5-hydroxymethylcytosine signatures of tissue as potential predictive markers for platinum resistance in epithelial ovarian cancer patients Authors: Yuan Xue, Zhanmiao Yi, Ping Yu, Cuiyu Huang, Hui Li, Mengzhu Zheng, Zhanqun Yang, Shengbo Xu, Hongyan Guo, Long Chen, Jian Lin, Libo Zhao Published: 2026 ABSTRACT Platinum-based systemic therapy is the cornerstone of treatment for epithelial ovarian cancer (EOC). Although most patients initially respond to platinum-based therapy, they eventually develop drug resistance after multiple relapses. Currently, there is no effective method to predict platinum resistance in EOC patients. We collected EOC tissues from 50 patients, including 15 drug-resistant (DR) and 35 drug-sensitive (unDR) individuals. Genomic DNA (gDNA) was extracted from the cancer tissues for 5-hydroxymethylcytosine (5hmC) sequencing. Concurrently, we performed 5hmC-Seal on the cisplatin-sensitive ovarian carcinoma cell line A2780 and its isogenic cisplatin-resistant derivative, A2780-DDP. Differentially hydroxymethylated regions (DhMRs) were identified separately, and the intersection of these differentially modified genes was used to construct platinum resistance and prognosis prediction models via logistic regression. We identified 27 upregulated and 19 downregulated differentially expressed intersection genes, totaling 46 differentially hydroxymethylated insection genes. Each patient group was divided into training and validation cohorts for model development and validation. In this predictive model, we selected 12 5hmC markers to distinguish platinum resistance in EOC patients and 10 5hmC markers to differentiate between patients with short overall survival (OS ≤ 36 months, n = 16) and long OS (OS > 36 months, n = 34). The area under the curve (AUC) values were 0.937 (95% CI 0.733–1.000) and 0.971 (95% CI 0.785–1.000), respectively. We used 5hmC-Seal to detect differential changes associated with platinum resistance and sensitivity in EOC on the tissue and cell line levels. These 5hmC signatures represented promising preliminary predictive markers for platinum resistance and clinical outcomes in patients with EOC.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-47912-w. Introduction Ovarian cancer (OC) is the most lethal female pelvic tumor, with an overall five-year survival rate barely exceeding 30%1. Epithelial ovarian cancer (EOC) is the predominant subtype among all histological subtypes and is characterized by an especially aggressive biological phenotype and poor clinical outcomes. Among these, high-grade serous carcinoma (HGSOC) is the most common form. Platinum-based chemotherapy is the primary treatment for EOC. Initially, patients exhibit high sensitivity to platinum-based chemotherapy, achieving robust and sustained clinical responses. However, with multiple courses of treatment and disease recurrence, patients develop increased resistance to chemotherapy2,3. Therefore, elucidating the molecular mechanisms underlying platinum resistance and devising rational strategies to overcome it represent a critical and actively pursued area of OC research. Additionally, reliable and clinically actionable predictors for the emergence of platinum resistance in OC are currently lacking. The application of epigenetic approaches offers potential for predicting platinum resistance and investigating its underlying mechanisms4. 5-Hydroxymethylcytosine (5hmC), a crucial epigenetic mark of DNA demethylation, is garnering increasing attention due to its emerging regulatory roles and translational significance across various malignancies5. 5hmC arises as an essential intermediate in active DNA demethylation, generated through the iterative oxidation of 5-methylcytosine (5mC) by TET (ten-eleven translocation) dioxygenases6. This modification regulates a wide range of fundamental biological processes, including cellular proliferation, lineage specification, and organismal development across multiple ontogenetic stages. It is also increasingly being applied in tumor detection7. The Chuan He team from the University of Chicago utilized the 5hmC-Seal method and next-generation sequencing (NGS) to construct a comprehensive human tissue map of 5hmC. They found that 5hmC exhibited high tissue specificity and was preferentially distributed in active regions of the genome, suggesting its potential significance in transcriptional regulation8. Research has shown that 5hmC exhibits an overall decreasing trend in both hematological and solid tumors, with significantly lower levels in tumor tissues compared to normal tissues9. This makes 5hmC a potential biomarker with significant clinical application value for tumor screening, diagnosis, and monitoring. Additionally, 5hmC demonstrates strong performance in for the non-invasive detection of hepatocellular carcinoma at its early stage10. Emerging evidence demonstrates that 5hmC-based signatures achieve exceptional discriminative accuracy, enabling the simultaneous differentiation of early-stage