copper and zinc, are also both necessary co-factors and potential toxins, and remediation pathways and transporters for both were observed consistently (copper resistance K07245; copper homeostasis K06201, K06079; zinc resistance K07803; and also many other metal transporters).

Although recent work has provided extensive insights into the mechanisms of bacterial interaction with the host immune system in the gut, much less is known about the relationship of the microbiota with host immunity for other body habitats and cell types. Two pathways observed in both the upper and lower digestive tracts and known to be involved in immunomodulation were hydrogen (H₂) and hydrogen sulfide (H₂S) production. Hydrogen production has been shown to be an important byproduct of acetogenic bacteria and also has an anti-inflammatory activity 58 . Enzymes both for utilization (for example, CoM methyltransferase, K14082) and for production (for example, hydrogenase-4, K12136) of hydrogen were identified specifically in the oral cavity (nearly completely absent from the gut), with potential bacterial contributors including Veillonella and Selenomonas species genomically 59 and, in one of the strongest links between genes and organisms in these data, an unclassified Pasteurellaceae clade in the oral cavity (Spearman correlation for K12136 >0.78, P-value <1e-15 in supragingival placque and tongue dorsum).

Hydrogen sulfide gas is involved in regulation of the host response at low concentrations and in host-cell toxicity and inhibition of short chain fatty acid production, specifically in the colon, at high concentrations 60 61 62 63 64 65 . H₂S may thus serve different purposes among the distinct bacterial communities of the digestive tract. The potential for its production was particularly enriched in stool (for example, by cystathione-beta-lyase, K14155), and somewhat enriched in the more anaerobic habitats, stool and plaques (for example, by methionine-gamma-lyase, K01761). A possible role in host-cell toxicity was strongly suggested by K01761's close association with the Treponema and Fusobacterium genera in plaque (Spearman correlation 0.74 and 0.82, respectively, P-values <1e-15), both of which include members specifically associated with periodontal disease (Additional file 11). These genes were again, however, present at low levels among all body sites analyzed here, consistent with a low-level immunomodulatory role for H₂S throughout the digestive tract.

Healthy adults 18 to 40 years old were recruited at two academic centers 10 . Fifteen and 18 body habitats were collected from enrolled males and females, respectively. The sites sampled included anterior nares, oropharynx (two specimens), oral cavity (seven specimens), skin (four specimens), stool, and vagina (three specimens per female) 10 . The Manual of Procedures and the Core Microbiome Sampling Protocol are available at the Data Analysis and Coordination Center for the HMP 99 , as well as dbGaP 100 . Genomic DNA was isolated from the collected samples using the MO Bio PowerSoil DNA Isolation Kit (MO BIO laboratories, Inc., Carlsbad, California, USA) 10 .

Detailed protocols used for 16S rRNA bacterial gene amplification and sequencing, using the 454 FLX Titanium platform and kits (Roche Diagnostic, Corp., Indianapolis, Indiana, USA), are available on the HMP Data Analysis and Coordination Center website 99 , and are also described elsewhere 10 . In brief, sequences were processed using a data curation pipeline implemented in mothur 10 101 starting with quality trimmed for homopolymer runs and a minimum 50 bp window average of 35. Any sequences not aligning against the appropriate subset of the SILVA database 102 were removed, as were chimeric sequences. Resulting sequences were processed using a data curation pipeline implemented in mothur 10 101 . Remaining sequences were classified with the MSU RDP classifier v2.2 29 using the taxonomy maintained at the RDP (RDP 10 database, version 6). Definition of a sequence's taxonomy was determined using a pseudobootstrap threshold of 80% 10 .

