Changes in Abundance of Oral Microbiota Associated with Oral Cancer

Brian L. Schmidt, Justin Kuczynski, Aditi Bhattacharya, Bing Huey, Patricia Corby, Erica Queiroz, K. Florence Nightingale, Alexander Ross Kerr, Mark D. DeLacure, Ratna Veeramachaneni, Adam B. Olshen, Donna G. Albertson, Muy-Teck TehView original
OverviewBalancedalloy voice
Let’s start with the stakes. Oral cancer doesn’t just nibble at the margins of public health; it bites. In the United States, about twenty-two thousand people are diagnosed each year, and nearly ninety percent of those tumors are squamous cell carcinomas. The five-year survival rate sits around forty percent and has barely budged in four decades. Globally, you’re looking at roughly three hundred fifty thousand to four hundred thousand new cases a year. Tobacco and alcohol loom large as risks. But here’s the twist: plenty of patients show up without those exposures, and human papillomavirus—which drives many throat cancers—plays only a minor role in the mouth. That mismatch pushes you to look beyond the usual suspects. Schmidt and colleagues did. They grounded their search in a clinical reality called field cancerization—the idea that genetically altered epithelial "fields" can extend centimeters beyond a visible lesion, sometimes up to about seven centimeters, and seed new tumors or recurrences. If cancer grows in a field, what else in that field changes? The oral microbiome is a prime candidate. Now, comparing one person’s mouth to another’s mouth is messy—diet, brushing, saliva, and the quirks of each person’s immune system. So, Schmidt’s team pulled a neat trick. They sampled each patient’s tumor and, within the same mouth, an anatomically matched, clinically normal spot on the contralateral side. Same person, same patch of the tongue or floor of the mouth, different biology. They swabbed noninvasively, profiled bacterial communities with sixteen S ribosomal RNA sequencing focused on the V4 region, and ran two cohorts to test robustness. The Discovery cohort used four hundred fifty-four pyrosequencing; the Confirmation cohort used Illumina MiSeq. The analytic backbone stayed consistent. They clustered sequences into operational taxonomic units at ninety-seven percent similarity, matched them to the Greengenes reference database in a closed-reference framework, assigned taxonomy with the mothur Bayesian classifier, and compared communities with UniFrac distances. Those are phylogenetically informed measures that ask, "How different are the branch lengths of these microbial trees?" They also brought in healthy participants and took left versus right swabs to benchmark what "no difference" looks like. They collected replicate swabs to check reproducibility. It’s a simple design with a powerful control. The Discovery cohort was small and deep: five patients, tumor versus contralateral normal. On the four hundred fifty-four platform, they generated about one hundred four thousand reads in a single run. After filtering, roughly ninety-three thousand sequences remained, with two hundred seventy-six distinct taxa at the ninety-seven percent threshold. Nearly all the reads fell into five familiar phyla—Firmicutes, Bacteroidetes, Proteobacteria, Fusobacteria, and Actinobacteria. The signal that popped was directional and coherent. Cancers showed a significant drop in Firmicutes, with a p-value of 0.004 that held up after controlling for multiple tests. They also saw a decrease in Actinobacteria that was nearly significant after correction. Fusobacteria trended higher but didn’t cross the usual cutoff. The team also asked a simple probability question: across the major phyla, what are the odds you’d see three or more move in the same direction by chance? Under a coin-flip null, that comes out to about two in a thousand. In other words, the community wasn’t wobbling randomly; it was shifting. Why does that matter? In a healthy mouth, Firmicutes and Actinobacteria include a lot of the go-to commensals—think Streptococcus and Rothia—that help keep the ecosystem balanced. A loss of those groups suggests the tumor microenvironment is less hospitable to the usual caretakers. Schmidt and colleagues then looked one taxonomic step down and found exactly that pattern. Streptococcus was significantly reduced in cancers relative to the matched normal sites, and Rothia dropped as well. Fusobacterium, often flagged in colorectal cancer studies, ran the other way and increased in cancers. Those genus-level changes also peeked into earlier disease. Precancerous lesions, when compared to their own contralateral normals, already showed a reduction in Streptococcus. It’s the kind of shift you want to see if you’re thinking about a continuum—from field changes, to pre-cancer, to frank tumor—rather than a single on or off switch. The team didn’t stop at cataloging winners and losers. They built a multivariate lens tuned to the most informative features—a dozen specific taxa. Eleven were from Actinobacteria and Firmicutes and were consistently decreased: Actinomyces, Rothia, and several Streptococcus groups. One Fusobacterium was increased. Then they asked, if you compare samples by the weighted UniFrac distance—which emphasizes differences in abundant lineages—does that twelve-member "microbial fingerprint" separate tumors from normals and pre-cancers? It did. A principal coordinates analysis using those distances pulled most cancers away from the rest along two axes that explained about half plus a quarter of the variation, respectively. Here’s a kicker: among patients whose cancers had already spread to lymph nodes, five of seven cases clustered tightly together. That tightness