DNA methylation profiling enables subclassification of mucinous ovarian carcinoma and distinguishes it from extraovarian mucinous metastases Authors: Teodor G. Calina, Elena Grafenhorst, Simon Schallenberg, Markus Möbs, Bjarne Daenekas, Alexandru A. Sabo, BaoQing Chen, Ines Koch, Tobias Janik, Michael Churchman, Charlie Gourley, Elizabeth L. Christie, Catalin Vasilescu, Vlad Herlea, Wolfgang D. Schmitt, Nicolae Leopold, Florian Heitz, Angelina Vlaski, Yuri Tolkach, Felix Bremmer, Frederick Klauschen, Reinhard Büttner, Philipp Ströbel, Eike Burandt, Guido Sauter, Frank L. Heppner, Eliane T. Taube, Jalid Sehouli, Ioana E. Braicu, Philipp Euskirchen, David Horst, Kylie L. Gorringe, David Capper, Mihnea P. Dragomir Published: 2026 ABSTRACT SummaryMucinous ovarian carcinoma (MOC) is an epithelial ovarian cancer subtype that is frequently misclassified as extraovarian mucinous metastasis (EOM) because of overlapping features. To address this diagnostic challenge, we perform genome-wide DNA methylation profiling of 58 MOCs, 38 EOMs, and 18 mucinous borderline ovarian tumors (mBOTs) collected from six institutions. Methylation analysis defines two mBOT groups, one epigenetically similar to normal ovary and one resembling MOC. Unsupervised clustering reveals two distinct MOC methylation subtypes with potential prognostic relevance in the internal cohort. Using these data together with 389 external profiles, we develop and validate a three-step machine-learning classifier that distinguishes MOC from EOM with 95.5% accuracy. External validation of this classifier on 21 MOCs and 24 EOMs yields an accuracy of 91.11% for differentiating MOC from EOM. These findings establish an epigenetic framework for mucinous ovarian tumors and provide a robust clinical classification tool. Introduction Mucinous ovarian carcinoma (MOC) is one of the rarest and least studied epithelial ovarian cancers. Two aspects make its surgical pathology particularly compelling: the unknown origin and the clinical challenge of distinguishing primary MOCs from metastatic mucinous tumors involving the ovary. The cell of origin remains uncertain. The prevailing hypothesis suggests that MOC arises from mucinous borderline ovarian tumors (mBOTs), which themselves develop from precursor benign mucinous lesions such as mucinous cystadenomas or adenofibromas.1 A key argument supporting this theory is the frequent presence of KRAS mutations across benign, borderline, and malignant mucinous lesions, indicating a molecular continuum of progression as well as the presence of all three morphological types within a tumor.1 However, despite this proposed pathogenesis, the true origin of mucinous ovarian lesions remains elusive2,3 and has not been explored from an epigenetic perspective. Given that the normal ovarian surface epithelium lacks mucin-secreting cells, several origins have been proposed. These include mucinous metaplasia of the ovarian surface epithelium,4 a germ cell origin—supported by associations between MOC and teratomas3,5—and Brenner tumors, which frequently contain mucinous epithelium.2,3,6 To support our approach of epigenetic characterization of MOC, we recently showed that mucinous cystic neoplasms (MCNs) of the liver and pancreas share an epigenetic profile with MOC,7 suggesting a possible common origin. This finding highlights the potential of DNA methylation to offer additional insights into the pathogenesis of MOC. In addition, we consider that an epigenetic characterization of MOC could also help the subclassification of these tumors. The ovary is a common site for extraovarian metastases (EOMs) of mucinous tumors, which account for approximately 10%–25% of malignant ovarian neoplasms.8,9 These metastases originate from various primary sites outside the ovary.8,9,10 Distinguishing primary MOC from EOM remains a significant diagnostic challenge. Retrospective studies have shown that, upon rigorous histopathological reassessment, 50%–70% of tumors initially diagnosed as MOC are reclassified as EOM.11 Current diagnostic approaches rely on clinical features, histopathology, and immunohistochemistry (IHC) but are limited in accuracy. Clinically, EOMs tend to be bilateral, smaller in size, and more infiltrative, whereas MOCs are typically unilateral, larger, and form solid and cystic masses. Histologically, EOMs often show stromal desmoplasia, infiltrative and nodular growth, ovarian surface involvement, and lymphovascular invasion.12,13 In contrast, MOCs display a more glandular architecture, often resembling mucinous cystadenomas or mBOTs.12,13 IHC analysis includes several markers that provide supportive evidence in differentiating MOCs from EOMs. However, a definitive diagnosis based solely on these parameters remains difficult, particularly for upper gastrointestinal primaries.14 Therefore, there is a pressing need for diagnostic tools that can distinguish MOCs from EOMs. Furthermore, in many cases, EOMs are often detected without an identifiable primary tumor, prompting a complex and time-consuming diagnostic workup to determine their origin and delaying treatment initiation. Thus, a tool capable of predicting the site of the primary tumor would not only streamline the diagnostic process but also accelerate treatment initiation. DNA methylation patterns are tissue-specific, making them an ideal tool for characterizing rare entities such as MOC and for distinguishing MOCs from EOMs. In recent years, DNA methylation profiling has revolutionized tumor classification, most notably in neuro-oncology, where it has improved the diagnosis of brain tumors.15 Moreover, methylation-based approaches have facilitated the development of algorithms capable of identifying the tissue of origin in cancers of unknown primary16,17 and have enabled machine-learning-based diagnostic tools that differentiate primary tumors from metastatic lesions with high accuracy.18,19 These advances highlight the potential of DNA methylation profiling to resolve the long-standing challenge of differentiating MOCs from EOMs. Results The DNA methylation profile of MOC in the context of ovarian tumors To obtain an epigenetic perspective on MOC, we established a multi-institutional cohort of 58 MOCs and performed a broad IHC and mutational analysis in parallel with genome-wide DNA methylation profiling (Figure 1A). Samples were collected from Charité-Universitätsmedizin Berlin, Germany (Charité; n = 34), Evangelische Kliniken Essen-Mitte, Germany (EKEM; n = 6), and the Peter MacCallum Cancer Center collective, Melbourne, Australia (PMCC; n = 18). The PMCC collective comprised samples from the Australian Ovarian Cancer Study (AOCS; n = 4), the Edinburgh Cancer Research Centre (ECRC; n = 12), the Mayo Clinic (Mayo; n = 1), and the Victorian Cancer Biobank (VCB; n = 1). All samples but one were from formalin-fixed, paraffin-embedded (FFPE) material, while the remaining sample was from fresh-frozen (FF) tissue. The mean patient age at diagnosis was 52 years, while the mean sample age was 12.1 years (Figure 1B).Figure 1The DNA methylation profile of MOC in the context of ovarian tumors(A) Workflow for DNA-methylation-based characterization and development of a machine-learning classifier to distinguish mucinous ovarian tumors (MOTs) from EOMs.Cohort characterization included 12-marker IHC and targeted DNA sequencing. Created with BioRender.com.