hepatocellular carcinoma from both healthy individuals and patients with chronic hepatitis B or cirrhosis. The 5hmC marker is associated with clinical staging and exhibits liver tissue specificity, further enhancing the understanding of liver cancer pathogenesis11. In addition, accumulating evidence indicates that global or locus-specific variations of 5hmC within neoplastic tissues are mechanistically linked to the development of chemotherapeutic resistance. Dysregulation of 5hmC may drive both the onset and progression of malignancies, thereby modulating the chemosensitivity of neoplastic cells12. In EOC, the 5hmC signal in serum is considered a predictor of chemotherapy resistance13. We speculate that there is a correlation between 5hmC differences in EOC tissues and drug resistance. In this study, we leveraged 5hmC-Seal technology to generate comprehensive, genome-wide 5hmC maps from tissue genomic DNA (gDNA) of 50 EOC patients. By applying machine learning and logistic regression algorithms to develop predictive models, we achieved promising preliminary predictive performance. Materials and methods Participants Between 2019 and 2020, surgically resected carcinoma specimens were prospectively collected from 50 patients with EOC treated at Peking University Third Hospital. During surgery, cancer tissues were removed, and all patients provided informed consent. All patients underwent surgery before developing drug resistance. Medical documents, including diagnosis and treatment records, were archived according to NCCN guidelines. Platinum resistance was defined as disease recurrence within six months after treatment with platinum-based drugs, indicating poor efficacy of further platinum-based therapy. Conversely, platinum sensitivity was defined as no disease recurrence for more than six months, suggesting better treatment outcomes with subsequent platinum-based therapy. Recurrence was assessed using the following criteria: (1) Radiologic evidence: CT or MRI showing new or enlarging lesions according to RECIST 1.1 criteria (≥ 20% increase in the sum of target lesions or unequivocal progression of non-target lesions); (2) Serum biomarker: Two consecutive CA125 values ≥ 2 × the upper limit of normalon two occasions at least one week apart, in patients with elevated baseline CA125 that normalized with treatment; (3) Clinical evidence: Physical examination or symptomatic progression consistent with disease recurrence. Patients’ conditions were closely monitored during hospitalization, followed by regular telephone follow-ups and outpatient visits to assess prognosis. This study complied with the Declaration of Helsinki. All samples’ clinical information was presented in Tables 1 and S1.Table 1Epithelial ovarian cancer (EOC) patient characteristics.CharacteristicsDRunDRP15 (30%)35 (70%)Age (years) ≤ 50 > 506 (40%)9 (60%)13 (37.14%)22 (62.86%)0.849Pathological pattern Serous carcinoma15 (100%)35 (100%)–Histological grade G315 (100%)35 (100%)–Clinical stage IIIA IIIB IIIC IV1 (6.67%)2 (13.33%)9 (60%)3 (20%)3 (8.57%)2 (5.71%)28 (80%)2 (5.71%)0.123(III/IV)Ascitic fluid Positive Negative13 (86.67%)2 (13.33%)23 (65.71%)12 (34.29%)0.131OS(months) ≤ 36 > 3611 (73.33%)4 (26.67%)5 (14.29%)30 (85.71%) < 0.0001****PFS(months) ≤ 36 > 3615 (100%)0 (0%)12 (34.29%)23 (65.71%) < 0.0001****Clinical outcome Death Survival9 (60%)6 (40%)3 (8.57%)32 (91.43) < 0.0001*****:P < 0.05; **: P < 0.01;***: P < 0.001;****: P < 0.0001 Table 1Epithelial ovarian cancer (EOC) patient characteristics.CharacteristicsDRunDRP15 (30%)35 (70%)Age (years) ≤ 50 > 506 (40%)9 (60%)13 (37.14%)22 (62.86%)0.849Pathological pattern Serous carcinoma15 (100%)35 (100%)–Histological grade G315 (100%)35 (100%)–Clinical stage IIIA IIIB IIIC IV1 (6.67%)2 (13.33%)9 (60%)3 (20%)3 (8.57%)2 (5.71%)28 (80%)2 (5.71%)0.123(III/IV)Ascitic fluid Positive Negative13 (86.67%)2 (13.33%)23 (65.71%)12 (34.29%)0.131OS(months) ≤ 36 > 3611 (73.33%)4 (26.67%)5 (14.29%)30 (85.71%) < 0.0001****PFS(months) ≤ 36 > 3615 (100%)0 (0%)12 (34.29%)23 (65.71%) < 0.0001****Clinical outcome Death Survival9 (60%)6 (40%)3 (8.57%)32 (91.43) < 0.0001*****:P < 0.05; **: P < 0.01;***: P < 0.001;****: P < 0.0001 Epithelial ovarian cancer (EOC) patient characteristics. ≤ 50 > 50 6 (40%) 9 (60%) 13 (37.14%) 22 (62.86%) IIIA IIIB IIIC IV 1 (6.67%) 2 (13.33%) 9 (60%) 3 (20%) 3 (8.57%) 2 (5.71%) 28 (80%) 2 (5.71%) 0.123 (III/IV) Positive Negative 13 (86.67%) 2 (13.33%) 23 (65.71%) 12 (34.29%) ≤ 36 > 36 11 (73.33%) 4 (26.67%) 5 (14.29%) 30 (85.71%) ≤ 36 > 36 15 (100%) 0 (0%) 12 (34.29%) 23 (65.71%) Death Survival 9 (60%) 6 (40%) 3 (8.57%) 32 (91.43) *:P < 0.05; **: P < 0.01;***: P < 0.001;****: P < 0.0001 Study design This study aimed to dissect 5hmC signature differences in tissues between platinum-resistant and platinum-sensitive EOC patients, as well as between patients with good and poor prognosis. We then sought to select predictive biomarkers using a logistic regression model. We performed 5hmC sequencing on tissues from 50 