hinted that some of the microbiome shift might track with metastatic behavior. It’s not proof; it’s a signal worth following. Replication matters, especially in small discovery sets. So the Confirmation cohort ramped up scale and switched instruments. Using MiSeq paired-end reads on the same V4 target, they sequenced eighty-three samples across cancers, carcinoma in situ, pre-cancers, and healthy controls, pulling in about four and a half million raw reads. After the same closed-reference pipeline, they identified just over two thousand taxa, and they normalized sequencing depth before comparing communities. Those five big phyla dominated again. Replicate swabs showed high reproducibility—correlations mostly above 0.8, with a median around 0.87. That reassured them that the swab-to-sequence process wasn’t lurching. And the core pattern held. Cancers showed reduced Firmicutes and Actinobacteria relative to their anatomically matched normal sites. Pre-cancers also showed reductions in both groups when compared within the same mouth. Healthy participants, sampled on the left and right sides of the same anatomic sites, showed no significant differences in those phyla. That last point is more than a sanity check. It says the contralateral control within a person is stable enough that the observed cancer-linked shifts aren’t just quirks of geography or swabbing. Zooming back to the genus level in this larger cohort, the same names came up. Streptococcus and Rothia were significantly decreased in cancers relative to contralateral normals, and Fusobacterium was increased. Pre-cancers again showed a dip in Streptococcus. These are not exotic microbes. They’re the everyday residents of a healthy mouth. Their coordinated loss, repeated across platforms and cohorts, reads like an ecological story. The tumor and its field seem to disfavor old commensals and make room for lineages often associated with inflammation and dysbiosis. What about that metastasis-tinged signal from the discovery analysis? When Schmidt and colleagues reran the weighted UniFrac ordination using the same twelve discriminating taxa, most cancers still peeled off from normals and pre-cancers. Node-positive cases again tended to cluster. That creates a visual that suggests compositional shifts could be stronger, or at least more consistent, in cancers that have already spread. The axes in that ordination captured a lot of the variance—more than half on the first axis, almost a quarter on the second. So the separation wasn’t a faint whisper. It was a readable pattern. The right interpretation, at this stage, is cautious: association, not causation, with an enticing hint of clinical stratification. All of this rests on a scaffold of choices that the team was transparent about. Closed-reference clustering at ninety-seven percent against Greengenes trades novelty for comparability. It means you only analyze reads that match known taxa at that threshold. The mothur Bayesian classifier, set with an eighty percent bootstrap cutoff, aims to keep assignments trustworthy. UniFrac distances—both unweighted, which treats all taxa equally, and weighted, which weights by abundance—offer complementary views. The weighted version carried the clearest separation here. And subsampling—down to a fixed read depth per sample—kept depth from masquerading as biology. The healthy left-right controls and replicate swabs were smart guardrails. Together, these choices made the cross-platform replication more convincing than a one-off pipeline would. So what do we have when we stand back? An oral cancer-linked microbiome signature that is surprisingly consistent. Losses in Firmicutes and Actinobacteria—pinpointed to Streptococcus and Rothia—and a rise in Fusobacterium are visible not just in tumors but already in pre-cancers when you compare each lesion to its partner site in the same mouth. A small set of taxa can pull cancers apart from normals in a multivariate space, and in node-positive disease that separation tightens. Healthy mouths don’t show left-right differences at the same sites, which makes the within-patient contrast believable. And that entire picture shows up on two sequencing platforms with different error profiles. There are limits, and Schmidt and colleagues said them out loud. The cohorts were small and heterogeneous. The design was cross-sectional, so we can’t say whether microbes drive cancer, ride along, or both. Sixteen S profiling compresses many near neighbors into single bins; that ninety-seven percent cutoff can blur species-level distinctions. Lifestyle factors—smoking especially—could confound. And as with any biomarker story, the real test is prospective and longitudinal: can a signature predict who progresses, who recurs, or who metastasizes? Even with those caveats, the framework is compelling. A non-invasive swab, a within-mouth control, and a handful of readouts that replicate across cohorts—that’s a recipe you can take into clinics for validation. You can imagine using these signatures to monitor a field after surgery, to flag a pre-cancer that’s drifting toward risk, or to complement pathology when two lesions look alike but behave differently. The biology underneath—how a tumor reshapes its microbial neighborhood, and whether those changes feed back on inflammation, metabolism, or immune tone—will take mechanistic work to untangle. But as Schmidt and colleagues showed back in 2014, the map is already readable. The neighborhood changes when cancer moves in. And now that we can measure it clearly, the next step is to learn when that change starts and how long it lasts. Then figure out how to use it to help patients.

