Colorectal cancer (CRC) is the third most common cancer and the second-leading cause of cancer death worldwide, with more than 1.9 million new cases diagnosed annually. When caught early, survival rates are high, but CRC typically develops slowly over 5 to 10 years from precancerous polyps, creating an important window for detection and intervention. The challenge is finding reliable, accessible ways to screen people during that window.
Current screening methods have significant limitations. Colonoscopy is the gold standard but requires bowel preparation, sedation, and carries small procedural risks, leading many patients to avoid or delay it. Stool-based tests such as fecal immunochemical tests (FIT) are easier to use but have reduced sensitivity for early-stage tumors and can produce false positives from non-cancer-related bleeding.
Standard blood markers like carcinoembryonic antigen (CEA) and CA19-9 are useful for monitoring known disease but lack the sensitivity and specificity needed for primary screening. This creates a genuine clinical need: a blood test that can reliably detect early CRC without the burden of colonoscopy.
The researchers in this study turned to metabolomics, the large-scale study of small molecules produced by the body's metabolic processes. Because cancer fundamentally alters how cells produce and consume energy and nutrients, the pattern of molecules circulating in plasma is expected to change in CRC patients. The goal was to identify which specific molecules are most informative and build a practical detection panel.
The study enrolled 172 CRC patients and 115 healthy controls in a discovery cohort from National Cheng Kung University Hospital in Taiwan. An independent validation cohort from E-Da Hospital included 47 CRC patients and 47 healthy controls. Plasma samples were analyzed using a combination of untargeted and targeted liquid chromatography-mass spectrometry (LC-MS), a highly sensitive technique for detecting and measuring thousands of molecules simultaneously.
The initial untargeted analysis produced 146,880 distinct spectral features across all samples, a number far too large to analyze manually. To handle this high-dimensional data without bias, the researchers used a random forest algorithm, a machine learning method that builds many decision trees and combines their outputs to robustly rank which features most reliably distinguish CRC patients from healthy individuals.
From the 1,007 features identified as reproducibly discriminatory by random forest analysis, a receiver operating characteristic (ROC) analysis further filtered the list to 18 high-confidence metabolite candidates with individual area under the curve (AUC) values above 0.90. Five of these were selected for targeted quantitative validation because commercial reference standards were available for them, enabling precise measurement.
After selecting the five metabolites, the team built a logistic regression model that combined all five markers into a single diagnostic score. This combined model was then evaluated for its ability to distinguish CRC patients from healthy controls in the independent validation cohort, ensuring the results were not simply artifacts of the original dataset.
All five selected metabolites, N-methylcytisine, 2-piperidone, theophylline, DL-norleucine, and linolenic acid, were consistently and significantly lower in the plasma of CRC patients than in healthy controls across both the discovery and validation cohorts. This consistent reduction across two independent groups of patients strengthens confidence that these differences reflect genuine cancer-associated changes rather than random variation.
Individual AUC values in the validation cohort ranged from 0.690 (N-methylcytisine) to 0.944 (2-piperidone). Three of the five metabolites achieved AUC values above 0.90, indicating excellent individual discriminatory power. The combined five-metabolite panel performed even better, achieving an AUC of 0.968 in the validation cohort, a 97.9% sensitivity, an 89.4% specificity, and an overall accuracy of 93.7%.
Critically, four of the five metabolites (2-piperidone, theophylline, DL-norleucine, and linolenic acid) remained significantly reduced in CRC patients at both early stages (I and II) and late stages (III and IV), demonstrating that the panel is not limited to detecting advanced disease. This is particularly important for screening purposes, where the goal is to find cancer before it has spread.
One metabolite, DL-norleucine, showed a progressive decrease with advancing cancer stage, suggesting it may carry additional value as a staging biomarker. The other four showed similar reductions regardless of stage, making them more appropriate for detection rather than staging. Age-stratified and sex-stratified analyses confirmed that the markers were robust across demographic groups, with only modest age-associated variation seen for N-methylcytisine.
To explore whether the reduced plasma levels of these metabolites have any biological meaning beyond being statistical markers, the researchers tested them on two colorectal cancer cell lines, HCT116 and T5088, at concentrations matching the levels found in healthy donor plasma. The goal was to see if restoring normal metabolite levels might affect cancer cell behavior.
At physiologically relevant concentrations, none of the five metabolites significantly reduced cancer cell proliferation in HCT116 cells, suggesting they do not directly kill cancer cells at normal blood levels. However, T5088 cells showed dose-dependent sensitivity to higher concentrations of N-methylcytisine and theophylline, pointing to possible cell-line-specific effects.
