Spatial Variation of the Native Colon Microbiota in Healthy Adults

The human colon is an ecosystem composed of numerous microenvironments that select for different microbiota. Concentrations of oxygen, water, and antimicrobial peptides change along the gut axis and influence which microbiota reside in each location. Microenvironments differ not only longitudinally along the colon, but also radially from the epithelium to mucosa to intestinal lumen, offering several sites for different microbial communities to flourish. The identity of these specific microbiota and communities is important for understanding the etiology of complex diseases such as colorectal cancer and inflammatory bowel disease (IBD). Colorectal cancer and IBD can be preceded or accelerated by perturbations of the structure of the gut microbiota (1–3). The manifestations of these diseases are known to vary based upon the location in which they occur. For instance, colorectal cancer that arises in the distal (left) colon is of hindgut origin and tends to have large chromosomal alterations indicative of chromosomal instability (1). In contrast, colorectal cancer arising in the proximal (right) colon is of midgut origin and tends to be sessile and microsatellite instable (MSI with BRAF and KRAS mutations; ref. 1). In addition to the environmental gradients within the colon, the distal and proximal sides of the colon differ in the amount of inflammation present and the genomic instability of precancerous cells, respectively (1, 4, 5). In IBD patients, disease occurring in the distal colon extending proximally is usually indicative of ulcerative colitis (UC), whereas Crohn disease (CD) can occur anywhere along the GI tract, most commonly in the ileum and the cecum (2). UC presents as continuous disease with only mucosal involvement, whereas CD has skip lesions and full thickness involvement that may cause abscesses, strictures, and fistulas (2). Thus, given the varied physiology of the proximal–distal axis of the colon and known differences in disease patterns at these sites, symbiotic microbiota and their metabolites likely vary as well, and may influence the heterogeneous disease prognoses of IBD and colorectal cancer. Because colorectal cancer can be a long-term complication of IBD, the distribution of microbiota is important to understanding the pathophysiology of both diseases.

Several recent findings have shown that development and progression of IBD or colorectal cancer can be attributed to specific molecular events as a result of interactions between the gut microbiota and human host (1, 3, 6). For instance, comparison of the bacteria present on colorectal cancer tumors with those found on nearby healthy tissue has identified specific species that are tumor associated (7). Specific bacteria have also been identified in fecal samples of patients with varying stages of colon tumorigenesis (8, 9). These species include the oral pathogens Fusobacterium nucleatum and Porphyromonas asacharolytica. F. nucleatum has also been found to be elevated in the stool and biopsies of patients with IBD as compared with healthy controls (10, 11). Furthermore, studies of F. nucleatum isolated from mucosal biopsies showed that more invasive F. nucleatum positively correlates with IBD disease level (10). Like many intestinal pathogens, the bacteria appear to have a high impact despite being lowly abundant in the community (2). The physiology of these rare taxa may contribute to the colonic disease state. These studies often examined only shed human stool or the small intestine, preventing fine-resolution analysis of paired samples from the proximal and distal sides of the colon. Similarly, comparisons of on- or off-tumor/lesion bacteria rarely have matched tissue from the other side of the colon from the same, disease-baring patient, limiting what conclusions can be drawn about the colonic microbiome overall, let alone at that specific site (12). Due to these limitations, the contribution of the gut microbiota to colorectal cancer and IBD disease location in the colon is largely undefined. Characterizing these communities in healthy individuals could provide needed insight into disease etiology, including how the disruption of the healthy community could promote the initiation or proliferation of the distinct proximal and distal colorectal cancer tumors or IBD flares.

The few existing profiles of the microbial spatial variation of the colon have been limited by sample collection methods. The majority of human gut microbiome studies have been performed on whole shed feces or on samples collected during colonoscopy or surgery (5). While invasive methods allow investigators to acquire samples from inside the human colon, typically these procedures are preceded by the use of bowel preparation methods, such as the consumption of laxatives to cleanse the bowel. Bowel preparation is essential for detecting cancerous or precancerous lesions in the colon, but complicates microbiome profiling as the chemicals strip the bowel of contents and disrupt the mucosal layer (13, 14). As such, what little information we do have about the spatial distribution of the microbiota in the proximal and distal colon is confounded by the bowel preparation procedure.