(B) Clinicopathological, immunohistochemical, and mutational features of mBOT and MOC.(C) Representative hematoxylin and eosin (H&E) images of MOC G1, G2, and G3 (scale bars, 100 μm; ×10 and ×40 magnification).(D) t-SNE representation of DNA methylation profiles for mBOT, MOC, HGSOC, normal ovary, and normal Fallopian tube, based on the top 20,000 most variable CpGs.(E) Hierarchical clustering and heatmap of DNA methylation profiles for the same tumor and normal samples as in (D), based on the top 20,000 most variable CpGs. The DNA methylation profile of MOC in the context of ovarian tumors (A) Workflow for DNA-methylation-based characterization and development of a machine-learning classifier to distinguish mucinous ovarian tumors (MOTs) from EOMs. Cohort characterization included 12-marker IHC and targeted DNA sequencing. Created with BioRender.com. (B) Clinicopathological, immunohistochemical, and mutational features of mBOT and MOC. (C) Representative hematoxylin and eosin (H&E) images of MOC G1, G2, and G3 (scale bars, 100 μm; ×10 and ×40 magnification). (D) t-SNE representation of DNA methylation profiles for mBOT, MOC, HGSOC, normal ovary, and normal Fallopian tube, based on the top 20,000 most variable CpGs. (E) Hierarchical clustering and heatmap of DNA methylation profiles for the same tumor and normal samples as in (D), based on the top 20,000 most variable CpGs. For all samples from Charité and EKEM, we performed a 12-marker IHC analysis to reevaluate and characterize the cases. All diagnoses were confirmed by the study pathologist based on histology, IHC, clinical data, and sequencing data. Regarding the PMCC samples, all diagnoses were confirmed based on similar criteria.1 Hence, all 58 samples represent bona fide MOCs. Twenty samples were G1 (well-differentiated), 27 G2 (moderately differentiated), and the remaining 11 were G3 (poorly differentiated) (Figures 1B and 1C). The morphological growth patterns were classified as either infiltrative or expansile subtypes, based on established criteria,20,21 resulting in 42 expansile and 11 infiltrative cases (Figure 1B). We stratified cases using IHC profiles following published algorithms.22,23,24,25 In total, 30 samples displayed an MOC-indicative profile and 6 were MOC-unspecific (Figure 1B; Figure S1A). Next, we analyzed a cohort of 18 mBOT Charité cases. All were FFPE, with a median sample age of 1.6 years and a median patient age of 54.1 years. All but one were stage T1, the other was T2. All diagnoses were confirmed by the study pathologist after assessment of clinical, histological, IHC, and mutational data (Figure 1B). To minimize heterogeneity, we extracted DNA exclusively from ovarian tumor blocks in all cases. For 37/40 samples in the Charité/EKEM subcohort, we performed targeted DNA sequencing, while for the 17/18 PMCC cases, mutations were retrieved from the prior publication.1 Because the targeted sequencing panel only partially covered TP53, we additionally employed IHC for p53 to infer the samples with a mutant phenotype. As expected,1,26,27,28 KRAS and TP53 were the most commonly mutated genes, with an enrichment of TP53 mutations in G2 and G3 samples (Fisher exact test, p = 0.0021) and an almost equal distribution of KRAS mutations across grades (Fisher exact test, p = 0.1203). Furthermore, four samples harbored PIK3CA mutations, one sample SMAD4 mutation, five BRAF mutations, and one sample GNAS mutation (Table S1). All these mutations were previously reported for MOC.26,27,28 KRAS was the most frequently mutated gene in mBOT, present in 16/18 samples. PIK3CA and BRAF were mutated in 2/18 each, while NRAS and GNAS appeared in 1/18 each. TP53 was mutated in 1/18 mBOTs. For all mBOT and MOC samples, we performed genome-wide DNA methylation profiling. Using the MOC cohort, the 18 mBOTs, 16 high-grade serous ovarian carcinoma (HGSOCs), 17 normal ovaries, and 12 normal Fallopian tubes, we created an epigenetic landscape of ovarian tumors. MOCs formed a distinct and mostly homogeneous subgroup, separate from HGSOC (Figure 1D). Furthermore, normal Fallopian tube, normal ovary, and approximately half of the mBOT samples formed three distinct groups, while the remaining mBOT samples grouped with MOC. Next, we examined parameters that could impact sample distribution and noted that institution, sample age, grade, T-stage, IHC-specificity profile, KRAS, TP53, BRAF mutation status, and precursor lesions had no impact on sample grouping (Figures S1B–S1J). Tumor purity, however, showed a gradient across MOCs in the t-SNE, and, for the mBOT samples, we clearly observed differences in tumor purity between the distinct group and the samples grouping with MOCs (Figure S1K). To validate purity estimates, we compared our results with values inferred by doubling the allele frequency of oncogenic mutations, most often KRAS (or BRAF/NRAS when KRAS was absent), and found a strong concordance (Spearman r test, p < 0.0001). Furthermore, six MOC samples grouped closer to HGSOC (Figure 1D), consistent with prior pan-ovarian DNA methylation analysis.29 These cases were not enriched in any characteristic except for being slightly older than the general cohort samples (Figures S1B–S1J). Using a second method for characterizing the samples based on DNA methylation, hierarchical clustering confirmed two mBOT groups: one intermingled with MOC and another forming distinct, near-normal samples. Additionally, MOC samples showed a distinct methylation profile relative to other tumors but separated internally into two distinct clusters, suggesting the existence of two potential epigenetic MOC subtypes (Figure 1E). DNA methylation defines molecular subtypes of MOC and mBOT The DNA methylation profiling of samples described above defined several groups: normal ovary, mBOTnormal_like, mBOTMOC_like, and MOC. By comparing the two mBOT subgroups, we did not observe any significant differences in patient age or mutational parameters. However, mBOTnormal_like showed significantly higher progesterone receptor (PR) and p16 expression versus mBOTMOC_like, whereas all other IHC parameters showed no difference between the two subtypes (Table S2). Morphological reanalysis revealed a more complex morphology and cytology for mBOTMOC_like compared to mBOTnormal_like (Figure S2A). This observation confirms our previous analysis of MCNs in the context of ovarian cancers, in which some mBOTs grouped with low-grade MCN of the pancreas/liver, while others aligned with high-grade and invasive MCNs.7 Furthermore, we performed unsupervised K-means clustering of MOC samples, identifying two clusters: MOCK1 and MOCK2 (Figure 2A; Figure S2B). Mapping these clusters onto the landscape t-SNE revealed specific spatial relations (Figure S2C) that were not easily discernible previously (Figure 1D). Notably, we observed that most (9/10) mBOTMOC_like grouped with MOCK1, forming a continuous progression.Figure 2DNA methylation defines molecular subtypes of MOC and mBOT(A) Principal-component analysis (PCA) of MOC samples colored by K-means clustering, separating MOCK1 and MOCK2 subtypes.The percentage of variance explained by PC1 and PC2 is indicated on the axes.