The Microbiomes of Pancreatic and Duodenum Tissue Overlap and Are Highly Subject Specific but Differ between Pancreatic Cancer and Noncancer Subjects

In 2018, an estimated 55,440 individuals will be diagnosed with pancreatic cancer in the United States, and only 8% of these individuals are expected to survive the next 5 years (1). Given this high fatality rate, and the silent progression of early disease, identifying risk factors for the prevention and early detection of pancreatic cancer is critical to reducing its mortality. To date, known risk factors for pancreatic cancer, including smoking, obesity, diabetes, heavy alcohol consumption, family history, and markers of genetic susceptibility, cannot, even collectively, be used for early detection and risk stratification of pancreatic cancer in the general population (2).

Studies have suggested a link between bacteria and pancreatic cancer risk (3), highlighting the need to more critically explore the underlying factors that affect the microbiome of the oral cavity and upper digestive tract in both cancer patients and cancer-free individuals. The current research on oral bacteria and pancreatic cancer risk stems from a number of observational studies that reported a higher risk of pancreatic cancer among individuals with periodontitis, when compared with those without periodontitis (3, 4). Periodontitis, an inflammatory disease of the gums, is largely driven by keystone pathogens and pathobionts (5). Two large prospective cohort studies have reported positive associations between periodontal disease pathogens and subsequent pancreatic cancer risk (6, 7); in these two studies, detection of elevated antibodies to Porphyromonas gingivalis, measured in blood collected prior to cancer diagnosis, was associated with a two-fold higher risk of pancreatic cancer (6), and presence (vs. absence) of P. gingivalis in saliva collected prior to cancer diagnosis was associated with a 60% increase in risk of pancreatic cancer (7). Aggregatibacter actinomycetemcomitans, another periodontal pathogen, was also associated with pancreatic cancer risk in the prospective study using saliva (7).

Few investigations to date have attempted to detect bacteria in pancreatic tissue. Earlier studies reported the presence of bacteria in pancreatic ducts of subjects with chronic pancreatitis or bile duct obstruction (8–10). Other studies have investigated the presence of specific bacterial DNA in the pancreatic tissue of pancreatic cancer subjects, namely species of Helicobacter (11) and Fusobacterium (12). The most comprehensive molecular microbiome studies to date reported the presence of a diverse bacterial populations in fluids collected from the bile duct, pancreas, and jejunum of subjects undergoing pancreaticoduodenectomy (13), and in pancreatic cyst fluid removed endoscopically from pancreatic cysts (14). In mice, bacteria have been shown to translocate from the mouth to the pancreas, and germ-free mice have reduced progression of pancreatic ductal adenocarcinoma (15).

Metagenomics studies on DNA isolated from tissue samples from cancer subjects have been conducted for lung (16), colorectal (17), esophageal (18), stomach (19), and breast cancer (20). These studies demonstrate that 16S rRNA gene sequencing can be effectively conducted on fresh tissue samples where the ratio of bacterial to human DNA is much lower than at other human sites (e.g., stool or oral cavity; ref. 21). Moreover, these studies have shown that bacterial profiles at different organ sites are often unique (16, 20) and that changes may be associated with cancer (18, 19). In two recent studies, bacterial DNA was measured in tumor tissue samples obtained from patients with pancreatic ductal adenocarcinoma (PDAC) using 16S rRNA gene sequencing (15, 22); however, comparison of microbiota in pancreas and different gastrointestinal tissue was not conducted in these patients.

To date, no study to our knowledge has characterized the overall microbiome in pancreatic and normal surrounding tissue samples, a critical step to understand whether and how bacteria may play a role in carcinogenesis. In an effort to address the specific question of whether the pancreas has its own microbiome, we recruited subjects from the Rhode Island Hospital (RIH; Providence, RI) with planned foregut surgery to obtain tissue samples for 16S rRNA gene microbiome analysis. In addition, for comparison to controls, we obtained pancreatic and duodenum tissue from National Disease Research Interchange (NDRI) from noncancer subjects.