Here, we address the limitations of previous studies and identify the microbes specific to the lumen and mucosa of the proximal and distal healthy human colon. We used an unprepared colonoscopy technique to sample the natural community of each location of the gut without prior disruption of the native bacteria in 20 healthy volunteers. To address the inherent interindividual variation in microbiota, we used a machine-learning classification algorithm trained on curated 16S rRNA sequencing reads to identify the microbiota that were specific to each location. We found that our classification models were able to separate mucosal and luminal samples as well as differentiate between sides of the colon based on populations of particular microbiota. By identifying the distinguishing microbiota, we are poised to ask if and how the presence or disruption of the microbiota at each site contributes to the development of the tumor subtypes of colorectal cancer in the proximal and distal human colon.

The procedures in this study and consent were approved by the Institutional Review Board at the University of Michigan Health System with protocol number HUM00082721. Subjects were recruited using the online recruitment platform and were prescreened prior to enrollment in the study. Exclusion criteria included: use of asprin or NSAIDs within 7 days, use of antibiotics within 3 months, current use of anticoagulants, known allergies to fentanyl, versed and benadryl, prior history of colon disease, diabetes, abdominal surgery, respiratory, liver, kidney or brain impairments, undergoing current chemotherapy or radiation treatment, and subjects who were pregnant or trying to conceive. Twenty subjects who met the criteria were selected and provided signed informed consent prior to the procedure. There were 13 female and 7 male subjects ranging in age from 25 to 64. Eighteen of the 20 subjects had not used antibiotics within a year prior to the collection date and 2 had not used antibiotics within 6 months. None of the subjects had medical conditions requiring frequent or extended antibiotic use.

At a baseline visit, subjects gave consent and were given a home collection stool kit (Zymo). One to 7 days prior to the scheduled colonoscopy, subjects collected whole stool at home and shipped the samples to a research coordinator on ice. Notably, subjects did not undergo any bowel preparation method prior to sampling. On the procedure day, subjects reported to the Michigan Clinical Research Unit at the University of Michigan Health System. Subjects were consciously sedated using fentanyl, versed and/or benadryl as appropriate. A flexible sigmoidoscope was first inserted about 25 cm into the colon and jumbo biopsy forceps used to collect the luminal contents. Two luminal samples were collected, and the contents were immediately deposited into RNAlater (Fisher) and flash-frozen in liquid nitrogen. The forceps were withdrawn, and new biopsy forceps were used to collect mucosal biopsies on sections of the colon that were pink and free of stool matter. Three mucosal biopsies were collected and flash-frozen in RNAlater. These samples comprised the distal colon samples. The sigmoidoscope was then withdrawn, and a pediatric colonoscope was inserted to reach the proximal colon. The proximal samples were taken from the ascending colon proximal to the hepatic flexure at 75 to 120 cm depending on the subject. Samples were then collected in the same manner as was done in the distal colon and the colonoscope was withdrawn. All samples were stored at −80°C.

DNA extraction was performed using the PowerMicrobiome DNA/RNA Isolation Kit (Mo Bio Laboratories). For tissue biopsies, Bond-Breaker TCEP solution (Fisher) and 2.8-mm ceramic beads (Mo Bio Laboratories) were added to the bead-beating step to enhance DNA recovery from mucosal samples. The resulting DNA was normalized to equal concentrations across all samples and used as a template for amplification of the V4 region of the 16S rRNA gene, and fragments were sequenced on an Illumina MiSeq as previously described (15). Sequences were curated using the mothur software as described previously (16). The sequences were assigned a taxonomic classification using a naïve Bayesian classifier trained using a 16S rRNA gene training set from the Ribosomal Database Project (RDP; ref. 17) and clustered into operational taxonomic units (OTU) based on a 97% similarity cutoff. Sequencing and analysis of a mock community revealed the error rate to be 0.018%. Samples were rarefied to 4,231 sequences per sample in order to reduce the effects of uneven sampling bias.

