Endometrial cancer is the most common gynaecological malignancy in high-income countries, with over 400,000 incident cases and 97,000 deaths reported globally in 2020. Its incidence is rising in tandem with growing rates of obesity. In the United States alone, there were an estimated 66,570 new cases in 2021. When diagnosed early, the cancer is amenable to curative hysterectomy, with over 90% of patients surviving at least five years. By contrast, those with advanced or metastatic disease face five-year survival estimates of approximately 15%. Early detection is therefore the single most important factor in improving outcomes.
The current diagnostic pathway: Over 90% of women with endometrial cancer present with postmenopausal bleeding, which triggers urgent investigation by sequential transvaginal ultrasound scan (TVS), outpatient hysteroscopy, and endometrial biopsy. These investigations are invasive, painful, and anxiety-provoking. Critically, only 5-10% of symptomatic women actually have an underlying malignancy, meaning the vast majority undergo distressing procedures unnecessarily. The development of a simple, non-invasive, and cost-effective detection tool has been identified as the top priority by the James Lind Alliance Detecting Cancer Early Priority Setting Partnership.
Why cervico-vaginal fluid? The anatomical continuity between the uterine cavity and the lower genital tract allows endometrial tumours to shed cancer-related biomolecules through the cervix into the vagina. This creates an opportunity to detect endometrial cancer using non-invasive sampling of cervico-vaginal fluid. Previous work by O'Flynn and colleagues demonstrated that cervico-vaginal fluid cytology could detect endometrial cancer with 89.6% sensitivity and 88.7% specificity in a study of 216 women, but cytology is labour-intensive and requires highly trained cytopathology specialists. Proteomics and machine learning offer a more scalable and reproducible alternative.
Blood plasma as a complementary biofluid: Plasma is attractive for cancer detection due to its simplicity and ease of collection. However, blood-based biomarkers may be limited by the low yield of cancer-derived signals in blood, especially in small and early-stage tumours. This study directly compares the diagnostic performance of protein biomarkers derived from cervico-vaginal fluid against those from matched plasma samples, providing a head-to-head evaluation of both biofluids.
Ethics and recruitment: The study received ethical approval from the North-West Greater Manchester Research Ethics Committee (reference 16/NW/0660). Women referred with postmenopausal bleeding, as well as those with known endometrial cancer, were recruited from the Gynaecology Outpatient Departments at St Mary's Hospital (Manchester University NHS Foundation Trust) and the Royal Oldham Hospital (Northern Care Alliance NHS Group) between April 2019 and March 2020. In total, 118 symptomatic post-menopausal women participated, including 53 (45%) with confirmed endometrial cancer and 65 (55%) controls with no evidence of cancer or atypical hyperplasia.
Cohort characteristics: The median age of the full cohort was 57 years (IQR 52-67), and the median BMI was 28 kg/m2 (IQR 24-34). Women with endometrial cancer were significantly older (median 67 years vs. 53 years, p < 0.0001) and had higher BMI (median 30.8 vs. 27.0 kg/m2, p = 0.048). The cohort was predominantly White British (86%), with 10% Asian and 4% Afro-Caribbean participants. Most cancer cases had low-grade (64% grade I/II), early-stage (77% FIGO stage I) tumours of endometrioid histological phenotype (79%). Eighteen women (38%) had lympho-vascular space invasion, and 21 (45%) had myometrial depth of 50% or greater.
Control group selection: Controls were symptomatic women who had completed routine diagnostic investigations (TVS, endometrial biopsy, and/or hysteroscopy) with no evidence of endometrial cancer or atypical hyperplasia. Women with benign conditions such as atrophic vaginitis and benign polyps were eligible as controls. This is an important strength: because the intended clinical use case is triaging symptomatic women, using symptomatic controls (rather than healthy volunteers) provides a more realistic estimate of diagnostic performance. Control participants were followed for 12 months to minimize the possibility of misclassification.
Power estimation: A power analysis confirmed that a sample size of 100 women (n = 50 cases, n = 50 controls) was required to identify a true endometrial cancer biomarker signature with an expected AUC of 0.90 (95% CI 0.84-0.96) at 95% confidence and greater than 90% power. The final cohort of 118 exceeded this requirement.