Seventy-seven subjects, enrolled between January 2014 and March 2016, were included in this study. Subjects were eligible if identified as candidates for surgery of the foregut by Dr. Charpentier (the lead surgeon at the RIH) and included those with pancreatic cancer, pancreatic cystic neoplasms, pancreatitis, bile duct, or small bowel diseases. All recruited subjects were between 31 and 86 years old (Table 1). Participants were asked to complete a self-administered questionnaire to provide data on demographic and behavioral factors, and included a question on past use of antibiotics; this variable was included in the statistical analysis to control for changes that may have occurred due to antibiotic use in recent past. Questions on family history of cancer, use of other over-the-counter medications were also included on the questionnaire. Stool collection kits with ethanol as a fixative [95% (w/w) ethanol] were provided prior to surgery (23); participants were asked to return the samples using a pre-paid box.

A protocol was established for processing tissue samples collected during surgery to reduce contamination. A technician from the Pathology Department was informed in advance of the surgery date and time, and was paged as soon as the specimens had been obtained. Surgical tissue samples were frozen within 1 hour of the surgery time, as well as tissues swab samples from the stomach, jejunum, and bile duct that were collected using DNA-free forensic sterile swabs whenever possible. During surgery, the surgeon also recorded (on a surgery form for the study) if the patient had received prior pre-OP endoscopic ultrasound (EUS), had previously had their gallbladder removed, or had received prior placement of a stent (for treatment of symptoms); all subjects received a single dose of perioperative antibiotics immediately prior skin incision at the time of the operation. Tissue samples (pancreatic tumor tissue, pancreatic cysts, normal pancreatic tissue, pancreatic ducts, and duodenums) were prepared by a Rhode Island Hospital pathologist; cancerous and noncancerous tissues were identified, separated and labeled. All samples were de-identified and stored at −80°C until processing.

Upon review of pathology records, ICD10 codes were assigned to each subject; 39 subjects had pancreatic cancer (ICD10 codes C25.0–C25.9; the majority of cases were adenocarcinomas, only two subjects had neuroendocrine tumors of the pancreas), 12 subjects had periampullary cancer (ICD10 codes C24.0–C24.1), 18 subjects had other pancreatic conditions (ICD10 codes K86.0–K86.3), and the remaining eight had other gastrointestinal conditions. The study was approved by Lifespan's Research Protection Office for recruitment at RIH, as well as the Institutional Review Boards for Human Subjects Research at Brown University, Tufts University, and the Forsyth Institute.

In addition, we obtained pancreatic specimens without known conditions of pancreatic diseases from the NDRI to serve as control samples in the absence of available healthy pancreatic tissue in noncancer subjects. Snap-frozen “control” whole-pancreas and duodenum (∼5 cm) human specimens from 34 deceased donors were obtained from NDRI with an average post-mortem recovery time of 13 hours. Control pancreas (head and tail), pancreatic ducts, and duodenums were dissected under sterile conditions, and stored at −80°C until processing. To remove additional contamination, we removed a thin tissue layer around each sample prior to extracting DNA. Details for DNA extraction and sequencing procedures are provided in the Supplementary Materials and Methods.

The 16S rRNA gene dataset consists of Illumina MiSeq sequences targeting the V3–V4 hypervariable regions. The DNA target sequencing was performed by the Forsyth Institute Sequencing Core. To evaluate effect of running samples on MiSeq runs at different times, we included bacterial mock community samples on each run and then compared their relative abundances across the MiSeq runs; the results for the mock communities were consistent across run, demonstrating minor fluctuations (Supplementary Fig. S1).

The MiSeq reporter analysis was used to discard low-quality sequences and to generate FASTQ files containing only filtered quality sequences, subsequently the overlapping paired-end reads were stitched together and further processed using used a multistage BLASTN-base search taxonomy read assignment pipeline that maximizes species level classification (24).

Sequences were BLASTN-searched against a combined set of 16S rRNA reference sequences that consist of the HOMD (version 14.5; ref. 25), Greengenes Gold (26), and the NCBI 16S rRNA reference sequence set. All assigned reads were subject to several down-stream bioinformatics analyses, including alpha and β diversity assessments, provided in the Quantitative Insights Into Microbial Ecology (QIIME; ref. 27) software package version 1.9.1.