Diversity analysis was performed using the Simpson diversity calculator and |\theta $|_YC calculator metrics in mothur version 1.39.5 (16). |\theta $|_YC distances were calculated to determine the dissimilarity between two samples. Random forest classification models were built using the AUCRF R package using a leave-one-subject-out approach (18). The random forest models were built using the full, non-rarefied, dataset as input. For each model, the data were split into a 19-subject training set and a 1-subject test set. The model was built and cross-validated using 10-fold k cross-validation (AUCRFcv) on the training set to estimate the prediction error of the model. The resultant model was then used to predict the outcome the left-out subject. This process was repeated iteratively for all subjects and results plotted as receiver operator characteristic curves using the pROC R package (19). Resultant models were used to identify the OTUs that were most important for classifying each location. Species-level information for sequences of interest was obtained by aligning the sequences to the GenBank nucleotide database using blastn. The species name was used only if the identity score was ≥99% over the full length of the contig and matched a single reference.

Differences in community membership at the phyla level were tested using the analysis of molecular variance (AMOVA) metric in mothur. Differences in |\theta $|_YC distances by location were tested using the Wilcoxon rank-sum test adjusted for multiple comparisons using the Benjamini–Hochberg procedure.

16S rRNA gene sequence reads and experiment metadata are available on the NCBI Sequence Read Archive (SRA) with accession numbers SRP124947 and PRJNA418115. A reproducible data analysis pipeline can be found at https://github.com/SchlossLab/Flynn_LRColon_CancPrevRes_2017.

Luminal and mucosal samples were collected from the proximal and distal colons of 20 healthy individuals who had not undergone bowel preparation (Fig. 1). Subjects also collected stool at home 1 week prior to the procedure. To characterize the bacterial communities present at these sites, 16S rRNA gene sequencing was performed on DNA extracted from each sample. As expected, each site was primarily dominated by Firmicutes and Bacteriodetes (Fig. 2A; ref. 20). Samples had varying levels of diversity at each site, irrespective of the individual (Fig. 2B). For example, the proximal mucosa was more diverse than the distal for some individuals while the opposite was true for others. Therefore, we could not identify a clear pattern of changes in microbial diversity along the gut axis.

To compare similarity between the proximal and distal sides and within the lumen and mucosa, we compared the community structure of these sites based on the relative abundances of individual Operational Taxonomic Units (OTU). Across all subjects, we observed wide variation when comparing sample locations (Fig. 3A). Those ranges did not follow a clear pattern on an individual basis. However, when comparing median dissimilarity between the communities found in the proximal lumen and mucosa, the proximal and distal lumen, the proximal and distal mucosa, and the distal lumen and mucosa, we found that the proximal lumen and mucosa were most similar to each other than to the other samples (P < 0.005, Wilcoxon, BH adjustment).

Next, we compared the luminal and mucosal samples to the fecal sample of each subject. Amidst the large intersubject variation, we did identify significantly less dissimilarity between the distal luminal sample and the feces (Fig. 3B; P < 0.05, Wilcoxon, BH adjustment). Furthermore, there was an even larger difference in the communities found in the distal mucosa compared with the fecal communities, indicating that the mucosa is as different from the stool as compared with lumen (P < 0.0005, Wilcoxon, BH adjustment). These results suggest that the contents of the distal lumen were most representative of the subjects' feces, and the mucosal microbiota are distinct from the fecal and luminal communities.

To determine what factors may have driven the differences seen among the samples, we compared the community dissimilarity between samples from all subjects (interpersonal) versus samples from within one subject (intrapersonal). We found that samples from one individual were far more similar to each other than to matched samples from the other subjects (Fig. 3C); this is consistent with previous human microbiome studies that have sampled multiple sites of the human colon (21–23). Thus, interpersonal variation drove the differences between samples more than whether the sample came from the proximal or distal side of the colon or from the lumen or mucosa.

