Study 1. Discovery cohort

We swabbed the oral cancer lesion and a corresponding anatomically matched clinically normal tissue area from the Discovery Cohort of five patients (Table 1). Using the Roche GS Junior instrument to perform pyrosequencing, we obtained, in a single run, a total of 104,380 sequences from amplicons that spanned the V4 hypervariable region of the bacterial 16S small ribosomal subunit (Table S7). The number of raw sequence reads varied by >10-fold across samples, ranging from 1,231 to a maximum of 17,682 raw reads. Quality filtered sequences were searched against the Greengenes reference database of 16S sequences, clustered at 97%, and Operational Taxonomic Units (OTUs) were assigned taxonomic classification using mothur's Bayesian classifier. Of the 92,987 sequences that passed quality filtering, 81,308 were similar to known bacteria and most could be classified to the genus level (65,037) with fewer classified at the species level (17,115). Sequence coverage was variable across samples; the number of reads per sample assigned to OTUs (excluding those filtered due to poor quality or lack of a related sequence in the Greengenes reference database) ranged from 1,038 to a maximum of 14,359, and comprised 76–85% of raw sequences (Table S7). A total of 276 OTUs were identified (per sample range, 37–161, (Table S8). Rarefaction analysis performed at the family level demonstrated a fairly wide range of α diversity with the number of detected families ranging from ∼15–28 (Figure S1a). Three patient samples, normal and cancer from patient 117 and the normal sample from patient 142, plateaued at fewer families than the other samples, indicating somewhat reduced diversity in these samples. All samples plateaued to some extent, though not entirely; further sequencing would likely reveal additional families. On the other hand, pyrosequencing noise and PCR errors could erroneously increase OTU numbers [31]. A similar effect was seen at the genus level (Figure S1b).

Diversity of microbiomes associated with anatomically matched oral cancer and normal samples

The OTUs of the cancer and clinically normal samples in the Discovery cohort were classified into 12 phyla (Table S7). The majority belonged to one of five phyla (99.2%, normal, 98.0% cancers) with the more abundant phyla being Firmicutes, Bacteroidetes, Proteobacteria, Fusobacteria, and Actinobacteria (Table 7, Figure 1). Although the distribution in cancer and clinically normal samples of these five common phyla varied amongst individuals (Figure 1), in all patients, we observed a significant reduction in the abundance of Firmicutes and Actinobacteria in cancer compared to the anatomically matched contralateral clinically normal patient sample, (p = 0.004, FDR adjusted p = 0.02 and p = 0.028, FDR adjusted p = 0.07, respectively). We also observed that the proportion of Fusobacteria was increased in all patients, but the change in abundance did not reach statistical significance (p = 0.074, Figure 2). We note, however, that to observe consistent changes in abundance of the three phyla in this small cohort is highly significant. Ignoring interactions among the relative abundances of different phyla, a binomial test yields p = 0.0022 as the probability (under the null hypothesis) of having three or more phyla where (all five patients share an increase) or (all five share a decrease).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Distribution of phyla in normal and cancer samples in Study 1.

The relative distribution of phyla (percent of sequences) is shown for each patient sample with clinically normal samples shown together on top and cancer samples on the bottom.

https://doi.org/10.1371/journal.pone.0098741.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Change in relative abundance of phyla associated with cancer compared to anatomically matched contralateral clinically normal samples in Study 1.