Samples with <500 total read counts were excluded from all analysis. In addition, only OTUs with a minimal read count of 100 sequences (across all samples) were included in the analyses. For QIIME analyses, we normalized the number of sequences in the different MiSeq runs by rarefying each library to 500 reads to account for differences in sequencing depth across runs (increasing rarefaction cutpoint to higher read number did not result in changes in alpha-diversity results or OTU numbers in samples; 500 reads was used as the cutpoint to reduce number of samples lost from the analysis). Range of sequencing counts for the different sample types are provided in Supplementary Fig. S2. Across samples, OTU relative abundance was computed as the ratio of an OTU's absolute abundance to the total number of reads for that sample.

To create relative abundance plots, we restricted bacterial taxa (at genus-level) present at >2% relative abundance and with >35% prevalence in both NDRI and RIH samples (this was done to simplify comparison between the RIH and NDRI samples). Jaccard Index was used for paired comparison of proportion of shared microbiota taxa present at >2% relative abundance in tissue/swab samples within subjects.

To examine the variation in the microbial profile across the different habitats/sites (Supplementary Table S1) among the NDRI and RIH subjects, we calculated the distance/dissimilarity between samples using the Bray-Curtis and Sorensen indices (28). Computed distances were subsequently used to generate principal coordinate analysis (PCoA) plots to visualize the arrangement of the samples in the ordination space. PERMANOVA (available in QIIME) was used to test whether the distances are more similar within a group of samples than that from other groups of samples.

To identify demographic and clinical correlates of pancreatic microbial composition, we fit a series of zero-inflated β regression models to examine associations between genus-level relative abundances and demographic [i.e., age, gender, race, body mass index (BMI)] and clinical [i.e., health status, chemotherapy, antibiotics use prior to surgery, anxiety medications, presence of stent prior to surgery, whether preoperative endoscopic ultrasound (pre-OP EUS) was conducted prior to surgery, tumor surgery classification by International Code of Disease (ICD10 code)]. In the results, we refer to relative mean abundance among nonzero observations (μ) merely as relative mean abundance. More details are provided in the Supplementary Materials and Methods.

We explored which factors obtained in the questionnaires and medical files in RIH subjects were associated with bacterial communities in the pancreatic tissue samples. The most influential factors were sequencing run, presence of stent, and chemotherapy prior to surgery (only five patients with available tissue/swab samples had chemotherapy in the past 6 months); each of these factors was significantly associated with a large number of genera tested in marginal models. Given that the mock bacterial communities were similar across runs (see Supplementary Materials and Materials), it is possible that “run” was associated with certain genera due to differences in number of samples per sequencing run. To adjust for potential confounding, we considered this covariate in the final models comparing cancer to noncancer subjects and the different ICD-codes among the RIH subjects. Similarly, age, BMI, and sex were adjusted for as these features were shared between the studies and were found to explain variation in the relative abundance of some of the genera. Smoking was not found to explain variation in relative abundance in our data.

The present analysis included a total of 246 pancreatic tissue and swab samples collected from 82 subjects [50 subjects from RIH providing 133 samples (57 swabs, 76 tissue) and 34 subjects from NDRI providing 113 tissue samples; Supplementary Table S1]. In addition, 12 RIH subjects provided stool samples. There were no significant differences in the distribution of age, gender, BMI, race, and smoking status between RIH and NDRI subjects (Table 1). The Illumina-based sequencing of V3–V4 hypervariable regions of the bacterial 16S rRNA gene resulted in a total of 19,498,743 high-quality sequences (with a median sequence length of 427 nucleotides).

