DNA methylation and transcription factor-driven immune subtypes in ovarian cancer

Abstract

Ovarian cancer (OC) remains one of the deadliest gynecological malignancies. Immune checkpoint blockade (ICB) inhibitors efficacy in OC has been minimal, highlighting the need for a deeper understanding of the immune microenvironment in OC. Recent studies suggest that DNA methylation and transcription factors may influence the response to immunotherapy. This study aims to classify ovarian cancer into distinct immune subtypes by integrating DNA methylation and transcription factor data through comprehensive bioinformatics analysis. Using data from The Cancer Genome Atlas (TCGA), we identified twelve differentially methylated genes (DMGs) associated with transcription factors and categorized OC into two immune subtypes, C1 and C2.The C1 subtype exhibited higher levels of immune infiltration and better prognosis, characteristic of immune "hot" tumors, whereas the C2 subtype was associated with lower immune infiltration and poorer prognosis, indicative of immune "cold" tumors. A prognostic prediction model based on four key genes—KRT81, PAPPA2, FGF10, and FMO2—was developed using the least absolute shrinkage and selection operator (LASSO) and Cox regression analyses. This model effectively stratified the TCGA OC cohort into high- and low-risk groups and was validated by predicting patient survival outcomes. Additionally, drug sensitivity analysis revealed potential therapeutic targets for different risk groups, offering new avenues for precision treatment in ovarian cancer. Immunohistochemical tests confirmed the potential of KRT81 as a prognostic marker for ovarian cancer. Our findings enhance the understanding of the molecular characteristics of the OC immune microenvironment, propose novel biomarkers for prognosis, which may potentially improve the prognosis of OC.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-02630-z.

Research highlights

• Identification of transcription factor-related DMGs as prognostic markers in ovarian cancer.

• Classification of OC into two immune subtypes with distinct prognoses.

• Construction of a predictive model using KRT81, PAPPA2, FGF10, and FMO2.

• Identification of immune "hot" and "cold" tumor subtypes in ovarian cancer.

• Drug sensitivity analysis reveals therapeutic targets for precision treatment in different risk groups.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-02630-z.

Introduction

Ovarian cancer is a highly lethal gynecological malignancy with a significantly high mortality rate. Although its incidence is markedly lower than that of breast cancer, which is the most prevalent female malignancy, the mortality rate associated with ovarian cancer is three times higher. Alarmingly, the death rate from ovarian cancer is projected to increase substantially by 2040 [1]. The principal factor contributing to the high mortality of ovarian cancer is the absence of effective screening methods and prognostic evaluation tools, leading to late-stage diagnoses and complicated treatment regimens. Platinum-based chemotherapy remains the standard first-line treatment for high-grade serous ovarian carcinoma (HGSOC), often yielding favorable initial outcomes. However, approximately 70% of HGSOC patients eventually develop chemoresistance within three years, resulting in cancer recurrence and, ultimately, death. This phenomenon is largely attributed to this cancer type's extensive intertumoral and intratumoral heterogeneity [2–7].

Currently, immune checkpoint inhibitors (ICIs), including antibodies that target the PD-1/PD-L1 axis, have demonstrated significant antitumor activity across various cancer types and have been approved by the FDA as recommended therapies for cancer treatment [8]. HGSOC is frequently associated with increased tumor-infiltrating lymphocytes (TILs), a biomarker indicative of a potential response to checkpoint inhibitor therapy. However, despite these promising biomarkers, monotherapy or combination ICI immunotherapy has not substantially improved progression-free survival or overall survival in HGSOC patients [9, 10]. While the efficacy of ICIs can sometimes be predicted by tumor PD-L1 expression or tumor mutational burden, these factors alone do not reliably indicate whether a patient will benefit from such treatments [11].

It is increasingly recognized that the progression to platinum-based chemotherapy resistance in ovarian cancer is associated with epigenetic alterations, including increased DNA methylation and the transcriptional silencing of key genes [12]. DNA methylation status can also influence the function of certain tumor immune cells and their response to immunotherapy [13–16]. In ovarian cancer, low doses of DNA methyltransferase inhibitors (DNMTis) can alter the tumor epigenome and transcriptome, primarily affecting the expression of immune reactivation pathways [17]. However, while methylation-related models hold predictive value, their accuracy remains limited, and their clinical utility beyond prediction is not yet fully established.

