Stool Test Detects 90% of Colorectal Cancers. The Real Question Is What That Means Next
The researchers who gave us Cologuard and its blood-test successors may soon have a new competitor: bacteria in your stool.
A team at the University of Geneva has developed a machine learning model that reads the subspecies-level signature of gut bacteria in a stool sample to detect colorectal cancer with 90% sensitivity, matching colonoscopy performance for cancer detection and outperforming all existing non-invasive alternatives. The findings appeared in Cell Host Microbe00287-2) in August 2025 under the title "Subspecies of the human gut microbiota carry implicit information for in-depth microbiome research," authored by Matija Tričković et al.
Colorectal cancer is the second leading cause of cancer death worldwide. When caught early it is highly treatable, but colonoscopies are costly, invasive, and widely avoided. The biology has been pointing toward the gut microbiome for years, but turning that knowledge into a diagnostic has been harder than it looks. The challenge, as lead researcher Mirko Trajkovski explains it, is that strains within the same bacterial species can behave in opposite ways. One subspecies of a given bacterium might feed cancer growth; another might do nothing. Looking only at species-level data, which is what most prior studies did, buries that signal.
Trajkovski's lab at UNIGE took a different tack. They built the first comprehensive catalogue of human gut bacteria at subspecies resolution, a level of detail they call operational subspecies units, or OSUs. They named it the HuMSub catalog. The work required developing a new computational method called panhashome, which can rapidly quantify and identify subspecies from metagenomic data without the massive computing overhead that usually comes with subspecies-level analysis.
By combining the HuMSub catalog with existing clinical datasets, first author Matija Tričković, a PhD student in Trajkovski's lab, built a machine learning classifier that could predict colorectal cancer from stool bacteria alone. The result was 90% sensitivity for cancer detection. The team was surprised by the number. "Although we were confident in our strategy, the results were striking," Tričković said. "Our method detected 90% of cancer cases, a result very close to the 94% detection rate achieved by colonoscopies and better than all current non-invasive detection methods."
The authors note a pending patent application related to the use of subspecies in diagnostics. The research is now moving toward the clinic: a clinical trial is being prepared in partnership with Geneva University Hospitals to better define which cancer stages and precancerous lesions the method can detect, and how it performs across different populations.
Here is where warranted skepticism kicks in. The gut microbiome has been linked to colorectal cancer in study after study, and the pipeline from promising finding to commercial diagnostic is long and littered with false starts. The CRC screening market is already crowded and consolidating fast. Exact Sciences received FDA approval for its next-generation Cologuard Plus test just weeks ago, reporting 95% sensitivity for colorectal cancer and 43% sensitivity for advanced precancerous lesions, with 94% specificity. Abbott finalized its $21 billion acquisition of Exact Sciences in March 2026, putting the full weight of a medtech giant behind the leading stool-based test. Freenome, whose blood-based CRC test is expected to receive FDA approval and launch later in 2026, has an exclusive license agreement with Exact Sciences worth up to $885 million. Guardant Health's Shield blood test became the first FDA-approved blood test as a primary screening option for colorectal cancer in 2024. All of these companies have run large validation studies, built distribution networks, and secured regulatory clearance. The Geneva team has a catalog, a paper, and a planned trial.
There is also a gap in the data worth flagging. The 90% sensitivity figure applies only to cancer detection. Pre-cancerous lesion sensitivity, which is the number that determines how many cancers the test might actually prevent by catching precursors early, is not reported anywhere in the available sources. Cologuard Plus catches 43% of advanced precancerous lesions. Until the Geneva team reports its own precancerous figure, the tests cannot be fully compared on the dimension that matters most for population-level screening impact.
What distinguishes this work is the subspecies resolution itself, not just the 90% figure. The paper's core claim00287-2) is that species-level analysis misses functionally relevant variation that exists at the subspecies level, and that the species which contains a cancer-associated subspecies may show no cancer association at all. That is a mechanistically interesting observation. It also means that reproducing the result is harder than running a standard metagenomic pipeline. The HuMSub catalog and panhashome method would need to be independently validated and replicated across diverse cohorts before this becomes a credible commercial claim.
Trajkovski has been careful about the scope of the claim. The subspecies catalog has applications far beyond cancer detection, he said. The same framework could support non-invasive diagnostics for any condition with a gut microbiome signature, which is a long list. That breadth is both the promise and the problem. Gut microbiome research has produced compelling biological insights that take years, sometimes decades, to translate into clinical tools.
But there is a bigger question the paper raises without answering, and it is the one worth sitting with: are the bacterial shifts this test detects a warning signal for cancer, or are they a cause of it?
The distinction matters more than it might seem. If altered gut bacteria are a downstream marker of the same processes that produce cancer, this is a better colonoscopy substitute. Useful, valuable, commercially significant. If they are a causal driver, the commercial logic extends far beyond diagnostics into prevention: probiotic interventions, dietary protocols, prophylactic supplements. An entirely different and potentially larger market. The field knows this is the open question. A 2024 review in Trends in Microbiology put it plainly: proving causality between gut microbiome and colorectal cancer remains challenging, in part because the microbiome varies enormously across individuals and across time. Animal model evidence for probiotic intervention reducing CRC exists; human evidence remains limited. Trajkovski's own phrasing in the available sources was careful, describing what the catalog "could" support rather than what it demonstrates.
What matters for the field is whether subspecies-level resolution proves to be a general principle. If it does, the commercial implications extend well past screening. If it doesn't, this remains a competitive landscape story in a market that Abbott just spent $21 billion to own.
