Methylomic Analysis of Ovarian Cancers Identifies Tumor-Specific Alterations Readily Detectable in Early Precursor Lesions.

The overarching goal of the study was to identify loci exhibiting highly-specific and sensitive differential methylation in HGSOC tissue that might be used as methylation biomarkers for detecting early or low-volume stage HGSOC, when intervention is most effective. We employed a supervised, three-stage process involving two different methylation assays and three patient cohorts ( Figure 1 ). In the first, discovery, stage, MethylationEPIC analysis was performed on 96 malignant and healthy gynecologic tissues. In the second stage, we used a combination of bioinformatics criteria to select candidate loci with a high-probability of robust differential methylation between healthy and cancerous tissues. Each of the top loci were then analytically validated for clinical sensitivity and specificity using sensitive melt after real-time methylation-specific PCR (SMART-MSP) ( 29 ) analysis in a 90-sample cohort derived primarily from the initial 96 samples. In the third and final stage, we sought to determine whether these loci exhibit differential hypermethylation at the earliest stages of HGSOC by employing the validated SMART-MSP assays to evaluate DNA derived from laser captured microdissected epithelia of 9 STIC precursor lesions and corresponding adjacent normal epithelia in comparison to 12 samples of fallopian tube epithelia derived from cancer-free women. We first sought to identify differences in DNA methylation between HGSOC in comparison with other gynecologic cancers and suitable control tissues by performing MethylationEPIC BeadChip analysis of 23 HGSOC, 37 non-HGSOC malignant and 36 healthy fresh-frozen gynecologic tissues ( Supplementary Tables S1-S2 ). Overall, the 37 non-HGSOC tumor samples comprised 23 endometrioid endometrial (EEC), 10 serous endometrial (ESC) and 4 other poorly-differentiated [possibly HGSOC], unclassified tumors. The normal-appearing mucosal epithelial tissues were isolated from respective tissues of 36 different women that underwent surgery for risk reduction or other benign conditions and comprised 11 fallopian, 13 endometrial and 12 endocervical mucosal epithelial samples. Initial unsupervised analyses of the EPIC BeadChip data ( Figure 2 ) indicated that the methylomes of HGSOC as well as other malignant gynecologic tissues were significantly altered and readily distinguishable from all three of the normal mucosa sample types. Unsupervised hierarchical clustering of the 5000 most variable CpG-sites ( Figure 2a ) demonstrated that the cancerous sample types were distinctly separated from the majority of healthy mucosae and, to a lesser extent, each other. The data also demonstrate that most of the differentially methylated regions (DMRs) lie outside CpG-island (CGI) regions, which is consistent with previous observations that the preponderance of cancer-specific DMRs lie in shores, shelves and the open sea ( 32 ). Assessment of the data by multidimensional scaling (MDS) ( Figure 2b ) further indicated very high similarity among the methylomes of normal mucosae derived from endometrial, fallopian tube and cervical tissues. Of the four cancer subtypes, HGSOC exhibited the most homogeneity in methylation between samples ( Supplementary Figure S1 ) and was clearly differentiable from those of cancers classified as endometrioid endometrial but partially overlapped with those of other serous and high-grade tumor types. Sample clustering remained largely unchanged if the unsupervised analysis was performed using only probes located within CpG-islands, shores and shelves ( Supplementary Figure S2 ), indicating that cancerous tissues can potentially be identified by employing locus-specific assay methods targeting these CpG-dense regions. We next sought to identify loci that were best suited to the development of more sensitive and clinically-attractive locus-specific methylation assays. However, even with the expanded genomic coverage of the EPIC BeadChip, fewer than 4% of CpG sites are probed genome-wide, with an average of 6 probes per CGI. This limited resolution of the BeadChip array is not congruent with identification of methylation biomarkers that can be readily-assessed by locus-specific methylation assays, such as MSP, which are generally