Let’s start with the stakes. Oral cancer doesn’t just nibble at the margins of public health; it bites. In the United States, about twenty-two thousand people are diagnosed each year, and nearly ninety percent of those tumors are squamous cell carcinomas.

The five-year survival rate sits around forty percent and has barely budged in four decades. Globally, you’re looking at roughly three hundred fifty thousand to four hundred thousand new cases a year. Tobacco and alcohol loom large as risks.

But here’s the twist: plenty of patients show up without those exposures, and human papillomavirus—which drives many throat cancers—plays only a minor role in the mouth. That mismatch pushes you to look beyond the usual suspects. Schmidt and colleagues did.

They grounded their search in a clinical reality called field cancerization—the idea that genetically altered epithelial "fields" can extend centimeters beyond a visible lesion, sometimes up to about seven centimeters, and seed new tumors or recurrences. If cancer grows in a field, what else in that field changes? The oral microbiome is a prime candidate.

Now, comparing one person’s mouth to another’s mouth is messy—diet, brushing, saliva, and the quirks of each person’s immune system. So, Schmidt’s team pulled a neat trick. They sampled each patient’s tumor and, within the same mouth, an anatomically matched, clinically normal spot on the contralateral side.

Same person, same patch of the tongue or floor of the mouth, different biology. They swabbed noninvasively, profiled bacterial communities with sixteen S ribosomal RNA sequencing focused on the V4 region, and ran two cohorts to test robustness. The Discovery cohort used four hundred fifty-four pyrosequencing; the Confirmation cohort used Illumina MiSeq.

The analytic backbone stayed consistent. They clustered sequences into operational taxonomic units at ninety-seven percent similarity, matched them to the Greengenes reference database in a closed-reference framework, assigned taxonomy with the mothur Bayesian classifier, and compared communities with UniFrac distances. Those are phylogenetically informed measures that ask, "How different are the branch lengths of these microbial trees?" They also brought in healthy participants and took left versus right swabs to benchmark what "no difference" looks like.

They collected replicate swabs to check reproducibility. It’s a simple design with a powerful control.

The Discovery cohort was small and deep: five patients, tumor versus contralateral normal. On the four hundred fifty-four platform, they generated about one hundred four thousand reads in a single run. After filtering, roughly ninety-three thousand sequences remained, with two hundred seventy-six distinct taxa at the ninety-seven percent threshold.

Nearly all the reads fell into five familiar phyla—Firmicutes, Bacteroidetes, Proteobacteria, Fusobacteria, and Actinobacteria. The signal that popped was directional and coherent. Cancers showed a significant drop in Firmicutes, with a p-value of 0.004 that held up after controlling for multiple tests.

They also saw a decrease in Actinobacteria that was nearly significant after correction. Fusobacteria trended higher but didn’t cross the usual cutoff. The team also asked a simple probability question: across the major phyla, what are the odds you’d see three or more move in the same direction by chance?

Under a coin-flip null, that comes out to about two in a thousand. In other words, the community wasn’t wobbling randomly; it was shifting.

Why does that matter? In a healthy mouth, Firmicutes and Actinobacteria include a lot of the go-to commensals—think Streptococcus and Rothia—that help keep the ecosystem balanced. A loss of those groups suggests the tumor microenvironment is less hospitable to the usual caretakers.

Schmidt and colleagues then looked one taxonomic step down and found exactly that pattern. Streptococcus was significantly reduced in cancers relative to the matched normal sites, and Rothia dropped as well. Fusobacterium, often flagged in colorectal cancer studies, ran the other way and increased in cancers.

Those genus-level changes also peeked into earlier disease. Precancerous lesions, when compared to their own contralateral normals, already showed a reduction in Streptococcus. It’s the kind of shift you want to see if you’re thinking about a continuum—from field changes, to pre-cancer, to frank tumor—rather than a single on or off switch.

The team didn’t stop at cataloging winners and losers. They built a multivariate lens tuned to the most informative features—a dozen specific taxa. Eleven were from Actinobacteria and Firmicutes and were consistently decreased: Actinomyces, Rothia, and several Streptococcus groups.

One Fusobacterium was increased. Then they asked, if you compare samples by the weighted UniFrac distance—which emphasizes differences in abundant lineages—does that twelve-member "microbial fingerprint" separate tumors from normals and pre-cancers? It did.

A principal coordinates analysis using those distances pulled most cancers away from the rest along two axes that explained about half plus a quarter of the variation, respectively. Here’s a kicker: among patients whose cancers had already spread to lymph nodes, five of seven cases clustered tightly together. That tightness hinted that some of the microbiome shift might track with metastatic behavior. It’s not proof; it’s a signal worth following.