EOC patients, including 15 who developed platinum resistance (DR group) and 35 who remained platinum-sensitive (unDR group). Based on a median overall survival (OS) of 36 months, patients were categorized into a shortOS group (OS ≤ 36 months) and a longOS group (OS > 36 months). Additionally, we conducted 5hmC sequencing on the cisplatin-resistant cell line A2780-DDP and its chemosensitive parental line, A2780. Differentially hydroxymethylated genes (DhMGs) identified from 5hmC-Seal sequencing were intersected, and these intersection markers were used to establish effective prediction models for platinum resistance and prognosis in EOC. The research and design workflow is illustrated in Fig. 1.Fig. 1Overview of research design. Collecting 50 tissues from EOC patients, and based on subsequent treatment and follow-up results, dividing the patients into drug-resistant (DR) group and drug sensitive (unDR) group. Performing 5hmC-seal on the tissues. Simultaneously, platinum resistant and sensitive ovarian cancer cell lines were underwent genome-wide 5hmC detection via 5hmC-Seal. The interaction DhMGs and differential genes were used to predict DR and prognosis patients using logistic regression model. Overview of research design. Collecting 50 tissues from EOC patients, and based on subsequent treatment and follow-up results, dividing the patients into drug-resistant (DR) group and drug sensitive (unDR) group. Performing 5hmC-seal on the tissues. Simultaneously, platinum resistant and sensitive ovarian cancer cell lines were underwent genome-wide 5hmC detection via 5hmC-Seal. The interaction DhMGs and differential genes were used to predict DR and prognosis patients using logistic regression model. Cell culture and treatment A2780 and A2780-DDP were obtained from the American Type Culture Collection and propagated in RPMI-1640 supplemented with 10% fetal bovine serum and 1% penicillin–streptomycin at 37 °C in a humidified atmosphere containing 5% CO₂. EOC tissues and cell lines collection and gDNA isolation We collected the EOC tissues removed during surgery into sterile tubes and stored them at -80 °C, digested well-growing EOC cells, centrifuged and washed them, then stored the cell pellet at − 80 °C. Cutting 25 mg of each tissue and collecting 1 million cells after thawing. Genomic DNA from tissues and cells was extracted with the Quick-DNA Miniprep Plus kit (ZYMO, D4069) and quantified on a Qubit 3.0 fluorometer (Thermo, Q33216). After electrophoretic confirmation of fragment sizes (~ 100–150 bp), 5hmC libraries were constructed and stored at − 80 °C. 5hmC library preparation followed by deep sequencing Construction of 5hmC libraries for all samples was performed according to previous instructions14,15. Genomic DNA was end-repaired, adapter-ligated (KAPA Hyper Prep), and subjected to 5hmC selective biotinylation: UDP-6-N3-Glc conjugation (37 °C, 2 h) followed by click-chemistry labeling with DBCO-PEG4-biotin (37 °C, 2 h), each step followed by Zymo purification. Biotinylated fragments were captured on streptavidin beads, PCR-amplified, and size-selected with AMPure XP beads. Libraries were sequenced 2 × 39 bp on the NextSeq 500. Mapping and identifying 5hmC-enriched regions Assessing the built library’s sequence quality with FastQC (version 0.11.5). Bowtie2 (version 2.2.9) was used to compare the raw 5hmC data acquired by sequencing with the human genome (version hg19)16, and SAMtools v1.3.1 was employed to retain only non-redundant, uniquely aligned reads, after which fragment coordinates were extended and converted to BedGraph with BedTools v2.19.1.17,18, and then converting to bigwig format for viewing using the bedGraphToBig Wig feature of Integrated Genomics Viewer. Every sample is examined for possible 5hmC enrichment regions using MACS2 (version 2.1.1)17. Combine less than 1000 bp areas from more than ten samples into a single directory for every patient. Genomic areas that frequently exhibit artifact signals are filtered out, according to ENCODE. A single peak call file and a merged peak file are intersected to create the 5hmC enrichment zone. Using the CHIP seeker package, we identified the 5hmC enriched region. The genes that were closest to this region were then employed in the subsequent analysis. 