To identify OTUs that were distinct at each site, we constructed several random forest models trained using OTU relative abundances. We built the first model to classify the luminal versus mucosal samples for the proximal and distal sides, independently (Fig. 4A). The models performed well when classifying these samples (proximal AUC = 0.716; distal AUC = 0.862). The OTUs that were most predictive of each site were identified by their greatest mean decrease in accuracy when removed from the model. For distinguishing the proximal lumen and mucosa, OTUs affiliated with the Bacteriodes, Actinomyces, Psuedomonas, and Enterobacteraceae were included in the best model (Fig. 5A). The model to differentiate between the distal lumen and mucosa included OTUs affiliated with the Turicibacter, Finegoldia, Peptoniphilus, and Anaerococcus (Fig. 5B). These results indicated that there were fine differences between the different sites of the colon, and that these could be traced to specific OTUs on each side.

Next, we built a random forest model to differentiate the proximal and distal luminal samples. The model performed best when distinguishing the proximal versus distal mucosa (Fig. 4B; AUC = 0.808), whereas the model to differentiate between the proximal versus distal lumen performed poorly (AUC = 0.599). The OTUs included in the model differentiating the distal and proximal mucosa included members of the Porphyromonas, Murdochiella, Finegoldia, Anaerococcus, and Peptoniphilus (Fig. 6A). The model that attempted to separate the proximal and distal lumen included OTUs affiliated with the Bacteroides, Clostridium IV, and Oscillibacter (Fig. 6B). Interestingly, Anaerococcus and Finegoldia were distinct between the mucosa and lumen and also helped to differentiate between the proximal and distal sides.

Given that specific bacterial species have been associated with colorectal cancer and IBD, we probed our sample set for these OTUs. Among our 100 samples, the most frequent sequence associated with the Fusobacterium genus was OTU179, which aligned via blastn to Fusobacterium nucleatum subsp animalis (100% over full length). This is the only species of Fusobacterium known to have oncogenic properties and be found on the surfaces of colorectal cancer tumors (24). There were 14 samples from 8 subjects with the F. nucleatum subsp. animalis sequences. Of the samples with the highest relative abundance of F. nucleatum subsp. animalis, four of the samples were from the proximal mucosa and three from the distal mucosa (Supplementary Fig. S1A). The second most frequent Fusobacterium sequence was OTU472, which aligned with 99% identity to F. varium. In addition to F. nucleatum, F. varium has been associated with IBD (25). Four subjects harbored F. varium and the samples were split evenly between the proximal and distal mucosa (Supplementary Fig. S1B). OTU152 was similar to the members of the Porphyromonas genus and the most frequent sequence in that OTU aligned to Porphyromonas asacharolytica (99% over full length), another bacterium commonly detected and isolated from colorectal tumors. OTU152 was only detected on the distal mucosa, and in fact was one of the OTUs the classification model identified as separating distal and proximal sides (Supplementary Fig. S1C). Among the 11 distal mucosa samples that were positive for P. asacharolytica, the relative abundances for this OTU ranged from 0.01% to 16%. Thus, disease-associated OTUs could be found in our sample set of 20 healthy individuals.

We identified bacterial taxa that were specific to the lumen and mucosa of the proximal and distal sides of the human colon using samples collected during an unprepared colonoscopy of healthy subjects. We found that all locations contained a range of phylum relative abundances and a range of diversity, but that there was a wide variability between subjects. Pairwise comparisons of each of the sites revealed that the proximal mucosa and lumen were most similar to each other. Further, comparison of colonoscopy-collected samples with fecal samples demonstrated that the distal lumen was most similar to feces. Random forest models built using OTU relative abundances from each sample identified microbiota that were particular to each location of the colon. Finally, we were able to detect some bacterial OTUs associated with colonic disease in our healthy cohort. Using unprepped colonoscopies and machine learning, we have identified bacterial taxa specific to the healthy proximal and distal human colon.