Replication matters, especially in small discovery sets. So the Confirmation cohort ramped up scale and switched instruments. Using MiSeq paired-end reads on the same V4 target, they sequenced eighty-three samples across cancers, carcinoma in situ, pre-cancers, and healthy controls, pulling in about four and a half million raw reads.

After the same closed-reference pipeline, they identified just over two thousand taxa, and they normalized sequencing depth before comparing communities. Those five big phyla dominated again. Replicate swabs showed high reproducibility—correlations mostly above 0.8, with a median around 0.87. That reassured them that the swab-to-sequence process wasn’t lurching.

And the core pattern held. Cancers showed reduced Firmicutes and Actinobacteria relative to their anatomically matched normal sites. Pre-cancers also showed reductions in both groups when compared within the same mouth.

Healthy participants, sampled on the left and right sides of the same anatomic sites, showed no significant differences in those phyla. That last point is more than a sanity check. It says the contralateral control within a person is stable enough that the observed cancer-linked shifts aren’t just quirks of geography or swabbing.

Zooming back to the genus level in this larger cohort, the same names came up. Streptococcus and Rothia were significantly decreased in cancers relative to contralateral normals, and Fusobacterium was increased. Pre-cancers again showed a dip in Streptococcus.

These are not exotic microbes. They’re the everyday residents of a healthy mouth. Their coordinated loss, repeated across platforms and cohorts, reads like an ecological story.

The tumor and its field seem to disfavor old commensals and make room for lineages often associated with inflammation and dysbiosis.

What about that metastasis-tinged signal from the discovery analysis? When Schmidt and colleagues reran the weighted UniFrac ordination using the same twelve discriminating taxa, most cancers still peeled off from normals and pre-cancers. Node-positive cases again tended to cluster.

That creates a visual that suggests compositional shifts could be stronger, or at least more consistent, in cancers that have already spread. The axes in that ordination captured a lot of the variance—more than half on the first axis, almost a quarter on the second. So the separation wasn’t a faint whisper.

It was a readable pattern. The right interpretation, at this stage, is cautious: association, not causation, with an enticing hint of clinical stratification.

All of this rests on a scaffold of choices that the team was transparent about. Closed-reference clustering at ninety-seven percent against Greengenes trades novelty for comparability. It means you only analyze reads that match known taxa at that threshold.

The mothur Bayesian classifier, set with an eighty percent bootstrap cutoff, aims to keep assignments trustworthy. UniFrac distances—both unweighted, which treats all taxa equally, and weighted, which weights by abundance—offer complementary views. The weighted version carried the clearest separation here.

And subsampling—down to a fixed read depth per sample—kept depth from masquerading as biology. The healthy left-right controls and replicate swabs were smart guardrails. Together, these choices made the cross-platform replication more convincing than a one-off pipeline would.

So what do we have when we stand back? An oral cancer-linked microbiome signature that is surprisingly consistent. Losses in Firmicutes and Actinobacteria—pinpointed to Streptococcus and Rothia—and a rise in Fusobacterium are visible not just in tumors but already in pre-cancers when you compare each lesion to its partner site in the same mouth.

A small set of taxa can pull cancers apart from normals in a multivariate space, and in node-positive disease that separation tightens. Healthy mouths don’t show left-right differences at the same sites, which makes the within-patient contrast believable. And that entire picture shows up on two sequencing platforms with different error profiles.

There are limits, and Schmidt and colleagues said them out loud. The cohorts were small and heterogeneous. The design was cross-sectional, so we can’t say whether microbes drive cancer, ride along, or both.

Sixteen S profiling compresses many near neighbors into single bins; that ninety-seven percent cutoff can blur species-level distinctions. Lifestyle factors—smoking especially—could confound. And as with any biomarker story, the real test is prospective and longitudinal: can a signature predict who progresses, who recurs, or who metastasizes?

Even with those caveats, the framework is compelling. A non-invasive swab, a within-mouth control, and a handful of readouts that replicate across cohorts—that’s a recipe you can take into clinics for validation. You can imagine using these signatures to monitor a field after surgery, to flag a pre-cancer that’s drifting toward risk, or to complement pathology when two lesions look alike but behave differently.

The biology underneath—how a tumor reshapes its microbial neighborhood, and whether those changes feed back on inflammation, metabolism, or immune tone—will take mechanistic work to untangle. But as Schmidt and colleagues showed back in 2014, the map is already readable. The neighborhood changes when cancer moves in.

And now that we can measure it clearly, the next step is to learn when that change starts and how long it lasts. Then figure out how to use it to help patients.

More in Dentistry