designed around short (< 150 bp) CpG-rich loci, as opposed to isolated CpG sites ( 3 ). In order to address this discrepancy, we employed bioinformatic algorithms to smooth methylation levels across adjacent CpG sites ( 25 ) as a means of identifying candidate DMRs that were best suited for translation into locus-specific methylation biomarker assays ( Figure 3 ). We approached this by selecting sites that lie within 1500 bp of the transcription start site (TSS), in more CpG-dense regions (islands, shores and shelves), with high-specificity methylation (β < 0.2 in healthy mucosae), and that demonstrated robust differential methylation (family-wise error rate < 0.1). Lastly, we selected for DMRs that contained at least two differentially-methylated probes (DM-probes) within the same region. These selective criteria provided increased confidence that the selected DM-probes were likely to represent the methylation state of the DMR as a whole and consequently improved the likelihood of successfully employing locus-specific assessment methods to achieve clinically-sensitive assays. In total, 294 probes, or 0.4% of all “high-specificity” probes, located in 91 DMRs were identified by the selection criteria ( Supplementary Table S3 ). Evaluation of the performance of these 91 “high-confidence” DMRs using the EPIC BeadChip data demonstrate that the collective methylation status of these CpG-sites retains excellent discrimination between the control tissue samples and gynecologic cancers ( Supplementary Figure S3 ). The imposed selection criteria, particularly for high-specificity, necessarily identified DMRs with increased homogeneous, low levels of methylation in all healthy mucosae, that would thereby facilitate the development of clinically-specific methylation assays. A plot of intra-sample-type homogeneity ( Supplementary Figure S4 ) shows that this comes at a cost of generally decreased homogeneity within the gynecologic cancer subtypes (as compared to the DMRs identified by unsupervised analysis) that might otherwise be expected to decrease downstream clinical sensitivity. Nonetheless, the majority of the high-confidence DMRs demonstrate exceptional sensitivity and specificity for discrimination of HGSOC when evaluated by 4 iterations of 3-fold cross-validation ( Supplementary Table S3 ). For further validation of the identified high-confidence DMRs, we assessed the performance of the corresponding BeadChip probes for distinguishing HGSOC in independent datasets obtained BeadArray analyses of ovarian tumors and fallopian tubes from HGSOC-positive and cancer-free women, respectively. We first assessed this performance in data from TCGA ( http://cancergenome.nih.gov/ ), consisting of 582 HGSOC of various stages and 12 normal fallopian samples. As previously mentioned, TCGA data were obtained using Illumina’s first-generation, HumanMethylation27 BeadChip, and thus validation was limited to the 16 probes (in 15 DMRs) in common between the two platforms. Analysis of TCGA data obtained with these 16 probes ( Supplementary Figure S5 and Supplementary Table S4 ) yielded an average area under the ROC curve of 0.70 for discrimination of HGSOC (compared with 0.82 in the EPIC dataset in the present study), significantly outperforming randomly selected probes (average AUC = 0.454). When combined into a single methylated signature, these sites yielded a composite AUC of 0.971. We next sought to further extend this validation by using BeadArray data obtained with the Illumina Infinium 450K BeadChip platform, which comprises the 279 of the 294 (95%) high-confidence differentially-methylated probes identified in this study. To achieve this, we assessed the performance of the probes using publicly-available Infinium 450K analyses of HGSOC tumors (GEO accession GSE72021 , described in ( 26 )), in comparison with data from cancer-free fallopian tubes (GEO accession GSE74845 , described in ( 9 )). ROC analyses of these data also confirmed these probes as highly differentially-methylated between various-stage HGSOC tumors and cancer-free fallopian tubes, yielding an average DMR AUC of 0.69, compared with 0.81 in the EPIC dataset of the present study ( Supplementary Table S3 ). While comparable, the discrepancy in performance of the high-confidence probes can likely be accounted for by significant differences