More notably, multiple metabolites significantly inhibited cancer cell migration, an early step in metastasis. In HCT116 cells, treatment with 2-piperidone, theophylline, DL-norleucine, and linolenic acid all reduced cell movement in Transwell migration assays. Complementary wound-healing (scratch) assays in HCT116 cells confirmed that all five metabolites slowed wound closure, a measure of collective cell migration ability.
These findings suggest that the lower levels of these metabolites in CRC patients may contribute to a more invasive tumor environment. When protective metabolites are depleted, cancer cells may be less restrained in their ability to migrate and spread. However, the researchers appropriately caution that these in vitro observations require in vivo confirmation and should not be interpreted as proof that restoring metabolite levels would directly block tumor spread in patients.
Three of the five metabolites, 2-piperidone, DL-norleucine, and linolenic acid, are produced or influenced by gut bacteria. This connects the findings to the well-established relationship between gut microbiome dysbiosis (imbalance) and colorectal carcinogenesis. The depletion of these metabolites in CRC patients may reflect a disrupted gut bacterial community that accompanies and may contribute to cancer development.
Theophylline is normally derived from dietary caffeine through metabolism by the liver enzyme CYP1A2. The researchers note that reduced theophylline in CRC patients is unlikely to simply reflect less coffee or tea drinking, because its reduction was observed even in Stage I patients who typically do not show the appetite changes seen in advanced cancer. Instead, this likely reflects systemic inflammation associated with CRC suppressing CYP1A2 enzyme activity.
N-methylcytisine, a plant-derived compound with anti-inflammatory properties, showed somewhat more variable behavior, including modest age-associated changes and partial attenuation after statistical adjustment for age. The researchers acknowledge this requires further investigation. Importantly, the other four metabolites remained independent predictors of CRC after adjusting for the age difference between patients and healthy controls in the discovery cohort.
The study's authors also compare their panel to existing tools: unlike CEA, which misses early-stage disease, the metabolite panel detected Stage I and II cancers effectively. Unlike FIT, the plasma test does not depend on tumor bleeding. The panel is positioned not as a replacement for colonoscopy but as a complementary risk stratification tool that could help identify which individuals most urgently need further diagnostic evaluation.
The five-metabolite plasma panel achieves a sensitivity of 97.9% in the validation cohort, meaning it correctly identifies nearly all true CRC cases. This high sensitivity is critical for a screening test, where missing a cancer case has serious consequences. With 89.4% specificity, the panel also avoids most false alarms in healthy individuals, reducing unnecessary follow-up procedures.
The panel was effective across all four CRC stages, demographic groups (age, sex, tumor location), and CRC primary sites (colon versus rectum). This broad applicability means that if further validated, the test could potentially serve a wide patient population without requiring adjustment for these variables.
Currently, this research represents discovery and early clinical validation. The next steps for translating these findings into clinical practice would include prospective studies with larger, more diverse, and age-matched cohorts. Comparison against FIT and CEA in head-to-head clinical trials would be needed to understand exactly where in a screening workflow the panel would add the most value.
Practically, plasma metabolomics requires mass spectrometry equipment and specialized laboratory expertise. The researchers acknowledge this as a limitation for widespread clinical deployment. However, with further development, targeted assays for just five specific metabolites could potentially be designed to be simpler and more affordable than full untargeted metabolomics platforms, eventually enabling routine clinical use.
This study demonstrates that a carefully selected panel of five plasma metabolites, identified using machine learning from nearly 147,000 candidate features, can accurately detect colorectal cancer with a combined AUC of 0.968 in an independent validation cohort. The approach is noninvasive, blood-based, and effective across cancer stages and patient demographics.
Three of the five biomarkers connect to gut microbiome metabolism, offering a biologically coherent explanation for why colorectal cancer, a disease tightly linked to gut health, alters circulating metabolite levels. This biological grounding, supported by in vitro functional evidence of metabolite effects on cancer cell behavior, strengthens the case that these markers reflect genuine disease biology rather than confounding factors.
The study also represents a methodological advance in how machine learning can be responsibly applied to biomarker discovery. By using random forest analysis only for feature prioritization and then building the final diagnostic model from quantified metabolites in a separate cohort, the researchers deliberately separated discovery from validation to reduce overfitting and improve generalizability.
Larger prospective, multicenter studies are needed before this panel can be used clinically. Nonetheless, these findings provide a strong scientific foundation for a novel, practical blood test that could improve early detection of colorectal cancer and potentially save lives by catching the disease while it is still treatable.