Endometrial cancer is classified into four molecular subtypes by The Cancer Genome Atlas (TCGA): POLE ultramutated (which carries the best prognosis), MSI-H (microsatellite instability-high, intermediate prognosis), CN-L (copy number-low, largely low-grade tumors), and CN-H (copy number-high, largely aggressive serous-type tumors with the worst prognosis). These molecular subtypes now guide adjuvant treatment decisions in clinical guidelines.
Despite this advance, the molecular subtypes are imperfect predictors of individual outcomes. The CN-L group is particularly heterogeneous - some patients do well while others experience early recurrence. Similarly, not all CN-H patients recur despite uniformly aggressive adjuvant treatment being recommended for this group. There is a clear need for additional biomarkers that can refine risk prediction within each molecular subtype.
This study used three types of molecular data from the TCGA database - DNA methylation (which genes are epigenetically silenced or activated), RNA sequencing (which genes are expressed), and genomic variants (which genes are mutated) - to search for recurrence-specific signatures within the CN-H and CN-L subtypes. Machine learning methods including decision trees and random forests were applied to identify the most informative biomarkers.
The researchers downloaded DNA methylation (378,278 CpG sites), RNA sequencing (53,409 genes), and genomic variant data (118,555 genomic locations) from 116 endometrial cancer patients in the TCGA-UCEC dataset. These patients were categorized into the four TCGA molecular subtypes and four histological grades. The key analysis compared patients who experienced cancer recurrence against those who did not.
For DNA methylation, the researchers identified differentially methylated regions (DMRs) - sites where the degree of methylation differs significantly between recurrence and non-recurrence groups. Methylation is an epigenetic modification where chemical groups attach to DNA and typically silence gene expression. For RNA sequencing, differentially expressed genes (DEGs) were identified. Variant analysis used Fisher's exact test to find mutations significantly associated with recurrence.
Two machine learning methods were applied: decision trees (which create a flowchart-like model of if-then rules to classify patients) and random forests (an ensemble method that builds many decision trees and aggregates their predictions). The top features were ranked using the Gini importance metric, which measures how much each variable contributes to the purity of the classification at each branch in the decision trees.
A striking early finding was that molecular clustering by recurrence worked well only in the CN-H subgroup (32 patients). When all 116 patients were analyzed together, or when CN-L and MSI subgroups were examined separately, the recurrence and non-recurrence groups could not be cleanly separated by molecular profiles alone. This suggests that recurrence in different molecular subtypes is driven by different molecular mechanisms.
In the CN-H group, 55 differentially methylated regions and 37 differentially expressed genes were identified between recurrent and non-recurrent cases. The recurrent CN-H patients were disproportionately classified as high-stage serous or grade 3 endometrioid type - consistent with CN-H being the most aggressive subtype. Decision tree analysis identified the methylation level of PARD6G-AS1 as the key branching variable distinguishing recurrent from non-recurrent CN-H patients.
In the CN-L group, the expression level of CD44 was identified as the key recurrence predictor. CD44 is a cell surface protein involved in cell adhesion, migration, and invasion - functions directly relevant to cancer spread. When survival analysis was applied to the full TCGA clinical dataset, PARD6G-AS1 hypomethylation in CN-H (p=0.006) and CD44 overexpression in CN-L (p=0.02) were the only two biomarkers with statistically significant recurrence associations.
PARD6G-AS1 hypomethylation (reduced methylation, which typically means the gene is more active) was significantly associated with advanced FIGO stage III/IV disease (p=0.018) and positive lymph node metastasis (p=0.037). PARD6G-AS1 is a long non-coding RNA gene in a region associated with maternal imprinting - a process where gene expression depends on which parent the copy was inherited from. This study is the first to show that its methylation status has prognostic significance in endometrial cancer.
CD44 overexpression was also significantly associated with advanced stage III/IV disease (p=0.014) and positive lymph node metastasis (p=0.013). CD44 is well established in cancer biology as a marker of invasive behavior - it helps cancer cells break free from their original location and spread to lymph nodes and distant organs. This finding is consistent with prior reports linking CD44 overexpression to myometrial invasion in endometrial cancer.
Institutional validation using surgical specimens from 16 patients at Korea University confirmed the PARD6G-AS1 methylation finding (p=0.002 in the CN-H group). The CD44 validation did not reach statistical significance in this small validation cohort, likely due to insufficient sample size, but the trend was consistent with the TCGA findings.
For the CN-H group, current treatment guidelines recommend aggressive adjuvant chemoradiation for essentially all patients, regardless of individual recurrence risk. This uniform approach means some patients may be overtreated while others may still be undertreated. The identification of PARD6G-AS1 hypomethylation as a CN-H-specific recurrence marker could enable more nuanced treatment decisions - patients with this epigenetic change could receive the most intensive treatment, while others might be candidates for de-escalation.
For the CN-L group, CD44 overexpression adds to an emerging toolkit for identifying which of these diverse, typically low-risk patients actually face elevated recurrence risk. The CN-L group currently includes patients with very variable outcomes, and the study results suggest CD44 testing could help identify the high-risk patients who need adjuvant treatment that is typically not given for standard-risk CN-L disease.
The study also demonstrated the value of integrating multiple omics data types for biomarker discovery. Neither methylation, expression, nor variant data alone was sufficient to identify recurrence patterns reliably. The combination of all three, analyzed through machine learning, revealed signals that single-modality analysis would have missed. This supports the broader direction of multi-omics precision medicine in cancer research.
This study demonstrates that integrating three molecular data layers - DNA methylation, gene expression, and genomic variants - with machine learning can identify subtype-specific biomarkers of endometrial cancer recurrence that are not visible when any single data type is used alone. The two identified biomarkers, PARD6G-AS1 and CD44, are independently validated and biologically plausible.
Several limitations should be noted. The study analyzed only 116 of the approximately 548 patients in TCGA-UCEC because complete data across all three omics platforms was available for only this subset. The retrospective design means treatment protocols were not standardized, which may influence recurrence patterns. Future prospective studies with larger, diverse patient cohorts are needed to confirm these findings and establish clinical utility.
The analytical framework developed here - integrating multi-omics data with machine learning decision trees and random forests to identify recurrence patterns within molecular subtypes - provides a template that could be applied to other cancer types. The methodology for combining methylation, transcriptomic, and variant data is generalizable, and the identified biomarkers warrant investigation as targets for potential therapeutic intervention in endometrial cancer.