5hmC markers functional analysis We used the website (Venny 2.1.0 (csic.es)) to obtain intersection deferentially 5hmC-enriched regions (DhMRs), and GO/KEGG annotation platform、sangerbox tool to analyze the functional differences19–23. We searched for the expression of major markers in the TGGA database and plotted the survival curves using Kaplan–Meier plotter. Feature selection and model construction We chose CV logistic regression (LR) to establish warning models. DhMR was detected by using DESeq2 1.30.0 in R 3.5.0. Due to the limited sample size and exploratory nature of this study, we used nominal p-values (P < 0.05) combined with fold-change thresholds (|log2FC| > 0.5) for initial marker selection (two-tailed t-tests), rather than stringent FDR correction. Overfitting was mitigated by performing five repeated tenfold cross-validations, and after randomizing the patients into 2 groups, with one of them were considered serving as the training set, with the remainder reserved for independent testing. The study then carried out 100 iterations for further filtering by using the elimination of recursive features of the Scikit-Learn module. Meanwhile, 100 iterations of ten-fold cross-validation were performed in each round, and the finalized prognostic model was established for the training cohort by taking the final markers recorded in at least 5 rounds. Then, characteristics selected from DhMR were used to train the LR CV model. Model performance was quantified by the area under the ROC curve (AUC) with 95% confidence intervals. The weighted predictive score (wp-score) was derived by summing the products of each 5hmC marker’s multivariate logistic coefficient and its corresponding abundance: wp-score = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sum }_{k=1}^{n}\beta k\times genek$$\end{document}, where genek is the 5hmC level for the kth marker gene and k denotes the multivariate logistic coefficient assigned to the kth marker gene. Statistical Analysis Using GraphPad Prism 9.0 to perform statistical processing on the data. The comparison of clinical variables between different groups was conducted using chi square test (Table 1 and Table S1). The correlation between variables and overall survival was analyzed using univariate and multivariate Cox regression analysis (Table S2). The percentile method was used to calculate the 95% confidence interval. The significance threshold was set at P < 0.05. Results Clinical characteristics of EOC patients The clinical information of 50 patients is shown in Table 1. Among them, 15 patients developed platinum resistance, while 35 were platinum-sensitive. There was no significant correlation between age and drug resistance. All patients were diagnosed with high-grade serous carcinoma and were in the middle to late stages of the disease. In the DR group, all patients were classified as stage III or IV (12 and 3, respectively), whereas in the unDR group, 33 patients were stage III and 2 were stage IV. Ascites was present in 13 patients in the DR group and in 12 patients in the unDR group. OS and progression-free survival (PFS) were significantly associated with drug resistance, with drug-resistant patients exhibiting notably shorter survival times compared to non-resistant patients. Mortality was markedly higher in the DR group compared to the unDR group (9 versus 3). Detailed OS grouping is provided in Table S1. 5hmC distribution in tissue gDNA of EOC patients We extracted gDNA from tissue samples of 50 HGSOC patients, which were subjected to rigorous quality control, and processed for 5hmC-Seal library construction and sequencing to generate genome-wide 5hmC maps. The sequencing results showed that a total of 37579 loci were detected. Global 5hmC abundance was markedly diminished in the DR cohort compared to the unDR cohort. Consistent with prior reports linking 5hmC deposition to active transcription, 5hmC exhibited pronounced enrichment at transcription start sites (TSS) and transcription termination sites (TTS), with reciprocal depletion in flanking regions (Fig. 2A)13. We then used HOMER21 to profile the genome-wide distribution of 5hmC. This analysis systematically identified and quantified DhMRs across all samples. The majority of DhMRs were located in intronic, distal intergenic, or promoter regions (Fig. 2B), which is consistent with prior observations22. Differential hydroxymethylation analysis (P < 0.05, |log2FC| > 0.5) identified 1769 hyper-hydroxymethylated and 837 hypo-hydroxymethylated regions in DR compared to unDR samples (Fig. 2C). Unsupervised hierarchical clustering of the 200 most variable regions clearly separated DR from unDR samples (Fig. 2D). The Kaplan–Meier curves for OS and PFS of the two groups are presented in Figs. 2E and FFig. 2Tissue-wide 5hmC landscape across EOC patients and survival curves. (A) Genome-wide 5hmC density along gene bodies in DR versus unDR ; adjacent bars quantify average hMRs per annotation class. (B) Pie-chart decomposition of 5hmC peaks across genomic features; promoters are defined as ± 2 kb around the TSS. (C) Volcano plot of DhMGs: blue, hyper-5hmC; yellow, hypo-5hmC (|log2FC| > 0.5, P < 0.05); gray, non-significant. (D) Unsupervised heatmap of the top 200 discriminative 5hmC markers; columns represent DR and unDR patients (age annotated), rows are clustered DhMRs. (E) Overall survival (OS) curves of DR and unDR patients. (F) Progression free survival (PFS) curves of DR and unDR patients. Tissue-wide 5hmC landscape across EOC patients and survival curves. (A) Genome-wide 5hmC density along gene bodies in DR versus unDR ; adjacent bars quantify average hMRs per annotation class. (B) Pie-chart decomposition of 5hmC peaks across genomic features; promoters are