in the purity of tumor samples of the cohorts used in the respective studies. Specifically, estimates obtained using InfiniumPurify ( 28 ) indicate that the median and mean purity of the tumor samples in the present (EPIC) study were 0.82 and 0.87, respectively, as compared to 0.29 and 0.11 for the tumor sample cohort of the external datasets. We next aimed to develop corresponding locus-specific methylation assays in order to validate the accuracy of the EPIC BeadChip results, verify the existence of differential methylation throughout the identified high-confidence DMRs and assess the potential use of the identified loci as methylation biomarkers for HGSOC. To achieve this, SMART-MSP ( 29 ) was employed because it provides excellent specificity (~0.1% epiallelic fractions, EAF) and somewhat greater flexibility for detecting heterogeneously-methylated epialleles when compared to MethyLight, which incorporates an additional stringency probe ( 33 ). For practical purposes, we focused our development of SMART-MSP assays on the top performing DMRs that yielded AUC’s ≥ 0.9 for discrimination of HGSOC from healthy mucosal tissues. Of these 23 loci, 11 were subsequently excluded due to suboptimal CpG density for MSP assay design (< 5 CpG-sites in a single 100 bp stretch between DM probes) or insufficient differential methylation (average Δβ < 0.2). MSP primers were designed and optimized for the remaining 12 loci and their respective analytical sensitivities and specificities were evaluated using bisulfite-treated (BST) unmethylated and methylated control genomic DNA. We set minimum QC criteria for adequate assay performance at an analytical sensitivity of 5 methylated genomic copies (i.e., 2 genomic copies per genome equivalent) and an [extrapolated] analytical specificity of 1 genomic copy in 500 BST-unmethylated genomic copies. In total, SMART-MSP assays for eight of the top-performing loci met our QC criteria: c17orf64 , c6orf174 , IRX2 , TUBB6 , PTPRN , OTX2 , LOC200726 and NEUROD1 ( Supplementary Table S5 ). When analyzed collectively, data from the EPIC BeadChip analysis indicated that the methylation status of these loci provide both very high sensitivity and specificity for the detection of HGSOC ( Figure 4 ), achieving an ROC curve AUC of 1.0 in that data set. Interestingly, the data further indicate that these DMRs would also distinguish healthy mucosae from EEC, USC and unclassified tumors (composite AUC = 0.98) with nearly the same proficiency as HGSOC; a finding once again in accord with previous studies showing that hypermethylation often occurs in the same regions in gynecologic cancers of different types ( 13 , 14 ). We next employed the eight SMART-MSP assays in 90 samples derived primarily from the initial cohort to analytically validate the corresponding DMRs ( Supplementary Tables S1 and S6 ) and to assess their performance and potential as biomarkers primarily for HGSOC and secondarily for other gynecologic cancers. Overall, the SMART-MSP assays were able to replicate the observed differences in methylation between healthy and cancerous tissue in the EPIC BeadChip analysis and validated the DMR selection paradigm ( Figure 5 and Supplementary Table S7 ). ROC analyses of the individual SMART-MSP assays ( Figure 6 ) show that, of the eight loci, only one ( PTPRN ) failed to achieve an AUC greater than 0.8 for discrimination of HGSOC. Furthermore, an ROC curve based on the unweighted average of the top three performing SMART-MSP assays, c17orf64 , IRX2 , and TUBB6 perfectly discriminated HGSOC from control tissues with 100% sensitivity and specificity in our sample cohort. The same three loci also achieved very high discrimination (AUC 0.991) of the other gynecologic cancer types from healthy mucosae, as well. The fallopian tube origin of HGSOC has previously been evidenced primarily through correlation between somatic TP53 mutations found within STIC lesions with those of HGSOC tumors in the same patient ( 15 – 17 ). As many epigenetic events have been shown to occur very early in, or even prior to carcinogenesis, we hypothesized that the highly prevalent methylation events observed in HGSOC tumor samples might also be present in precursor STIC lesions. To test this hypothesis, we performed LCM on fallopian tube epithelia from nine formalin-fixed paraffin-embedded (FFPE) STIC