Methylation-based biomarkers from thin-prep cytology test samples for early detection of cervical cancer: application and clinical potential Authors: Zhihan Liu, Guorong Liu, Shuanping Liu, Xiuxiu Wang, Yunshi Liang, Yuanyuan Guo, Baochen Du, Nabo Miao, Wenli Shi, Qiuyun Zhu, Liping Yu, Cui Deng, Guo Chen, Lijuan Huang, Lingling Zhan, Xiangyu Zheng, Jie Zhao, Yongjun Li, Xiaoliang Han, Yang Cheng, Jing Yang Published: 2026 ABSTRACT AbstractBackgroundAlthough human papillomavirus (HPV) testing and Thin-Prep Cytology Test (TCT) are currently adopted as the first-line screening methods for cervical cancer, both of them have limitations. This study aimed to develop and validate new methylation-based biomarkers from formalin-fixed paraffin-embedded (FFPE) tissue and TCT samples in a Chinese population to improve early detection of cervical cancer (CC).MethodsWe performed a differential methylation analysis of cervical FFPE tissues and TCT samples through targeted methylation sequencing. Quantitative methylation-specific PCR (qMSP) was performed on cervical tissues, TCT samples and cell lines to construct a diagnostic model containing three highly methylated sites, namely JAM3, VSTM2B and EPHA7. The clinical effectiveness was evaluated on overall 279 TCT multiple group samples, comprising 33 squamous cell carcinomas (SCC), 14 adenocarcinoma(ADC), 43 high-grade squamous intraepithelial lesion (HSIL), 49 low-grade squamous intraepithelial lesion (LSIL), 78 high risk-HPV (HR-HPV)–positive/cytology-normal (NILM) cases, 32 healthy controls (NC), and 30 benign gynecological samples. The model performance was assessed using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).ResultsIn an independent validation cohort (n = 108), the panel of three methylation-based biomarkers showed outstanding diagnostic accuracy. It achieved sensitivities of 100.0% for SCC (n = 10), 80.0% for ADC (n = 5), and 86.7% for HSIL (n = 15). Specificity was 90.6% in HR-HPV–positive women (n = 32), 100.0% in NC (n = 12), and 93.3% in other benign conditions (n = 15), with 89.5% specificity in LSIL (n = 19). Overall sensitivity and specificity were 90.0% and 92.3%, respectively. HPV subtype–stratified analysis revealed 90.9% sensitivity and 85.7% specificity for HPV16/18–positive samples, with consistent performance for non-HPV16/18 subtypes (85.7% sensitivity, 89.5% specificity).ConclusionsThe JAM3/VSTM2B/EPHA7 methylation model (JAM3m/VSTM2Bm/EPHA7m) has shown excellent performance in the detection of cervical cancer and precancerous lesions. It provides a new direction for overcoming the limitations of the traditional HPV and TCT combined screening method.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13148-026-02154-5.Keywords: DNA methylation, Cervical cancer, HR-HPV Background Cervical cancer (CC) ranks as the fourth most frequently diagnosed cancer among women worldwide [1]. Globally, there are approximately 661,000 new cases and 348,000 deaths annually, with 89.9% of new cases and 91.6% of deaths occurring in low- and middle-income countries [2]. The pathological progression of CC proceeds through a series of well-characterized precursor stages, classified as cervical squamous intraepithelial neoplasia grades 1–3 (CIN1, CIN2, and CIN3), which may ultimately advance to invasive carcinoma. Clinical management strategies are different substantially at these different stages. CIN1 often regresses spontaneously and usually requires no treatment, with clinical management typically involving follow-up observation, while CIN2 requires age-dependent intervention considering reproductive status. CIN3 and invasive carcinoma necessitate immediate therapeutic intervention. Histologically, cervical malignancies predominantly manifest as SCC (80%) and ADC (20%) [3, 4]. Prognostic outcomes depend on clinical stage strongly. The 5-year survival rate for patients with FIGO 2018 criteria stage IB cervical cancer is approximately 90–95%. However, this rate declines to between 86.4 and 92.4% upon progressing to stage IIA and further decreases to 56.3% in advanced-stage IIIC2 disease. These stark survival differences underline the critical importance of early detection through effective screening protocols, which will significantly improve therapeutic success and survival outcomes [5, 6]. Persistent infection with HR-HPV is the main cause of cervical cancer. Among all HR-HPV, HPV 16 and 18 are responsible for most of the cervical cancers. Current epidemiological data reveal that approximately half of cases of cervical cancer worldwide can be attributed to HPV 16 infection, while HPV 18 accounts for 10%. The current strategies for early screening of cervical cancer include TCT and HPV testing. However, both of these methods have their own limitations so as to affect their screening efficacy. The TCT aims to screen for cervical cancer and precancerous lesions by identifying abnormal cells through their cellular morphology. It has the characteristics of high specificity (75–85%) and low sensitivity (60-70%).In addition, the quality of cytology is greatly reliant upon sample quality and remains a largely subjective analysis dependent upon cytologists.As for HPV test, despite