Over 99% of the reads from RIH pancreatic samples were attributed to five bacterial phyla (45.9% Proteobacteria, 35.6% Firmicutes, 9.5 % Bacteroidetes, 4.3% Fusobacteria, and 3.9% Actinobacteria). The remaining low abundance phylotypes (0.6% of the total) belonged to six bacterial phyla (Synergistetes, TM7, Deinococcus-Thermus, Verrucomicrobia, Spirochaetes, and Tenericutes). 99.6% of the reads observed among the NDRI pancreatic samples belonged to the same five bacterial phyla as observed in RIH subjects. The phylum Tenericutes (Bacteria) was present only in RIH samples, and the phylum Euryarchaeota (Archaea) was present only in NDRI samples, but both of these phyla were uncommon.

Although the microbial communities in the pancreatic tissues were dominated by the phyla Firmicutes and Proteobacteria, substantial interindividual variability was observed. In RIH samples, Proteobacteria relative abundance ranged from 2% to 99%, and similarly, Firmicutes relative abundance ranged from 0.6% to 84%. Large interindividual variability was also observed in the NDRI samples.

Mean relative abundance for bacterial taxa (mostly at the genus-level) in the pancreatic tissue samples (duct, head, tail, normal, and tumor), duodenum tissue samples, and jejunum, bile duct and stomach swabs are presented for each subject with more than one available sample in the RIH Fig. 1 and NDRI in Fig. 2 (males and females are presented separately for ease of comparison—no major differences were observed by sex). Three striking patterns emerge: (i) bacterial profiles in the pancreas are subject-specific rather than site-specific, (ii) bacterial profiles in duodenum tissue are remarkably similar to those in pancreatic tissue in the same subjects, (iii) concordance of paired comparisons of bacterial profiles in cancer subjects (RIH) are slightly lower across tissue type or site than those for noncancer subjects (NDRI; Fig. 3). Subjects from RIH with only one sample available (n = 5) demonstrate similar bacterial profiles as those with multiple samples. Bacterial taxa commonly recognized as oral bacteria, including Fusobacterium spp., Prevotella spp., Dialister spp., Veillonella spp., and Haemophilus spp. were identified in many of the tissue samples, both cancer and noncancer subjects (Fig. 1). Other oral bacterial taxa, including Parvimonas micra, Selenomonas noxia, Capnocytophaga spp., Peptostreptococcus spp., and Solobacterium moorei were also identified in tissue samples but were less common (present in 20%–35% of all samples).

With the exception of the stool and the jejunum, all the bacterial communities were characterized as habitats with low bacterial richness including the pancreatic sites, duodenums, and the bile ducts (Fig. 4A). Among RIH subjects, the microbial communities of the stool samples were represented by higher richness than the microbial communities in the tumors of the pancreas (P = 0.007), duodenums (P = 0.013), and bile duct swabs (P = 0.017). Likewise, the stool bacterial communities had higher richness than the NDRI pancreatic heads (P = 0.012), pancreatic ducts (P = 0.020), and duodenums (P = 0.005). The microbial communities in the jejunum swabs showed more richness than the communities in the RIH pancreatic head (P = 0.014) and duodenums (P = 0.028). In general, the bacterial communities in the pancreas of RIH subjects had slightly higher richness when compared with those from the pancreas of the NDRI matching sample types. Similar results were observed using additional alpha diversity measures of the bacterial communities (Fig. 4B–D). As expected, the stool samples were the most diverse with a Shannon index ≥4 (Fig. 4B). As the number of phyla represented in high abundance in these samples was relatively low (∼5), we observed relatively low levels of phylogenetic distances across all samples (Fig. 4D).

The ordination β-diversity analysis revealed that the majority of samples belonged to a single cluster, without any visually apparent groupings by the nature of the sample, health status, or anatomical site (Fig. 5A–C). However, the PERMANOVA tests revealed statistically significant differences between NDRI and RIH samples (P < 0.001), and for the swab samples obtained from the bile duct, jejunum, and stomach (compared with pancreas tissue samples). Differences between sites within the pancreas (i.e., head, tail, duct), and compared with the duodenum (for NDRI and RIH, separately), were not statistically significant (after accounting for multiple comparisons). The principal component analyses of both Bray-Curtis and Sorensen distances between all samples (tissues and swabs) showed that both RIH and NDRI samples clustered mostly by subject.