(B) Kaplan-Meier OS curves for patients with MOCK1 versus MOCK2 (n = 52). p value from log rank test.(C) Forest plot summarizing multivariate Cox proportional hazards analysis for OS.Dots indicate hazard ratios, and horizontal lines show 95% confidence intervals; p values from Wald tests.(D) Distribution of DMPs identified in pairwise comparisons among normal ovary, mBOTnormal_like, mBOTMOC_like, MOCK1, and MOCK2.Bars show numbers of hypermethylated (orange) and hypomethylated (green) CpGs.(E) Genome-wide copy-number alteration profiles for mBOTnormal_like (n = 8), mBOTMOC_like (n = 10), MOCK1 (n = 38), and MOCK2 (n = 20).The frequency of chromosomal gains (red) and losses (blue) is shown across the genome.(F) Inferred fractions (%) of fibroblasts, cancer cells, and CD8+ T cells across mBOTnormal_like, mBOTMOC_like, MOCK1, and MOCK2, with normal ovary as reference.Each dot represents a unique patient.Statistical comparisons were made using Kruskal-Wallis or one-way ANOVA, as appropriate. ns, not significant; ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001. DNA methylation defines molecular subtypes of MOC and mBOT (A) Principal-component analysis (PCA) of MOC samples colored by K-means clustering, separating MOCK1 and MOCK2 subtypes. The percentage of variance explained by PC1 and PC2 is indicated on the axes. (B) Kaplan-Meier OS curves for patients with MOCK1 versus MOCK2 (n = 52). p value from log rank test. (C) Forest plot summarizing multivariate Cox proportional hazards analysis for OS. Dots indicate hazard ratios, and horizontal lines show 95% confidence intervals; p values from Wald tests. (D) Distribution of DMPs identified in pairwise comparisons among normal ovary, mBOTnormal_like, mBOTMOC_like, MOCK1, and MOCK2. Bars show numbers of hypermethylated (orange) and hypomethylated (green) CpGs. (E) Genome-wide copy-number alteration profiles for mBOTnormal_like (n = 8), mBOTMOC_like (n = 10), MOCK1 (n = 38), and MOCK2 (n = 20). The frequency of chromosomal gains (red) and losses (blue) is shown across the genome. (F) Inferred fractions (%) of fibroblasts, cancer cells, and CD8+ T cells across mBOTnormal_like, mBOTMOC_like, MOCK1, and MOCK2, with normal ovary as reference. Each dot represents a unique patient. Statistical comparisons were made using Kruskal-Wallis or one-way ANOVA, as appropriate. ns, not significant; ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001. We compared MOCK1 and MOCK2 by patient age, histology, IHC marker expression and score, mutational profile, and precursor lesions, observing no significant statistical difference (Table S3). However, consistent with the tumor purity distribution, a significant difference between MOCK1 and MOCK2 was observed (unpaired t test, p < 0.0001) (Figure S1J; Table S3). Next, as there are very few prognostic parameters for MOC,30 and in order to check the potential clinical relevance of MOCK1 and MOCK2, we assessed whether these epigenetic subtypes were associated with differences in overall survival (OS). The median cohort OS was 9.1 years. Patients with MOCK1 exhibited longer OS compared to those in MOCK2 (log rank test, p = 0.0059; Figure 2B). Patient age, all available pathological, IHC, and molecular parameters had no impact on the OS, except for T-stage, which showed a notable significance (log rank test, p = 0.0006; Figure S2D; Table S4). In a multivariate Cox proportional hazards analysis, both advanced tumor stage and DNA methylation cluster K2 were independently associated with reduced survival. T2+ tumors were associated with a 4.38-fold increased risk of death, and MOCK2 patients had a 3.22-fold increased risk (hazard ratio [HR] 3.22, 95% confidence interval [CI] 1.23–8.42, p = 0.017), independent of tumor stage (Figure 2C). Therefore, we focused on understanding the epigenetic characteristics of the newly observed mBOT and MOC subgroups and their epigenetic relationship to normal ovary. We hypothesized that mBOTMOC_like have already acquired most of the molecular characteristics of an MOC. First, we performed differentially methylated probe (DMP) analysis pairwise among the four groups and normal ovary. We observed a clear crescendo in the absolute number of DMPs, progressing from mBOTnormal_like versus normal ovary (2,966 DMPs), to mBOTMOC_like versus normal (31,986 DMPs), to MOCK1 versus normal (44,894 DMPs), and finally to MOCK2 versus normal (96,810 DMPs) (Figure 2D; Figure S2F). This analysis further revealed minimal differences between mBOTMOC_like and MOCK1 (only 19 DMPs) (Figure 2D). Using DMPs, we built an epigenetic relatedness network, where each edge represents the number of DMPs between two groups. While the t-SNE and clustering analysis were performed using the most standard deviated CpGs, this analysis is based on a completely different CpG prioritization, further showing the closest epigenetic relationship between mBOTMOC_like and MOCK1, as well as close relationships between the two types of mBOTs, including mBOTnormal_like with MOCK1, mBOTMOC_like with MOCK2, MOCK1 with MOCK2, and normal ovary with mBOTnormal_like (Figure S2E). By contrast, HGSOC and normal Fallopian tube were more distant from mucinous ovarian tumors (MOTs) (Figure S2E). Second, to further investigate the relationship between dysregulated methylation and tumor progression, we analyzed the differentially methylated pathways across the groups (Table S5). Pathways related to apoptosis and cell death were activated in mBOTnormal_like compared to normal ovary. Comparing mBOTnormal_like to mBOTMOC_like, we identified a few differentially methylated pathways, all of which were inactivated, including O-linked glycosylation of mucins. Aberrant glycosylation has previously been hypothesized to play a role in cancer progression.31 Interestingly, the epigenetic differences between mBOTMOC_like and MOCK1 were minimal. In contrast, a larger number of dysregulated pathways distinguished mBOTMOC_like and MOCK2, including activation of O-linked glycosylation and inhibition of G-protein-coupled receptor (GPCR)-related signaling (Table S5) in MOCK2. GPCR dysregulation has been previously reported in MOC and was associated with poorer survival outcomes.32 Third, we performed summary copy number analysis (CNA) plots for the two subtypes of mBOT and MOC. mBOTnormal_like presented an almost flat copy-number profile, with only one sample showing a clear KRAS gain. mBOTMOC_like showed CDKN2A/B loss in 4/10 samples, which