high sensitivity (HR-HPV >95%) and negative predictive values (>99%), HR-HPV test is associated with sub-optimal specificity (60–70%). Most of HPV infections are usually eliminated spontaneously by the body, even persistent HR-HPV infections rarely progress to cervical cancer. A positive HR-HPV test result can easily lead to excessive anxiety, overdiagnosis, overtreatment and waste of medical resources. Therefore, it is urgent to find a molecular biological screening marker that does not rely on traditional morphological observation, to improve the coverage and accuracy of screening, and it is simple and easy to implement. DNA methylation is a well-known epigenetic marker that plays a crucial role in maintaining normal cellular functions, genetic imprinting, embryonic development and the occurrence of human tumors. The abnormal methylation state is manifested by a decrease in the genomic methylation level and high methylation of the CpG islands in the promoters of certain genes (such as tumor suppressor genes). Abnormal DNA methylation has been confirmed in many studies, contributing to the induction of proto-oncogenes, genomic instability and tumorigenesis. It was found that the frequency of certain gene methylation increases as the severity of cervical lesions increases, and the number of methylated genes also increases as the degree of cervical intraepithelial lesions worsens [7]. Nowadays, an increasing number of researchers are paying attention to the methylation detection of cervical lesions, especially in the screening of cervical cancer and precancerous lesions. Approximately 100 abnormal methylation genes have been detected in cervical cancer and nearly 20 genes have been reported in different studies. Among them, genes such as SOX1, JAM3, PAX1, CDKN2A, CADM1, MAL, miR-124, FAM19A4 and EPB41L3 were all hypermethylated in cervical high-grade lesions and cervical cancer [8]. A methylation analysis of DLX1/ITGA4/RXFP3/SOX17/ZNF67/ASTN1 revealed that it has a good detection performance for cervical high-grade lesions. The sensitivity and specificity were 64.5–76.2% and 76.0–94.6% respectively [9]. However, the screening of methylation-based biomarkers mostly comes from literature reports or related databases, and only few methylation-based biomarkers have been identified through high-throughput DNA methylation sequencing from clinical tissue samples and TCT samples. Additionally, most of the methylation-based biomarkers are focused on SCC. The methylation-based biomarkers for ADC lacks large-sample clinical studies. In this study, we aim to identify, optimize, and clinically validate methylation-based biomarkers for the early detection of cervical cancer and precancerous lesions. To achieve this, we conducted a comprehensive differential methylation analysis on clinical tissue and cytology samples. Through systematic screening, we selected three hypermethylated DNA regions strongly associated with cervical cancer. A predictive model was established using a training cohort and rigorously validated in an independent cohort. The validated model demonstrated superior performance for early-stage detection compared to conventional methods. This work established a promising methylation-based diagnostic framework and enabled more precise risk stratification in clinical management. Materials and methods Study design and participants Study design The investigation was systematically conducted through three phases: (1) biomarker discovery, (2) marker screening, and (3) clinical validation. In the first phase, the targeted methylation sequencing data of FFPE tissues of cervical, TCT samples and FFPE tissues of EC (used to exclude nonspecific methylation signals) were analyzed to determine the differentially methylated sites. Subsequently, the DMRs were identified between normal cervices and cervical cancer. In the second phase, among the methylation candidates, our in-house data were combined with published methylation hotspot genes associated with cervical cancer (JAM3, ASTN1, DLX1, ITGA4, RXFP3, SOX17, ZNF582) [10–12], the top 9 was selected for verification by qMSP on FFPE tissues, TCT samples and CC cell lines. We evaluated the detection sensitivity and specificity of the markers to select the optimal methylation-based biomarker In the third phase, a clinical cohort (n = 279) was used to construct and validate a methylation-based diagnostic model. The diagnostic performance of the biomarker panel for cervical cancer and precancerous lesions was assessed using metrics including sensitivity and specificity. Study populations and samples This investigation was conducted in strict compliance with the Declaration of Helsinki and received ethical approval from the Institutional Review Board of Guangzhou First People’s Hospital (Approval No.: K-2021-109-05). All patients provided their consent before enrollment. The inclusion criteria consisted of the following: aged 18 or older; performed TCT and HPV test; definitive cervical histological pathology findings; no history of cervical cancer or precancerous lesions, and no history of treatment for related lesions; no history of medical radiotherapy; negative HIV results and no history