Using multiple regression analyses, we examined presence or absence, and relative mean abundance of bacterial taxa (at the genus and species level) among present (nonzero) observations using all tissue and swab samples comparing RIH subjects to NDRI subjects. Table 2 presents the bacterial taxa (at the genus level) that were present in at least 20% of all tissue and swab samples; each model includes both a zero-inflated component (testing for differences in presence/absence of bacterial taxa) and a relative mean abundance comparison. Lactobacillus taxa were present in almost all noncancer tissue samples (estimated proportion of presence [P1] = 0.98), but were much less likely to be present in cancer tissue samples (P1 = 0.58, P < 0.0001), and mean relative abundance was higher in noncancer subjects (μ = 0.06 vs. μ = 0.02 in RIH subjects, P < 0.0001; Table 2). In contrast, a number of bacterial taxa, including Porphyromonas, were present in higher mean relative abundance in cancer subjects than noncancer subjects (Table 2, and species level data presented in Supplementary Table S2). Oral bacteria Fusobacterium spp. and Prevotella spp. had higher mean relative abundance in cancer subjects than noncancer subjects (P-values < 0.0001 according to Wald tests for μ). Although these two bacteria do not appear in Table 2 because they were not significant according to the joint permutation (based test for prevalence and mean relative abundance at the genus level), a number of Fusobacterium species, for example, Fusobacterium nucleatum _subsp._vincentii, were much more prevalent in RIH samples and are significant in the species-level models (Supplementary Table S2).

Table 3 presents the bacterial taxa (at genus level) for which statistically significant associations remained after multiple comparison correction (at P < 0.10) when comparing bacterial taxa in tumor tissue (RIH) by ICD code to those identified in normal pancreatic head tissue from NDRI subjects (labeled as “controls” in Table 3; given that the bacterial profiles were highly similar by subject, we included only pancreatic head tissue for this analysis). In the marginal models (prior to adjusting for other covariates), a total of 16 bacterial genera were identified as being significantly associated with disease status prior to correction for multiple comparisons (Supplementary Table S3); a number of these taxa have representative strains in the Human Oral Microbiome Database (http://www.homd.org; e.g., Fusobacterium, Capnocytophaga, Prevotella, Porphyromonas, Parvimonas, Selenomonas, and Haemophilus). Mean relative abundances for some of these taxa (namely, Capnocytophaga, Prevotella, Selenomonas) were higher in samples coming from subjects diagnosed with pancreatic cancer (ICD C25) compared with NDRI samples. The model with Porphyromonas had the strongest association overall (P = 4.5 × 10⁻⁷); the relative mean abundance for periampullary cancer tissue samples was substantially higher than that of NDRI samples (P = 5.8 × 10⁻¹⁹), as were the IPMNs (K86.2) samples (P = 3.6 × 10⁻⁷). The associations with Porphyromonas remained elevated in multiple regression models (Table 3).

The multivariable regression models for the pancreatic tissue samples identified bacterial taxa (at the genus level) that had not been significant in the marginal regression models, including Simonsiella, Helicobacter, and Bilophia (Table 3 vs. Supplementary Table S3). Helicobacter was commonly identified in periampullary pancreatic tumors (C24) but at very low levels; in contrast, Helicobacter was infrequently identified in the NDRI samples, but was a dominant genus when present (relative mean abundance 47%; Table 3).

We further examined the RIH pancreatic tumor tissue samples without the NDRI samples given the difference in source of tissue and to account for clinical factors such as prior chemotherapy. Porphyromonas were also strongly associated with ICD code in both marginal (Supplementary Table S4) and multiple regression models suggesting clinical covariates were not confounding the main findings for these bacteria.

To test whether the associations would be similar using pancreatic duct tissue samples (vs. tumor tissue), we repeated the analysis using RIH and NDRI samples obtained from the pancreatic ducts. The associations for Porphyromonas remained detectable and statistically significant in these analyses (P = 1.53 × 10⁻¹¹).