(a – e) Relative abundance of each of the five more abundant phyla in cancers compared to clinically normal samples from each of five patients. Note, that data are shown on different scales, reflecting the abundance of the phyla. The magnitudes of the changes in abundance are clearly greater than the statistical counting noise, as indicated by the error bar estimates, which are based on the square root of the actual number of reads. (f) Change in relative abundance shown as the difference in abundance of phyla associated with cancers compared to anatomically matched contralateral clinically normal samples. In cancers, decreases in the relative abundance of Firmicutes and Actinobacteria were seen in all patients, while the relative abundance of Fusobacteria was elevated in cancers from all patients.

https://doi.org/10.1371/journal.pone.0098741.g002

Study 2. Confirmation cohort

To confirm these initial observations, we performed Study 2. As before, we swabbed both the lesion (cancer or pre-cancer) and a contralateral anatomically clinically normal matched site. We swabbed the left and right sides of the lateral tongue and floor of mouth of the healthy individuals. Using the Illumina MiSeq instrument, we sequenced amplicons spanning the 16S rDNA V4 hypervariable region. We obtained 4,486,196 raw sequence reads with a range of 31,109 to 125,847 reads per sample after excluding two samples that failed in sequencing (Tables S9 – S13). We assigned taxonomic classification to the OTUs as before. Of the sequences that passed quality filtering, 4,444,432 were similar to known bacteria and most could be classified to the genus level (4,148,785) with fewer classified at the species level (1,650,037). We identified a total of 2,107 OTUs (per sample range, 90–482, Tables S9 – S13). Rarefaction analysis demonstrated a fairly wide range of α diversity with almost all samples plateauing to some extent (Figure S2). As discussed above, sequencing noise and PCR errors may increase OTU numbers [32].

Diversity of microbiomes associated with anatomically matched samples from oral cancer, pre-cancer and healthy normal individuals in Study 2

We first determined that data collected on nine of the 10 Discovery Cohort samples that were successfully profiled in Study 2 (Table S9) were highly correlated with the original data obtained by 454 pyrosequencing (Figure S3). We then considered the Confirmation Cohort comprised only of samples from cancer patients not included in the Discovery Cohort, i.e., we excluded the nine Discovery cohort samples (Table 2, Table S9). We again found that the majority of OTUs (61–100%) belonged to one of the five more abundant phyla (Firmicutes, Bacteroidetes, Proteobacteria, Fusobacteria, and Actinobacteria) (Figure 3a, Table S14). We then asked if there were significant reductions in the abundance of Firmicutes and Actinobacteria in cancer compared to the anatomically matched contralateral clinically normal patient sample. Indeed, as we had observed in the Discovery cohort (Study 1), there was a significant reduction in abundance of these two phyla in cancers compared to the anatomically matched contralateral clinically normal patient samples (p = 0.042 and p = 0.004, respectively), confirming our initial observations (Figure 3b, Figure S4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Distribution of phyla in cancer, pre-cancer and healthy normal samples in Study 2.

(a) Shown is the relative distribution of phyla (percent of sequences). For cancers, we included only patients for which both the cancer and contralateral clinically normal samples were available. (b) Change in relative abundance of phyla shown as the difference in abundance of phyla associated with cancers or pre-cancers compared to anatomically matched contralateral clinically normal samples. For healthy normal samples, we compared left and right sides of the lateral tongue or floor of mouth.

https://doi.org/10.1371/journal.pone.0098741.g003

In Study 2, we also observed that the majority of OTUs in samples from pre-cancer patients and healthy normals (Figure 3a) belonged to one of the five phyla Firmicutes, Bacteroidetes, Proteobacteria, Fusobacteria, and Actinobacteria (pre-cancers  = 99–100%, healthy normal  = 87–100%). We again asked if significant reductions in abundance of Firmicutes and Actinobacteria might also be present when comparing pre-cancers and anatomically matched contralateral clinically normal patient samples. We did find that abundance of Firmicutes and Actinobacteria was reduced (p = 0.048 and p = 0.037, respectively), suggesting that changes in abundance of these two phyla may occur early (Figure 3b). By contrast, we did not find significant differences in abundance of these two phyla or any of the five more abundant phyla when comparing the left and right sides of the lateral tongue and floor of mouth of healthy normal individuals (Figure 3b).