of organ transplantation or immunosuppressive therapy. The study utilized a total of 381 samples across three phases. For biomarker discovery, 52 cervical FFPE tissues (comprising 5 SCC, 6 ADC, 11 HSIL, 15 NC and 15 EC) and 31 TCT samples (3 SCC, 2 ADC, 10 HSIL, 4 benign uterine lesions and 12 NC) were included. The screening phase employed 50 cervical FFPE tissues (comprising 5 SCC, 4 ADC, 19 HSIL, 22 NC), 30 healthy TCT samples and 13 CC cell lines. Clinical validation was performed on 279 TCT samples (comprising 33 SCC, 14 ADC, 43 HSIL, 49 LSIL, 78 HR-HPV positive only, 32 NC and 30 potential confounding cases such as cervicitis or endometrial hyperplasia). Cell line and clinical sample collection The FFPE tissues and residual TCT samples were collected from the Department of Pathology of Guangzhou First People's Hospital. TCT samples were obtained using a disposable cervical specimen collection brush and stored in Cell Prep Plus LBC Solution (LBP, Guangzhou, China) at − 80 °C. FFPE tissues were sectioned into 8-μm paraffin sections, collected into sterile microcentrifuge tubes, and stored at 4 °C. A panel of thirteen human CC cell lines (H1HeLa, HT-3, HeLa, SiHa, ME-180, HelaChangLiver, HeLa [ChangLiver], Hela, C33A, CaSki, HeLa229, HeLaS3, MS751) was used in this study. All cell lines were obtained from BNCC commercial. Detailed information, including the BNCC catalog numbers, cultivation conditions, and culture media, is provided in the supporting information. To confirm cell line identity and exclude potential cross-contamination, each cell line was authenticated by short tandem repeat (STR) loci prior to experimentation. The corresponding STR loci are also documented in the supporting information. Experiments and analysis DNA isolation and bisulfite modification Genomic DNA was extracted from frozen cervical exfoliated cell pellets using the Blood/Cell/Tissue Genomic DNA Extraction Kit (DP304, Tiangen Biotech, China), strictly following the manufacturer’s instructions. For FFPE tissue sections, genomic DNA was extracted using the QIAamp DNA FFPE Tissue Kit (Cat. No. 56404, QIAGEN, Germany). Due to potential DNA damage caused by formalin fixation, DNA quality was evaluated using the Agilent 2200 TapeStation system. Only samples with a DNA Integrity Number (DIN) > 6 and a total DNA amount > 1.5 μg were selected for subsequent procedures. For each qualified sample, 500 ng of DNA extracted from either FFPE tissues or TCT samples was fragmented to approximately 300 bp using a focused ultrasonicator (M220, Covaris, USA). The fragmented DNA was subjected to bisulfite conversion using the EZ DNA Methylation-Gold Kit (D5005, ZYMO RESEARCH, USA). Bisulfite-converted DNA was used for library construction with the manufacturer’s optimized library preparation kit. Quantitative methylation-specific PCR (qMSP) Bisulfite-converted DNA was amplified using the SLAN-96S Fully Automated Medical PCR Analysis System (ShangHong Medical Technology Co., Ltd., China). The PCR protocol was as follows: initial denaturation at 94 °C for 20 min; 45 cycles of 93 °C for 30 s and 57 °C for 35 s; and a final step at 40 °C for 5 s. Each reaction was performed in a 30 μL volume with primers and probes at a concentration of 200 nM. ACTB was used as a methylation-independent reference gene for quality control. Additionally, a positive control (PC) was included in each run to monitor assay performance. All input volumes and sample types are detailed in Supplementary Table 1. In this study, 187 differentially methylated regions (DMRs) were initially identified. Subsequently, 132 primer–probe sets were designed targeting these DMRs for qMSP validation. Fully methylated (Takara 3522) and unmethylated (Takara 3521) control DNA were used as positive and negative controls, respectively. A cycle threshold (Ct) cutoff of 45 was applied, with Ct < 45 defined as positive and no amplification as negative. Based on these criteria, 80 primer–probe sets were selected as candidate markers, demonstrating 100% amplification efficiency for the methylated control and no amplification for the unmethylated control. Bioinformatics analysis for biomarker discovery First, stringent filtering criteria were established based on sequencing depth, retaining only methylation detection regions where at least 80% of samples had an average sequencing depth >10X, thereby effectively reducing the impact of sequencing bias on the false-positive rate. Subsequently, the Mann-Whitney U non-parametric test (implemented using the Python package SciPy) was applied to statistically compare methylation levels between the case and control groups (Supplementary Fig. S1), and candidate regions with a p-value <0.05 were selected. A principal screening criterion mandated higher methylation levels in cervical cancer tissues than in endometrial cancer tissues, thereby ensuring specific detection of cervical cancer and excluding endometrial carcinoma. Through group-wise comparisons, it was observed that the mean methylation levels and inter-group differences (case – control) in the FFPE sample group were higher than those in the TCT sample group. Therefore, differential screening thresholds were