Cervico-vaginal fluid collection: Samples were collected using the Delphi screener (Rovers, Netherlands), a sterile plastic syringe-like device approximately 20 cm in length. The device was inserted into the posterior fornix of the vagina, a reservoir of saline was expelled, and the fluid was re-aspirated by taking pressure off the plunger while slowly rotating and retracting the device. The Delphi screener has demonstrated superiority in terms of reproducibility, sample quality, and patient acceptability when compared with other collection methods, and reports lower mean pain scores than diagnostic hysteroscopy or endometrial biopsy. It also has the advantage of being a self-sampling device that can be used in community settings by practice nurses. Following collection, cervico-vaginal samples were centrifuged at 1000 g for 10 minutes to separate cellular pellets from supernatant fractions.
Plasma collection and preparation: Blood samples were collected simultaneously in standard EDTA tubes, centrifuged at 1500 g for 15 minutes, and the plasma supernatant was stored at -80 degrees C. Plasma samples were immunodepleted of the 12 most abundant proteins using Pierce Top 12 Abundant Protein Depletion Spin Columns to reduce the dynamic range and improve detection of lower-abundance, potentially cancer-related proteins. Both plasma and cervico-vaginal fluid samples were then purified and concentrated using Amicon Ultra-0.5 centrifugal filter devices and Agilent spin concentrators, respectively.
SWATH-MS profiling: Sequential window acquisition of all theoretical mass spectra (SWATH-MS) is a data-independent acquisition proteomics platform with high reproducibility, precision, and accuracy. All samples were analysed on a 6600 TripleTOF mass spectrometer (Sciex) using a 120-minute liquid chromatography gradient, with duplicate injections per sample. The spectral data were searched against a published bespoke consensus spectral library comprising 19,394 peptides and 2,425 cervico-vaginal fluid proteins (for cervico-vaginal data) and the human plasma library (for plasma data) using OpenSWATH version 2.0.0. Peptide matches were assessed using pyProphet (version 0.18.3) and aligned using the TRIC tool. Researchers were blinded to clinical data and histopathological findings during sample preparation and mass spectrometric analyses.
Protein quantification results: Across all sample types, 941 unique proteins were identified. The cervico-vaginal fluid supernatant yielded 597 proteins, matched vaginal cell pellets yielded 310, and plasma yielded 533. Of these, 302 proteins were exclusive to plasma, 203 to vaginal supernatant, and 29 to vaginal cell pellets. Only 90 proteins were quantified across all three sample types, underscoring the distinct proteomic landscapes of these biofluids.
Differential protein expression: The proportion of differentially expressed proteins (log2 fold change greater than 1.0) varied dramatically by sample type: 32.5% in cervico-vaginal supernatant versus 13.3% in cell pellets versus only 1.3% in plasma (p < 0.05, chi-square test). This immediately suggested that cervico-vaginal fluid supernatant would be the most informative biofluid for endometrial cancer detection. In the supernatant, fold changes ranged from -4.4 to +3.6, while plasma showed a much narrower range of -1.9 to +1.1. Principal component analysis (PCA) using all cervico-vaginal supernatant proteins showed good visual separation between cancer and control samples.
Random forest for feature selection: The study used the random forest (RF) algorithm (randomForest package in R) with a random seed set at 1000 for feature selection. An approximate 80/20 bootstrap split was used for training and testing. The model was first tuned to determine the optimal mtry parameter (the number of features considered at each decision tree split) on the training set, then trained using 1,000 trees. Discriminatory proteins were ranked by their contribution to the Mean Decrease in Accuracy metric. PCA using just the top 10 discriminatory proteins showed a much stronger degree of separation between cancers and controls than using all proteins.
Nested logistic regression models: Forward stepwise regression modelling, adjusting for age and BMI as continuous covariates, was used to build nested logistic regression models of increasing complexity. Proteins were incorporated sequentially based on their RF ranking. Model performance was assessed through receiver-operator characteristic (ROC) curves, AUC with 95% confidence intervals (using 2,000 bootstrap replicates), and Akaike information criteria (AIC). Likelihood ratio tests compared nested model performances, and the parsimonious model was defined as the one that best balanced diagnostic accuracy with model simplicity.