To investigate changes in abundance at the genus level for the five more abundant phyla, we considered patient matched samples from the Confirmation Cohort and also included four of the five Discovery Cohort cases that were successfully analyzed in Study 2 (14 cancers from 13 patients, Table S9). We normalized the number of OTUs to one million counts, (Table S15) and for each phylum, we determined changes in abundance of the genera that represented >10% of OTUs in more than 20% of samples (Figure S5 and S6). Significant reduction in abundance was observed for Streptococcus (p = 0.003) and Rothia (p = 0.021) in cancers relative to anatomically matched clinically normal samples (Table S16). By contrast, we observed increased abundance of Fusobacterium (p = 0.044) relative to matched clinically normal samples from the cancer patients. In pre-cancers, we observed significantly reduced abundance of Streptococcus (p = 0.042) (Table S16). We found no significant differences in abundance of these more common genera when comparing abundance in samples taken from the left and right sides of the lateral tongue and floor of mouth of healthy individuals (Table S16).

We also noted that although we found no consistent changes in abundance of Bacteroidetes when comparing within individuals (Figure 3b), samples from cancer and pre-cancer patients (both lesion and anatomically matched contralateral clinically normal tissue) were associated with increased abundance of Bacteroidetes compared to samples from healthy normal individuals (Figure 3a). Prevotella species, in particular, differed and included, for example, OTUs corresponding to P. intermedia, P. melaninogenica, P. nanceiensis, P. oris, P. tannerae and unclassified species (Figure S5, Table S15). Elevated levels of P. melaninogenica have been reported previously as a potential salivary biomarker of oral cancer [33] and P. intermedia is a periodontal pathogen [34]. Further studies will be required to understand the contributions of general (Bacteroidetes) and lesion-specific changes (Actinobacteria, Firmicutes) in the microbiome of oral cancer and pre-cancer patients.

OTUs distinguishing cancer from anatomically matched contralateral normal patient samples

To address the need for biomarkers to predict behavior of oral cancers that could be assayed by non-invasive tests, we asked whether cancer samples could be distinguished by bacterial composition. Considering cancer/normal paired samples, we identified 11 OTUs from the phyla Actinobacteria (Actinomyces and Rothia, 2 OTUs from each genus) and Firmicutes (Streptococcus, 7 OTUs) that were significantly decreased in cancers and one OTU from the phylum Fusobacteria (Fusobacterium) that was increased in cancers compared to anatomically matched contralateral normal patient samples (Table S17). Weighted UniFrac principal coordinates analysis (PCoA) based on this set of OTUs separated most cancers from normal and pre-cancer samples with five of the seven lymph node positive cases forming a tight cluster in the lower right corner of the plot (Figure 4). Metastasis to the cervical (neck) lymph nodes is a major determinant of oral cancer patient survival [35]. The current lack of reliable methods to assess risk of metastasis results in patients being routinely subjected to additional surgery to remove the lymph nodes, even though the majority will not benefit from the procedure. The tight clustering of samples from node positive patients in Figure 4 suggests that shifts in the composition of the oral cancer microbiome may also hold promise as a tumor associated biomarker of risk of metastasis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Distinguishing cancer and normal samples.

PCoA based on Weighted UniFrac distance between samples given abundance of 12 OTUs. Axis 1 (PCoA1): 54% of variation explained. Axis 2 (PCoA2): 24% of variation explained. N0 and N+ indicate the nodal status of the cancer patient, N0  =  node negative, N+  =  node positive. Cancer control and pre-cancer control are contralateral clinically normal patient samples. Other identifies samples from healthy normal individuals.

https://doi.org/10.1371/journal.pone.0098741.g004

Whole micrbiome – β diversity

To investigate sample-to-sample dissimilarity, we subsampled the data by randomly selecting 31,109 sequences from each community to adjust for variation in sequencing depth (Table S18). When considering the three sample groups (Cancers, Healthy Normals, and Pre-cancers), we observed significant microbiome differences for patient identity based on both weighted UniFrac (abundance) and Unweighted UniFrac (presence/absence) metrics (Table S19, Figures S7 and S8). No other comparisons, such as left vs. right, number of sequences or lesion (cancer or pre-cancer) vs. control normal sites revealed significant differences. These observations are consistent with other studies that have highlighted the inter-individual differences in the oral microbiome [24]. Further, they support our study design, which measures changes in the microbiome within individuals, i.e., using each patient as his or her own control.