lesions derived from eight patients who had undergone bilateral salpingo-oophorectomy for various medical reasons ( Figure 7a and Supplementary Table S8 ). As a control, we also performed LCM on morphologically-normal, p53-immunohistochemically-negative fallopian tube epithelia from FFPE tissue derived from 12 healthy, cancer-negative women that had also undergone bilateral salpingo-oophorectomy. All samples were also sequenced to assess for mutations within TP53 , which were only detected within epithelia derived from STIC lesions ( Supplementary Table S8 ). Analysis using the SMART-MSP panel (excluding PTPRN due to poor performance as described above) showed that DNA hypermethylation was detectable in all nine STIC samples and, with the exception of TUBB6 , was completely absent in fallopian tube epithelia from healthy women ( Figure 7b and Supplementary Table S8 ). Interestingly, TUBB6 methylation was detectable in most of the samples tested, however at much lower estimated EAFs, on average, in the normal fallopian epithelia controls than in the STIC lesions. This could be due the presence of a small subpopulation of epithelia in which the TUBB6 promoter is normally methylated or due to a limitation in the analytical specificity of the TUBB6 SMART-MSP assay itself. We set a PMR threshold to account for the background TUBB6 methylation signal and compared the presence of positive detectable hypermethylation in each of the loci in the LCM epithelial samples ( Figure 7b , solid bars). Overall, all of the tested LCM-epithelia derived from STIC lesions exhibited detectable hypermethylation in at least two, and up to five of the seven loci. The presence of positive methylation was also highly specific to women with fallopian tube STIC lesions, with no detectable hypermethylation in any of the control samples without cancer or STIC. In addition to assessing the clinical specificity of hypermethylation with fallopian tube epithelia from healthy control subjects, we also assessed the phenotypic specificity of hypermethylation by analyzing LCM-extracted DNA from morphologically-normal epithelia located distally to (away from) each STIC lesion within the same fallopian tube ( Figure 7b , dotted bars). SMART-MSP analyses revealed that in five of eight of the women, hypermethylation was also present within one to three loci from DNA extracted from these normal fallopian tube epithelia. Importantly, the loci in which methylation was detected in each of these epithelial samples was also present (and in a greater extent) in the corresponding STIC lesions from the same fallopian tube, potentially indicating an epigenetic field effect and evolutionary progression of cancer-primed epithelia leading to formation of the STIC lesions themselves ( Supplementary Figure S6 ). This observation corroborates and further extends the previous report of precancerous epigenetic reprogramming in fallopian tube fimbriae of BRCA mutation carriers ( 9 ). In particular, our results indicate that the accumulation of aberrant methylation may be a general phenomenon that occurs prior to and throughout the formation of precursor STIC lesions. Furthermore, it is notable that while some of the methylation alterations in certain loci ( IRX2 , OTX2 and NEUROD1 ) are all but exclusively observed within the STIC region, others ( c17orf64 , c6orf174, LOC200726 and TUBB6 ) extend to p53-negative epithelia that lie outside the immediate lesions. That methylation within the promoters of IRX2 and NEUROD1 was not observed in epithelia outside the STIC-lesion may indicate that, in these samples, methylation of these loci likely occurred concomitantly with or after acquiring TP53 mutations and ensuing hyperplasia during the development of the respective lesion. On the other hand, the observation of methylation within the promoters of c17orf64 , c6orf174 , LOC200726 and TUBB6 in normal epithelia beyond the STIC lesion is consistent with a cancerizing epigenetic field effect that has been observed directly in other cancer types ( 34 , 35 ) and as an “extended field effect” within the endometrium of ovarian cancer patients ( 36 – 38 ), although further mechanistic studies will be required to investigate this hypothesis.

