Colorectal cancer is the second leading cause of cancer death worldwide, yet it is one of the most preventable cancers if detected early. The key to prevention is finding - and removing - advanced precancerous lesions (APLs) before they become cancer. APLs are growths such as large polyps or polyps with abnormal cells that have a significant chance of turning malignant if left untreated.
Despite the availability of screening tests, only about 59% of eligible adults in the United States undergo colorectal cancer screening. The main screening options are colonoscopy (highly accurate but invasive, requiring bowel preparation and sedation) and the fecal immunochemical test (FIT) (a stool-based blood test that is noninvasive and easy to perform at home). FIT is widely used but has a major limitation: it misses a large fraction of APLs and early cancers, particularly at the lower sensitivity thresholds used in many national screening programs.
A highly sensitive noninvasive test that outperforms FIT - especially for detecting precancerous lesions before they become cancer - could dramatically increase both screening participation and the preventive impact of existing programs. This study evaluated a new approach that combines stool-derived RNA biomarkers with FIT and a machine learning algorithm.
The test evaluated in this study - called the multimodal stool RNA (mm-stRNA) test - works by measuring the activity levels of five specific genes in cells shed from the lining of the colon into stool. These genes produce messenger RNA (mRNA), which reflects what the cell is currently doing, not just what its DNA blueprint says it could do. Cancer and precancerous cells express these genes at abnormal levels, leaving a detectable signature in stool samples.
The five mRNA biomarkers are: CEACAM5 (a cancer-associated cell surface protein), PTGS2 (an inflammation enzyme linked to early tumor development), ITGA6 (an integrin protein whose loss promotes colon cancer), MACC1 (a key driver of cancer growth and metastasis), and S100A4 (a protein that promotes tumor progression). Each has been independently linked to colorectal cancer in prior research.
Rather than simply adding up these signals, the researchers combined the mRNA biomarker measurements with FIT results using a machine learning (ML) algorithm. The algorithm was trained to recognize patterns in the combined data that distinguish cancer and APLs from normal findings - creating a composite score more powerful than any single marker alone.
The study enrolled 265 subjects through the eAArly DETECT study, a multi-site observational study conducted at 21 US gastroenterology centers. The cohort included 34 people with confirmed colorectal cancer, 68 with APLs, and 163 controls (people with normal colonoscopies or non-advanced polyps). The average age was 60 years, with nearly equal proportions of men and women.
APLs were defined using standard high-risk criteria: adenomas with high-grade dysplasia or carcinoma in situ, villous adenomas, adenomas 1 cm or larger, and serrated lesions 1 cm or larger or with cellular abnormalities. These are the lesion types most likely to progress to cancer if not removed. The study also included a subset of average-risk individuals representative of routine screening populations, alongside a diagnostically enriched group suspected of having cancer or APLs.
RNA was extracted from stool samples preserved in stabilizing collection tubes, then measured by quantitative real-time PCR. FIT was evaluated at two clinical thresholds: 5 micrograms of hemoglobin per gram of stool (the manufacturer's recommended threshold) and 17 micrograms per gram (a commonly used threshold in US screening programs). The ML algorithm was developed using a stratified three-way data split - training, selection, and a blinded test set - to prevent overfitting and ensure unbiased performance estimates.
The mm-stRNA test achieved 97.1% sensitivity for colorectal cancer and 83.8% sensitivity for APLs, with 95.7% specificity. In comparison, FIT at the lower 5 microgram threshold detected 76.5% of cancers and only 45.6% of APLs at 84% specificity. At the higher 17 microgram threshold used in many screening programs, FIT detected 70.6% of cancers and just 36.8% of APLs at 90.8% specificity. The mm-stRNA test substantially outperformed FIT on every metric.
Particularly striking was the mm-stRNA test's performance by APL subtype. For adenomas with high-grade dysplasia, it achieved 100% sensitivity versus 75% for FIT at either threshold. For villous adenomas, it detected 100% versus 41.7 to 50% for FIT. For adenomas 1 cm or larger - the most common APL type - it detected 82.4% versus 29.4 to 41.2% for FIT. Even for serrated lesions, which are notoriously difficult to detect because they rarely bleed, the mm-stRNA test reached 50% sensitivity while FIT caught only 10 to 20%.
For colorectal cancer stages, the mm-stRNA test detected 90.9% of stage I cancers (the most curable stage), compared with just 45.5 to 54.5% for FIT. For stages II through IV, both approaches performed well, but the mm-stRNA test maintained an advantage at all stages. In the younger-age subgroup (45 years and under), the test achieved 83.3% sensitivity for cancer and 100% for APLs - a particularly important finding given rising colorectal cancer rates in younger adults.
Unlike DNA-based stool tests that detect somatic mutations or methylation changes (which may appear only in established cancers), mRNA reflects active gene expression changes that occur even during the earliest stages of tumor initiation. This gives mRNA biomarkers a potential head-start in detecting precancerous changes before DNA alterations are fully established.
Blood-based cancer screening tests are also under development but have been shown in multiple studies to be inferior to stool-based multi-marker tests for detecting APLs and early-stage colorectal cancers. The stool provides a direct sampling of the colon's inner lining, where shed cells from polyps and tumors carry a concentrated molecular signal that may be diluted or absent in blood.
The machine learning component of the mm-stRNA test also offers an advantage over simple biomarker cutoffs: rather than treating each marker independently, the algorithm identifies patterns in the combined data that are more informative than any individual input. The same core feature set - FIT plus the five RNA markers - was consistently selected across multiple algorithm training runs, suggesting the combination is robust and not an artifact of data fitting.
For patients and their families, this study describes a potential future where a simple stool sample collected at home could reliably detect colorectal cancer at its most treatable stages and remove dangerous precancerous polyps before they ever become cancer. The mm-stRNA test requires no bowel preparation, no sedation, and no clinic visit for sample collection - addressing the main barriers that prevent millions of people from getting screened.
The study authors are candid about limitations: the cohort was partially enriched with high-risk subjects to ensure enough cancer and APL cases for algorithm development, which means the sensitivity and specificity numbers may not directly translate to population-level screening among average-risk individuals. A second, larger phase of the eAArly DETECT study has already been launched with a more representative average-risk cohort to address this gap.
If confirmed in that larger study, the mm-stRNA test concept could represent a meaningful advance for colorectal cancer prevention programs globally. By catching more APLs - and catching them earlier - the test could enable clinicians to remove high-risk lesions before they ever become invasive cancer, reducing both colorectal cancer incidence and the deaths it causes.