established for the two sample types: For FFPE samples: methylation level in the case group > 0.1, in the control group < 0.035, and inter-group difference > 0.08. For FFPE samples: methylation level in the case group > 0.1, in the control group < 0.035, and inter-group difference > 0.08. For TCT samples: methylation level in the case group > 0.05, in the control group < 0.03, and inter-group difference > 0.05. For TCT samples: methylation level in the case group > 0.05, in the control group < 0.03, and inter-group difference > 0.05. Considering the exponential amplification effect of PCR, additional thresholds were set to ensure effective signal amplification: the 70th percentile of the case group was required to be at least 3 times (for FFPE) or 2.5 times (for TCT) higher than the 80th percentile of the control group. To ensure that the candidate markers exhibited good predictive performance, further sample-level filtering was performed. Under the condition that the specificity in the control group remained ≥ 80%, the sensitivity in the case group was required to reach at least 70% for FFPE samples and 60% for TCT samples, thereby identifying candidate markers with diagnostic potential. Through this multi-tiered screening approach, 377 and 1090 significantly differential DMRs were identified in FFPE and TCT samples, respectively. Finally, the intersection of DMRs obtained from FFPE and TCT samples was taken, resulting in 180 consistently significant DMRs. Based on a systematic review of existing literature on cervical cancer-related hotspot genes, 11 candidate genes were initially selected. Comparison with the 180 DMRs revealed that 4 of these genes were already covered within our DMR set. After deduplication, 7 candidate genes with research value were retained. Methylation profiling of these 7 genes was performed using an in-house TCT sequencing dataset, ultimately identifying 7 DMRs associated with cervical cancer. We have provided full details of all software used in Supplementary Table S2. Statistical analysis bioinformatic and statistic analysis In the differential analysis of relevant indicators in this study, we conducted the following three comparisons. First, we compared the Ct values of the three candidate biomarkers between the case group and the control group. Second, we compared the methylation levels of the JAM3 gene in FFPE samples and TCT samples between the case group and the control group. Third, we analyzed the scores of LSIL, HSIL, and Cancer in both the training set and the validation set. To ensure the reliability of the analytical results, Student’s t-test was employed to assess the significance of differences between the comparison groups. A p-value of less than 0.05 was considered statistically significant, indicating a notable difference between the compared groups. Use clinical samples to model and validate markers This study employed qMSP to assess methylation levels of three biomarkers. A diagnostic model was developed and validated using 279 clinical samples to evaluate its diagnostic performance. During the model development phase, 171 TCT samples (60CC/HSIL cases; 111 non-cancer controls) were analyzed. A methylation score model was established through logistic regression analysis, calculating diagnostic metrics including sensitivity and specificity. For independent validation, 108 samples (30 CC/HSIL specimens; 78 non-cancer controls) were subsequently evaluated using the same parameters to confirm model robustness. Result Discovery of methylation-based biomarkers for cervical cancer The experimental workflow of this study is schematically illustrated in Fig. 1. Three-phase workflow for methylation biomarker development: Discovery, prioritization, and clinical validation. Abbreviations: CC, cervical cancer; EC, endometrial carcinoma; qMSP, quantitative methylation-specific PCR; DMR, differentially methylated region In the biomarker discovery phase, comparative methylation profiling was conducted between 22 HSIL/CC FFPE tissues and 15 adjacent normal tissues, as well as among TCT samples comprising 10 HSIL, 5 CC cases, 4 benign cervical lesions, and 12 healthy controls. Analytical results revealed 10,693 FFPE tissue-associated DMRs, comprising 8,175 hypermethylated and 2,518 hypomethylated loci in cancer specimens. Similarly, 15,272 TCT-associated DMRs were detected, with 7,514 hypermethylated and 7,758 hypomethylated regions in cervical cancer samples (Fig. 2A). To establish robust qPCR-compatible biomarkers, candidate loci were prioritized through a dual-filter strategy: (1) exclusion of endometrial cancer-associated methylation patterns to ensure tissue specificity, and (2) stringent selection based on inter-group methylation divergence. This process yielded 377 FFPE-derived and 1,090 TCT-derived hypermethylated candidates, with intersection analysis retaining 180 consensus DMRs for subsequent validation. Preliminary screening of methylation biomarkers. A Comparative analysis of DMRs between FFPE and TCT sequencing datasets. B Heatmap illustrating methylation levels of 180 identified DMRs in tissue (left panel) and TCT (right panel) samples. Distinct methylation patterns are observed across