Validation with Boruta algorithm: To verify the robustness of the RF-selected biomarkers, the authors performed a parallel feature selection analysis using the Boruta algorithm. Unlike standard RF, the Boruta algorithm compares each candidate feature's classification performance against randomly created "shadow features" and incorporates information from all collinearly related proteins rather than selecting one arbitrarily. For cervico-vaginal fluid, the Boruta algorithm confirmed 38 proteins as important, and these were consistent with the top discriminatory biomarkers identified by the RF model. For plasma, 6 Boruta-confirmed proteins aligned with the RF top biomarkers.
The parsimonious cervico-vaginal model: Sequential addition of discriminatory proteins improved model performance, with a relative plateau reached between the 5th and subsequent models. The 5-biomarker panel combining HPT (haptoglobin), LG3BP (galectin-3-binding protein), FGA (fibrinogen alpha chain), LY6D (lymphocyte antigen 6D), and IGHM (immunoglobulin heavy constant mu) had the lowest AIC value (83) and was selected as the parsimonious model. It predicted endometrial cancer with an AUC of 0.95 (95% CI 0.91-0.98), sensitivity of 91% (83%-98%), and specificity of 86% (78%-95%). The negative predictive value, adjusted for a 9% disease prevalence in symptomatic post-menopausal women, was 99% (97%-100%).
Building up to the 5-protein panel: The first model (HPT alone) achieved an AUC of 0.89 (0.84-0.95) with 66% sensitivity and 85% specificity. Adding LG3BP (model 2) improved the AUC to 0.90 (0.85-0.96) with sensitivity rising to 74%. The addition of FGA (model 3) pushed the AUC to 0.91 (0.87-0.96) with 83% sensitivity. Model 4 (adding LY6D) reached AUC 0.94 (0.91-0.98) and 87% sensitivity. The 5-protein model (adding IGHM) hit the final AUC of 0.95 with 91% sensitivity. Adding a sixth protein (FN1) did not improve any accuracy metric, confirming the 5-marker panel as the optimal balance of performance and simplicity.
The plasma model comparison: In contrast, the best plasma model was a 3-marker panel combining APOD (apolipoprotein D), PSMA7 (proteasome subunit alpha type-7), and HPT (haptoglobin). This panel predicted endometrial cancer with an AUC of 0.87 (0.81-0.93), sensitivity of 75% (64%-86%), and specificity of 84% (75%-93%). HPT was the only protein shared between the cervico-vaginal and plasma panels. No significant improvement was observed with incorporation of additional plasma protein classifiers beyond the 3-marker panel.
Why cervico-vaginal fluid outperformed plasma: The superior performance of cervico-vaginal fluid is consistent with its proximal nature, as it derives from or has been in direct contact with endometrial tumours. The significantly reduced protein dynamic range in cervico-vaginal fluid compared to plasma allows for better sensitivity in detecting clinically relevant biomarkers. In plasma, cancer-derived signals are diluted into a much larger circulating proteome, reducing the signal-to-noise ratio for early-stage detection.
Early-stage detection (FIGO stage I): Because early detection drives survival, the authors specifically tested biomarker performance for stage I disease (n = 41 cases vs. n = 65 controls). A 3-marker cervico-vaginal panel of HPT, LY6D, and C5 (complement component 5) predicted stage I endometrial cancer with an AUC of 0.92 (0.87-0.97). The corresponding plasma model, a 4-marker panel of CNDP1, CDC5L, APOD, and PRDX6, achieved an AUC of 0.88 (0.82-0.95). These results confirm that protein biomarkers can detect the disease at its most treatable stage, though validation in larger cohorts is needed.
Advanced-stage detection (FIGO stages II-IV): A clinically useful biomarker assay must also identify cancers that have spread beyond the uterus. A 5-biomarker cervico-vaginal panel of APOE, GGCT, CFAB, LY6D, and CEAM5 predicted advanced-stage endometrial cancer with an AUC of 0.96 (0.92-1.00). The plasma counterpart, a 6-biomarker panel of SERPINA4, APOA2, TTR, CRP, CLEC3B, and APOD, achieved an AUC of 0.93 (0.85-0.99). Interestingly, the plasma panel performed relatively well here, which is consistent with the higher yield of cancer-derived signals in blood in advanced versus early-stage disease.