When examining the relative abundance of the dominant phyla at each site (i.e., Bacteriodes and Firmicutes), there was a wide amount of variation. This likely reflects not only the variability between human subjects, caused by differences in age, sex, and diet, but may also reflect spatial patchiness in the gut microbiome within a subject. Patchiness refers to inconsistent distribution of microbial populations due to fluctuations in local resources (26). One study noted that the bacteria recoverable from the same mucosal sample location can be vastly different when the samples are taken just 1 cm away from each other (27). Similar patchiness was also observed in luminal contents and fecal samples themselves; there was separation of different interacting microbes along the length of a stool sample, for instance (28). A third study that sampled six mucosal sites along the colon observed such patchiness in two of the three study subjects (21). While our subjects were not sampled frequently enough to draw specific conclusions about patchiness along the unprepped colon, we did observe some specific differences in mucosal versus luminal samples at the phylum level. The mucosal samples harbored more Proteobacteria, consistent with previous studies comparing mucosal swabs to luminal content in humans (4). However, we must still consider that the results from phyla analysis may have been impacted by intersubject patchiness.

To get around the noisiness from a diverse set of samples, we built random forest classification models to identify the microbiota that were specific to each side and in the lumen and mucosa. For each comparison we identified the top five OTUs that were strongly predictive of one site or another. Generally, OTUs identified in each location were consistent with known physiological gradients along the gut axis (5). For instance, the proximal mucosa contains the highest oxygen concentrations of the colon and harbored mucosa-associated facultative anaerobes such as Actinomyces and Enterobacteraceae and aerobic Psuedomonas. The distal mucosa was far more likely to host strictly anaerobic species such as Porphyromonas, Anaerococcus, Finegoldia, and Peptoniphilus. Thus, the gut microenvironment of each location likely enriches for these specific microbiota.

In addition to identifying features that are specific to each side of the gut, the ability of the random forest to classify samples can serve as a proxy for similarity. That is, a higher AUC value indicates the samples are more efficiently classified (and thus more different) than a model with a lower AUC value. For instance, the model separating the proximal and distal mucosa had an AUC of 0.599, whereas the model for classifying the proximal and distal lumen had a much lower AUC of 0.580. Further, the latter model required 44 OTUs to best separate the samples, whereas the models separating the mucosa only needed 10 OTUs. The much lower AUC and need for a high number of features compared with other models suggest that these locations are the most similar of the comparisons tested. We speculate that the model was less effective at classifying the proximal and distal luminal contents because the mucosal microenvironments have more variable selective pressure along the colon than the luminal microenvironments.

We detected F. nucleatum and P. asacharolytica in 8 and 5 of our subjects, respectively. These bacteria have been shown to be predictive of colorectal cancer in humans (9) and have oncogenic properties in cell culture and in mice (29). Though the bacteria are known to colocalize on colorectal cancer tumors, in our study F. nucleatum was found on both sides of the colon while P. asacharolytica was detected only in the distal mucosa. Not much is known about the distribution of P. asacharolytica along the healthy colon, but given its anaerobic lifestyle and asacharolytic metabolism, it is perhaps not surprising that our study detected the bacteria primarily in the less-oxygen-rich and protein-rich distal mucosa (4). In studies examining bacteria on colorectal cancer tumors, F. nucleatum was more commonly detected on proximal-sided tumors, and distribution of F. nucleatum decreased along the colon to rectum (30). Of the 8 (40%) individuals positive for F. nucleatum in our study, the bacterium was spread across the proximal mucosa, distal lumen, and distal mucosa. The Fusobacterium species nucleatum and varium have been commonly isolated from mucosal biopsies of patients with IBD and UC (25, 31). In our study, F. varium was only detected in three subjects, and two of those samples were isolated from the proximal mucosa (Supplementary Fig. S1B). F. varium is most commonly isolated from UC patient biopsies from the ileum or cecum (adjacent to the proximal colon; ref. 32), suggesting this species may exhibit preference for the higher oxygen content of these gastrointestinal sites.