sample types. C Pathway enrichment analysis of candidate biomarkers, highlighting biologically relevant pathways associated with the identified DMRs. D Genomic distribution characteristics of DMRs, presented as a pie chart to delineate their chromosomal localization preferences Following the identification of 180 consensus DMRs, we systematically evaluated their methylation profiles in both FFPE tissue and TCT sample datasets. The left panel of Fig. 2B illustrates the methylation levels of 180 candidate DMRs derived from panel-targeted sequencing of FFPE tissues, while the right panel displays corresponding methylation patterns observed in TCT samples through identical sequencing methodology. Cross-platform comparative analysis demonstrated concordant hypermethylation of these loci across both specimen types, confirming robust tissue-cytology consistency in methylation signatures. Pathway enrichment analysis of genes harboring the identified DMRs revealed predominant associations with cellular proliferation and metabolic regulation (Fig. 2C). Genomic annotation of DMR loci identified 71.1% (128/180) of these regions localized to promoter-proximal regulatory elements (±2 kb from transcriptional start sites), as quantified in Fig. 2D. Refined screening strategy for cervical cancer-specific methylation-based biomarkers In this study, a multi-stage screening strategy was employed, leading to the identification of nine robust biomarkers based on detection data from 13CCcell lines and 30 healthy controlTCT samples. The identified biomarkers include ZNF671, JAM3_1, RGS7, VSTM2B, EPHA7, JAM3_2, KHDRBS2, KCNK12, and ST6GALNAC5 (Fig. 3A). ROC analysis revealed superior diagnostic performance for JAM3(JAM_1), VSTM2B, and EPHA7, with AUC values exceeding 0.915 (Fig. 3B). Comparative Ct value distributions demonstrated marked differential methylation: cancer cell lines exhibited median Ct <40 versus ≥45 in healthy controls (Fig. 3C). A combinatorial detection model integrating JAM3, VSTM2B, and EPHA7 was subsequently established. Bisulfite sequencing validation of the top-performing biomarker, JAM3, confirmed its ability to discriminate between cancerous and non-cancerous tissues in both FFPE and TCT samples (Fig. 3D). The qMSP primer and probe sequences used in this study are given in supplementary table S3. Selection of the optimal methylation–based biomarkers. A Performance evaluation of nine candidate biomarkers across 13 CC cell lines and 30 healthy TCT samples. Bar Chart visualization quantifies biomarker sensitivity (red, methylation frequency in cancer cell lines) and specificity (green, preservation of unmethylated status in healthy controls). Analyses were performed using qMSP. B ROC curves comparing the discriminatory power of individual biomarkers for distinguishing cancer cell line DNA from healthy TCT samples. AUC values are annotated for each candidate. C Distribution profiles of threshold cycle (CT) values for three validated methylation-based biomarkers. Boxplots contrast methylation levels between malignant specimens (red, n = 43) and non-cancerous controls (green, n = 61), with whiskers representing 5th-95th percentiles. D Bisulfite sequencing validation of JAM3_1 methylation status in independent sample sets: FFPE tissues (22 cancers vs. 15 benign) and TCT samples (15 cancers vs. 16 controls). Methylation boxplot plots demonstrate locus-specific epigenetic alterations across sample types. Left panel: FFPE cohort; Right panel: TCT cohort Development and validation of a three biomarker-based diagnostic model for cervical cancer using clinical cohorts data Subsequently, we performed qMSP assays for the three selected biomarkers using 279 TCT samples. A logistic regression model was developed based on Ct values, with a score threshold of 0.366 established through analysis of the 171-sample training cohort. Samples yielding a score below this threshold were classified as negative, while those exceeding it were recommended for confirmatory colposcopy. The combined biomarker panel demonstrated high diagnostic accuracy in the training set, achieving 93.3% sensitivity and 93.7% specificity. Remarkably, the model maintained robust performance in an independent validation cohort(n=108), showing 90.0% sensitivity and 92.3% specificity. ROC analysis revealed exceptional discriminative capacity, with AUC values exceeding 0.950 in both training and validation sets (Fig. 4A). To mitigate the risk of overfitting, all samples were evaluated using 5-fold cross-validation (Supplementary Table S4). Comparative analysis confirmed the logistic regression model’s superior diagnostic performance over individual biomarker assessments. Demographic and clinical characteristics of the training and validation cohorts are detailed in Table 1. The age distribution of the study population is summarized in Supplementary Table S5. Development and validation of the three-marker diagnostic model for cervical cancer. A ROC curves of markers in training and validation cohorts, displaying AUC values for individual markers (JAM3, VSTM2B, EPHA7) and the combinatorial logistic regression model. The training set comprised case groups (30 invasive cancers, 28 HSIL