could also be observed in most MOCK1 and MOCK2. In MOCK1, we noticed a gain of chromosome 17q that included the ERBB2 gene (16/38 samples, 42.11%) and a loss of 17p that included TP53 (22/38 samples, 57.89%). These chromosomal changes were also present in MOCK2, ERBB2 gain (8/20, 40%), and TP53 loss (11/20, 55%) and were previously reported especially in high-grade MOCs.1,33 Finally, we noticed a large gain of most of chromosome 19, which was more pronounced in MOCK2 compared to the other groups (Figure 2E). This region includes the CCNE1 gene that was recently associated with poor prognosis in multiple epithelial ovarian cancers34 and was gained in 10/38 (26.32%) MOCK1 and 6/20 (30%) MOCK2 cases. Finally, we deconvoluted the cellular composition using MethylCIBERSORT.35,36 We observed a gradual decrease in fibroblasts and a gradual increase in cancer cells from mBOTnormal_like toward MOCK2 (Figure 2F). Specifically, fibroblast proportions were highest in mBOTnormal_like and decreased noticeably in both MOCK1 (one-way ANOVA test, p < 0.0001) and MOCK2 (one-way ANOVA test, p < 0.0001). Conversely, cancer cell proportions were significantly higher in MOCK1 (Kruskal-Wallis test, p = 0.0056) and MOCK2 (Kruskal-Wallis test, p < 0.0001) compared to mBOTnormal_like samples. In addition, we observed significantly higher levels of CD8+ T cells in MOCK1 versus MOCK2 (Kruskal-Wallis test, p = 0.0414) (Figure 2F). Spatial IHC analysis confirmed these CD8+ differences: MOCK1 tumors exhibited more intra-tumoral infiltration, whereas in MOCK2, CD8+ positivity was more dominant in the stromal compartment (Mann-Whitney test, p = 0.0452) (Figure S2H). This suggests that for MOCs, an immune-enriched subtype (MOCK1) shows better survival rates, similar to HGSOC.37 For the other immune cells, we found no significant differences (Figure S2G). DNA methylation profiling of EOM in the context of ovarian tumors To develop a machine-learning algorithm that differentiates MOC from EOM, we assembled a cohort of mucinous metastases to the ovary. We collected 38 EOM samples (Figure 3A), including appendiceal adenocarcinomas (ApAD, n = 3), low-grade appendiceal mucinous neoplasm (LAMN, n = 3), cholangiocarcinoma (CHOL, n = 2), colorectal adenocarcinoma (COAD, n = 9), esophageal adenocarcinoma (EAC, n = 1), lung adenocarcinoma (LUAD, n = 2), and stomach adenocarcinoma (STAD, n = 18) (Figure 3B). All samples were FFPE, and all but one were from Charité. The mean age of the patients at diagnosis was 53.7 years, while the mean sample age was 7.8 years. For each case, we performed IHC analysis (35 showed IHC profiles consistent with EOM and 3 were unspecific) and corroborated the results with the clinical history, thus we have carefully confirmed the EOM diagnosis.Figure 3DNA methylation profiling of EOM in the context of ovarian tumors(A) Clinicopathological, immunohistochemical, and mutational features of the EOM cohort.(B) Representative H&E images of EOMApAD, EOMCHOL, EOMCOAD, EOMEAC, EOMLAMN, EOMLUAD, and EOMSTAD (low magnification, ×3.8; high magnification, ×20).(C) t-SNE representation of DNA methylation profiles for EOM, mBOT, MOC, HGSOC, normal ovary, and normal Fallopian tube, based on the 20,000 most variable CpGs.(D) Genome-wide copy-number alteration profiles of MOC (n = 58) and EOM (n = 38). Frequencies of chromosomal gains (red) and losses (blue) are shown across the genome.(E) Inferred fractions (%) of fibroblasts, cancer cells, and CD8+ T cells in EOM and MOC, with normal ovary as reference. Each dot represents a unique patient.Statistical comparisons were made using unpaired two-tailed t test or Mann-Whitney U test, as appropriate. ns, not significant; ∗∗p < 0.01; ∗∗∗∗p < 0.0001. DNA methylation profiling of EOM in the context of ovarian tumors (A) Clinicopathological, immunohistochemical, and mutational features of the EOM cohort. (B) Representative H&E images of EOMApAD, EOMCHOL, EOMCOAD, EOMEAC, EOMLAMN, EOMLUAD, and EOMSTAD (low magnification, ×3.8; high magnification, ×20). (C) t-SNE representation of DNA methylation profiles for EOM, mBOT, MOC, HGSOC, normal ovary, and normal Fallopian tube, based on the 20,000 most variable CpGs. (D) Genome-wide copy-number alteration profiles of MOC (n = 58) and EOM (n = 38). Frequencies of chromosomal gains (red) and losses (blue) are shown across the genome. (E) Inferred fractions (%) of fibroblasts, cancer cells, and CD8+ T cells in EOM and MOC, with normal ovary as reference. Each dot represents a unique patient. Statistical comparisons were made using unpaired two-tailed t test or Mann-Whitney U test, as appropriate. ns, not significant; ∗∗p < 0.01; ∗∗∗∗p < 0.0001. Importantly, all analyses were performed on tissue from the ovarian metastasis site. Targeted DNA sequencing and IHC revealed TP53 mutations or mutant p53 phenotype in 20/37 (54.05%) cases. Other recurrent mutations included KRAS (n = 4), APC (n = 4), SMAD4 (n = 2), and single cases with PIK3CA, NRAS, or GNAS (Figure 3A). When comparing the different parameters between MOCs and EOMs, we found a significantly higher histologic grade and increased CDX2 expression in EOMs, while MOCs more frequently expressed CK7, MUC5AC, PAX8, and PR. At the mutational level, KRAS mutations were significantly more common in MOC, whereas TP53 frequency was similar (Table S6). We then performed genome-wide DNA methylation profiling for all 38 EOMs and projected them with the existing ovarian tumor t-SNE (Figure 3C). EOMs tended to form spatially distinct groups apart from MOCs, with EOMCOAD samples most separated (Figure 3C). Sample distribution was not influenced by patient age, tumor primary site, or mutational spectrum but was partially influenced by tumor purity (Figures S3A–S3F). Because EOMSTAD samples frequently exhibit signet ring morphology, this lower purity was somewhat expected (Figure S3F; Figure 3B). CNA profiling highlighted key differences between MOC and EOM groups. As already reported in the literature, MOC showed more frequent CDKN2A/B losses (MOC: 55.17% vs. EOM: 36.84%),26 and, in addition, we observed an enrichment in losses of the TP53 region (MOC: 56.9% vs. EOM: 31.58%) and a gain of ERBB2 (MOC: 41.38% vs. EOM: 10.53%). On the other hand, we observed more frequent gains of entire chromosomes (chrs. 7, 13, 19, and 20) in the EOM group versus MOC group (Figure 3D). We also compared the cellular composition after DNA methylation deconvolution. We observed a higher proportion of fibroblasts (Mann-Whitney test, p < 0.0001) and lower level of cancer cells (Mann-Whitney test, p = 0.0014) in EOM versus MOC, and regarding the immune infiltrate we noticed higher levels of neutrophils (Mann-Whitney test, p = 0.0384) and lower levels of B cells (Mann-Whitney test, p = 0.0271) in MOC versus EOM, while all other immune cells showed no differences (Figure 3E; Figure S3G). DMP analysis identified 293 significant DMPs between EOM and MOC (Figure S3H). This degree of relatedness resembled that seen between mBOTnormal_like