Transcription factors (TFs) are a critical class of biological molecules that play a significant role in tumorigenesis and tumor progression [18, 19]. DNA methylation has been described as a transcriptional repressor; methylation at a gene’s promoter region can obstruct TF binding, leading to gene silencing [20, 21]. Recent advancements in our understanding of tumor heterogeneity and complexity have revealed that TFs are deeply involved in regulating the tumor immune microenvironment, immune evasion, and antitumor immunity [22–25].

DNA methylation not only directly affects gene expression, but also indirectly regulates tumor immune escape by affecting the activity or expression of transcription factors. For example, silencing of DNA methylation may affect the expression of transcription factors, such as c-Myc or AP-1, which directly affect the response to immunotherapy by regulating immune escape and tumor cell proliferation and migration [26, 27]. In addition, some transcription factors (such as NF-κB and STAT3) themselves are also involved in regulating the expression of DNA methylases, thereby affecting the methylation status of genes [28, 29]. This bidirectional regulatory mechanism makes the interaction between DNA methylation and transcription factors in ovarian cancer immunotherapy more complicated. The interaction between DNA methylation and transcription factors in ovarian cancer immunotherapy provides a new perspective for the study of tumor immune escape, the shaping of immune microenvironment and the efficacy of immunotherapy.

This study integrates transcriptome and DNA methylation data from a large cohort of ovarian cancer patients, utilizing resources from The Cancer Genome Atlas (TCGA). We analyzed differential methylation levels and correlations with transcription factors by combining information from TRRUST, TransMir, AnimalTFDB transcription factor databases, and the CHEA Transcription Factor Targets dataset. This approach allowed us to identify key genes and categorize ovarian cancer into two distinct subtypes. The analysis revealed significant differences in the immune microenvironment of these subtypes, which we defined as the immune infiltration subtype and the immunosuppression subtype. We further explored the differences between these subtypes and developed a predictive prediction model based on four genes closely associated with ovarian cancer prognosis, providing potential biomarkers for predicting patient outcomes.

Materials and methods

Data sources

All datasets involved in our study were obtained from The Cancer Genome Atlas (TCGA) and gene expression omnibus (GEO) summary data that are publicly Accessible. For this study, gene expression data (n = 429), clinical information (n = 410), and corresponding DNA methylation data (Infinium HumanMethylation27 platform, Tumor = 601, Normal = 12) related to ovarian serous cystadenocarcinoma were obtained using the TCGAbiolinks package. Additionally, normal ovarian tissue RNA-Seq data (n = 88) were retrieved from the Genotype-Tissue Expression (GTEx) database as a control group for differential expression analysis, encompassing gene expression data across 53 tissue types, including associated RNA sequences and eQTL data. The GSE13876 dataset and the GPL7759 platform files containing 157 ovarian serous cancer tissue samples were downloaded from the Gene Expression Omnibus (GEO) database, this cohort was used as an independent validation set to validate clinical prediction models. Details of the clinical information of the TCGA cohort and the GSE13876 cohort are provided in Supplementary Table (Table S1-2).

Data processing and differential analysis

To identify differentially methylated genes (DMGs), methylation sites were mapped to their corresponding genes, with an absolute methylation level difference greater than 10% considered significant (Padj < 0.05). The gene expression data from TCGA and GTEx were merged using the"limma"and"sva"packages, with batch correction applied to identify intersecting genes from both datasets. Principal component analysis (PCA) was performed to validate the reliability of batch correction. Differentially expressed genes (DEGs) between ovarian cancer (OC) samples (n = 429) and normal ovarian tissue samples (n = 88) from the TCGA dataset were identified using the DESeq2 package, with a cutoff value of |log2 FoldChange|> 1 and p < 0.01. All statistical analyses and visualizations were conducted using R software, version 4.0.2.