This study was performed after approval by institutional review board (IRB) and conducted in accordance with the U.S. Common Rule. All tissues were collected or retrieved from JHU GYN tissue bank with written consent given by (all) the participant subjects. The discovery cohort ( Supplementary Tables S1 and S2 ) comprised thirty-six gynecologic mucosae of fallopian tube (11 cases), endometrium (13 cases) and endocervix (12 cases) that were obtained as control groups from women who received surgery for benign gynecologic disease at The Johns Hopkins University from 2016 to 2017. Sixty malignant tissues were collected from women who had surgery for ovarian or uterine cancers from the Johns Hopkins Hospital and the Shimane University Hospital, Japan. The indications for benign gynecologic disease surgeries included symptomatic fibroids, endometriosis, adnexal mass or cysts, endometrial polyps, severe bleeding/pelvic pain desiring surgical intervention, and risk reducing surgery due to family history of cancer. Among 60 cancer cases, 23 endometrioid endometrial carcinoma (EEC), 10 uterine serous carcinoma (USC), 23 high-grade serous ovarian carcinoma (HGSOC), and 4 high grade tumors named “unclassified cohort” (including carcinosarcoma of uterine and ovary, and poorly differentiated ovarian tumor) were selected. The epithelial cells from normal mucosa were enriched by carefully scrapping the mucosal surface on fresh tissues. To validate the Mullerian duct origin of the mucosa specimens, we performed Western blot with a PAX8 antibody and found that more than 95% of the specimens express PAX8 at comparable level to an endometrium epithelial cell line, hEM3.The histopathology from individual tissues was examined by two pathologists (SFL and IMS) based on H&E stained slides prepared from a fragment of the same tissues to be analyzed. For MSP analysis, DNA from nine malignancies were depleted after Methylation EPIC BeadChip analysis and were replaced with DNA from specimens of the same tumor type (as described in Supplementary Tables S1 and S6 ). For MSP analysis, DNA from nine malignancies were depleted after Methylation EPIC BeadChip analysis and [partially] replaced with DNA from other tumor specimens (as described in Supplementary Tables S1 and S6 ). For all samples, genomic DNA was extracted and purified by QIAamp FFPE DNA tissue kit (Qiagen, Hilden, Germany). Sample quality assessment and microarray analysis were conducted at The Sidney Kimmel Cancer Center Microarray Core Facility at Johns Hopkins University, supported by NIH grant P30 CA006973 entitled Regional Oncology Research Center. Genomic DNA quality was assessed by low concentration agarose gel (0.6%) electrophoresis and fluorescent spectrometry with PicoGreen DNA Kit (Life Technologies). DNA bisulfite conversion was carried out using EZ DNA Methylation Kit (Zymo Research) by following manufacturer’s manual with modifications for Illumina Infinium Methylation Assay. Briefly, 400 ng of genomic DNA was first mixed with 5 μl of M-Dilution Buffer and incubate at 37°C for 15 minutes and then mixed with 100 μl of CT Conversion Reagent prepared as instructed in the kit’s manual. Mixtures were incubated in a thermocycler with 16 thermal cycles at 95°C for 30 seconds and 50°C for one hour. Bisulfite-converted DNA samples were loaded onto 96-column plates provided in the kit for desulphonation and purification. Concentration of eluted DNA was measured using Nanodrop-1000 spectrometer. Bisulfite-treated (BST-) DNA was analyzed using Illumina’s Infinium Human MethylationEPIC BeadChip Kit (WG-317–1002) by following manufacturer’s manual. Briefly, 4 μl of bisulfite-converted DNA was added to a 0.8 ml 96-well storage plate (Thermo-Fisher Scientific), denatured in 0.014N sodium hydroxide, neutralized and amplified with kit-provided reagents and buffer at 37°C for 20–24 hours. Samples were fragmented using kit-provided reagents and buffer at 37°C for one hour and precipitated by adding 2-propanol. Re-suspended samples were denatured in a 96-well plate heat block at 95°C for 20 minutes. 