Ethics statement

The study was approved by the Institutional Review Board of New York University College of Dentistry and all patients provided written informed consent.

Study population and biospecimen collection

In Study 1, samples from cancer and anatomically matched contralateral clinically normal regions of the oral cavity were obtained from five patients with oral cancer who were referred to the Bluestone Center for Clinical Research, New York University College of Dentistry in the period July 2011 to March 2012. Individuals enrolled in Study 2 included cancer and pre-cancer patients who were referred to the Bluestone Center for Clinical Research during the period April 2011 to August 2012 and individuals with no history of oral cancer (healthy normal). For this study, we used a well-defined clinical protocol for swabbing the oral lesion and a contralateral normal site. In addition to sampling the microbiome, the procedure provides a tumor genomic DNA sample with a DNA copy number profile that, when tested, reflects the profile obtained with the genomic DNA isolated from an incisional biopsy of the lesion; however, the procedure is not designed to optimally sample the entire oral microbiome.

Specifically, to collect samples from the cancer or pre-cancer, the lesion was dried by blotting with gauze and then the lesion was stroked with an Isohelix SK-2 swab (Cell Projects Ltd., Harrietsham, UK). The swab was held at an angle of approximately 20° relative to the surface of the lesion and one side of the swab was stroked across the lesion 10 times applying gentle downward pressure. The swab was then rotated 180° and the other side of the swab was stroked 10 times across the lesion in the same manner. Anatomically matched contralateral normal tissue and tissues from healthy normal individuals were sampled using an Isohelix swab, the brush from the OralCDx Brush Test (CDx Diagnostics, Suffern, NY) or a Rovers Orcellex brush (Rovers Medical Devices B.V., Oss, The Netherlands). The swabs and the brushes provided with the OralCDX Brush Test were placed into the tube provided as part of the Isohelix SK-2 swab or a microfuge tube, respectively and kept on ice for no more than 30 minutes prior to adding the Isohelix cell lysis and DNA stabilization solution (LS solution, 500 µL) and proteinase K solution (20 µL) both provided in the Isohelix DSK-2 kit. Samples were stored at room temperature and subsequently shipped to the University of California San Francisco for nucleic acid extraction.

To assess the reproducibility of the swab procedure, we included six pairs of replicate swabs taken at the same visit as the original swab as part of Study 2 (Tables S6 and S13). Five pairs of samples showed good concordance (Pearson correlation, minimum  = 0.628, first quartile  = 0.801, median  = 0.868, mean  = 0.843, third quartile  = 0.912, maximum  = 0.962), indicating that in most cases the procedure reproducibly samples the microbial communities of oral lesions and normal sites. Sample p113N2, however, showed little concordance with p113N1 (R²  = 0.116), but good concordance with p113C2 (R²  = 0.702). In other analyses, p113N2 behaved similarly to the cancer samples from patient 113 (data not shown), suggesting an error had occurred at some point during collection and processing of this sample.

DNA extraction

Swabs in Isohelix tubes with 500 µL solution were vortexed, then centrifuged briefly and the liquid removed to a fresh 1.5 mL microfuge tube. This process was repeated 2–3 more times before transferring the swab to a fresh 1.5 mL microfuge tube and centrifuging at 14,000 rpm for 1 minute to extract the remaining cell lysate from the swab. The DNeasy blood and tissue kit (Qiagen Corp.) was used to extract nucleic acid from the solution recovered from the swab, following the manufacturer's protocol and including an initial incubation with Proteinase K (addition of 20 µL solution provided as part of the DNeasy blood and tissue kit) at 56°C for 10 minutes. The DNA concentration was determined using the Qubit v2.0 fluorometer.