defined as ± 2 kb around the TSS. (C) Volcano plot of DhMGs: blue, hyper-5hmC; yellow, hypo-5hmC (|log2FC| > 0.5, P < 0.05); gray, non-significant. (D) Unsupervised heatmap of the top 200 discriminative 5hmC markers; columns represent DR and unDR patients (age annotated), rows are clustered DhMRs. (E) Overall survival (OS) curves of DR and unDR patients. (F) Progression free survival (PFS) curves of DR and unDR patients. Enrichment analysis and functional exploration of DhMRs Functional enrichment analysis was performed on all DhMGs in HGSOC patients. The results highlighted several biological processes and signaling pathways, including cancer, diseases and infections, various signaling pathways, hormones and metabolism, among others (Fig. 3A). Additionally, we identified the central nodes of the GO functional interaction network (Fig. 3B).Fig. 3DhMGs enrichment analysis19–23. (A) Bar plot of GO term enrichment. (B) GO enrichment and Gene-Concept Network. DhMGs enrichment analysis19–23. (A) Bar plot of GO term enrichment. (B) GO enrichment and Gene-Concept Network. Analysis of 5hmC sequencing of EOC cell line The 24-h half-maximal inhibitory concentration (IC50) of cisplatin was determined for parental A2780 and cisplatin-resistant A2780-DDP cells. The IC50 values were 12–16 μM for A2780 cells and 33–38 μM for A2780-DDP cells, confirming the higher cisplatin resistance of A2780-DDP cells (Fig. 4A). Most DhMRs were enriched in intronic, distal intergenic, and promoter regions (Fig. 4B). A heatmap displays the hydroxymethylation levels of the top 200 DhMRs in the two cell lines (Fig. 4C). A volcano plot shows 222 upregulated and 207 downregulated DhMRs between the DR and unDR groups (Fig. 4D). The functional enrichment analysis of these DhMRs is presented in Fig. 4E.Fig. 4Characteristics and differential enrichment analysis of 5hmC in EOC cell lines. (A) 24-h IC50 determination of cisplatin in OC A2780 cells and the isogenic cisplatin-resistant derivative A2780-DDP. (B) Genomic distribution of 5hmC peaks across chromatin compartments in the cell lines. (C) Unsupervised heatmap of the top 200 differential 5hmC loci stratifies A2780-DDP and parental A2780 cells; DhMRs and samples were ordered by hierarchical clustering. (D) The volcano plots of DhMG with significant changes in A2780-DDP and A2780 gDNA (|log2FC| > 0.5, P < 0.05) are marked in pink (upward) and green (downward), respectively. Black dots represent genes with no differential expression. (E) GO enrichment bar plot of DhMGs. Characteristics and differential enrichment analysis of 5hmC in EOC cell lines. (A) 24-h IC50 determination of cisplatin in OC A2780 cells and the isogenic cisplatin-resistant derivative A2780-DDP. (B) Genomic distribution of 5hmC peaks across chromatin compartments in the cell lines. (C) Unsupervised heatmap of the top 200 differential 5hmC loci stratifies A2780-DDP and parental A2780 cells; DhMRs and samples were ordered by hierarchical clustering. (D) The volcano plots of DhMG with significant changes in A2780-DDP and A2780 gDNA (|log2FC| > 0.5, P < 0.05) are marked in pink (upward) and green (downward), respectively. Black dots represent genes with no differential expression. (E) GO enrichment bar plot of DhMGs. 5hmC biomarkers for predicting DR in EOC We identified the intersection of upregulated and downregulated DhMRs detected in EOC tissues and cell lines, respectively (Fig. 5A, B), resulting in a total of 46 intersecting genes (Fig. S1). 50 patients were randomly assigned to training and validation cohorts (25 each). A logistic regression-driven recursive feature elimination procedure refined the 5hmC panel from 46 to 12 optimal loci (Fig. 5C), which effectively predicted drug resistance in EOC patients in the training set (Fig. 5D,E). The AUC was 1.000 (95% CI 0.889–1.000), with both sensitivity and specificity at 100% (Fig. 5F). The predictive performance of these 12 markers was well validated in the validation cohort (Fig. 5G,H), with an AUC of 0.937 (95% CI 0.733–1.000), achieving 85.7% sensitivity and 88.9% specificity. The individual AUCs for each marker in the training group (Fig. S2) and validation group (Fig. S3) were also calculated. Additionally, drug-resistant patients exhibited significantly elevated wp-scores compared to drug-sensitive patients (Fig. S6A).Fig. 5Intersection of tissues and cell lines 5hmC markers for predicting DR in EOC. (A) Intersection of DhMGs upregulation in tissues and cell lines. (B) Intersection of DhMGs downregulation in tissues and cell lines. (C) Overview of DR prediction model design. (D) ROC curve for a 12-marker 5hmC classifier in the training set, illustrating sensitivity versus 1-specificity. (E) 12 5hmC markers for unsupervised hierarchical clustering for training-set of the predictive model. (F) Confusion table displaying the model efficacy within the training cohort. (G) 12-marker 5hmC classifier ROC curve in the validation cohort. (H) 12 5hmC markers for unsupervised hierarchical clustering for