Using tissue samples obtained from the duodenum, we compared relative abundance of bacterial taxa in NDRI and RIH subjects to examine whether any bacteria from the pancreatic tissue analyses were also noticeably different in the duodenum samples. Of the significant associations noted in the pancreatic tissues, Selenomonas was also elevated in the duodenum tissue of pancreatic cancer subjects compared with duodenum tissue from NDRI subjects (P = 3.9 × 10⁻¹²). A weak association was also observed for Gemella for the duodenum samples, consistent with an overall elevated mean relative abundance in the RIH samples compared with the NDRI samples (Table 2); other associations were either not significant or not consistent in direction of differences.

We only had one pancreatic duct stent to examine microbial community; the bacterial taxa from this stent were characterized as the members of the genera Klebsiella and Enterobacter.

Using pancreatic and duodenum tissue samples from subjects with pancreatic cysts or pancreatic cancer, and comparing them to pancreatic tissue samples obtained from donors who died of noncancer causes, we were able to demonstrate that pancreatic tissue contains a number of different bacterial taxa, including taxa that are known to inhabit the oral cavity. Our findings provide evidence that the pancreas is not a sterile organ and that there is substantial between-person variability in relative abundance of bacterial taxa at the genera level in the pancreas, but we also observed marked within-person stability across site (Figs. 1 and 2); bacterial composition at different sites in the pancreas (i.e., duct, head and tail) as well as the duodenum were highly similar in the same individuals. Finally, we noted lower presence and relative abundance of Lactobacillus in cancer subjects compared with noncancer subjects, and a significant increase in the mean relative abundance of periodontal-related pathogens in the tissue of pancreatic subjects when compared with noncancer subjects.

Dissemination of oral bacteria to different parts of the body has been well reported, and oral bacteria have been linked to a number of chronic diseases, including cardiovascular diseases (29, 30). Fusobacterium nucleatum has been associated with colon cancer in a number of cross-sectional studies (31, 32). Mouse models of colorectal cancer provide some support for a causal link (17, 33), demonstrating how this bacterium has the ability to initiate recruitment of tumor-infiltrating immune cells. Moreover, a recent study demonstrated similar microbiome profiles in primary colon cancer tumors and liver metastases from the same individuals (resected at a later time point), especially for Fusobacterium positive tumors (17), suggesting stability in the microbiome as the tumor progresses and metastasizes. Given the findings from this study, where multiple tissue specimens were examined in the same subjects, it may also be plausible that each individual has a unique microbiome profile that exists in different gastrointestinal tissue and that certain profiles increase cancer susceptibility by impacting the immune environment to allow for tumor promotion and growth. Bacterial taxa found in this study were highly consistent with those reported in a microbiome study on colon cancer; enriched bacterial taxa associated with Fusobacterium nucleatum positive tumors were similar to those we identified in this study (e.g., Bacteroides, Prevotella, Selenomonas, and Leptotrichia; ref. 17).

Presence of Lactobacillus spp. was significantly reduced in both periampullary and PDAC cancers compared with noncancer patients (including those with pancreatic cysts). Certain strains of this bacterium have been identified as playing a key role in mediating antiinflammatory pathways in calorie-restricted mice (34). Further research on the role of these bacteria in pancreatic cancer should be conducted.

Previous studies have reported associations between periodontal disease pathogens and pancreatic cancer risk, especially Porphyromonas gingivalis (6, 7). Periodontal disease is an inflammatory disease of the gums that can, in advanced conditions of periodontitis, result in systemic inflammation. In this study, we observed significantly higher mean relative abundance levels (at the genus level) for two bacterial taxa previously associated with periodontitis in pancreatic tissue, including Porphyromonas and Selenomonas (35–38); however, only Porphyromonas remained statistically significant after adjusting for age, sex, BMI, and library size. Porphyromonas was also elevated in the pancreatic duct tissue of periampullary pancreatic cancers, but no statistically significant associations were noted for the other oral bacterial taxa. Whether Porphyromonas play a role in pancreatic carcinogenesis will need to be further examined in other studies and confirmed in animal models. Proposed mechanisms for carcinogenesis include the ability of certain bacteria to induce a proinflammatory response in the tumor microenvironment (33); inhibit the immune response targeted at eliminating tumor cells (39); and modulate key cellular pathways associated with cell division (40).