Identification of DMGs associated with transcription factors

To identify transcription factor-associated DMGs, data from the TRRUST, TransMir, and AnimalTFDB databases were integrated, yielding 2028 transcription factors. These were intersected with DEGs, resulting in 99 differentially expressed transcription factors. Univariate Cox regression analysis was performed on the identified DMGs, and 80 DMGs were significantly associated with ovarian cancer prognosis. The correlation between these 80 DMGs and the differentially expressed transcription factors was assessed using the cor. Test function led to the identification of 12 DMGs associated with transcription factors. The CHEA Transcription Factor Targets dataset confirmed the interaction between these DMGs and transcription factors.

Consensus clustering analysis

Consensus clustering analysis was performed using the"ConsensusClusterPlus"R package to determine whether transcription factor-related DMGs could significantly classify ovarian cancer subtypes. Cluster values (k) from 1 to 9 were evaluated, and the k value with the highest cluster stability was selected based on the clustering results. PCA was subsequently employed to validate the clustering accuracy. Kaplan–Meier survival analysis, genotyping analysis, immune infiltration analysis, somatic mutation analysis, and immune checkpoint analysis were then conducted to characterize the identified subtypes.

Functional enrichment and immune infiltration analysis

Functional enrichment analysis was conducted using the"GSVA"package with the"c2.cp.kegg.v7.4.symbols.gmt"gene set as a reference. This analysis determined the biological pathways that differed between the identified subtypes, with a threshold of P < 0.05 considered statistically significant. Immune infiltration in ovarian cancer subtypes was quantified using the single-sample enrichment analysis (ssGSEA) algorithm within the"GSVA"package, while the"estimate"package was used to score the tumor microenvironment of the subtypes. Further functional enrichment analysis between the ovarian cancer subtypes was performed using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses via the"clusterProfiler"R package.

Analysis of somatic mutations

Somatic mutation data were retrieved from the TCGA GDC database. The"maftools"R package was employed to perform somatic mutation analysis, and waterfall plots were generated to visualize the mutation profiles of the identified subtypes.

Development and evaluation of prognostic models for ovarian cancer

A predictive model was developed using the"Caret"R package. 389 OC patients with complete clinical data were randomly divided into training (70%) and test (30%) groups. Differentially expressed genes (DEGs) between OC clusters were identified using the"limma"package, with the selection criteria of |log2 FC|> 1 and FDR < 0.05. Univariate Cox regression analysis was applied to screen for prognostic DEGs, followed by multivariate Cox regression and LASSO regression analyses to construct the predictive risk model. The final model was validated using Kaplan–Meier survival analysis, time-dependent ROC curves (generated using the"timeROC"package), and various graphical tools, including nomograms, calibration, and risk curves, to evaluate and validate the model's predictive performance.

Drug-susceptibility analysis

The'calcPhenotype'function in the R package'oncoPredict'was utilized to assess drug susceptibility in high- and low-risk groups. Gene expression data were combined with drug response data from the GDSC2 database for model training, and TCGA dataset expression data were used as the test set to predict drug sensitivity (IC50 and AUC values) in the samples. Empirical Bayesian methods were applied for batch correction to minimize batch effects, and power transformation was used to optimize model performance. During prediction,'removeLowVaryingGenes = 0.2'was set to eliminate genes with low variability, enhancing model robustness. To ensure analysis reliability, a minimum sample number threshold ('minNumSamples = 10') was set. The Wilcoxon rank-sum test was then employed to assess differences in drug sensitivity between high- and low-risk groups, identifying potential candidate drugs with p-values < 0.001.