Non-endometrioid (aggressive) histology: Biologically aggressive, non-endometrioid tumours carry the worst prognosis. Cervico-vaginal fluid proteins PLTP, A2MG, APOE, FIBB, CO5, and FIBA each individually achieved AUCs above 0.97, and a combined 6-protein signature predicted non-endometrioid cancers with a near-perfect AUC of 0.99 (0.98-1.00). A 5-protein plasma panel (SELL, CCT3, IGF2, IFGALS, IGFBP3) achieved an AUC of 0.88 (0.77-1.00) for the same task. This suggests that aggressive cancers shed protein biomarkers into cervico-vaginal fluid more readily than less aggressive tumours.
Clinical significance: The protein signatures identified in this study outperformed clinical risk predictors such as age, BMI, and even endometrial thickness (the current gold-standard triage tool) in predicting endometrial cancer. These protein panels thus have the potential to replace or complement transvaginal ultrasound for triaging symptomatic women, and could potentially offer screening tools for women with a genetic predisposition to endometrial cancer, such as those with Lynch syndrome.
Galectin-3-binding protein (LG3BP): LG3BP was significantly increased in endometrial cancer cases and was the second most discriminatory cervico-vaginal biomarker. It plays a crucial role in integrin-mediated cell adhesion and stimulates the host defence against viruses and tumour cells. Studies have consistently demonstrated pro-tumorigenic properties for LG3BP via its regulation of cell proliferation, apoptosis, cell adhesion, angiogenesis, and metastasis. LG3BP has also been reported as upregulated in cancers of the colorectum, central nervous system, stomach, lung, and breast.
Lymphocyte antigen 6D (LY6D): LY6D identified endometrial cancer with an individual AUC of 0.89 and showed strong performance for early-stage tumour detection. LY6D plays an important role in lymphocyte differentiation, cell adhesion, cancer progression, and immune escape. It has been associated with distant metastasis in oestrogen receptor-positive breast cancer and shown to be prognostic in hepatocellular malignancies and cancers of the head and neck.
Immunoglobulin heavy constant mu (IGHM): IGHM is an antibody produced by B lymphocytes that plays a crucial role in the body's primary defence mechanisms, involved in early recognition and elimination of precancerous and cancerous lesions. The upregulation of IGHM in cervico-vaginal fluid of women with endometrial cancer likely results from an immunological response to the malignancy. There is also evidence that endometrial cancers harbour more mutations in the IGHM protein compared to cancers of other sites.
HPT and FGA: HPT (haptoglobin) is an acute-phase glycoprotein that regulates the immune response. Serum HPT levels have been reported as elevated in lung, breast, and ovarian malignancies. It was the only protein shared between the cervico-vaginal and plasma panels. FGA (fibrinogen alpha chain) is a cell adhesion molecule and a cancer-related gene that has been previously reported as a biomarker in endometrial cancer. Functional pathway analysis confirmed that the top discriminatory proteins have inflammatory, immune, and protein regulatory functions, supporting their biological plausibility as cancer biomarkers.
DNA methylation-based tests: Herzog and colleagues used cervical smear specimens from 726 women and validated in 562 cervico-vaginal fluid samples, identifying a 3-marker assay for endometrial cancer detection based on DNA methylation changes in GYPC and ZSCAN12 gene regions. That test detected endometrial cancer with sensitivities of 97.2%, 90.1%, and 100% across cervical, self-collected genital, and vaginal swab samples, respectively. However, their study was limited by its case-control design and low DNA yield in up to 12% of self-collected samples. The PapSEEK test, incorporating assays for mutations in 18 genes plus aneuploidy in Pap brush samples from 382 women, demonstrated an 81% detection rate (93% when a Tao brush was used).
Single-analyte approaches: He and colleagues reported elevated levels of cervico-vaginal cancer antigen 125 (CA125) in women with endometrial cancer (n = 148) compared to controls (n = 77). However, a pilot study by Calis et al. found that cervico-vaginal CA125 detected endometrial precancer or cancer with only 78% sensitivity at a threshold of 575 micrograms per millilitre. This suboptimal sensitivity carries real clinical risk, including false reassurance and delayed diagnosis. Notably, the current study did not identify cervico-vaginal CA125 as an important biomarker for endometrial cancer detection.
Uterine aspirate proteomics: Martinez-Garcia and colleagues explored 52 proteins in uterine aspirates from 69 endometrial cancer cases and 47 controls, finding that MMP9 and KPYM combined detected endometrial cancer with 94% sensitivity and 87% specificity. Bakkum-Gamez and colleagues used tampon-collected vaginal fluid and showed that 28 methylated DNA markers discriminated cancer from controls with 76% sensitivity at 96% specificity (AUC 0.88). However, vaginal tampons are generally unappealing to post-menopausal women and may be inadequate for detection in women without bleeding symptoms.