validation-set of the predictive model. (I) Confusion table displaying the performance of the model in validation. Intersection of tissues and cell lines 5hmC markers for predicting DR in EOC. (A) Intersection of DhMGs upregulation in tissues and cell lines. (B) Intersection of DhMGs downregulation in tissues and cell lines. (C) Overview of DR prediction model design. (D) ROC curve for a 12-marker 5hmC classifier in the training set, illustrating sensitivity versus 1-specificity. (E) 12 5hmC markers for unsupervised hierarchical clustering for training-set of the predictive model. (F) Confusion table displaying the model efficacy within the training cohort. (G) 12-marker 5hmC classifier ROC curve in the validation cohort. (H) 12 5hmC markers for unsupervised hierarchical clustering for validation-set of the predictive model. (I) Confusion table displaying the performance of the model in validation. 5hmC biomarkers for predicting prognosis in EOC Patients were divided into a short OS group (n = 16) and a long OS group (n = 34) based on the cohort’s median overall survival of 36 months (Fig. 6A, B). Both OS and PFS differed significantly between the two groups (P < 0.0001) (Fig. 6C, J). The cohort was randomly split 1:1 into training and validation sets. Using recursive feature elimination with cross-validated logistic regression in the training cohort, we developed a predictive model that achieved an AUC of 1.000 (95% CI 0.899–1.000) with 100% sensitivity and specificity (Fig. 6D,F). This model refined the initial 5hmC biomarker panel to a concise set of 10 discriminative loci (Fig. 6E). These 10 markers also performed well in the validation cohort, with an AUC of 0.971 (95% CI 0.785–1.000), 87.5% sensitivity, and 94.1% specificity. The AUC for each individual marker in the training and validation sets is provided in Figs. S4 and S5. Patients in the short OS group exhibited a significantly higher wp-score than those in the long OS group (Fig. S6C). Among the 10 markers, RBMS3 and GBF1 demonstrated the strongest association between tissue 5hmC levels and OS, and had the highest hazard ratios (HR) among all markers analyzed (Fig. 6K,L). Cox regression analysis suggested that RBMS3 and GBF1 were independent prognostic factors (Table S3).Fig. 6Intersection of 5hmC markers for predicting prognosis in EOC. (A, B) The number and distribution of patients in short OS group and long OS group. (C) OS curves of patients in different survival time. (D) Training-set ROC curve of the 10-5hmC-signature classifier. (E) 10 5hmC markers for unsupervised hierarchical clustering for training-set of the predictive model. (F) Confusion table displaying the the model efficacy within the training cohort. (G) Validation-set ROC curve of the 10-5hmC-signature classifier. (H) 10 5hmC markers for unsupervised hierarchical clustering for validation-set of the predictive model. (I) Confusion table displaying the model efficacy within the validation cohort. (J) PFS curves of short OS and long OS patients. (K, L) Survival curves related to 5hmC levels for the two highest hazard ratio (HR) markers (RBMS3, GBF1). Intersection of 5hmC markers for predicting prognosis in EOC. (A, B) The number and distribution of patients in short OS group and long OS group. (C) OS curves of patients in different survival time. (D) Training-set ROC curve of the 10-5hmC-signature classifier. (E) 10 5hmC markers for unsupervised hierarchical clustering for training-set of the predictive model. (F) Confusion table displaying the the model efficacy within the training cohort. (G) Validation-set ROC curve of the 10-5hmC-signature classifier. (H) 10 5hmC markers for unsupervised hierarchical clustering for validation-set of the predictive model. (I) Confusion table displaying the model efficacy within the validation cohort. (J) PFS curves of short OS and long OS patients. (K, L) Survival curves related to 5hmC levels for the two highest hazard ratio (HR) markers (RBMS3, GBF1). Discussion 5hmC, an emerging epigenetic biomarker, has attracted significant interest in cancer research. In tumors, the levels and distribution of 5hmC are often altered, contributing to tumorigenesis, progression, and therapy response24. This study systematically profiled the whole-genome 5hmC landscape associated with platinum resistance and poor prognosis in EOC at both the tissue and cellular levels. Furthermore, it preliminarily constructed an accurate prediction model based on a small set of tissue 5hmC markers, with the validation set achieving AUCs of 0.937 and 0.971. This exploratory research not only provide novel biomarkers for predicting platinum resistance in EOC but also offer new insights into its underlying epigenetic regulatory mechanisms. First, our study confirms that 5hmC distribution in the EOC tissue genome is preferentially enriched in functional genomic regions. We observed significant 5hmC accumulation at TSS and TTS, which is generally associated with gene activation. Specifically, marked 5hmC enrichment