A similar study using swab specimens from the pancreas, bile, and jejunum was conducted on subjects with pancreatic cancer undergoing pancreaticoduodenectomy (13). In that study, many bacterial taxa were present in fluids obtained from the pancreatic ducts and the common bile duct, including Prevotella, Haemophilus, Aggregatibacter, and Fusobacterium (13). Consistent with our findings, microbial communities in the pancreas, bile, and jejunum fluids were similar within individuals (13). Mean relative abundance for the bacterial genus Klebsiella was high in the samples from pancreatic cancer subjects in that study (13); in our study, we found Klebsiella to be one of two taxa on a swab taken from the stent itself. Placement of stent prior to surgery may impact the type of bacteria present in the pancreas, as observed in our study. In a separate study, metagenomics was conducted on freshly frozen duodenum samples from five normal and five obese individuals; Streptococcus (30%–32%) and Actinomyces (12%–17%) were the most common bacterial taxa identified in those samples, and relatively higher counts of Gemella were also identified in all 10 subjects (41). Porphyromonas were not identified in the duodenal samples (41).

In a recent study examining tumor resistance to the drug gemcitabine, bacteria were found in tumor tissues of 65 PDAC patients (out of 113), and 51.7% of bacterial taxa belonged to the class Gammaproteobacteria (22), which is highly consistent with our findings (Fig. 1). Similar to our study, there was large inter-individual variability in relative abundance of bacteria in each tumor, but in contrast to our study, only 3 of 20 organ donors were found to be positive for bacteria, and no normal tissue samples were included from the same patients (22). In addition, a high number of reads for Porphyromonas was found in one (of 65) pancreatic cancer tissue specimens (mean relative abundance of 0.123; Supplementary Materials and Material; ref. 22). In our study, read counts for Porphyromonas spp. were also extremely high in two RIH subject. In a separate study, 408 genera of bacteria were identified in pancreatic cyst fluids obtained from patients through endoscopy (14); many of the taxa found in pancreatic cysts were similar to those in tissue from our study, including the presence of Fusobacterium. Furthermore, Porphyromonas was present in 33% of fluid samples and relative abundances for those taxa were similar to those in our study (nonzero cysts mean relative abundance: 0.00178, range 0.0001–0.004; ref. 14).

In a recent study, Bifidobacterium spp. was found to increase in abundance in the feces of mice with Kras mutations (genetically modified to increase pancreatic cancer) as disease progressed, compared with wild-type mice (15). Furthermore, gut repopulation of the germ-free (Kras) mouse with Bifidobacterium pseudolongum increased T-cell infiltration and tumor growth (15). Similarly, we also noted a higher prevalence for the genus Bifidobacterium in cancer subjects compared with noncancer subjects (Table 2).

Several studies have looked at the involvement of bacteria in biliary and pancreatic diseases and have observed a high number of bacterial taxa present in the calcified pancreatic duct epithelium and in pancreatic abscess (8, 42–45). Anaerobic bacterial taxa have been found at a variable rate in pancreatitis; the results depend on the process for bacterial identification (8, 42, 43). Previous studies have also reported the presence of bacteria in bile (46, 47). In a study of six subjects with gallstones, 16S rRNA gene sequencing identified high relative abundances of Escherichia, Klebsiella, and Pyramidobacter in the bile, and the bacterial profile of the bile was very similar to the duodenum in the same subjects (47). Pyramidobacter species was originally isolated from the oral cavity (48) and was also found in our study samples, but at low levels (<20% of all samples).

Several bacterial taxa we observed with elevated relative mean abundance in RIH samples have been previously identified in immunocompromised patients and are largely believed to be opportunistic pathogens, including Acinetobacter (49) and Kluyvera (50). The genus Gemella, which was found at higher relative abundance in pancreatic cancer subjects when compared with NDRI samples, has been previously associated with a number of infections, including endocarditis, soft-tissue abscesses, empyema, bloodstream infection, and bone infections (51–54). Because our analysis was based on a cross-sectional study design, we expected to identify bacteria that were present as a result of opportunistic nosocomial infections given that the majority of RIH subjects were likely immunocompromised from their cancer. However, our results show that even normal pancreatic tissue harbors a microbial community.