Immunohistochemical (IHC) assay

The 57 tumor tissue sections utilized in this study were obtained from patients diagnosed with high-grade serous cystadenocarcinoma of the ovary who underwent primary debulking surgery (PDS) at the Affiliated Tumor Hospital of Harbin Medical University between 2019 and 2022. The control group comprised a total of 30 cases of normal ovarian tissue. Clinical pathological data, including age, FIGO staging, lymph node metastasis, differentiation grade, and CA125 levels, were extracted from the hospital's Health Information System (HIS). All patients had no prior history of other primary malignant tumors. This research was approved by the Ethics Committee of the Affiliated Tumor Hospital of Harbin Medical University (KY2024-91). The primary antibody employed was polyclonal rabbit anti-KRT81. Ovarian tissue blocks were sectioned to a thickness of 5 μm and three slides per specimen were prepared following established immunohistochemical protocols. The specimens underwent processing in a sequential manner: baking, xylene deparaffinization, graded alcohol rehydration, distilled water rinsing, incubation in 3% H2O2 for 15 min to inhibit endogenous peroxidase activity; antigen retrieval using citrate buffer; PBS washing; blocking non-specific binding with normal goat serum; antibody incubation; DAB chromogenic reaction; hematoxylin counterstaining; thorough rinsing with running tap water; differentiation using 1% hydrochloric acid alcohol solution; lithium carbonate bluing treatment; subsequent graded alcohol dehydration and xylene clearing before being mounted with neutral resin after drying. Immunohistochemical results for KRT81 were assessed based on intensity scoring combined with positive cell frequency evaluation. Staining intensity scores were categorized as follows: negative (0 points), weak (1 point), moderate (2 points), and strong positive (3 points). Positive cell frequency was classified as follows: less than 5% (0 points); between 5 and 25% (1 point); between 26 and 50% (2 points); between 51 and 75% (3 points); greater than 75% (4 points).

Result

Screening of transcription factor-related DMGs in HGSOC

Differential methylation analysis of 601 tumor samples and 12 normal samples from the TCGA database identified 1264 significant differentially methylated genes (DMGs), with 297 genes upregulated and 967 genes downregulated (Fig. 1A, B). A list of DMGs is provided in the Supplementary Table (Table S3). Subsequently, univariate Cox regression analysis was performed on these DMGs, yielding 80 prognostic DMGs for further investigation. Differential expression analysis of 429 tumor samples from the TCGA database and 88 normal samples from the GTEx database identified 1217 differentially expressed genes (DEGs) (Fig. 1E). A list of DEGs is provided in Supplementary Table (Table S4).Utilizing transcription factor data from the TRRUST, TransMir, and AnimalTFDB databases, 99 differentially expressed transcription factors were identified (Fig. 1F). Correlation analysis between the 80 DMGs and the 99 differentially expressed transcription factors identified 12 DMGs associated with transcription factors. These DMGs were confirmed to interact with transcription factors using the CHEA Transcription Factor Targets dataset.

Fig. 1.

Identification of 12 transcription factor-related DMGs. A Volcano plot of differentially methylated genes. B Heatmap of differentially methylated genes. C, D TCGA data and GTEX principal component analysis diagram before and after batch removal. E Volcano plot of differentially expressed genes. F Intersection of DEGs and transcription factor

Tumor subtypes based on DMGs associated with transcription factors

The ConsensusClusterPlus package was employed to classify ovarian cancer based on the methylation values of the identified DMGs. The optimal clustering was achieved at K = 2, where the intra-cluster correlation was highest, and the inter-cluster correlation was lowest (Fig. 2 A–C). To validate the accuracy of this classification, principal component analysis (PCA) was performed, confirming the two subtypes (Fig. 2D). Survival analysis using the"survival"and"survminer"R packages revealed significant differences in prognosis between the two subtypes (P < 0.05), with the Type 1 subtype generally exhibiting a better prognosis than Type 2 (Fig. 2E). These findings suggest that HGSOC patients can be classified according to transcription factor-related DMGs.

Fig. 2.

Identification of ovarian cancer subtype. A Sample clustering heatmap for the TCGA cohort. B CDF curve for the TCGA cohort. C Scores of different cluster numbers. D Result of PCA between two subtypes. E KM curves showing the OS of the OC subtypes in the TCGA cohort