was detected at gene promoters and enhancers. This pattern may facilitate transcription factor binding and subsequent protein expression. Mechanistically, the presence of 5hmC could attenuate the repressive effect of DNA methylation on gene expression, thereby promoting transcription factor recruitment and gene transcription8. Regarding the TTS region, 5hmC enrichment may be linked to transcriptional termination and the regulation of polycistronic gene clusters25. For example, in some organisms, 5hmC is crucial for proper termination at the ends of such clusters. More importantly, we found that global 5hmC levels were significantly lower in platinum-resistant tumors than in platinum-sensitive ones. This observation aligns with the widespread downregulation of 5hmC reported in various solid tumors, further supporting the view that global loss of 5hmC may be a common epigenetic feature in cancer. This loss could contribute to the chemoresistance phenotype by affecting the expression stability of a large number of genes26. Based on the comprehensive genome-wide 5hmC differential analysis of DR versus unDR patient tissues, we identified 2606 differentially hydroxymethylated genes (DhMGs). Functional enrichment analysis revealed that these genes were significantly enriched in cancer-related pathways, multiple signal transduction pathways (such as Hippo, Wnt, etc.), hormone and metabolic pathways, as well as disease- and infection-related pathways. This finding underscores the extreme complexity of platinum resistance mechanisms in EOC, which represents a multidimensional biological process involving coordinated alterations across multiple pathways. For instance, lipid metabolic reprogramming has been established as a critical mechanism underlying platinum resistance27. The hydroxymethylation alterations we observed in metabolic pathway–related genes align with this perspective. Furthermore, upregulation of the Wnt/β‑catenin signaling pathway is a well‑established driver of drug resistance28, and activation of the Hippo pathway effector YAP/TAZ not only promotes tumorigenesis but also contributes to chemotherapy resistance29. 5hmC may act as an upstream epigenetic regulator in acquired platinum resistance in EOC, possibly by modulating the expression of key resistance‑associated genes. To enhance the reliability of our findings and identify core resistance-related markers, we integrated 5hmC‑seq data from clinical tissues and an isogenic platinum‑resistant cell line pair (A2780/A2780‑DDP). Intersection analysis yielded 46 genes exhibiting consistent differential hydroxymethylation in both systems. Although A2780 is derived from endometrioid ovarian carcinoma and thus differs histologically from the HGSOC samples in our cohort, using of this isogenic pair offers a advantage: it minimizes confounding genetic background differences, thereby allowing a clearer focus on epigenetic changes directly linked to platinum resistance. This integrated approach reduces technical and biological noise, selecting epigenetic events conserved across distinct biological contexts and improving the specificity of subsequent marker screening. We acknowledge, however, that the biological differences between this cell model and HGSOC constitute a key limitation of the present study. Future studies should verify the function of these markers in HGSOC-derived cell lines, patient-derived organoids, or orthotopic mouse models. Using these 46 intersecting genes, we successfully built two highly predictive logistic regression models via machine learning. The first model, based on 12 5hmC markers (PMAIP1, RNF19A, NTM, NRARP, ZNF44, PPP1CB, LIPC, RBMS3, ARHGEF7, LINC01014, SFSWAP, SNTB1), effectively discriminated between platinum-resistant and platinum-sensitive patients. In the validation set, it achieved an AUC of 0.937, demonstrating excellent diagnostic performance. Among these 12 biomarkers, RNA‑binding motif protein 3 (RBMS3) showed the highest single‑gene predictive efficacy (AUC = 0.795). This aligns closely with prior evidence that loss of RBMS3 activates the miR‑126‑5p/β‑catenin/CBP axis, thereby promoting platinum resistance in EOC30. Our data show that in DR patients, the 5hmC level in the RBMS3 gene region undergoes specific changes, which may be the epigenetic basis for its downregulation and promotion of drug resistance. In addition, in other cancers, such as breast cancer, the down-regulation of RBMS3 has also been confirmed to be related to gemcitabine resistance. Our data further reveal that drug-resistant patients exhibit specific 5hmC alterations at the RBMS3 locus, which may represent the epigenetic basis for its down‑regulation and role in driving resistance31. Collectively, these findings establish RBMS3 as a central player in cancer chemoresistance, and our study provides novel epigenomic evidence supporting its key role in EOC platinum resistance. Beyond RBMS3, other markers in the predictive model are also noteworthy. For instance, PMAIP1 