cases) and control groups (30 CIN1 cases, 46 HR-HPV-positive/colposcopy-negative subjects, 20 healthy controls, 15 other benign pathologies). The validation set included case groups (15 cancers, 15 HSIL cases) and control groups (19 LSIL cases, 32 HR-HPV-positive/colposcopy-negative subjects, 12 NC, 15 benign pathologies). B Diagnostic performance comparison across detection approaches: HPV testing alone(blue), HPV/TCT co-testing (green), and the methylation model (red). Left panel: training cohort (n = 171); Right panel: validation cohort (n = 108). Sensitivity/specificity metrics are annotated for each method. C Stacked histograms stratifying detection outcomes by HPV genotypic risk: Upper panels (HPV 16/18-positive subgroups), Lower panels (non-16/18 HPV-positive subgroups). Left: training cohort; Right: validation cohort. D Methylation score assessment across histological categories (LSIL, HSIL, invasive carcinomas) in both cohorts, presented via boxplots with median/interquartile ranges Table 1Demographic and clinical characteristics of the training and validation cohortCharacteristicAllModel constructionIndependent validationN (%)PerformanceN (%)PerformanceNumbers279171108Age (years)Mean (range)45.07(17–79)45.88(17–79)43.79(17–78)Histopathological diagnosis, N(%) SCC3323 (13%)100.0% (23/23)(85.2–100.0)10 (9%)100.0% (10/10)(69.2–100.0) ADC149 (5%)88.9% (8/9)(51.8–99.7)5 (5%)80.0% (4/5)(28.4–99.5) HSIL4328 (16%)89.3% (25/28)(71.8–97.7)15 (14%)86.7% (13/15)(59.5–98.3) LSIL4930 (18%)90.0% (27/30)(73.5–97.9)19 (18%)89.5% (17/19)(66.9–98.7) HR HPVinfection7846 (27%)95.7% (44/46)(85.2–99.5)32 (30%)90.6% (29/32)(75.0–98.0) Healthy3220 (12%)95.0% (19/20)(75.1–99.9)12 (11%)100.0% (12/12)(73.5–100.0) Others3015 (9%)93.3% (14/15)(68.1–99.8)12 (11%)93.3% (14/15)(68.1–99.8)Sensitivity (%)93.3% (56/60)(83.8–98.2)90.0% (27/30)(73.5–97.9)Specificity (%)93.7% (104/111)(87.4–97.4)92.3% (72/78)(84.0–97.1) Demographic and clinical characteristics of the training and validation cohort 45.07 (17–79) 45.88 (17–79) 43.79 (17–78) 100.0% (23/23) (85.2–100.0) 100.0% (10/10) (69.2–100.0) 88.9% (8/9) (51.8–99.7) 80.0% (4/5) (28.4–99.5) 89.3% (25/28) (71.8–97.7) 86.7% (13/15) (59.5–98.3) 90.0% (27/30) (73.5–97.9) 89.5% (17/19) (66.9–98.7) HR HPV infection 95.7% (44/46) (85.2–99.5) 90.6% (29/32) (75.0–98.0) 95.0% (19/20) (75.1–99.9) 100.0% (12/12) (73.5–100.0) 93.3% (14/15) (68.1–99.8) 93.3% (14/15) (68.1–99.8) 93.3% (56/60) (83.8–98.2) 90.0% (27/30) (73.5–97.9) 93.7% (104/111) (87.4–97.4) 92.3% (72/78) (84.0–97.1) The methylation model demonstrated distinct diagnostic advantages over conventional approaches. In the training cohort comprising HR-HPV-infected cases with colposcopy-negative findings and LSIL specimens, the model achieved significantly higher specificity(95.7% and 90.0%, respectively) compared to HR-HPV testing(0% specificity in both groups) and combined HR-HPV/TCT testing(47.8%and 16.7%). Although the model’s sensitivity 93.3% was marginally lower than conventional methods 94.8% for HR-HPV alone and 100.0% for combined HR-HPV/TCT), it exhibited superior overall diagnostic accuracy as evidenced by the integrated discrimination improvement(IDI) of 0.42. Notably, the validation cohort demonstrated strong diagnostic consistency across all specimen categories, with performance metrics closely aligning with training set observations (Fig. 4B). In clinical practice, HPV testing plays a pivotal role in cervical cancer screening, with HPV16/18 demonstrating the highest carcinogenic potential. We conducted stratified analyses comparing HPV16/18-positive specimens (n = 72) with HPV-positive non-16/18 subtypes(n = 135) (Fig. 4C). Among the HPV16/18-positive cases, the methylation model achieved superior specificity of 93.3% in the training set and 85.7% in the validation set, whereas both HR-HPV testing and combined HR-HPV/TCT screening demonstrated 0% specificity across both cohorts. Similarly, for non-HPV16/18 subtypes, the model maintained high specificity levels of 93.3% and 89.5% in training and validation sets respectively, significantly outperforming conventional methods (HR-HPV testing: 0% specificity; combined testing: 45.0% training, 50.0%validation). Comprehensive analysis revealed statistically significant improvements in specificity for the methylation model compared to conventional screening methods. In this study, a methylation score was developed based on a logistic regression model to comprehensively reflect the overall methylation level and its association with disease characteristics, using Ct values derived from three methylation-based biomarkers. The analysis revealed a significant correlation between the methylation score and histopathological staging. Specifically, methylation levels exhibited a stepwise increase with disease progression: samples from low-grade squamous intraepithelial lesions (LSIL) showed methylation levels close to baseline; those from high-grade squamous intraepithelial lesions (HSIL) demonstrated a significant elevation in methylation; and invasive carcinoma samples displayed the highest methylation levels (see Fig. 4D). Furthermore, methylation levels of these biomarkers remained stable across patient