26 μl of each sample was loaded onto an 8-sample chip and the chips were assembled into hybridization chamber as instructed in the manual. After incubation at 48°C for 16–20 hours, chips were briefly washed and then assembled and placed in a fluid flow-through station for primer-extension and staining procedures. Polymer-coated chips were image-processed in Illumina’s iScan scanner. All data analysis for DNA methylation marker discovery was carried out using the R statistical software suite ( 19 ). DNA methylation data was processed from Illumna EpicArray iDat files using the functional normalization algorithm ( 20 ), as implemented in the minfi package( 21 ) from the Bioconductor bioinformatics software project ( 22 ). The ilm10b2.hg19 package ( 23 ) from Bioconductor provided up-to-date hg19 annotations for the probe sequences represented on the array, which were remapped to hg38 based on the UCSC liftOver mapping ( 24 ). CpGs were filtered to include island, shore and shelf sites within 1500 bases of the transcription start site for the closest gene, before applying the bumphunter algorithm ( 25 ) with a bootstrap null, cutoff of 0.2 and 100 iterations, to identify regions showing significant, differential methylation between HGSC tumors and normal ovarian samples. The regions flanking each putative DMR were examined to ensure that the differential methylation call would be robust to the selection of specific probes, after which the remaining candidates were ranked according to area under ROC curve (AUC) calculated on HGSC tumors and normal ovarian samples. DNA methylation levels in other tissue types including fallopian tube and endometrium were not formally incorporated into selection of regions but were evaluated to ensure that unusual methylation patterns in adjacent tissues would not confound marker calls. Level 3 ovarian methylation data from the TCGA was downloaded from the Broad Institute ( https://gdac.broadinstitute.org/ ). Additional, publicly-available DNA methylation profiles of gynecological tumors and healthy fallopian tube controls were obtained from the NCBIs Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/ ). Processed Illumina Human 450K data (GEO accession GSE72021 , described in ( 26 )) from 221 tumor samples (171 serous, 18 endometrioid, 14 clear cell, 9 mucinous and 9 other histological cancer subtypes) was downloaded, along with accompanying clinical annotations, using the GEOquery package from Bioconductor ( 27 ). Clinical annotations for a related set (GEO accession GSE74845 , described in ( 9 )) of 215 normal Fallopian tube samples, hybridized to the same platform, was downloaded in the same way, although we were unsuccessful in obtaining methylation profiles using the GEOquery package, and downloaded them directly from the GEO website as a csv file. We excluded BRCA1 and 2 carriers from our analysis of the Fallopian tubes, leaving 60 non-carrier controls. The InfiniumPurify R package ( https://cran.r-project.org/web/packages/InfiniumPurify/index.html ) was used to estimate tumor purity levels based on DNA Methylation Profiles ( 28 ). Tumor purity estimates for GSE72021 tumors were calculated with respect to the 60 non-BRCA carrier controls from the GSE74845 study. EPIC array data from JHU samples were converted to 450K methylation profiles using the convertArray function in the minfi package, before calculating purity estimates. DNA sequences including 100 bp upstream and downstream from the top performing (highest AUC) differentially methylated loci were obtained from the UCSC genome browser ( 24 ). Primers for Sensitive Melting Analysis after Real-Time Methylation-Specific PCR (SMART-MSP) assays ( 29 ) were designed according to MSP guidelines and obtained from Integrated DNA Technologies. Initial assessment and subsequent assay optimization was performed using bisulfite-converted CpG-Methylated HeLa Genomic DNA (New England BioLabs) and Human Male Genomic DNA (Promega) as an unmethylated control. The performance of each SMART-MSP assay was evaluated using standard curves consisting of 500 total genomic copies of varying ratios of over an epiallelic fraction (EAF) range of 0% to 100%. The minimum specificity/sensitivity for successful assay development was set at an EAF of 0.2% (1 in 500) as determined using linear regression of the standard curve to estimate the lowest EAF that could be detected over the unmethylated control. All primer sequences and corresponding annealing temperatures of the successfully-developed assays are listed in Supplementary Table S5 . Bisulfite conversion of extracted tissue-DNA was performed using the EZ-96 DNA Methylation-Lightning™ Kit (Deep-Well; Zymo Research) according to manufacturer’s instructions. BST-DNA