encodes the pro‑apoptotic protein NOXA, a transcriptional target of p53. Multiple reports indicate that PMAIP1 functions as an apoptosis sensor in chemotherapy resistance32. NRARP, a key regulator of the NOTCH pathway, is frequently discussed in the context of cancer stem cell function and drug resistance33. ARHGEF7 is thought to contribute to treatment resistance by regulating cytoskeletal rearrangement and cell migration34. The second model successfully predicted patient OS using 10 5hmC biomarkers (RBMS3, LINC00861, ZNF44, HLF, RC3H1, NRARP, PPP1CB, CPAMD8, GBF1, MAP3K5). This prognostic model, with a median OS cutoff of 36 months, accurately distinguished between long- and short-survival patients in the validation set (AUC = 0.971). Notably, RBMS3 again emerged as a single-gene biomarker demonstrating superior predictive performance, a finding further confirmed by multivariate Cox regression analysis (Table S2). It is an independent predictor of OS, regardless of traditional clinical factors such as age, clinical stage, and ascites status (HR = 0.830, 95% CI 0.772–0.956). The expression of RBMS3 is downregulated in many cancers, and its low expression has been correlated with poor prognosis in several cancer types35,36.In addition, the Golgi-based Brefeldin A resistance factor 1 (GBF1) is highly correlated with patient survival, as higher tissue 5hmC levels at its locus indicate shorter overall survival. The role of GBF1 in cancer prognosis is currently underreported; our findings suggest it may be a novel epigenetically regulated target associated with poor prognosis in EOC, which warrants further investigation. We obtained expression profiles of these markers in ovarian cancer from the TCGA database and plotted survival curves using the Kaplan–Meier plotter (Fig. S7). We found that certain biomarkers (MAP3K5, NTM, RBMS3, RNF19A, SNTB1) are specifically expressed in ovarian cancer, and some are significantly correlated with survival rates (MAP3K5, ARHGEF7, GBF1, CPAMD8, NTM), suggesting that our research offers valuable insights for identifying potential disease targets. This study demonstrates clear potential for clinical translation. First, the predictive models are based on 5hmC‑Seal technology, which offers high sensitivity, specificity, and requires only a low amount of DNA input. This suggests that the biomarker panel could be applied to minute tumor tissue samples obtained from surgical resection or biopsy, allowing individualized assessment of resistance risk and prognosis before or early in treatment. Such assessment could help guide clinicians toward more precise therapeutic strategies. We acknowledge that the main limitations of this study are its single-center design and relatively small sample size (n = 50). The limited cohort reduces statistical power, increases the risk of model overfitting, and constrains the generalizability of the findings. Although we employed rigorous training-validation splitting and cross-validation to evaluate model performance—achieving excellent metrics—the reported high AUC values must be interpreted as preliminary. In addition, the lack of an independent external validation queue still limits the strength of conclusions and and validation needs to be conducted in larger, multicenter, prospective independent cohorts before clinical translation can be considered. Furthermore, the predictive models presented here should be regarded as exploratory proof-of-concept rather than clinically validated diagnostic tools. Considering that this study is an exploratory biomarker discovery investigation, and strict screening was conducted through cell line data cross-validation and machine learning feature selection, we used uncorrected p-values in this preliminary screening to minimize the risk of Class II errors (false negatives) and avoid overlooking potential biomarkers. We acknowledge that this approach increases the risk of Class I errors (false positives), which is a limitation of this study. In future validations involving larger cohorts, more rigorous multiple testing corrections must be applied. Finally, while this work focuses on biomarker discovery and predictive modeling, the exploration of gene function is mainly based on data-driven rather than mechanistic supportthe. The precise mechanisms by which 5hmC alterations at key genes regulate their transcription and, through which downstream signaling networks, specifically influence platinum sensitivity remain to be elucidated through further gain‑ and loss‑of‑function experiments. Conclusion We have constructed a predictive model with a panel of 5hmC biomarkers comprising a small set of genes that can preliminarily predict platinum resistance status and survival outcomes in EOC patients. These findings provide a foundation for developing novel epigenetic biomarkers for precision medicine in EOC and offer a new avenue for exploring the epigenetic mechanisms underlying platinum resistance. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1