Colorectal cancer (CRC) remains one of the most common and deadly cancers worldwide. Even patients with similar tumor stages can have very different outcomes, which suggests that standard staging alone cannot fully predict prognosis. Researchers have long sought biological markers that capture the true nature of a tumor.
One such property is tumor stemness -- the tendency of cancer cells to behave like stem cells, retaining abilities like self-renewal, flexibility, and dedifferentiation (reverting to a less specialized state). Stemness-high tumors are harder to treat and more likely to spread or return after therapy.
Stemness also affects the immune environment around tumors. Stemness-high colorectal cancers have been linked to altered immune cell presence, disrupted immune signaling, and the ability to evade immune attack. Understanding this relationship could unlock better ways to predict who will respond to immunotherapy.
Researchers used bulk RNA sequencing data from the TCGA (The Cancer Genome Atlas) CRC dataset as their main training cohort. Gene expression values were normalized and transformed, and patients with incomplete survival data were excluded. External validation datasets included GSE38882 for prognostic testing and IMvigor210 and GSE78820 for immunotherapy response validation.
Each patient was assigned a stemness score using ssGSEA (single-sample Gene Set Enrichment Analysis), which estimates how strongly a tumor expresses known stem cell-related genes. Patients were split into high- and low-stemness groups at the median score.
To identify survival-linked genes and build a compact model, the team performed univariate Cox regression to find stemness-associated genes tied to overall survival, then applied LASSO Cox regression (a statistical method that selects only the most informative variables) to finalize the gene signature. Single-cell RNA sequencing (scRNA-seq) data from GSE196964 were analyzed in parallel to pinpoint which cell types carry the stemness signal.
LASSO Cox regression identified an optimal 10-gene risk signature: SLC2A3, SERPINE1, MT-ND2, LTBP1, CPNE7, SFRP2, CCN1, MMP1, SMIM22, and CPA3. Using this signature, each patient received a risk score calculated as a weighted combination of gene expression levels.
In the TCGA cohort, high-risk patients had significantly worse overall survival than low-risk patients (p less than 0.0001). Time-dependent ROC analysis showed the model maintained strong predictive accuracy with AUC values around 0.74 at 1 year, 0.70 at 3 years, and 0.70 at 5 years.
The model was independently validated in the GSE38882 external cohort, where high-risk patients again had significantly worse survival (p = 0.041). Among the 10 genes, SLC2A3, SERPINE1, LTBP1, and CPNE7 were risk-increasing factors, while MMP1 and CPA3 acted as protective factors, highlighting the biological complexity of stemness-related pathways.
To test whether the stemness risk score adds value beyond conventional clinical information, the researchers performed multivariate Cox regression adjusting for age, sex, metastatic status, lymph node involvement, and pathological stage. The risk score remained independently associated with survival even after these adjustments, while most clinical variables lost statistical significance.
A comparison of time-dependent AUC values showed that the stemness risk score outperformed individual clinical factors including tumor stage (T, N, M classifications) at nearly every time point. When combined into a nomogram model that merges molecular and clinical information, predictive accuracy was highest of all.
Interestingly, stemness scores were significantly higher in tumor tissues than in normal colon tissue, but did not differ significantly by age, sex, survival status, or pathological stage. This suggests that stemness captures an intrinsic molecular property of tumors, rather than a consequence of advanced disease spread.
Using CIBERSORT, a computational method that estimates immune cell populations from gene expression data, the researchers found significant differences in immune cell composition between high- and low-risk groups. High-risk tumors showed altered levels of activated T cells, macrophage subtypes, and dendritic cells, indicating that stemness-based risk classification reflects genuine immune landscape differences.
Several immune checkpoint genes -- including PDCD1 (PD-1), CD274 (PD-L1), CTLA4, LAG3, TIGIT, HAVCR2 (TIM-3), and IDO1 -- were positively correlated with higher stemness scores. This stepwise increase across stemness groups suggests that stemness-high CRCs activate multiple inhibitory immune pathways simultaneously, potentially explaining their resistance to immune attack.