yield for each sample was quantified using a control β-Actin assay with synthetic DNA targets (Integrated DNA Technologies) used as standards. Samples were then standardized by dilution to yield final DNA concentrations of 1000 β-Actin genomic copies (500 genomic equivalents, ~2.5 ng BST-DNA) per 2 μl. The 10X master mix for all SMART-MSP assays consisted of 16.6 mM (NH 4 ) 2 SO 4 , 67 mM Tris pH 8.8, 6.7 mM MgCl 2 and 100 mM β-mercaptoethanol. SMART-MSP assays were performed in 25 μl PCR mixtures consisting of 1X master mix, 2 μl (~2.5ng) of BST-DNA, 1X Evagreen dye (Biotium), 300 nM forward and reverse primers, 10 nM of fluorescein reference dye (Life Technologies), 200 μM (each) dNTPs (Life Technologies), and 1 Unit of Platinum Taq® DNA Polymerase (Invitrogen). Amplification reactions were performed in duplicate using 96 well-plates using a CFX96 Touch Real-time PCR Detection System (Bio-Rad). Thermocycling conditions were: 5 minutes at 95°C, followed by 50 cycles of (95°C for 30 seconds, T A for 30 seconds and 30 seconds at 72°C, where T A is the primer annealing temperature) followed by a high-resolution melt (HRM) step using a temperature range of 65°C-90°C at 0.2°C increments and a 10-second hold time before each measurement. SMART-MSP results were analyzed using the CFX Manager v3.1 (Bio-Rad) using regression to obtain the mean Cq for each respective sample. A percent methylated reference (PMR) value was calculated for each sample-assay using standard curves for each respective SMART-assay with the formula: [# methylated copies marker] / [# copies β-Actin = 1000] × 100%. PMR values were then used to generate receiver-operating characteristic (ROC) analysis and AUCs (area under the curve) for each SMART-MSP assay. These were calculated using R statistical software suite ( 19 ) with linear regression models. Nine pairs of STIC and normal tubal epithelium formalin-fixed and paraffin-embedded (FFPE) samples were retrieved from Johns Hopkins Hospital under institutional review boards approval and reviewed by pathologists. The diagnostic criteria for STIC was based on morphological and immunohistochemical criteria previously described ( 30 , 31 ). The purity of each sample was estimated according to the proportion of STIC epithelial (p53+-staining) cells among total dissected epithelial cells including STIC and adjacent normal. For some samples, we intentionally dissected the STIC lesion with adjacent normal tubal epithelium in order to obtain enough DNA quantity for both methylomic and sequencing analysis. Another 12 frozen fallopian tube tissue samples were also retrieved and processed as FFPE sample and confirmed as normal tubal epithelium by H&E and p53 immunohistochemical stains. Details for each patient and sample are listed in Supplementary Table S8 . All STIC and normal tubal samples were sectioned on the MembraneSlide 1.0 PEN slides (Zeiss) for laser capture micro-dissection (Leica LMD7000) of STIC and its adjacent normal epithelium. Genomic DNA was purified by using the QIAamp FFPE DNA tissue kit (Qiagen, Hilden, Germany). Genomic DNA (10–25 ng) samples from selected cases were used to prepare libraries by Accel-Amplicon 56G oncology panel v2 (Swift Biosciences, Ann Arbor, NI, USA), which was designed for hotspot coverage of 56 oncology-related genes by 263 amplicons. The samples were prepared according to the manufacturer’s protocol and quantified by NEBNext Library Quant Kit for Illumina (New England Biolabs). The libraries were sequenced on MiSeq instrument (Illumina) by using MiSeq Reagent kit v2 Micro (300 cycles). Results for each sample are shown in Supplementary Table S8 . Bisulfite conversion of DNA extracted from fallopian and control tissues was performed using the EZ DNA Methylation-Lightning™ Kit (Zymo Research) according to manufacturer’s instructions. All samples were standardized with DNA Elution Buffer to 40 μl total volume. The BST-DNA yield for each sample was quantified using the control β-Actin assay described above. Final BST-DNA concentrations ranged from 0.343–0.878 ng/μl, with an average of 0.567 ± 0.175 ng/μl. SMART-MSP was performed in duplicate, as described above, using 2 × 1 μl of BST-DNA from each sample. Results were analyzed using the CFX Manager v3.1 (Bio-Rad) using regression to obtain the mean Cq of each SMART-MSP assay for each respective