Definition and validation of ovarian cancer subtypes

To determine whether the two subtypes reflect the tumor immune microenvironment of HGSOC patients, an immune infiltration analysis of 429 HGSOC tumor samples from the TCGA database was conducted using the ssGSEA method. As illustrated in the accompanying figure, except for effector memory CD8 + T cells, the C1 subtype exhibited higher immune cell infiltration than the C2 subtype (Fig. 3 A). Additionally, the immune score of the tumor microenvironment indicated that the C1 subtype, characterized by a higher immune score, had a better prognosis with increased immune cell infiltration (Fig. 3B). Conversely, the C2 subtype, associated with a worse prognosis, exhibited lower levels of immune cell infiltration, suggesting that an immunosuppressive tumor microenvironment may contribute to decreased overall survival (OS). Consequently, the C1 subtype was defined as the immune infiltration subtype, while the C2 subtype was designated as the immunosuppression subtype. This classification was further supported by Gene Set Variation Analysis (GSVA), which showed that the C1 subtype was enriched in pathways related to antigen processing and presentation, chemokine signaling, natural killer cell-mediated cytotoxicity, and complement and coagulation cascades—pathways closely linked to tumor immune regulation (Fig. 3C, D). A list of hallmark/KEGG pathways is provided in Supplementary Table (Table S5).

Fig. 3.

Tumor microenvironment analysis and functional enrichment analysis. A ssGSEA evaluated two ovarian cancer subtypes based on transcription factor-related DMGs clusters (The * above the immune cells indicates the difference between the two types of immune infiltrating cells, where *** represents P < 0.001, ** represents P < 0.01, and * represents P < 0.05). B Relationship between tumor microenvironment and risk score. C, D GSVA analysis of cluster clusters. E Volcano plot of differentially expressed genes between two subtypes. F Heatmap of the top 50 significantly different genes between two subtypes. G GO enrichment analysis. H KEGG pathway enrichment analysis

In contrast, the C2 subtype was primarily enriched in RNA synthesis-related pathways. Differential gene expression analysis of the 429 HGSOC tumor samples from the TCGA database was performed to further explore the differences between these subtypes, identifying 308 significantly different genes (Fig. 3E). The top 50 significantly different genes were visualized using the"pheatmap"R package, generating heatmaps and volcano plots (Fig. 3F). Enrichment analysis of these differentially expressed genes revealed that the KEGG pathway was primarily enriched in the IL-17 signaling pathway (Fig. 3H), closely associated with tumor development, proliferation, metastasis, and drug resistance under pathological conditions [30]. Gene Ontology (GO) enrichment analysis also indicated that the biological processes were predominantly associated with antigen-binding pathways (Fig. 3G).

Somatic mutation analysis and immune checkpoint analysis

Tumor heterogeneity is intricately linked to somatic mutation patterns, and different subtypes may exhibit distinct mutation profiles. Somatic mutation analysis was conducted to explore the mutation characteristics of each subtype, with the results visualized using waterfall plots. Somatic mutations were detected in 96.03% of samples from the immune infiltration subtype and 95.83% from the immunosuppression subtype. The top 20 mutated genes in the immune infiltration subtype were TP53, TTN, NF1, RYR2, CSMD3, MUC16, BRCA1, USH2 A, APCB, DNAH3, HMCN1, SI, DNAH5, FLG2, ADGRV1, CSMD1, FAT3, MUC17, DNAH10, and TOP2 A. The full list of genes is provided in Supplementary Material (Table S6). In the immunosuppression subtype, the top 20 mutated genes were TP53, TTN, CSMD3, USH2 A, FAT3, AHNAK, HMCN1, CDK12, DNAH8, DST, FCGBP, FLG, KCNH7, LRP1, LRP1B, LRP2, MACF1, MDN1, MUC16, and MYH4. The full list of genes is provided in Supplementary Material (Table S7). Notably, BRCA1 mutations were exclusively present in the immune infiltration subtype (Fig. 4A, B). Recent studies have highlighted the close association between BRCA mutations and the development of ovarian cancer, suggesting that BRCA mutations may influence the immune status of the ovarian cancer microenvironment. To further investigate the potential differences in immunotherapy efficacy between the two subtypes, we analyzed the expression of several immune checkpoints. Significant differences between the two subtypes were observed in the expression of BRAF, BTLA, CD70, CD19, EGFR, and CTLA4 (Fig. 4C–H) [31–33].

Fig. 4.