16S rRNA amplicon preparation and 454 pyrosequencing

The V4 region of the small subunit ribosomal RNA gene (16S rRNA) was selected for study, because it is suited to analysis on multiple high throughput platforms yielding short reads. It has been reported to give low error rates when assigning taxonomy [51], [52] and to be suitable for community clustering [53]. For Study 1, the region was amplified using the primer set 515F (5′-GTGCCAGCMGCCGCGGTAA-3′) and 806R (5′-GGACTACVSGGGTATCTAAT-3′) [32], [54]. The complete forward primer construct (5′-3′) consisted of the Roche 454 Life Sciences Sequencing FLX Adaptor A (Roche Applied Science, Branford, CT, USA), a 12 bp Golay nucleotide barcode, and a GT linker followed by the 515F primer sequence. The 806R was similarly constructed but incorporated the Roche 454 Life Sciences Sequencing FLX Adaptor B, and a GC linker followed by the 806R primer sequence.

For each patient sample in Study 1, three separate amplifications were carried out in 25 µL reaction volumes. Each reaction contained 1 7L each of forward and reverse primers at a 10 µM concentration, 10 µL of 5 Prime HotMasterMix (5 Prime Inc., Gaithersburg, MD, USA), and 1 µL of extracted genomic DNA from each patient sample (concentration range  = 2.3–64.2 ng/ µL). After denaturation at 94°C for 3 minutes, 35 cycles of incubation at 94°C for 45 seconds, 50°C for 30 seconds, and 72°C for 1.5 minutes were carried out, followed by a 10 minute final elongation step at 72°C. The replicate reactions from each patient sample were pooled together by sample and quantified using the Quant-IT PicoGreen dsDNA Assay (Invitrogen, Carlsbad, CA, USA). Barcoded samples were then normalized to equimolar amounts to ensure equal sequencing depth for each sample, and finally pooled into one combined sample, which was further purified using the UltraClean PCR Clean-Up Kit (Mo Bio Laboratories, Inc., Carlsbad, CA). A final quantification was performed using the Quant-IT PicoGreen dsDNA Assay. Pooled amplicon libraries were sequenced unidirectionally in a single run on the Roche GS Junior instrument (100,000 reads per run) using the A beads for emulsion PCR.

For Study 2, the 16S V4 region was amplified using 515F and 806R fusion primers tailed with sequences to incorporate Illumina flow cell adapters with indexing barcodes [32]. Each sample was PCR amplified in a 10 µL reaction. The reaction contained 1 µL of a mixture of forward and reverse primers at a 5 µM concentration, 5 µL of Qiagen HotStar Master Mix (Qiagen Inc., Gaithersburg, MD, USA), 0.5 µL genomic DNA from each patient sample (20–50 ng) and nuclease free water. After incubation at 94°C for 15 minutes, 35 cycles of incubation at 94°C for 30 seconds, 50°C for 30 seconds and 72°C for 30 seconds were carried out, followed by incubation at 72°C for 10 minutes and storage at 4°C. Primer dimers and low molecular weight products were removed using Agencourt Ampure Beads (Beckman Coulter, Inc., Indianapolis, IN) and samples were quantified and quality checked for amplicon size using the Agilent Bioanalyzer. Amplicons (1×10¹⁰ molecules) were pooled. The pooled sample was diluted to 3.5 pM and 8 pM phiX DNA was spiked in to a final concentration of 2 pM. Amplicons were sequenced from both ends for 250 cycles using primers designed for paired-end sequencing avoiding the PCR amplification primers. The indexing barcode was sequenced using a third sequencing primer and an additional 13 cycles. Sequence data are available at the European Bioinformatics Institute, accession number PRJEB4953.