A table of read counts from the 16S rRNA bacterial gene pipeline was created by summing clade counts from the three regions and was further processed for removing low-coverage samples. Firstly, those taxa not supported in the whole dataset by at least two sequences in at least two samples were removed. Then, the quality control procedure compared, for each sample, the read count of the most abundant taxon t and the highest abundance value that the same taxon t achieved in the entire dataset. If the former term of the comparison is <1% of the latter, the sample was discarded. Second, third, and fourth time-point samples from the same subjects were discarded. The resulting dataset of read counts containing 2,105 samples is reported in Additional file 12, which represents 210 ± 7 samples per body site. Further analysis of the dataset was performed using the per sample normalization to relative abundances. In the text, mean values are presented with standard deviation. The number of subjects with samples in the digestive tract retained for the 16S rRNA-based analysis was 209 post-quality control, from which 147 had sample data for all 10 body sites post-quality control. Unless otherwise noted, only first visit samples were used in all analyses.

The characterization of functional and organismal features differentiating the microbial communities specific to different body sites in the digestive tract was performed using our method for biomarker discovery and explanation called LEfSe 30 . LEfSe, publicly available 103 , couples a standard test for statistical significance with a quantitative test for biological consistency, finally ranking the results by effect size. Briefly, it first uses the non-parametric factorial Kruskal-Wallis test to detect features (taxonomic clades or metabolic pathways) with abundances that differ below a significance threshold among groups of samples. Biological consistency is subsequently tested using the unpaired Wilcoxon rank-sum test among all pairs of sample groups; in our case this occurred between single body habitats. Finally, linear discriminant analysis (LDA) with bootstrapping is then used to rank differentially abundant features based on their effect sizes. A significance alpha of 0.05 and an effect size threshold of 2 were used for all biomarkers discussed in this study. Organismal and functional biomarkers are graphically represented here on hierarchical trees reflecting the RDP taxonomy 29 for 16S rRNA gene data and the KEGG BRITE hierarchy 49 on KEGG modules for metagenomic functional data.

For assessing bacterial community structure similarities between different samples and body sites, we compared the relative abundances of every pair of samples in our dataset using the Bray-Curtis measure of beta diversity 31 . The comparisons have been summarized in terms of within- and between-group averages as reported in Table 1; moreover, statistical significance has been tested for within versus between group distances, providing strong support (all P-values <10^-20) for the clustering of all four groups in distinct community structures. A multidimensional scaling analysis was then performed on the Bray-Curtis diversity matrix and the four groups were denoted with different colors for highlighting the clustering structure (Additional file 4).

Whole genome shotgun sequencing employed the Illumina GAIIx platform (Illumina, Inc.) as previously described 10 . The number of samples and nucleotide content from 98 subjects is summarized in Additional file 12. The abundances and presence (or absence) of pathways in these metagenomic data were inferred using the HUMAnN pipeline (HMP Unified Metabolic Analysis Network) 48 . Briefly, the metabolic and biomolecular potential of each sample was profiled starting from the 100 bp Illumina sequences after quality and length filtering. Reads were mapped to KEGG v54 orthologous gene families (KEGG KOs 49 ) using MBLASTX (MulticoreWare, St. Louis, MO, USA), an accelerated translated BLAST implementation, using default parameters and a maximum E-value of 1. Hits were mapped to abundances of each KO using up to the 20 most significant hits, weighted by the quality of each hit (inverse blastx P-value) and normalized by the length of the hit gene. Pathway information was then recovered by assigning KO gene families to KEGG modules (representing small pathways of approximately 5 to 20 genes) using a combination of MinPath 104 , filtering of pathways not consistent with the BLAST-derived taxonomic composition of the community, and gap filling of likely missing enzymes. The resulting KO and KEGG module relative abundances were used in the presented analysis. Further details of the HUMAnN methodology, its software implementation, and an extensive validation of each computational step appear in 48 .

The datasets used in these analyses were deposited by the NIH Common Fund Human Microbiome Consortium at the Data Analysis and Coordination Center (DACC) for the Human Microbiome Project. Specifically, the downloadable packaged datasets are the 16S rRNA gene dataset 105 , phylotype-classification of the 16S rRNA gene dataset 106 107 , Human Microbiome Illumina whole genome shotgun reads 108 , and the metabolic reconstruction tables 109 . The phylotype classification processed for normalization and quality control is available in Additional file 7.