Spatial organization of Fusobacteria and other bacterial species into polymicrobial biofilms that can invade the gut mucosa has been linked to colorectal cancer (33). The biofilms promote tumorigenesis by allowing bacteria to grow near the epithelium, inducing inflammation, genotoxicity, and metabolic changes that favor tumor cell growth (33). In one study of colorectal cancer biofilms and tumors, all of the proximal tumors examined contained a polymicrobial biofilm on the tumor mucosa (7). Further examination of these tumors identified Fusobacteria species as members of the biofilm. These results indicate that it is not only the presence of the bacterium but the tumor community as a whole that contributes to tumorigenesis (7). A “driver-passenger” model has been proposed as a mechanism for biofilm assembly in the gut (34, 35). In this model, “driver” species such as Fusobacterium spp and Porphyromonas spp exert tumorigenic effects locally and create a niche for adherence of “passenger” species that comprise the rest of the biofilm (34, 35). Thus, the distribution of these disease-associated microbes in healthy patients is of interest as their presence can be predictive of disease prior to the onset of symptoms (9). A better understanding of the early microbial changes in the gut microbiome is essential for elucidating a mechanism for development of colorectal cancer or IBD subtypes in the proximal or distal colon.

Specific comparisons of our findings to previously published studies of spatial variation are confounded by the use of bowel preparation methods. A rare report of a matched-colonoscopy study sampled 18 patients’ colonic mucosa and luminal contents prior to and after bowel cleansing (36). This study found that mucosal and luminal samples were distinguishable prior to bowel cleansing, but that bowel preparation resulted in an increase in shared OTUs between each site (36). After 7 days, bowel cleansing not only made the samples more difficult to distinguish, but it also decreased the diversity observed across sites. Bowel preparation clearly biases the representation of microbiota recovered from sampling the lumen or mucosa.

By revealing specific differences in microbial populations at each location in the gut via sampling an unprepared bowel, we can begin to form hypotheses about how specific host–microbe interactions can affect disease progression of proximal and distal colorectal cancer and IBD subtypes. Future investigation of these samples using metagenomics and metatranscriptomics would illuminate the microbial activities in these gut microenvironments. Further, combining this approach with a more comprehensive sampling strategy along the unprepped colon could enhance microbiome-based screening and treatment modalities for colon disease.

P.D. Schloss reports receiving a commercial research grant from Janssen Pharmaceuticals and has received and expert testimony from Blood Hurst & O’Reardon, LLP. No potential conflicts of interest were disclosed by the other authors.

Conception and design: K.J. Flynn, M.T. Ruffin IV, D.K. Turgeon, P.D. Schloss

Development of methodology: K.J. Flynn, M.T. Ruffin IV, D.K. Turgeon, P.D. Schloss

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K.J. Flynn, M.T. Ruffin IV, D.K. Turgeon, P.D. Schloss

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K.J. Flynn, M.T. Ruffin IV, P.D. Schloss

Writing, review, and/or revision of the manuscript: K.J. Flynn, M.T. Ruffin IV, D.K. Turgeon, P.D. Schloss

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K.J. Flynn, D.K. Turgeon, P.D. Schloss

Study supervision: D.K. Turgeon, P.D. Schloss

We thank all the individuals who volunteered for the study. This work was supported by the Rose and Lawrence C. Page Foundation (D.K. Turgeon). We would also like to thank Brian Kleiner, Chelsea Crofoot, and Kirk Herman for their roles in study coordination, subject recruitment, sample collection, and sample processing. Thanks also to GI Fellows Drs. Amy Hosmer, Alexander Larson, and Anna Tavakkoli, who assisted Dr. Turgeon with the colonoscopies.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.