and mBOTMOC_like (Figure 2D), underscoring the difficulty of distinguishing MOC and EOM. Not surprisingly, genes with hypomethylated CpGs in EOM versus MOC included transcription factors linked to specific tumor types, such as SATB2, a marker of COAD. A DNA methylation classifier for differentiating EOM from MOC With the assembled MOC and EOM cohorts, we next tackled one of the most difficult clinical challenges of mucinous carcinomas of the ovary—differentiating between MOC and EOM. To build an accurate and safe DNA methylation classifier, we developed a three-step machine-learning pipeline: step 1 to detect primary MOTs (defined as mBOT+MOC), step 2 to identify EOM by separating mucinous carcinomas (MCs) from all other lesions (anomaly detection), and step 3 to predict the tissue of origin of EOM. Samples not classified as MOTs in step 1 proceed to step 2, while those labeled as MCs in step 2 advance to step 3 (Figure 4A).Figure 4A DNA methylation classifier for differentiating EOM from MOC—step 1 and step 2(A) Schematic of the three-step classifier workflow.Step 1 detects MOT. Samples not classified as MOT proceed to step 2, where EOMs are detected and other tumor types are filtered out. Step 3 further resolves the tissue of origin for EOM. Created with BioRender.com.(B) Overview of training and test cohorts used in each classifier step.(C) Comparison of step 1 accuracies using LR versus SVM across tumor groups: MOC, mBOT, EOM, primary mucinous carcinomas that can metastasize to the ovary (pMC), other lesions (non-mucinous neoplasms), and normal ovary.(D) Confusion matrix showing step 1 LR classification results.(E) Comparison of step 2 accuracies (LR versus SVM) across EOM, pMC, other lesions, and normal ovary.(F) Confusion matrix showing step 2 SVM classification results.(G) Confusion matrix showing combined step 1 (LR) and step 2 (SVM) classification results.(H) t-SNE representation of DNA methylation profiles (as in Figure 3C) showing MOT and EOM samples classified by the step 1 + step 2 pipeline. Correctly classified samples (green), misclassified samples (red). A DNA methylation classifier for differentiating EOM from MOC—step 1 and step 2 (A) Schematic of the three-step classifier workflow. Step 1 detects MOT. Samples not classified as MOT proceed to step 2, where EOMs are detected and other tumor types are filtered out. Step 3 further resolves the tissue of origin for EOM. Created with BioRender.com. (B) Overview of training and test cohorts used in each classifier step. (C) Comparison of step 1 accuracies using LR versus SVM across tumor groups: MOC, mBOT, EOM, primary mucinous carcinomas that can metastasize to the ovary (pMC), other lesions (non-mucinous neoplasms), and normal ovary. (D) Confusion matrix showing step 1 LR classification results. (E) Comparison of step 2 accuracies (LR versus SVM) across EOM, pMC, other lesions, and normal ovary. (F) Confusion matrix showing step 2 SVM classification results. (G) Confusion matrix showing combined step 1 (LR) and step 2 (SVM) classification results. (H) t-SNE representation of DNA methylation profiles (as in Figure 3C) showing MOT and EOM samples classified by the step 1 + step 2 pipeline. Correctly classified samples (green), misclassified samples (red). For step 1, we used 20 MOC samples and 172 other lesions—including primary MCs (pMCs) that can metastasize to the ovary and show mucinous morphology—as the training set. For step 2, we used 133 pMCs (origin of EOM) and 116 non-mucinous lesions. For step 3, the training data contained the same 133 pMCs from step 2, divided into three classes: COAD (n = 20), STAD (n = 20), and other MCs (n = 93). Internal model validation was performed on an independent cohort of 311 samples, including 38 MOCs, 18 mBOTs, and 38 EOMs, none of which were included in the training or parameter tuning phases (Figure 4B). In a t-SNE analysis containing all samples mentioned above, MOCs, mBOTs, and EOMs largely grouped together and separately from the other tumors, including the tumors of origin of the EOMs. The only exception was EOMCOAD, which grouped near the primary COAD group (8/9 samples). Overall, this analysis revealed differences between MOC and EOM, while also highlighting similarities that may partly explain the challenges in their diagnosis (Figure S4A). For step 1, we selected the top 5,000 most variable CpGs of the training set and compared two algorithms: logistic regression (LR) and support vector machine (SVM) (Figure 4A). LR outperformed SVM for MOT detection, with accuracies of 94.44% versus 83.33% for mBOT and 92.11% versus 89.47% for MOC. For EOMs, SVM outperformed LR: 94.74% compared to 89.47% (Figure 4C). Overall, LR achieved a higher accuracy of 97.11% than SVM 96.78% (Figure 4C). We therefore selected the LR classifier for further analysis. Per-class analysis showed misclassification of three MOCs, one mBOT, four EOMs, and one pMC (Figure 4D). Analyzing the factors associated with the classification results revealed lower accuracy for FFPE versus FF samples, EPICv1 versus EPICv2 arrays, and in-house samples versus other datasets (in-house samples included metastases) (Figure S4B). We also tested if different probability thresholds would improve accuracy, specificity, and sensitivity and observed that the best results are reached without using any threshold, ensuring predictions for all samples (Figures S4C and S4D). The goal of step 2 was to detect the MC samples, both primary (pMC) and metastatic (EOM), and filter out all other tumor types. This anomaly detection layer serves as a safeguard, ensuring that only mucinous carcinomas progress to downstream classification. We again developed and compared LR and SVM classifiers. Here, SVM outperformed LR (98.04% vs. 94.51%), especially for pMCs (99.13% vs. 89.57%; Figure 4E). Hence, SVM was chosen as the classification method for step 2. The investigation of misclassified samples showed no EOM cases were incorrectly classified; only one pMC was misclassified, and four non-mucinous lesions were incorrectly labeled as MC (Figure 4F). The SVM performed better for FFPE vs. FF, EPICv2 vs. EPICv1, and for the datasets containing mainly EOMs (in-house) and pMC samples (Figure S4E). Again, we observed that the highest accuracy and the best balance between specificity and sensitivity is achieved if no thresholds are applied (Figures S4F and S4G). These two steps jointly answer the main clinical question: is a mucinous ovarian lesion primary or metastatic? The overall accuracy achieved was 95.5% (Figure 4G), with most misclassified samples being MOCs (3/38) and EOMs (4/38) (Figure S4H). t-SNE mapping showed two misclassified MOC cases grouping near HGSOC, while the rest of the samples were distributed heterogeneously (Figure 4H). Finally, we compared our data with the results of an IHC classification of the in-house samples. Based on IHC, we reached an accuracy of 87.84% (Figure S4I), with only one sample overlapping between IHC and DNA methylation