Inside every cell, thousands of tiny RNA molecules called microRNAs (miRNAs) play a crucial role in controlling which genes are active at any given time. These molecules - each only 19 to 24 genetic letters long - work by silencing or dampening the activity of specific protein-coding genes. In cancer cells, the normal patterns of miRNA expression are often disrupted: some miRNAs become overactive (promoting cancer growth) while others are silenced (removing normal braking mechanisms on cell proliferation).
Because miRNA expression patterns differ so consistently between cancer cells and normal cells, researchers have proposed using them as biomarkers - molecular signals that could identify cancer early or predict how it will behave. For endometrial cancer, current diagnosis depends on clinical symptoms and tissue biopsy examined under a microscope. While effective, this approach misses the approximately 15-20% of tumors that are at high risk of recurrence despite appearing relatively benign under standard pathological evaluation.
A 2020 study published in the Journal of Cellular and Molecular Medicine used gene expression data from two large public databases to identify miRNA markers that could serve two distinct purposes: first, distinguishing cancerous endometrial tissue from normal tissue (diagnostic function); and second, predicting which patients will survive longer (prognostic function). Using machine learning to sift through hundreds of candidate molecules, they identified panels of nine diagnostic miRNAs and five prognostic miRNAs.
The study used miRNA and mRNA expression data downloaded from The Cancer Genome Atlas (TCGA), the large federally-funded US database of cancer genomic profiles. A total of 441 endometrial cancer samples were available for miRNA analysis and 419 for mRNA analysis. The cohort was split into training (two-thirds) and testing (one-third) sets. An additional independent dataset, GSE35794 from the Gene Expression Omnibus database, provided a third cohort for further validation of the diagnostic results.
To identify diagnostic miRNAs, the researchers first used differential expression analysis to find all miRNAs that differed significantly between cancer and normal tissue (417 were identified). From this large candidate list, they applied LASSO regression - a feature selection method that compresses the candidate list by retaining only the most informative variables - to narrow down to nine final diagnostic markers. These nine markers were then combined into a single classifier whose accuracy was evaluated by confusion matrices and ROC (Receiver Operating Characteristic) curves.
The prognostic miRNA analysis followed a different path. The starting point was identifying miRNAs associated with overall survival using Cox regression (a statistical method for analyzing survival data). This produced a list of candidates that was then further filtered using a machine learning method called variable hunting implemented in Random Forest - an algorithm that identifies which variables most consistently split patients into different outcome groups. Multivariate Cox regression then confirmed which miRNAs were independent predictors of survival after accounting for age, cancer stage, and tumor grade.
The nine diagnostic miRNA markers divide into two groups: four tumor suppressor miRNAs (hsa-miR-542-3p, hsa-miR-152-3p, hsa-miR-24-1-5p, and hsa-miR-374b-5p) that are reduced in cancer tissue, and five oncomiRNAs (hsa-miR-183-5p, hsa-miR-1307-3p, hsa-miR-429, hsa-miR-200b-3p, and hsa-miR-183-3p) that are elevated in cancer tissue. The combined diagnostic classifier that used all nine markers achieved 100% accuracy in both the training and testing cohorts, and 95.45% accuracy in the independent GSE35794 cohort.
Among individual markers, hsa-miR-542-3p showed the strongest single-marker performance with an AUC of 0.997, followed by hsa-miR-183-5p with an AUC of 0.992. However, the combined nine-marker classifier consistently outperformed any single marker, demonstrating the advantage of combining complementary signals into one model. Hierarchical clustering of the samples by these nine markers visually confirmed that cancer and normal tissue samples grouped together correctly in the heatmap.
Several of these markers have known biological roles in endometrial cancer or related cancers. hsa-miR-542-3p is known to be reduced in endometrial serous adenocarcinoma and suppresses tumor blood vessel growth by targeting a protein called angiopoietin-2. hsa-miR-152-3p has been shown to be silenced by DNA methylation (chemical modification of DNA) in endometrial cancer. hsa-miR-200b and hsa-miR-429 are known oncomiRNAs that target the PTEN tumor suppressor gene in endometrial carcinoma - a gene whose loss is one of the most common early events in endometrial cancer development.