The strength of this study was the collection of specimens specifically for the purpose of microbiome analysis, with precautions made to reduce contamination during collection and processing of samples. Moreover, multiple types of samples were collected on each patient at RIH, including obtaining tissue or swabs from multiple sites, to allow for inter- versus intraindividual differences at different sites. Finally, the multivariable regression analyses was conducted to adjust for potential confounding by known pancreatic cancer risk factors, including BMI and smoking, as well as other factors that may cause bias, including pre-OP EUS and prior chemotherapy.

The major limitation of this analysis was the small number of subjects with pancreatic cysts and pancreatic cancer; despite recruiting 77 subjects, not all subjects had tissue resections during surgery (as more advanced pancreatic cancer patients are often not operable). We did not have sufficient power to examine in great detail the differences in bacterial composition between different pancreatic cancer subtypes, including IPMNs; however, we were the first to include ICD 24 tumors and to explore differences with ICD 25 tumors. Moreover, cancer versus noncancer comparisons of bacterial presence/absence and relative abundances were based on subjects spread across two different data sources (i.e., RIH and NDRI). Differences in microbiota between these two sources may have been due to differences in collection methods and collection times; DNA was extracted from frozen tissue using the same protocol and methods, but tissue samples were either collected during surgery (RIH) or from organs that were rapidly frozen after death (NDRI samples had a mean time of 13 hours to processing of samples). Consequently, it is possible that the identified genera (and overall differences in bacterial taxonomy) merely reflect study-specific differences, rather than real cancer-specific differences.

In this culture-independent study, we detected many bacterial taxa in pancreatic tissue from cancer subjects as well as noncancer subjects. Furthermore, the bacterial profiles in the pancreas were more similar within individuals across different sites of the pancreas (i.e., head, tail, ducts) and duodenum than between individuals at each site. Bacterial taxa known to inhabit the oral cavity were common in the pancreas microbiome and several periodontal pathogens were also identified in pancreatic tissue samples. Further research is needed to address if and how bacteria may be related to pancreatic carcinogenesis or disease progression.

K.T. Kelsey is a consultant/advisory board member at Celintec. No potential conflicts of interest were disclosed by the other authors.

The datasets analyzed for this manuscript will be available through the NCBI under the BioProject accession no.: PRJNA421501. R codes generated to analyze these data will be available upon request.

Conception and design: J. Izard, D.S. Michaud

Development of methodology: E. del Castillo, K.T. Kelsey, J. Izard

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E. del Castillo, B.J. Paster, K.P. Charpentier, K.T. Kelsey, J. Izard, D.S. Michaud

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E. del Castillo, R. Meier, M. Chung, D.C. Koestler, T. Chen, B.J. Paster, K.T. Kelsey, J. Izard, D.S. Michaud

Writing, review, and/or revision of the manuscript: E. del Castillo, R. Meier, M. Chung, D.C. Koestler, B.J. Paster, K.P. Charpentier, K.T. Kelsey, J. Izard, D.S. Michaud

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E. del Castillo, K.T. Kelsey, D.S. Michaud

Study supervision: K.T. Kelsey, J. Izard, D.S. Michaud

Other (obtained funding): D.S. Michaud

The research reported in this publication was supported by the NIH/NCI grants R01 CA166150 and P30 CA168524.

We thank the participants who graciously enrolled in this study. We thank Ms. Priyanka Joshi for her tremendous help with recruitment of subjects at the RIH, and Dr. Ross Taliano for assisting with the preparation of the tissue specimens in the Pathology Department at the RIH. We also thank Drs. Murray Resnick, Kara Lombardo, Alexis Kokaras, Emily Walsh, and Ms. Laura Gantt and Naisi Zhao for their help with this project.

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.