misclassifications, underlining the complementarity of the two tools (Figure S4J). Exploring the origin of EOM Next, we analyzed the possibility of expanding our classifier to predict the primary site of EOM. Because sample availability was limited, we trained and tested this step only for the two most frequent origins: gastric (EOMSTAD) and colon (EOMCOAD), which account for the majority of mucinous metastases to the ovary.12 For step 3, all samples labeled as MC in step 2 represented the training set for the tissue of origin detection and were divided into three classes: STAD, COAD, and other MC (Figure 5A). Similar to the previous steps, we trained two binomial classifiers to identify STAD and COAD origin, respectively, with other MC classification resulting from the combination of the two models. We again compared LR and SVM models for each classifier and observed higher or equal accuracies for LR when compared independently (STAD: 91.5% [LR] vs. 88.24% [SVM]; COAD: 98.04% [LR] vs. 96.73% [SVM]). We therefore selected LR for both models in step 3. The two classifiers were arranged sequentially, the COAD classifier running first due to the higher accuracy, followed by the STAD classifier. Across the whole testing set, step 3 reaches an accuracy of 92.81%, with an accuracy of 78.94% on just the EOM subset (Figure 5B). Per-sample analysis showed 8/9 EOMCOAD correctly labeled as COAD, 11/18 EOMSTAD classified as STAD, 11/11 other EOM classified as “other” (Figure 5C).Figure 5Exploring the origin of EOM(A) Schematic of step 3 of the DNA-methylation-based classifier, which predicts the tissue of origin for EOM.(B) Comparison of step 3 classification accuracies using LR and SVM across tumor groups: EOM, pMC, other lesions, and normal ovary.(C) Confusion matrix showing step 3 LR classification results. EOM and pMC cases are displayed separately.(D) Sankey plot illustrating per-sample classification flow through steps 1–3 of the pipeline. Exploring the origin of EOM (A) Schematic of step 3 of the DNA-methylation-based classifier, which predicts the tissue of origin for EOM. (B) Comparison of step 3 classification accuracies using LR and SVM across tumor groups: EOM, pMC, other lesions, and normal ovary. (C) Confusion matrix showing step 3 LR classification results. EOM and pMC cases are displayed separately. (D) Sankey plot illustrating per-sample classification flow through steps 1–3 of the pipeline. Next, we analyzed the entire diagnostic pipeline, which yielded an overall accuracy of 92.28% with very few samples misdiagnosed. Of the focal samples, MOTs and EOMs, only three MOCs were misassigned as COAD or STAD, and one mBOT as “no match.” Four EOM samples were labeled as MOTs, one EOMCHOL and three EOMSTAD. All remaining EOMs were correctly identified as “other” in step 1 and confirmed as MC in step 2 (Figure 5D; Figure S5A). We also applied the crossNN classifier38 using rapid Nanopore sequencing of FF tissue for potential intraoperative diagnosis. The Nanopore sequencing showed limited accuracy in classifying MOC FF samples and showed better results on FFPE tissue, even after downsampling EPIC CpGs to match Nanopore coverage (Figures S5B–S5E). Overall, Nanopore-based DNA methylation profiling for intraoperative MOC diagnosis remains challenging, as accurate classification depends critically on selecting tumor-rich regions for DNA extraction. External validation of the classifier and prognostic tool To validate our classifier, we analyzed 45 FFPE samples (21 MOCs and 24 EOMs) sourced from three independent institutions (Figure 6A). Hybridization was performed using the Illumina EPICv2 array, a next-generation array not utilized during the classifier development. For step 1 of the classification pipeline, 18 MOC samples were correctly identified as MOT, while 27 were labeled as “other,” yielding an initial accuracy of 93.33% (Figure 6B). The step 2 analysis of EOM samples resulted in a classification accuracy of 95.83% (Figure 6C). Critically, the combined step 1 and step 2 pipeline, which addresses the primary clinical diagnostic challenge, achieved an accuracy of 91.11% (Figure 6D). Only four samples were misclassified: three MOC cases, two of which exhibited sarcomatoid morphology with osteoclast-like giant cells, and a single EOM sample, a gastric metastasis, incorrectly labeled as “no match” due to low tumor purity. Step 3 correctly indicated the origin of EOM in 17 out of 23 cases that reached this step, yielding an accuracy of 73.91% (Figure S6A). The lower accuracy was induced by some atypical EOM origins, such as an ileum adenocarcinoma and an appendiceal goblet cell adenocarcinoma, all of which were labeled as STAD. The classifier achieved an overall accuracy of 77.78% in the external validation cohort (Figure S6B), mainly because of step 3. When compared to the MOC and EOM samples of the internal validation cohort, accuracy was similar, although modestly lower than that observed in the full cohort due to the significant number of primary tumors in the internal validation cohort that serve as both positive and negative controls. Step 3 is designed specifically to indicate metastatic origins and narrow the search for the primary tumor and does not represent a pan-adenocarcinoma classifier.Figure 6External validation of the classifier and DNA methylation subtypes of MOC(A) Clinicopathological, immunohistochemical, and classification and subclassification features of the external validation cohort. MOC samples that were analyzed twice are marked with bold rectangles.(B) Confusion matrix showing step 1 classification results in the external validation cohort.(C) Confusion matrix showing step 2 classification results in the external validation cohort.(D) Confusion matrix showing combined step 1 + 2 classification results in the external validation cohort.(E) Distribution of DMPs identified in pairwise comparisons among normal ovary, MOCK1 in the external validation cohort (n = 13), and MOCK2 in the external validation cohort (n = 8).Bars show numbers of hypermethylated (orange) and hypomethylated (green) CpGs.(F) Inferred fractions (%) of cancer cells across MOCK1 and MOCK2 in the external validation cohort, with normal ovary as reference. Statistical comparisons were made using unpaired t test. ∗∗p < 0.01.