Somatic mutation analysis and Immune checkpoints analysis. A, B The somatic mutation statuses of two subtypes. C–H Expression of BRAF, BTLA, CD19, CD70,CTLA4 and EGFR between two subtypes

Development and validation of prognostic models based on two immune subtypes in HGSOC

The TCGA dataset was randomly divided into a training set and a test set in a 7:3 ratio to facilitate the construction and validation of predictive risk models. The training set was used to construct the risk models, while the test set was employed for model validation, with the GSE13876 dataset serving as an independent external validation set. Univariate Cox regression analysis was conducted to identify prognostic DEGs from the two subtypes, which were then subjected to cross-validation using the"glmnet"package. The Lasso and multivariate Cox regression models with the least error were selected to identify four prognostic genes for constructing the risk score model (Fig. 5A, B).Collinearity analysis confirmed the independence of the genes within the model (Fig. 5C). The differential expression of these four genes between the high-risk and low-risk groups was subsequently examined, revealing that KRT81 expression increased with patient risk (Fig. 5J), suggesting its potential as a prognostic marker. The final prognostic model termed the"TFDMGs score,"was constructed as follows [34, 35]:

$$\text{TFDMGs score}\text{=}\text{Coef*exp}\text{(KRT81)}\text{+}\text{Coef*exp}\text{(PAPPA2)}\text{+}\text{Coef*exp}\text{(FGF10)}\text{+}\text{Coef*exp}\text{(FMO2).}$$

Fig. 5.

Construction and validation of prognostic risk model for patients with ovarian cancer. A Lasso regression analysis of 12 significantly different genes associated with ovarian cancer prognosis. B The model obtained from Lasso regression analysis was cross-validated and the optimal model was selected. C Independence verification of model genes. D–F K-M survival graphs for training sets, test sets, and external validation sets. G EOC Prognosis nomogram. H Cumulative risk curve. I Correction curve analysis. J Heat map of model gene expression in high-low risk group

Patients were categorized into high-risk and low-risk groups based on the median risk score of the training group. Kaplan–Meier survival analysis was performed on both groups within the HGSOC cohort, demonstrating that the risk model accurately predicted the survival prognosis of HGSOC patients (P < 0.05). Patients in the low-risk group exhibited a survival advantage, a consistent trend across both the test and external validation sets (Fig. 5D–F). In order to predict the 1 -, 3- and 5-year survival of patients with EOC, a nomogram was constructed based on multivariate Cox analysis, including age, Stage stage and risk score into the model (Fig. 5G).The cumulative risk curve shows that the total survival risk of EOC patients in the Normo chart increases year by year, and the risk of samples in the high-risk group is higher than that in the low-risk group (Fig. 5H).The calibration curve of the column chart also shows that the predicted values are consistent with the observed values (Fig. 5I).

Drug susceptibility analysis of high- and low-risk groups

Antineoplastic drugs remain a cornerstone of clinical treatment for HGSOC. However, due to the heterogeneity of HGSOC, the sensitivity to these drugs varies among patients. Understanding the differential drug sensitivity between high-risk and low-risk HGSOC patients can provide insights for personalized clinical treatment. As depicted in the figure, the IC50 of most antitumor drugs was lower in low-risk HGSOC patients compared to high-risk patients. Drugs such as ulixertinib, sorafenib, KRAS G12 C inhibitor, and PRIMA-1MET—an emerging small molecule targeting TP53 gene mutations in ovarian cancer—demonstrated lower IC50 values in low-risk patients (Fig. 6A–D). PRIMA-1MET, currently in Phase I/II clinical trials, has shown good tolerability with low toxicity [36].

Fig. 6.

