Within any tumor, a small population of cells called cancer stem cells (CSCs) behaves like the root system of the cancer. These cells are especially resistant to chemotherapy and radiation, capable of self-renewal, and thought to be responsible for cancer recurrence and spread. Traditional prognostic tools focus on tumor size and stage, but they do not directly measure how "stem-like" a tumor is.
Scientists have developed a metric called mRNAsi (mRNA expression-based stemness index) that scores how similar a tumor's gene expression pattern is to stem cells. A higher mRNAsi score indicates a more stem-like tumor. This study used mRNAsi to classify endometrial cancers into high-stemness and low-stemness subtypes, then built a machine learning model to predict outcomes based on genes associated with stemness biology.
The researchers analyzed data from 514 endometrial cancer patients in TCGA-UCEC (The Cancer Genome Atlas), one of the largest publicly available genomic cancer datasets. The goal was to identify a small set of genes that could serve as a practical prognostic signature without needing full genomic profiling.
Patients were divided into high-mRNAsi and low-mRNAsi groups using consensus clustering - an unsupervised method that groups patients purely based on molecular similarity. The two stemness subtypes differed significantly in survival outcomes, with the high-stemness group showing worse prognosis. This validated that mRNAsi captures biologically meaningful risk stratification in endometrial cancer.
Next, researchers identified genes that were both differentially expressed between the two stemness groups and individually associated with patient survival. Weighted gene co-expression network analysis (WGCNA) further narrowed the field by identifying modules of genes that track together and correlate with stemness. This multi-step filtering produced a candidate gene list.
Finally, machine learning - specifically LASSO Cox regression - selected the minimal set of genes with the strongest independent prognostic value. The resulting 7-gene signature includes: FOXD3, LMO1, ART3, FRMPD2, TMEM114, C1orf64, and IHH. A risk score formula was derived from the weighted expression levels of these genes, placing each patient on a continuous risk spectrum.
In the TCGA training dataset, patients with high risk scores (top 50%) had significantly worse overall survival and progression-free survival compared to low-risk patients. The signature remained prognostic after adjusting for age, stage, histological grade, and other established clinical variables - confirming it captures information beyond standard clinical factors.
Validation in an independent hospital cohort achieved AUC of 0.82 for overall survival and 0.85 for progression-free survival at 5 years. These are strong values suggesting the signature is clinically informative in real-world patient data, not just in the publicly available TCGA dataset where it was originally developed.
Among the 7 genes, FOXD3 and LMO1 carry positive coefficients (higher expression associates with worse prognosis), while ART3, FRMPD2, IHH, and TMEM114 carry negative coefficients (higher expression is protective). This suggests the signature captures a balance between cancer-promoting and cancer-suppressing biological pathways linked to stemness biology.
High-risk patients (by the stemness signature) showed a distinct immune microenvironment profile. Their tumors had lower overall immune scores and significantly fewer tumor-infiltrating immune cells - particularly CD8+ cytotoxic T cells that normally attack cancer. This suggests that high-stemness tumors have found ways to evade or suppress immune surveillance.
Conversely, low-risk patients had tumors with more active immune infiltration. This pattern has important implications for immunotherapy: if high-stemness, high-risk tumors are immunologically "cold" (few immune cells present), they may respond poorly to immune checkpoint inhibitors like pembrolizumab, which work by releasing existing immune cells that are being suppressed.
The authors also found that high-risk tumors were more likely to be copy-number high (chromosomally unstable) and less likely to be microsatellite unstable - consistent with known molecular subtypes of endometrial cancer. The POLE-mutated (ultramutated) subtype, which has excellent prognosis, was enriched in the low-risk group, further validating the signature's biological meaning.
The 7-gene stemness signature represents a practical advance because it is compact enough to be measured with standard gene expression assays currently used in pathology labs. Unlike complex genomic profiling, a 7-gene panel is more affordable and interpretable, making it potentially translatable to routine clinical care.
The study's strength lies in combining multiple lines of evidence: clustering by stemness, WGCNA module analysis, machine learning feature selection, and external clinical validation. Each step filtered the data more rigorously, reducing the risk of overfitting and improving confidence in the final signature.
Future work should validate this signature in larger, diverse patient cohorts from multiple countries and explore whether targeting stemness pathways (for example, through IHH - the Indian Hedgehog signaling gene - which has known inhibitors) could improve treatment outcomes for the high-risk subgroup identified by this signature.