(G) Evaluation of intra-tumoral morphological heterogeneity regarding DNA-methylation-based prognostic and diagnostic classifications.Spatial analysis of two histologically distinct regions from three representative cases: sample 9: mBOT and invasive MOC; sample 10: G1 and G2 MOC; sample 15: expansive and infiltrative growth patterns (magnification, ×10; scale bars, 100 μm). External validation of the classifier and DNA methylation subtypes of MOC (A) Clinicopathological, immunohistochemical, and classification and subclassification features of the external validation cohort. MOC samples that were analyzed twice are marked with bold rectangles. (B) Confusion matrix showing step 1 classification results in the external validation cohort. (C) Confusion matrix showing step 2 classification results in the external validation cohort. (D) Confusion matrix showing combined step 1 + 2 classification results in the external validation cohort. (E) Distribution of DMPs identified in pairwise comparisons among normal ovary, MOCK1 in the external validation cohort (n = 13), and MOCK2 in the external validation cohort (n = 8). Bars show numbers of hypermethylated (orange) and hypomethylated (green) CpGs. (F) Inferred fractions (%) of cancer cells across MOCK1 and MOCK2 in the external validation cohort, with normal ovary as reference. Statistical comparisons were made using unpaired t test. ∗∗p < 0.01. (G) Evaluation of intra-tumoral morphological heterogeneity regarding DNA-methylation-based prognostic and diagnostic classifications. Spatial analysis of two histologically distinct regions from three representative cases: sample 9: mBOT and invasive MOC; sample 10: G1 and G2 MOC; sample 15: expansive and infiltrative growth patterns (magnification, ×10; scale bars, 100 μm). Lastly, we extended our analysis of K1 and K2 MOC clustering to the external cohort. The clustering of the external MOC samples resulted in 13 MOCK1 and 8 MOCK2 (Figure 6A). Due to the recent nature of the external validation cohort, the limited follow-up duration precluded a statistically powered survival analysis. Hence, for the analysis of K1 and K2 MOC subtypes, we performed DMP analysis in MOCK1 and MOCK2 versus normal ovary, along with CNA profiling and cellular decomposition using the DNA methylation data. The absolute number of DMPs between MOCK1 and normal ovary was 46,781 and between MOCK2 and normal ovary was 69,647, while only four DMPs were identified between MOCK1 and MOCK2, indicating minimal differences between the two subtypes (Figure 6E). This is in line with the internal cohort where we observed a smaller number of DMPs between MOCK1 and normal ovary than between MOCK2. CNA analysis showed the same trends as in the internal set, with MOCK1 revealing higher levels of TP53 losses (76.9% versus 25%) and ERBB2 gains (61.53% versus 12.5%) compared to MOCK2 (Figure S6C). As in the internal cohort, cancer cell proportions were significantly higher in MOCK2 compared to MOCK1 (unpaired t test, p = 0.0093) (Figure 6F), while fibroblast proportions were higher, but not statistically significant in MOCK1 compared to MOCK2 (Figure S6D). Regarding the CD8+ T cells proportions, no differences were observed between the two groups; this could be explained in part by the much lower sample number (Figure S6E). Statistical comparisons between MOCK1 and MOCK2 were non-significant for most available parameters and showed borderline significance for grade (Fisher exact test, p = 0.0442) (Table S7). Because MOCs are very heterogeneous tumors, we wanted to see if different morphologies impose diagnosis and stratification problems. We selected three heterogeneous MOC samples and microdissected spatially distinct tumor regions (mBOT versus MOC, G1 versus G2, and expansile versus infiltrative growth), which were analyzed separately. For the diagnosis analysis, all six regions were correctly classified as MOT (Figure 6G). However, the stratification into MOCK1 and MOCK2 showed some degree of heterogeneity. In one sample, the region with infiltrative growth was assigned MOCK1, whereas the region with expansile growth was assigned MOCK2. The other two samples showed concordant subclassification across the analyzed regions (Figure 6G). Discussion In this study, we address two important aspects of MOCs: their stratification and relationship to mBOTs, as well as the differential diagnosis with EOMs. We believe these aspects hold significant clinical relevance. To explore these topics, we employed DNA methylation profiling on a large cohort of MOCs, mBOTs, and EOMs samples. Genome-wide DNA methylation analysis has previously been applied to smaller MOC datasets within the broader context of ovarian carcinomas, demonstrating that MOCs represent a distinct ovarian carcinoma subclass.3 Our findings provide further evidence for the evolutionary relationship between mBOTs and MOCs, demonstrating a close spatial relationship in the DNA methylation t-SNE analysis between mBOTs and MOCs and important similarities in the DNA methylation relatedness network. Furthermore, we observe two epigenetic patterns of mBOT: one type closer to normal tissue (mBOTnormal_like) and another type closer to MOC (mBOTMOC_like), both exhibiting specific spatial relationships and epigenomic relatedness with these tissue types. We made a similar observation in our previous study, which analyzed the epigenetic relationship of MCN to MOT and noted that some mBOT samples tend to share DNA methylation similarities with MCN high-grade/invasive tumors and MOCs, while others tend to resemble MCN low-grade tumors.7 This suggests that the mBOTMOC_like could represent a more advanced intraepithelial stage. However, as our study does not include the full spectrum of precursors (benign cystadenomas, Brenner-tumor-associated lesions, and teratoma-associated lesions), these results should be interpreted as supportive of a stepwise transformation model rather than a definitive conclusion. To confirm this hypothesis, a larger cohort with longer follow-up is needed. Such studies will be essential to determine if these molecular signatures have clinical utility, such as avoiding overly conservative surgery like unilateral salpingo-oophorectomy,39,40 potentially adding therapy for the mBOTMOC_like subgroup to prevent recurrence39,40 or increasing surveillance measures. Similarly, the DNA methylation analysis characterized two distinct types of MOC, which showed differing survival associations in the internal cohort, with MOCK2 exhibiting more aggressive outcomes. Except for the T-stage, no other parameter demonstrated prognostic value, further supporting the notion that grading lacks prognostic power for MOC.21 The clinical course of MOC is highly variable; some cases exhibit aggressive progression, while most remain indolent.41 Consequently, there is a need for molecular markers to better understand patient risk. If validated in larger, prospective studies, such molecular stratification could help identify the patients who may require more intensive treatment, such as the addition of targeted therapies, including the evaluation of emerging therapies like KRAS-inhibitors.42 Furthermore, our characterization of MOC identifies distinct molecular features that may warrant further investigation into personalized treatment strategies. For instance, the higher levels of T-lymphocytes observed in MOCK1 could suggest a rationale for exploring immune checkpoint inhibition. Similarly, MOCK1 samples harboring ERBB2 gain might represent a population for investigating HER2 inhibitors/antibody-drug conjugates.42 On the other hand, MOCK2 cases, with more frequent CCNE1 gains, suggest these tumors could be evaluated