Drug sensitivity analysis in high and low risk groups. A–E IC50 of ulixertinib, sorafenib, PRIMA-1MET, KRAS G12 C inhibitor and JAK_8517 between high-low risk groups

Interestingly, the IC50 of JAK inhibitors was higher in low-risk patients than in high-risk patients (Fig. 6E), suggesting that JAK inhibitors may be more effective in the high-risk group. Studies by Zak, Jaroslav et al. found that JAK inhibitor ruxolitinib can remodel myeloid immunity and synergize with immune checkpoint inhibitors to enhance anti-tumor immune responses, providing a new strategy for cancer immunotherapy [37]. Mathew, Divij et al. found that itacitinib can improve ICB efficacy, and this combination therapy has a high response rate and durable remission in patients with non-small cell lung cancer. JAK inhibitors alter CD8 T cell differentiation dynamics and improve their plasticity, but are less effective in patients with persistent inflammation [38]. This suggests that JAK inhibitors and immune checkpoint inhibitors may improve the poor response to immunotherapy in patients with ovarian cancer. These results should be interpreted with caution due to small sample size, and clinical application requires confirmation in larger cohorts.

The clinical value of KRT81 was verified by immunohistochemistry

In the TCGA ovarian cancer cohort, the genes incorporated in the prognostic prediction model were all aberrantly expressed in ovarian cancer tissues (Fig. 7A). Among them, KRT81 was closely associated with the stage of ovarian cancer (Fig. 7B). Therefore, by using ovarian cancer tissues and normal ovarian tissues from our hospital, we performed immunohistochemical detection of KRT81 expression. The results indicated that the expression of KRT81 was significantly higher in ovarian cancer tissues compared to normal ovarian tissues (Fig. 7C, D). It is worth noting that there was a significant disparity in KRT81 expression between early and late-stage ovarian cancer tissues, with higher expression in late-stage ovarian cancer tissues (Fig. 7E, F), which is consistent with the results of the TCGA dataset. This implies that KRT81 may be one of the markers for predicting the prognosis of ovarian cancer. Though further mulicenter trials are needed to validate this observation.

Fig. 7.

Immunohistochemical experiments. A Differential expression of prediction model genes in TCGA cohort. B KRT81 was differentially expressed in different periods in the TCGA cohort. C Immunohistochemical image of KRT81 in ovarian cancer and normal tissue. D Differences in expression of KRT81 in normal tissues and ovaries. E Immunohistochemical images of KRT81 in different stages of ovarian cancer. F The difference of KRT81 expression in different stages of ovarian cancer

Discussion

OC is a highly heterogeneous disease, with significant variability in patient survival rates, even among those at similar stages. Current treatment modalities have had limited success in improving patient prognosis. Epithelial ovarian cancer (EOC), which accounts for over 95% of ovarian cancer cases, is recognized as an immunogenic cancer, with spontaneous anti-tumor immune responses observed in approximately 55% of patients [39]. Despite the promise of immunotherapy for HGSOC, several challenges remain, including low response rates, potential side effects, and the absence of robust biomarkers for patient stratification.

The initiation and progression of cancer are closely associated with alterations in DNA methylation patterns. Yizhak et al. demonstrated that gene methylation detection is particularly suitable as a marker for cancer screening [40]. Furthermore, an increasing body of research indicates that DNA methylation is intricately linked to the development and progression of ovarian cancer and to mechanisms of immune escape [41, 42].

Recent studies have explored the potential of using DNA methylation-based immune subtyping to improve the efficacy of ovarian cancer immunotherapy, promote personalized treatment, and enhance survival rates. However, the field of molecular typing remains in its early stages. For instance, Fang et al. conducted a comprehensive analysis of DNA methylation and transcriptome data, initially identifying differences in immune characteristics between two risk groups. This work lays the foundation for further exploration of synergistic targets to improve the efficacy of immunotherapy in OC patients [43]. Nonetheless, current studies on DNA methylation often focus downstream on differentially expressed genes, neglecting the upstream role of transcription factors. DNA methylation levels are spatially correlated across the genome, typically higher in repressed regions and lower in transcription factor (TF) binding sites and active regulatory regions. A study by Martina et al. revealed that DNA methylation increases as TF occupancy decreases during evolution, suggesting a coordinated evolutionary process [44].

