When most people think of genes, they think of genes that code for proteins - the molecular machines that do the work of the cell. But a large portion of the human genome is transcribed into RNA molecules that do not make protein. Among these, long noncoding RNAs (lncRNAs) are sequences of RNA more than 200 base pairs long with a variety of regulatory roles - though many of their specific functions remain unknown.
In recent years, researchers have discovered that lncRNAs can behave differently in cancer cells compared to normal cells, making them potentially useful as biomarkers - molecular indicators that can identify whether cancer is present. This is particularly valuable for endometrial cancer (cancer of the uterine lining), which currently lacks reliable early detection biomarkers in routine clinical use.
A 2020 study published in BioMed Research International used publicly available gene expression data from The Cancer Genome Atlas (TCGA) - a large national database of cancer genomic profiles - combined with machine learning algorithms to identify a set of lncRNAs capable of distinguishing endometrial cancer tissue from normal adjacent tissue. The researchers used both supervised learning (Support Vector Machine or SVM) and unsupervised learning (hierarchical clustering) to validate their findings.
The researchers downloaded RNA sequencing data from TCGA for 23 patients who had both tumor tissue (TP) and adjacent normal tissue (NT) samples available, yielding 46 total samples. Using both tissue types from the same patients is methodologically important because it controls for individual variation - any gene expression differences found are more likely to be genuinely cancer-related rather than patient-specific quirks.
After normalization and filtering to remove noise and low-quality reads, 8,700 transcripts were analyzed for differential expression. The study used a stringent cutoff: genes had to show a log fold change of 5 or greater (meaning the expression level was at least 32-fold different between tumor and normal tissue) and a false discovery rate below 1%. This high threshold was intentional - by focusing only on the most dramatically changed genes, the researchers maximized the likelihood of identifying candidates with genuine biomarker potential.
From the 191 transcripts that met these criteria, 17 lncRNAs were identified as differentially expressed. These were then evaluated individually using ROC (Receiver Operating Characteristic) analysis, which measures how well each single lncRNA can distinguish cancer from non-cancer samples. The five lncRNAs with AUC greater than 0.7 were considered high-confidence individual biomarkers. The full set of 14 remaining candidates (excluding 3 with high variability) was then used in machine learning models.
Of the 17 identified lncRNAs, 16 were downregulated in tumor compared to normal tissue - their activity was suppressed in cancer. Only one, LINC01376, was upregulated (with a log fold change of 5.88, meaning it was expressed at roughly 58-fold higher levels in tumor than in normal tissue). This pattern of widespread suppression with one notable activator is characteristic of many cancer gene expression profiles, and the sole upregulated candidate stands out as particularly important.
LINC01376 achieved the best individual AUC of 0.913 - meaning that using LINC01376 expression alone, the model correctly identified 91.3% of the time whether a sample came from tumor or normal tissue. It also showed the strongest correlation (r = 0.91) with HDAC7, a histone deacetylase enzyme whose inhibition has been shown to cause cell death and growth arrest in endometrial cancer cells. The high expression of both LINC01376 and HDAC7 in endometrial cancer tissue suggests these may be mechanistically linked in supporting the tumor environment.
BRWD1-AS1 was the second-ranked lncRNA with an AUC of 0.820. It is an antisense RNA - transcribed from the opposite strand of a protein-coding gene called BRWD1, which is involved in cell cycle regulation and programmed cell death (apoptosis). The downregulation of BRWD1-AS1 in cancer may therefore disrupt normal apoptosis regulation, potentially allowing cancer cells to survive longer than they should. Together, LINC01376 and BRWD1-AS1 form a complementary pair - one upregulated, one downregulated - that proved highly effective at discrimination.
The machine learning analysis tested three different lncRNA panels using Support Vector Machine (SVM) - a classification algorithm that finds the mathematical boundary that best separates two groups. SVM was trained on the 23 patients with matched tumor and normal samples, then evaluated on an independent set of 24 additional TCGA patients who had only tumor or only normal samples available.
The results were impressive. The full set of 14 lncRNAs achieved an accuracy of 95.83% - meaning the model correctly classified almost 96 out of 100 samples. The top 5 lncRNAs (those with individual AUC above 0.7) achieved 91.67% accuracy. Perhaps most strikingly, a panel of just 2 lncRNAs (LINC01376 and BRWD1-AS1) also achieved 91.67% - matching the performance of the 5-lncRNA panel with four fewer biomarkers. This suggests that a simple two-biomarker test could potentially be developed for clinical use.
Unsupervised hierarchical clustering - which groups samples based purely on expression similarity, without being told the group labels - confirmed these findings. Normal tissue (NT) samples clustered together very tightly and consistently, while tumor (TP) samples were more dispersed. This pattern reflects the well-known characteristic of cancer: while normal cells behave in relatively uniform ways, cancer cells show more variable gene expression due to the genetic instability of tumors. Nevertheless, the clustering still largely separated the two groups.
Beyond their classification value, the researchers explored what pathways these lncRNAs might be involved in. A coexpression network was built showing which genes tend to be activated or silenced in tandem with the candidate lncRNAs. Several interesting connections emerged. CHD5 - a tumor suppressor gene known to be lost in multiple cancers - was strongly correlated with three of the candidate lncRNAs (TCF4-AS1, LINC02249, and LINC02475). These lncRNAs had not been previously associated with cancer, suggesting they may regulate CHD5 activity in endometrial cancer.
Pathway enrichment analysis identified connections to viral carcinogenesis pathways (through HDAC7 and MAPKAPK2), histone methylation (through CHD5 and WRB), and the Hedgehog signaling pathway (through GPR161). Aberrant Hedgehog signaling has been specifically linked to cell proliferation in endometrial cancer in multiple studies, so this connection is biologically meaningful. The coexpression of GPR161 with MIR545, a microRNA associated with cancer cell proliferation in other cancer types, adds further context.
Notably, the survival analysis performed for each lncRNA did not find that any of them were significantly associated with better or worse patient prognosis. This means they should be considered diagnostic biomarkers (tools for identifying whether cancer is present) rather than prognostic ones (tools for predicting how the cancer will behave). The practical implication is that these lncRNAs may be most useful in early detection testing rather than in guiding treatment intensity decisions.
This study demonstrates proof-of-concept that a small panel of lncRNAs - potentially as few as two - can reliably distinguish endometrial cancer tissue from normal tissue using machine learning. The rigorous use of paired tumor and normal samples, high statistical thresholds, an independent test set, and multiple machine learning approaches (supervised SVM plus unsupervised clustering) strengthens confidence in the findings.
The study has important limitations to acknowledge. The dataset of 23 paired patients is small, and the analysis was performed entirely in silico (computationally, using existing data) rather than on new clinical samples. Before these biomarkers can be considered for clinical use, they would need to be validated in fresh tissue samples from a broader patient population, ideally in a prospective study. The study also used tissue samples rather than blood or urine, which limits immediate application as a minimally invasive screening test - though future work could explore whether lncRNA expression patterns are detectable in less invasive sample types.
The identification of LINC01376 as the sole upregulated lncRNA with the highest individual discriminatory power, combined with its strong correlation with HDAC7 (a therapeutically relevant gene already being studied as a drug target in endometrial cancer), makes it a particularly compelling candidate for further investigation. Future studies building on this foundation could help develop the first clinically useful molecular markers specifically for endometrial cancer detection.