Five miRNAs were identified as independent predictors of patient survival: hsa-miR-18b-3p, hsa-miR-128-3p, hsa-miR-106a-5p, and hsa-miR-7706 were all associated with worse outcomes when expressed at high levels, while hsa-miR-455-5p was associated with better outcomes at higher expression. Together, these five miRNAs form a risk score calculated by combining each miRNA's expression level with its weighted coefficient from Cox regression analysis.
The prognostic model achieved a concordance index (C-index) of 0.74 in the training cohort and 0.70 in the testing cohort. The C-index measures how often the model correctly ranks two patients by survival - a score of 1.0 would be perfect, while 0.5 would be random chance. For the 5-year overall survival prediction, the area under the ROC curve was 0.749 in training and held up in testing. Using the model, patients were divided into high-risk and low-risk groups; the hazard ratio between these groups was 5.25 (p less than 0.001), meaning high-risk patients were more than five times more likely to die during follow-up than low-risk patients.
Crucially, the five-miRNA model outperformed the current gold standard for endometrial cancer prognostic staging - the FIGO staging system - in head-to-head comparison (AUC 0.737 vs. 0.614 for 5-year survival prediction). When the two were combined, performance improved further to an AUC of 0.777. This suggests the miRNA model captures prognostic information that is not fully captured by clinical staging alone, making it a potentially valuable complement to existing risk assessment tools.
Beyond the classifiers themselves, the researchers mapped out which genes each diagnostic and prognostic miRNA controls. They identified 375 genes regulated by the nine diagnostic miRNAs and 367 genes regulated by the five prognostic miRNAs. These gene sets were then analyzed for enriched biological pathways - collections of genes that work together in the same molecular process - to understand what cancer-related functions are being disrupted when these miRNAs behave abnormally.
For the diagnostic miRNA gene targets, the enriched pathways included the MAPK signaling pathway (which regulates cell proliferation and inflammatory responses), Wnt signaling (which controls cell fate and is mutated in 20-50% of endometrioid endometrial carcinomas), and focal adhesion pathways involved in how cells attach to and invade surrounding tissue. These pathways are well-established drivers of endometrial cancer, validating that the identified miRNAs are genuinely connected to disease biology rather than being statistical coincidences.
For the prognostic miRNA gene targets, 23 genes were specifically found to be associated with patient survival in this dataset. Notable among these were PTEN (one of the most commonly mutated tumor suppressor genes in endometrial cancer) and PDPN (podoplanin, a protein used as a biomarker and therapeutic target in multiple cancers). The FoxO, Ras, and Wnt signaling pathways were all enriched in the prognostic target gene set - consistent with the aggressive biology underlying poor-prognosis endometrial tumors.
This study demonstrates that a small panel of miRNA markers - nine for diagnosis, five for prognosis - can provide clinically meaningful information about endometrial cancer that currently requires expensive genomic testing. The 100% diagnostic accuracy in controlled TCGA conditions, while remarkable, should be interpreted carefully: real-world samples from fresh tissue, blood, or other sources will have additional variability that the model must handle. The external validation in GSE35794, where accuracy was 95.45%, provides a more realistic benchmark for clinical performance.
The prognostic model's advantage over FIGO staging is particularly interesting because FIGO staging is a well-established system that already incorporates substantial clinical information. The fact that miRNA patterns provide additional prognostic information beyond what staging captures suggests these molecular signals reflect distinct biological processes that correlate with outcomes. This is consistent with the broader movement toward molecular profiling in cancer management, where treatment decisions are increasingly guided by tumor biology rather than anatomic staging alone.
Key limitations include the use of tissue samples rather than blood-based (non-invasive) testing, the relatively small normal tissue sample in the training cohort (21 normal samples vs. 258 cancer samples), and the fact that all data came from TCGA - which represents mainly patients treated at major US cancer centers. Before clinical use, the markers need validation in fresh clinical samples from diverse patient populations, and ideally in a form that can be measured non-invasively such as from blood plasma. If these hurdles can be cleared, a combined miRNA diagnostic-and-prognostic panel could offer a practical molecular complement to current endometrial cancer management.