ESTIMATE analysis (a tool measuring non-cancer cell content in tumors) found that high-risk tumors had elevated immune scores but no significant difference in stromal scores. This pattern suggests a complex state of immune activation combined with immune suppression, rather than simple absence of immune cells.
To move beyond bulk analysis, the team examined single-cell RNA sequencing (scRNA-seq) data from paired normal colon and colorectal tumor samples (GSE196964). Using UMAP visualization, normal and tumor cells separated clearly, showing that malignant transformation produces distinctive gene expression states.
Five major cell types were identified: epithelial cells, enterocytes (mature intestinal absorptive cells), endothelial cells, fibroblasts, and T cells. Compared to normal tissue, CRC tumors had more epithelial cells and fibroblasts but fewer enterocytes and T cells, indicating substantial remodeling of the cellular environment.
CytoTRACE analysis, which predicts how undifferentiated (stem-like) each cell is based on gene expression diversity, revealed that malignant epithelial cells had the highest stemness and the lowest differentiation of all cell types. Normal epithelial cells, fibroblasts, and T cells showed progressively more differentiated profiles. Genes associated with stemness in these malignant cells included ribosomal and translational regulators, while loss of mature intestinal markers (FABP1, CA1, CA2, SLC26A3) confirmed their dedifferentiated state.
The risk model was tested in two real-world immunotherapy cohorts. In the IMvigor210 cohort (bladder cancer patients treated with anti-PD-L1 therapy), high-risk patients had significantly worse survival, and patients with progressive or stable disease had higher risk scores than those achieving complete or partial responses. Low-risk patients were enriched for responders, while high-risk patients were predominantly non-responders.
Identical patterns emerged in the GSE78820 cohort: high-risk patients had inferior survival, and non-responders carried significantly higher risk scores than responders. These cross-cancer validations demonstrate that the stemness-based signature captures biology relevant to immunotherapy response more broadly.
In silico drug sensitivity analysis estimated IC50 values (the drug concentration needed to inhibit 50% of cell growth) for multiple agents. High-risk tumors showed reduced sensitivity to KIN001-135, Sunitinib, and Imatinib. Conversely, low-risk tumors were more sensitive to CP466722, CGP-60474, and Roscovitine, suggesting that stemness-based classification could help identify patients likely to benefit from specific drugs.
To confirm that the RNA-based findings translate to actual protein expression, the researchers examined immunohistochemical (IHC) staining data from CRC tissue microarrays. Five signature genes were evaluated: SLC2A3, SERPINE1, LTBP1, CPNE7, and CPA3.
SLC2A3 showed moderate protein expression in CRC tissues, consistent with its transcriptional upregulation in stemness-high tumors. CPNE7 showed high protein expression, aligning with its role as a key risk-associated gene. SERPINE1, LTBP1, and CPA3 showed low or undetectable protein expression, which may reflect post-transcriptional regulation or variation between individual tumors.
These mixed findings highlight that gene expression and protein abundance do not always align perfectly due to cellular regulatory complexity. Nevertheless, the confirmable protein-level signals for SLC2A3 and CPNE7 reinforce the biological relevance of the stemness signature.
This study establishes tumor stemness as a central biological axis linking CRC prognosis, immune regulation, and therapeutic vulnerability. By integrating bulk transcriptomics, single-cell analysis, immunotherapy cohort data, and protein-level evidence, the authors built a comprehensive picture of how stemness shapes tumor behavior and clinical outcomes.
The 10-gene stemness risk model reliably separates CRC patients into groups with meaningfully different survival probabilities, immune landscapes, and responses to immunotherapy. Crucially, it provides prognostic information that standard clinical staging cannot, and its performance improves further when integrated with clinical variables in a nomogram.
The authors acknowledge limitations: the model was derived from retrospective datasets, and functional laboratory experiments are still needed to prove causal links between stemness programs, immune modulation, and drug resistance. Prospective clinical validation will be essential before this tool can be used in clinical decision-making. Future research should focus on directly targeting stemness pathways to enhance the effectiveness of existing treatments for CRC.