Endometrial cancer (EC) is the most common gynecological malignancy worldwide, responsible for approximately 661,000 new cases and 348,000 deaths in 2022 alone - numbers that continue to rise. Despite this prevalence, there is currently no established screening test for EC in the general population.
Most EC cases are diagnosed only after symptoms such as abnormal vaginal bleeding appear, which typically occur at early stages and allow surgical cure - but for some patients, the disease is already more advanced. Existing tests have limitations: cervical cytology (a Pap smear-type approach) detects EC with only 45% sensitivity, and transvaginal ultrasound, while more sensitive (77-85%), has relatively low specificity, meaning it often generates false alarms about benign conditions.
Liquid biopsy - analyzing biological signals in blood rather than in tissue - represents a promising non-invasive alternative. When tumors grow, they shed fragments of their DNA into the bloodstream. These fragments, called cell-free DNA (cfDNA), carry signatures of the tumor's genomic activity and can be captured and analyzed with a blood draw.
Recent research has shown that cfDNA fragments from cancer cells have distinct fragmentation patterns - differences in size, distribution, and context compared to cfDNA from healthy cells. This field, known as fragmentomics, has already shown strong results for liver cancer, colorectal cancer, and breast cancer detection. This study applies fragmentomics to endometrial cancer for the first time in a comprehensive way.
The DECIPHER-UCEC-2 study enrolled patients with confirmed endometrial cancer and healthy female volunteers from two Chinese cancer centers. Preoperative blood samples were collected and processed to extract cfDNA from plasma. The cfDNA was then sequenced using low-pass whole-genome sequencing (WGS) - a cost-effective approach that reads the genome at shallow depth but across its full length, providing enough signal for fragmentomic analysis.
Five distinct types of fragmentomic features were analyzed from each sample: copy number variation (CNV) (gains or losses of large DNA segments), fragment size score (FSS) and fragment size distribution (FSD) (how big the DNA pieces are), mutation context and mutational signature (MCMS) (patterns of DNA damage characteristic of specific cellular processes), and nucleosome footprint (NF) (imprints of where nucleosomes - the protein spools DNA wraps around - were positioned, revealing gene regulation patterns). Together, these captured 7,230 raw features per sample.
Multiple machine learning algorithms - gradient boosting machine (GBM), generalized linear model (GLM), random forest (RF), and deep learning (DL) - were applied to each feature type, generating 60 individual base models. These were combined into a single ensemble model (averaging predictions across well-performing base models), which was evaluated in an independent prospective test cohort and further validated in a separate external cohort from Chongqing.
The model was tested for three tasks: early detection (cancer vs. healthy), clinicopathological subtyping (predicting tumor stage, histological type, grade, and microsatellite instability status), and recurrence-free survival prediction (estimating which patients are at high risk of the cancer returning after treatment).
The cancer detection model achieved an AUC (area under the curve) of 0.96 in the independent test cohort - meaning it could distinguish EC patients from healthy women with very high accuracy. In practical terms, this translated to 75.8% sensitivity (of all cancer patients, this fraction tested positive) and 96.8% specificity (of all healthy women, this fraction correctly tested negative).
In the external validation cohort from Chongqing, the model achieved an even stronger performance: AUC of 0.93, with 88% sensitivity and 98.7% specificity. These consistently high results across two independent datasets provide confidence that the model generalizes to new patients and hospital settings.
A key strength of this test is its performance across all disease stages. Sensitivities were broadly consistent across FIGO stages I through IV (74.4%, 85.7%, 75%, and 75% respectively), meaning the assay detected early-stage cancer (which is most common and most curable) nearly as well as late-stage cancer. This is critical for an early detection tool.
Among the five feature types, MCMS (mutation context and mutational signature) was the single most predictive feature, consistently achieving the highest individual AUC scores. The mutational signature SBS10a - linked to mutations in the polymerase epsilon exonuclease domain, a known EC-related defect - was the key signature differentiating cancer from healthy samples.
Beyond detecting cancer presence, the researchers tested whether the same blood-based fragmentomic signatures could reveal information about the cancer's characteristics. Predicting cancer stage (early vs. late) achieved an AUC of 0.72 in the independent test cohort; predicting histological subtype (Type I - estrogen-dependent endometrioid, vs. Type II - more aggressive non-endometrioid) achieved an AUC of 0.73.
The model also predicted microsatellite instability (MSI) status with an AUC of 0.77. MSI is a key molecular feature of some endometrial cancers that affects treatment decisions - patients with MSI-high tumors may respond to immunotherapy (checkpoint inhibitors), making pre-treatment identification valuable for personalizing therapy without waiting for biopsy results.
While these AUCs (0.72-0.77) indicate moderate rather than exceptional predictive power, the significance lies in the fact that this information comes from a simple blood draw before any surgery or biopsy. Having even partial information about stage, subtype, and molecular features preoperatively could help surgeons and oncologists plan treatment strategies earlier and more efficiently.
Predicting which patients are at high risk of cancer recurrence after treatment is one of the most clinically important - and most difficult - challenges in oncology. The researchers built a recurrence-free survival model using LASSO Cox regression (a statistical method designed for time-to-event outcomes) applied to the fragmentomic features.
Patients classified as high-risk by the model had a hazard ratio (HR) of 8.6 for cancer recurrence compared to low-risk patients (P less than 0.001) - meaning high-risk patients were more than 8 times more likely to experience cancer recurrence. This striking difference confirms that cfDNA fragmentomic signals captured before treatment genuinely encode information about disease biology and future outcomes.
An additional analysis using similarity network fusion (a method that clusters patients based on their combined fragmentomic profiles) further stratified patients into high- and low-risk groups, with high-risk patients showing an HR of 6.2 for recurrence (P = 0.049). Crucially, patients flagged as high-risk by both methods simultaneously showed an HR of 10.1 (P less than 0.0001) - a very strong signal that combining approaches improves risk stratification.
This study demonstrates that a single blood draw, processed through low-pass whole-genome sequencing and analyzed with an ensemble machine learning model, can simultaneously detect endometrial cancer, provide information about its biological characteristics, and identify patients at high risk of recurrence - all before any surgical intervention.
The combination of high specificity (reducing unnecessary anxiety and procedures from false alarms) and consistent sensitivity across stages makes this assay potentially suitable as a population-level screening tool, though the absolute sensitivity of 75.8% may not yet be high enough to serve as the sole screening method and would benefit from further refinement.
The study has limitations: samples were collected at two Chinese institutions, and broader multi-ethnic validation is needed. The prognostic model was built on relatively small numbers of recurrence events, and longer follow-up data would strengthen conclusions. The authors note that future work should focus on expanding validation and exploring whether serial monitoring of cfDNA over time (tracking changes during and after treatment) could provide additional clinical value.
The broader vision is a comprehensive, non-invasive liquid biopsy platform that could replace or supplement multiple current diagnostic steps - from initial screening to pre-treatment staging to post-treatment surveillance - using only blood samples, making cancer management less burdensome and more accessible.