age distributions, with no significant age-related fluctuations detected in subsequent analyses (Supplementary Fig. S2, Supplementary Table S6). Discussion In this study, we confirmed the effectiveness of the triplex combination of JAM3m /VSTM2Bm /EPHA7m in the detection of cervical cancer and precancerous lesions through the systematic identification and validation of methylation-based biomarkers for cervical cancer. Our findings indicate that the selected methylation-based biomarkers are significantly enriched in multiple key tumor suppression pathways. All of the 3 genes that we identified have previously been reported to be methylated in cancer. JAM3, a member of the adhesion molecule family, is critically involved in diverse biological processes, including angiogenesis, spermatogenesis and leukocyte migration [13, 14]. Abnormal methylation of JAM3 has been widely observed linked to tumorigenesis and tumor metastasis across diverse malignancies [15–17]. In cervical cancer, aberrant hypermethylation of the promoter region of the JAM3 leads to transcriptional silencing, which is a kind of mechanism closely linking to tumorigenesis and progression of cervical cancers [18, 19]. JAM3 is implicated in promoting cervical cancer metastasis via the activation of HIF-1α/VEGFA pathway [20]. A prospective cohort study involving 307 patients found that the sensitivity of JAM3/EPB41L3 dual-gene methylation detection for CIN2 + was 68.7%, and the specificity reached 96.7% [21]. In line with these findings, our data demonstrated high methylation levels of JAM3 in patients with CIN3 + . Another biomarker in our panel, EPHA7, belongs to the Eph receptor tyrosine kinase family and regulates cell proliferation and differentiation through intercellular signaling. EPHA7 exhibits a context-dependent dual role, acting as either a tumor promoter or suppressor, and is involved in modulating cell proliferation, migration, apoptosis, tumor differentiation, staging, and prognosis [22]. For example, in lung cancer, miR-448 and miR-944 enhance the proliferation and invasion of lung cancer cells by upregulating the expression of EPHA7 [23, 24]. On the contrary, EPHA7 is a tumor suppressor gene in tongue squamous cell carcinoma (TSCC). Li et al. used immunohistochemistry to analyze the expression of EPHA7 in 54 cases of TSCC tissues and surrounding normal tissues, and found that patients with high EPHA7 expression prolonged the survival, and the lower the expression, the more likely to occur lymph node metastasis and vascular infiltration [25]. Furthermore, the hypermethylation of EPHA7 exhibited the most significant difference between SCC and ADC compared to normal cervix (△βvalue = 0.464, P < 0.05), yet there was no significant difference between SCC and ADC [26]. These findings provide compelling evidence supporting the potential clinical value of EPHA7 in the diagnosis and screening of ADC. In contrast, VSTM2B has been less studied in the context of cervical cancer. However, existing evidence indicates its promoter is highly methylated in HPV-related oropharyngeal squamous cell carcinoma (OPSCC) [27], which suggesting a potential epigenetic regulatory role in HPV-associated malignancies. Overall, our findings align with and extend recent researches on methylation-based biomarkers for cervical cancer. The consistency of these results across different populations and detection platforms strengthens the evidence for their reliability as diagnostic tools. Consistent with prior evidence [7], our study corroborates the strong discriminatory power of DNA methylation profiling across the cervical lesion spectrum, encompassing LSIL, HSIL, and invasive carcinoma. Specifically, we observed a positive correlation between lesion severity and both the level and frequency of methylation events within our panel. Besides, the combination of HPV testing with JAM3m/VSTM2Bm/EPHA7m effectively improved both the sensitivity and specificity of cervical lesion detection. The integrated approach achieved a sensitivity of 90.0% (95% CI: 82.4%–94.8%) and a specificity of 92.3% (95% CI: 84.7%–96.5%), yielding an AUC of 0.950 (95% CI: 0.913–0.987). For patients infected with HPV16/18, immediate referral to colposcopy remains clinically indication due to the significantly elevated risk of precancerous lesions [28, 29]. However, the conventional co-testing strategy of HPV and TCT shows limited specificity (approximately 50%) in non‑16/18 HPV‑positive cases, whereas the JAM3m/VSTM2Bm/EPHA7m panel increased specificity to 89.5%.This objective molecular method overcomes the subjection of TCT diagnosis and compensates for the limited specificity of HPV testing. Consequently, it enhances the detection of potential high‑grade cervical lesions, reduces over‑diagnosis and unnecessary colposcopy referrals resulting from transient HPV infections, and minimizes overtreatment in low‑risk individuals. As of October 2024, more than 50 patents associated with DNA methylation detection for cervical cancer have been authorized in China. Over 30 kits utilizing methylation detection technology have received regulatory approval, of which 4 are specifically designed for cervical cancer