For years, researchers have searched for gut microbial signatures of colorectal cancer by asking: How do people with cancer differ from healthy people? It seems like the obvious comparison, but perhaps it is not the most useful one to answer; those with cancer differ from those without in many ways. One of our most effective ways of fighting cancer is detecting and treating it early on, before it has the chance to spread. For colorectal cancer, the most common way of organizing such screening is based on detection of minute traces of blood in the stool, since we know that premalignant and malignant tumors are prone to bleed. The issue is that most people with blood in their stools do not have cancer or cancer precursors. So, shifting our perspective from identifying differences between those with cancer and healthy individuals, we instead asked whether it is possible to identify the few people developing cancer among the many who test positive for hidden blood in their stool.
We were in a unique position to answer this question. Norway had recently launched a large colorectal cancer screening trial, and crucially, all stool samples had been carefully stored. This gave us access to thousands of samples collected before diagnosis, creating an exceptional resource for studying the gut microbiome in screening. This allowed us to focus on the group that matters most for screening: people who tested positive for blood in stool, but among whom only a minority actually had cancer or precancerous lesions. Over five years, we carefully designed the study, collected samples, generated shotgun metagenomes, and assembled detailed clinical and lifestyle data to account for the many factors that shape the gut microbiome in a real-world screening population.
Setting out with this shift in perspective, we expected our results to resemble studies that distinguish cancers and controls using microbial profiles with high accuracy. However, even with similar methodology, our results showed only moderate accuracy, or, as one collaborator said, our predictions were unimpressive. It took us a while and a bit of frustration to appreciate that our new perspective implied that seemingly unimpressive accuracy could be meaningful. Beyond the immediate prediction results, we were uncovering findings that helped explain the role of the gut microbiome in colorectal cancer screening and, more broadly, in disease development. Our findings also made us look more critically at how previous studies had selected their comparison groups. In many cases, participant inclusion criteria were unclear, making it difficult to know whether reported microbial signatures reflected cancer itself or other differences between groups.
Changing the comparison group changed not only what we could predict, but also what biology we were able to see. We saw that many bacteria identified in previous studies as being associated with cancer, were less common in cancer in our dataset. These included bacteria such as Clostridium symbiosum, Flavonifractor plautii, Eggerthella lenta, and Hungatella hathewayi.
Our first reaction was the same as everyone else's: Could this be a technical artifact since we used the stool samples contained in the screening tests (FIT) themselves rather than standard sampling tubes for microbial profiling? After careful analyses, the answer was no; the microbial communities looked very similar to those reported from conventional stool samples, adding this information to our growing supplementary information. Instead, the discrepancy appeared to stem from the populations being compared. When we focused on individuals who all had a positive FIT test, several bacteria previously thought to be cancer-associated no longer distinguished cancer and precancer from non-cancer, suggesting that they may reflect poor gut health rather than cancer itself.
Other bacteria linked to colorectal cancer were confirmed though, including
the famous Escherichia coli. Some E. coli strains produce colibactin, which has been implicated in colorectal cancer. By separating E. coli strains carrying the colibactin gene cluster (pks) from those lacking it, we found two contrasting patterns: pks-negative E. coli was associated with a lower likelihood of colorectal neoplasia, whereas pks-positive E. coli showed the opposite trend. This suggests that the presence of E. coli and its pks status may be linked not only to the development of disease, but could potentially contribute to future screening approaches. It is a finding that we think will shape the next phase of our research.
Importantly, the microbiome did add information beyond the blood signal alone. By combining FIT measurements with microbial profiles, we improved the identification of premalignant changes in the colon. Although these models are not yet ready for clinical use, they demonstrate the potential of combining traditional screening approaches with biological information from the microbiome. Further improvement of the discriminatory power is then crucial, along with validation across different screening programs and development of a cost-effective test.
Improving microbiome-based prediction models requires more than adding more samples, new variables, and using more advanced machine learning tools. It requires understanding the biology behind the signals. That is why we set out to do much more than optimize prediction accuracy. We integrated detailed colonoscopy findings, clinical characteristics, lifestyle information, and shotgun metagenomic profiles to identify the factors shaping the colorectal cancer–associated microbiome.
The opportunities for exploration seemed endless. Some analyses grew into stories of their own, some found a home in the supplementary material, and others became key pieces of the paper. For example, we examined whether microbial signatures varied with lesion size and location. Surprisingly, lesion size had little impact, whereas location mattered. The microbial signal became stronger in the distal colon, raising intriguing questions about whether this reflects biological differences in carcinogenesis or simply the gradual dilution of microbial signatures as stool travels through the colon. Only by understanding these biological signals can we build screening models that are both accurate and biologically meaningful.
This study reminded us that biomarkers are not simply waiting to be discovered. They are shaped by the questions we ask, the populations we study, and the perspective we choose. Changing the comparison group changes the biology we are able to see