Data analysis

We used QIIME [55] and custom scripts to process the sequencing data. Sequences were quality filtered and de-multiplexed using exact matches to the supplied DNA barcodes. Paired end sequencing performed on the MiSeq resulted in forward and reverse reads with some overlap in the V4 hypervariable region of the 16S. Each forward and reverse pair were stitched together using pandaseq with parameters "-F -l 220 -L 300 -t 0.5". Read pairs without sufficient overlap to allow stitching were discarded. For both MiSeq and 454 sequences, the resulting sequences were then searched against the Greengenes reference database of 16S sequences from 4 February 2011 [56], clustered at 97% by uclust. Sequences not matching any Greengenes reference sequence at >97% sequence identity were discarded (closed-reference OTU picking). The longest sequence from each Operational Taxonomic Unit (OTU) thus formed was then used as the OTU representative sequence, and assigned taxonomic classification using mothur's bayesian classifier, trained against the Greengenes reference database of 16S sequences from 4 February 2011 [56], clustered at 98% (bootstrap confidence level  = 80%). We note that ambiguous sequence identifications between species, for example, were addressed by not assigning any species identification to such reads, rather they were identified at broader levels (e.g., phylum, genus) where confident assignments could be made.

After pooling sequences into OTUs as described above, we generated a table listing the abundance of each OTU in each microbial community in this study (the OTU table). To obtain a summary statistic representing the overall dissimilarity between any two microbial communities in this study, we used the phylogenetically aware dissimilarity measures Unweighted UniFrac [57] and Weighted UniFrac [58]. Because some OTUs are closely related, while others are more distantly related, these two community-wide dissimilarity measures provide a more informative assessment of community resemblance than a more naive approach of e.g., counting shared OTUs between communities. Unweighted UniFrac considers only presence/absence of taxa and counts the fraction of the branch length unique to either community. In weighted UniFrac, the branch distance is weighted by difference in abundance. In both cases, the result is a scalar measure of dissimilarity between each pair of communities in this study, and is based on the information contained in the OTU table as well as the phylogeny relating the OTUs. We used the supplied Greengenes phylogeny (which is a curated version of a FastTree constructed phylogeny built from full length Greengenes 16S/SSU sequences). Because these measures are sensitive to differences in sequencing depth amongst samples, we randomly selected 1,038 (Study 1) or 31,109 (Study 2) sequences from each sample (subsampled dataset), and generated the distance matrix by a pair-wise inter-comparison of the profiles in the subsampled dataset. We used PCoA to compare samples based on the Unweighted and Weighted UniFrac distances. We clustered samples by applying the average neighbor (HC-AN) method as implemented in hclust (R package http://www.R-project.org) to the distance matrix.

Statistical analysis

Statistical tests with p-values less than 0.05 were considered significant. We used a binomial test to evaluate the significance of finding consistent changes in abundance of three phyla in the five Discovery cohort patients (Study 1). In the absence of any cancer effects, we expect that in a single patient, the chance that a microbial phylum would increase in the cancer sample is 0.5, as is the chance that the phylum would decrease. Although our counts are discrete, the probability of "no change in abundance at all" is so improbable that we can neglect it. Therefore we can calculate the probability that (all five patients show an increase or all five patients show a decrease) – a two-sided binomial with p  = 2*(1/2)∧5  = .0625. In our data, three of the five major phyla show such a unanimous shift (all five patients show an increase or all five patients show a decrease). The probability that (three or more phyla have such an unanimous shift) absent any cancer effects is thus the CDF of a binomial with B(n = 5,p = 0.0625) to x = 3. p = 0.0022 in this case.

The adonis function implemented in the vegan package [59] was used to test for whole microbiome differences between sample groups.

For comparisons of specific sample groups, we used the relative abundance of OTUs after normalizing samples to one million counts. The normalization was done only for ease of display – no imputation of missing sequences was performed. A paired t-test was performed to identify OTUs significantly increased or decreased across cancer samples relative to the normal samples from the same patient, and Welch's t test (a two tailed, unequal variance test) was used to identify OTUs significantly increased or decreased across samples from two groups of patients.