In this study, we utilized ovarian cancer DNA methylation data from the TCGA database, integrating it with transcription factor databases to identify key DMGs significantly associated with ovarian cancer prognosis. The analysis enabled the accurate classification of HGSOC into distinct subgroups. The C1 subgroup was characterized by enriched biological processes related to immune pathways, higher immune infiltration levels, and better prognosis, leading us to define it as the immune infiltration subtype. In contrast, the C2 subgroup, with lower immune infiltration and poorer prognosis, was designated as the immunosuppression subtype. To further delineate the differences between these subtypes, we analyzed transcriptome data, dividing tumor samples from the TCGA database into training and test sets in a 7:3 ratio. A risk score for prognostic prediction was constructed, and survival analysis revealed significant differences between high-risk and low-risk groups. The external validation cohort GSE13876 further confirmed the prognostic value of this risk score in independent samples.

Among the four genes used to construct the prediction model—KRT81, PAPPA2, FGF10, and FMO2—each has been implicated in immune function. Yan et al. found that high KRT81 expression in triple-negative breast cancer (TNBC) is associated with increased immune evasion, CD8 + T cell function inhibition, and promoting TNBC cell proliferation, migration, invasion, and adhesion [45]. Dong et al. reported that PAPPA2 mutations could serve as new indicators for stratifying beneficiaries of immune checkpoint inhibitor (ICI) treatment in non-small cell lung cancer (NSCLC) and cutaneous melanoma (SKCM) [46]. Wu et al. discovered that in breast cancer, patients with high FMO2 levels were more responsive to anti-programmed cell death protein 1, anti-programmed cell death ligand 1, and anti-cytotoxic T lymphocyte antigen 4 immunotherapy [47]. These studies suggest that the identified prognostic genes may represent potential synergistic targets for modulating immune cell function. However, the precise mechanisms by which these genes influence ovarian cancer warrant further investigation in prospective studies.

In conclusion, the TF-DMGs score provides a promising means of predicting survival in ovarian cancer patients. Cross validation showed that the TF-DMGs score was effective on different datasets and showed high robustness in survival prediction. Considering these facts, it is reasonable to suggest that patients with ovarian cancer should be classified according to TF-DMGs score to select an appropriate treatment strategy, such as antineoplastic drugs or immunotherapy. In addition, studies have found that KRT81 is closely related to the clinical stage of ovarian cancer, which is also verified by immunohistochemical experiments, suggesting that KRT81 may be a potential biomarker.

Our study also has several limitations. First, further experimental studies and clinical trials are required to validate how the relationship between transcription factor-related DMGs and immune cells can be leveraged to enhance the efficacy of immunotherapy. Second, constructing a prognostic prediction model based on the four genes quantified by microarray presents challenges for clinical application. The specific mechanisms of action of these predictive genes in ovarian cancer must be elucidated through in vivo and in vitro experiments. Third, although the prediction model was verified in an independent queue, current models lack validation in multi-center cohorts and diverse populations. Finally, microarray-based quantification may miss low-abundance transcripts critical for immune regulation.

Author contributions

Jingshu Hu wrote the main manuscript text. Su Mu participates in processing data; Zhijun Qin, Hengyu Wang, Jiayu Li participated in the collection of clinical information; Kexin Chang, Guosheng He participated in text proofreading; Yan Zhang is responsible for reviewing the manuscript; Xiuwei Chen was in charge of reviewing the original manuscript. All authors commented on the manuscript. All authors contributed to the article and approved the submitted version.

Funding

This study was supported by Foundation of Harbin Medical University (PDYS2024-08).

Data availability

All datasets involved in our study were obtained from The Cancer Genome Atlas (TCGA), Gene Expression Omnibus (GEO) and Genotype-Tissue Expression (GTEx) summary data that are publicly Accessible. Sequence data that support the findings of this study have been deposited in the Gene Expression Omnibus with the primary accession code GSE13876.

Partial serous ovarian cancer sample RNA-Seq data were deposited into the Gene Expression Omnibus database under accession number GSE13876 and are available at the following URL: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE13876.

Declarations

Ethics approval and consent to participate

All patients gave written informed consent for using clinical data in the study in accordance with the Declaration of Helsinki. This study was approved by the Ethics Committee of Harbin Medical University Cancer Hospital, affiliated with Harbin Medical University (KY2024-91). All methods were performed in accordance with the guiding principles of the Ethics Committee of Harbin Medical University Cancer Hospital.

Competing interests

The authors declare no competing interests.