screening. Methylation-based biomarkers include, but are not limited to, ASTN1, DLX1, ITGA4, RXFP3, SOX17, ZNF671, and JAM3 (https://www.nmpa.gov.cn). To objectively evaluate the performance of our triplex methylation panel (JAM3m/VSTM2Bm/EPHA7m), we compared its diagnostic accuracy with two commercially available kits, Qiasure (FAM19A4, miR124-2) and GynTect ® (ASTN1, DLX1, ITGA4, RXFP3, SOX17, ZNF671). The comparison focused on two core performance metrics: sensitivity and specificity. From the data [30], our panel maintained comparably high specificity relative to GynTect (92.3% vs. 91%) while substantially improving sensitivity (90.0% vs. 59%). Compared with Qiasure, our panel demonstrated superior performance in both sensitivity (90.0% vs. 65%) and specificity (92.3% vs. 72%). Collectively, these results indicated that the JAM3m/VSTM2Bm/EPHA7m panel have achieved a favorable balance between both high sensitivity and high specificity, underscoring its overall diagnostic advantage for cervical cancer detection. Several strengths should be acknowledged. First, targeted methylation sequencing was performed on cervical tissues, endometrial cancer tissues, and TCT samples from a Chinese population to screen for candidate methylation biomarkers. This strategy avoided the potential bias inherent in candidate‑gene‑based studies, thereby yielding more robust and reliable data. Additionally, VSTM2B and EPHA7 have not been widely reported were identified. Secondly, by constructing a multi‑gene biomarker panel rather than relying on a single marker, we significantly improved the diagnostic reliability and accuracy. Despite these encouraging results, several limitations should be noted. (1) The use of qMSP provided a cost-effective approach, but it may pose limitations for high-through clinical implementation. Emerging technologies such as digital PCR and next‑generation sequencing offer enhanced sensitivity, multiplexing capability, and scalability, which may facilitate broader adoption of methylation‑based biomarker testing in clinical practice [31]. Additionally, the biological mechanisms driving the hypermethylation of these biomarkers and their functional consequences in cervical carcinogenesis remain incompletely understood. This knowledge gap constrains a comprehensive interpretation of their clinical relevance and potential role in disease pathogenesis. (2) Although the sample size was sufficient for model construction and preliminary validation, it remains relatively modest, which may constrain the generalizability of our findings. Future work should focus on validating the stability of the biomarker panel in larger, multi‑center cohorts and evaluating its long‑term predictive value in cervical cancer screening programs. Conclusion Our study found that the three methylation-based biomarkers have both high sensitivity and high specificity in the screening of cervical precancerous lesions and early-stage cervical cancer. In the stratified management of HR-HPV-positive individuals, our methylation panel demonstrates superior detection accuracy compared to conventional cytology testing. Based on our findings, DNA methylation testing has the potential to replace cytology testing as a screening method for HR-HPV-positive individuals. Supplementary Information Additional file 1 Additional file 2 Additional file 3 Acknowledgements Not applicable. Author contributions SL, GL and YG performed the experiments and data curation; ZL, QZ and LZ contributed to the methodology, resources, and writing—original draft preparation; XW and QZ performed the data visualization; YL, CD, GC performed the formal analysis; NM, WS, BD, and LY performed the investigation; JZ, XZ and LH contributed to writing—review and editing; JY, YC, YL and XH were responsible for conceptualization and project administration. All authors read and approved the final manuscript. Funding This work was supported by the Basic and Applied Basic Research Project of Guangzhou Science and Technology Bureau (Grant No. 2025A03J4160). Availability of data and materials The original contributions presented in this study are included in the article. Further inquiries can be directed to the correspond-ing authors. Declarations Ethical approval and consent to participate This investigation was conducted in strict compliance with the Declaration of Helsinki and received ethical approval from the Institutional Review Board of Guangzhou First People’s Hospital (Approval No.: K-2021-109-05). Consent for publication All authors have reviewed and approved the manuscript and consent to its publication. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Zhihan Liu, Guorong Liu and Shuanping Liu are contributed equally to this work. Contributor Information Yongjun Li, Email: li.yongjun@biochainbj.com. Xiaoliang Han, Email: sean.han@biochainbj.com. Yang Cheng, Email: eychengy@scut.edu.cn. Jing Yang, Email: eyjingyang@scut.edu.cn. References Associated Data Supplementary Materials Additional file 1 Additional file 2 Additional file 3 Data Availability Statement The original contributions presented in this study are included in the article. Further inquiries can be directed to the correspond-ing authors.