Colorectal cancer is the second leading cause of cancer-related death worldwide. If detected early, it is often highly treatable. However, the main screening method used today, colonoscopies, can be expensive and uncomfortable, making many people reluctant to undergo the test on time.
Researchers at the University of Geneva (UNIGE) have developed a new approach that could change this. They used machine learning to create the first detailed catalog of all human gut bacteria with enough precision to reveal how different microbial subgroups function in the body. This information was then used to detect colorectal cancer based on bacteria in a simple stool sample, providing a non-invasive, low-cost alternative. The findings, published in Cell Host & Microbe, could also help scientists better understand how the gut microbiome influences overall health and disease.
Why we need better screening tools
Colorectal cancer is often diagnosed late and treatment options are limited. This highlights the urgent need for simpler, less invasive screening methods, especially as cases continue to rise among young adults for reasons that remain unclear.
Scientists have long known that the gut microbiome is involved in colorectal cancer. However, converting that knowledge into actual medical tools has been difficult. One major challenge is that different strains within the same bacterial species can behave very differently. Some contribute to the development of cancer, while others have no effect at all.
Focus on microbiome subspecies
“Rather than relying on analyzes of the various species that make up the microbiome, which do not capture all the meaningful differences, and the bacterial strains that vary widely from individual to individual, we focused on an intermediate level of the microbiome: subspecies,” explains Mirko Trajkovsky, full professor in the Department of Cell Physiology and Metabolism and Diabetes Center at UNIGE Faculty of Medicine, who led the study.
“Vanant resolution is specific and can capture differences in how bacteria function and contribute to diseases such as cancer, but at the same time remains general enough to detect changes between different groups of individuals, populations, or countries.”
Deciphering your gut using machine learning
This study required analysis of large amounts of biological data. “As bioinformaticians, our challenge was to come up with innovative approaches for analyzing large amounts of data,” says Matija Triković, a doctoral student in Trajkowski’s lab and lead author of the study.
“We have successfully developed the first comprehensive catalog of human gut microbiota subspecies and developed accurate and efficient methods to use it in both research and clinical practice.”
A stool test comparable to a colonoscopy
By combining the bacterial catalog with existing clinical datasets, the research team built a model that can identify colorectal cancer using only stool samples. The results exceeded expectations.
“We were confident in our strategy, but the results were surprising,” says Matija Triković. “Our method detected 90% of cancer cases, which is very close to the 94% detection rate achieved with colonoscopy and better than all current non-invasive detection methods.”
Additional clinical data could make the model even more accurate and ultimately match colonoscopy performance. In practice, this type of test can be used for routine screening, and colonoscopies are only used to confirm positive cases.
Expanding beyond cancer detection
To better define the cancer stages and lesions that can be detected with this method, clinical trials are currently being prepared in collaboration with the University Hospitals of Geneva (HUG).
Its impact extends far beyond colorectal cancer. By examining differences between subspecies within the same bacterial species, researchers can begin to uncover how the gut microbiome influences a wide range of health conditions.
“Using the same method, it will soon be possible to develop non-invasive diagnostic tools for a wide range of diseases based on a single microbiome analysis,” concludes Mirko Trajkowski.