Somatic mutation detection: Pelegrina and colleagues found that 73% of cervico-vaginal fluid samples from women with endometrial cancer had detectable somatic mutations, adding to the growing body of evidence for minimally invasive detection. Taken together, this landscape of emerging approaches highlights the competitive performance of the current study's proteomic panel (AUC 0.95), which rivals or exceeds most alternative strategies while using a sampling method (the Delphi screener) that is more acceptable and practical than several alternatives.
Small sample size and overfitting risk: With only 53 cancer cases and 65 controls, the study acknowledges that accuracy estimates for the parsimonious biomarker models may be biased by overfitting, especially for the advanced-stage and non-endometrioid tumour analyses where subgroup sizes were even smaller. The near-perfect AUC of 0.99 for non-endometrioid cancers, while striking, must be interpreted with particular caution given the limited number of non-endometrioid cases in the cohort. Validation in a larger independent cohort is essential before any clinical translation can proceed.
Case-control design limitations: The study is prone to verification bias and other inherent limitations of a case-control design, where controls are selected based on known disease status rather than sampled from a prospective screening population. While the authors mitigated misclassification risk by following controls for 12 months, the study remains subject to potential residual confounding. Although models were adjusted for age and BMI, other confounding factors cannot be excluded.
Demographic and molecular gaps: The cohort was 86% White British, limiting the ability to extrapolate findings to women from other ethnic backgrounds and nationalities. The study also lacked data on the molecular classification of endometrial cancer (the four TCGA subtypes: POLE ultramutated, microsatellite instability-high, copy-number low, and copy-number high), which precluded exploration of biomarkers for more specific molecular phenotypes. Additionally, the study did not assess performance in asymptomatic women or premenopausal women with genetic predisposition, such as Lynch syndrome.
SWATH-MS not directly translatable: The use of SWATH-MS for protein quantification, while excellent for biomarker discovery, is not easily deployable in routine clinical settings. The authors acknowledge that clinical translation will require development of assays based on more practical platforms such as ELISA, Lumipulse technology, or lateral flow test technology for point-of-care testing. This transition from discovery-phase mass spectrometry to validated clinical immunoassays represents a significant and time-consuming step.
Validation in larger, diverse cohorts: The most immediate next step is confirmatory studies using immunoassays or targeted proteomics in a larger and more diverse study cohort. These studies need to include women from a wider range of ethnic backgrounds, premenopausal women at high genetic risk (e.g., Lynch syndrome), and asymptomatic women to understand how the biomarker panels perform outside the symptomatic post-menopausal population. Prospective study designs, rather than case-control, will be essential for generating reliable estimates of sensitivity, specificity, and predictive values in a real-world clinical population.
Platform translation: The authors envision that the identified protein signatures will be translated to clinically actionable assays based on ELISA, Lumipulse technology, or lateral flow test technology. A lateral flow test would enable point-of-care testing, potentially allowing rapid triage of symptomatic women in primary care or even self-testing at home, similar to what has been achieved with COVID-19 lateral flow assays. A validated cervico-vaginal fluid multi-protein biomarker assay could lead to rapid discrimination between symptomatic women with and without endometrial cancer, with substantial cost-saving implications for health service providers.
Clinical pathway integration: The protein panels identified in this study outperformed endometrial thickness measured by transvaginal ultrasound, the current gold-standard triage tool. If validated, these panels could replace or complement TVS in the diagnostic pathway for symptomatic women, reducing the number of unnecessary hysteroscopies and endometrial biopsies. The Delphi screener's self-sampling capability also opens the possibility of community-based or home-based testing, removing the need for an initial clinical visit and reducing pressure on specialist gynaecology services.
Elucidating mechanistic roles: Beyond clinical translation, the authors highlight the need for further studies to elucidate the mechanistic roles of the identified biomarkers in endometrial carcinogenesis. Understanding why these proteins are differentially expressed could provide insights into tumour biology, identify new therapeutic targets, and inform the development of more refined multi-omics diagnostic panels that combine proteomic, genomic, and methylation-based biomarkers for maximum accuracy.