for sensitivity to cyclin E1 pathway inhibitors.43,44 However, given the observational nature of our study, these associations remain speculative and require functional validation to determine their relevance. The second aspect of MOC addressed in this study is its differential diagnosis from EOM. We have employed DNA methylation profiles, combined with machine learning, to tackle this challenge. We developed a three-step DNA methylation classifier: in the first step, it identifies MOC; in the second step, it detects EOM and filters out other lesions through an anomaly detection layer; and in the final step, it determines the origin of EOM. Currently, the final step can differentiate between STAD, COAD, and other pMCs. We plan to expand this step to incorporate additional origins of EOM. The clinical relevance of our classifier is significant, as the diagnosis of an MOT often requires a labor-intensive process to exclude the possibility of an occult primary mucinous carcinoma from other sites.12,45 The classifier is available for other clinicians online at https://app.epiresolve.com/. Our classifier achieves an accuracy of 95.5% (internal validation) and 91.11% (external validation) in detecting MOC and EOM. These results are generally superior to previous IHC and molecular-based classifiers. Previous studies that focused on morphology and IHC for differentiating MOCs from EOMs report accuracy ranges from 71.3% to 96.1%,22,46 with only one study surpassing the accuracy of our classifier.47 In that study, the diagnosis was based on tumor size, bilaterality, and the presence of surface nodules, although its high accuracy was not confirmed by other groups. The mutations in MOC and EOM are similar, with only APC and SMAD4 mutations potentially contributing to the diagnostic algorithm, both suggesting EOM.3 Transcriptomic studies focusing on the differential diagnosis between MOCs and EOMs are limited and largely unsuccessful. For example, a 19-gene panel failed to distinguish MOCs from EOMs.48 An alternative approach to improve the diagnosis of MOC versus EOM involves digital pathology. For instance, Wu et al. developed an artificial intelligence (AI) tool to differentiate ovarian carcinoma subtypes using cytological images, achieving an accuracy of 78.2%.49 Furthermore, we explored the potential of predicting the origin of the EOMs using DNA methylation data. The classification accuracy was lower compared to steps 1 and 2: 78.95% (internal cohort) and 73.91% (external cohort). Given the current limitations in the diversity of the other MC category, which currently lacks rare mimics such as gastric-type endocervical adenocarcinoma and specific pancreatobiliary neoplasms, we consider this analysis preliminary and not ready for clinical implementation. By contrast, the primary objective of this diagnostic framework is the robust differentiation of MOC from EOM, where the classifier achieves high diagnostic performance. We explored moving the diagnosis process into the intraoperative setting. Nanopore technology is advancing as a reliable tool for performing intraoperative DNA methylation classification of cancers.38,50 In this context, our initial results were suboptimal. The preliminary data indicated that the issue was not the lower number of CpG probes obtained by Nanopore sequencing, but rather the tumor purity of FF samples, which were collected blindly during surgery. Therefore, our findings suggest that intraoperative DNA methylation diagnosis remains an open problem and will require strategies that ensure better sampling. In conclusion, using genome-wide DNA methylation profiling, we were able to subclassify mBOTs and MOCs, identifying two MOC subtypes with prognostic implications. Furthermore, these two MOC subtypes exhibit distinct molecular characteristics and tumor microenvironment changes, which could have clinical relevance. Finally, we developed a DNA methylation classifier that can differentiate MOC from EOM, potentially enhancing and accelerating the diagnostic workflow for ovarian mucinous lesions. Limitations of the study Moreover, it is important to acknowledge several limitations. First, we did not validate the prognostically relevant observations of the MOC subclassification because of the limited follow-up duration of these recently diagnosed samples. Second, our classifier can pinpoint only two origins of EOM, and rare metastases to the ovary are all grouped together as “other.” In addition, the classifier has not been tested against rare but diagnostically critical mimics such as gastric-type endocervical adenocarcinoma. Consequently, step 3 should be regarded as a proof-of-concept analysis. The main purpose of our classifier is to differentiate MOTs from EOMs, and addressing the origin of adenocarcinoma metastases will represent a future work. Third, our study is primarily based on DNA methylation data. It would be valuable to incorporate additional omics technologies to further refine the subclassification of MOCs. Fourth, we did not investigate the full spectrum of MOC precursors, such as Brenner tumors, teratomas, and mucinous cystadenomas. This limitation arises partly because DNA methylation analysis requires large tissue samples, and alternative methods, such as spatially resolved proteomics,51 are needed for such analyses. Future work could also integrate the ovarian surface epithelium and Walthard cell nests for a more comprehensive analysis. Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Mihnea P. Dragomir (mihnea.dragomir@charite.de). Materials availability This study did not generate new unique reagents. Data and code availability •DNA methylation data generated in this study are deposited in the Gene Expression Omnibus (GEO) under accession number GEO: GSE310580.•The machine learning classifier code and analysis scripts have been deposited at GitHub (https://github.com/teocalina/moc-eom-classifier-model) and are publicly available at https://doi.org/10.5281/zenodo.20408260.•Any additional data reported in this work is available from the lead contact upon reasonable request. •DNA methylation data generated in this study are deposited in the Gene Expression Omnibus (GEO) under accession number GEO: GSE310580. DNA methylation data generated in this study are deposited in the Gene Expression Omnibus (GEO) under accession number GEO: GSE310580. •The machine learning classifier code and analysis scripts have been deposited at GitHub (https://github.com/teocalina/moc-eom-classifier-model) and are publicly available at https://doi.org/10.5281/zenodo.20408260. The machine learning classifier code and analysis scripts have been deposited at GitHub (https://github.com/teocalina/moc-eom-classifier-model) and are publicly available at https://doi.org/10.5281/zenodo.20408260. •Any additional data reported in this work is available from the lead contact upon reasonable request. Any additional data reported in this work is available from the lead contact upon reasonable request.