sample. Positive wells were subject to secondary confirmation, as previously described ( 29 ), by ensuring that the product melt temperature fell within the Tm range experimentally-determined by SMART-MSP of BST- normal human male DNA (Promega) and BST- CpG Methylated HeLa Genomic DNA (New England Biolabs). Results of the experimentally-determined positive melt temperature (Tm) ranges for each SMART-MSP assay are provided in Supplementary Table S5 . Percent methylation ratios (PMRs) were calculated using standard curves for each marker, according to the formula: [# methylated copies marker] / [# copies β-Actin] × 100%. Except for TUBB6 , samples were considered positive by SMART-MSP if HRM analysis of both replicates demonstrated a Tm that fell within the predetermined range and an average methylated copy number ≥ 1 per replicate. For TUBB6 , a PMR threshold of [average healthy epithelia PMR] + 2.5*σ = 4.61% was used to assess positivity. A summary of the results, as well as the patient and sample characteristics are listed in Supplementary Table S8 .

The present work identifies a novel panel of alterations in DNA methylation that commonly occur in ovarian cancers and can be used to distinguish HGSOC from normal or benign gynecologic tissues with high specificity. Furthermore, we also showed that these alterations occur prevalently in precursor STIC lesions and might be used for detection of the disease at its earliest stages, possibly even during the seven-year window that is on average required for the lesions to progress and seed into the ovary ( 16 ). While the vast majority of STIC lesions occur in the fimbrial end of the fallopian tube, we also found that, in many samples, hypermethylation in these loci extends beyond the STIC lesions themselves, occurring in morphologically-normal fallopian tube epithelia outside the immediate lesions. While these observations support the previous report of epigenetic reprogramming in morphologically-normal fallopian tubes of women at high-risk for developing HGSOC ( 9 ), they also indicate that cancer-specific alterations in methylation may not be confined to the fimbrial end of the fallopian tube and that further epigenetic evolution likely continues to occur during the development of precursors lesions. However, additional mechanistic studies will be necessary to determine the extent to which these epigenetic alterations actively contribute to their emergence. Relatedly, our results indicate that careful attention should be given when comparing samples from different regions of the fallopian tube. Specifically, suitable controls are required in order to mitigate differences in methylation due to cellular composition between samples that might otherwise confound biological interpretation. With the exception of TUBB6 , none of the validated high-confidence DMRs in our study ( c17orf64 , IRX2 , c6orf174 , OTX2 , LOC200726, and NEUROD1 ) have previously been reported as epigenetic biomarkers for HGSOC. Several of these loci have, however, been reported as methylation biomarkers for other cancer types, including breast ( NEUROD1 , IRX2 , c6orf174 ) ( 39 – 41 ), lung ( IRX2 ) ( 42 ), and colorectal ( NEUROD1 ) ( 43 ) cancers, among others. Hypermethylation of TUBB6 likely contributes to reduced expression of the TUBB6 protein, as has been observed in HGSOC ( 44 ) and several other cancer types( 45 ). The high specificity and sensitivity of the validated DMRs, as well as redundancy in their ability to detect malignant and premalignant gynecologic tissues, underscores their individual and collective potential as high-value biomarkers for HGSOC. And while our cohort was ostensibly designed to identify HGSOC-specific DMRs, the data also demonstrate that these DMRs are exceptional biomarkers for the other gynecologic malignancies included in our study. This is perhaps not unexpected as previous studies have documented the existence of altered methylation patterns shared among various gynecologic cancers ( 13 , 14 ). Nonetheless, the observation of methylation biomarkers in common between the cancer types, coupled with the common occurrence of synchronous primary gynecologic cancers and common etiologies ( 46 ), has potential implications for use of a single biomarker panel for detection of pan-gynecologic cancers. The employment of the latest-generation, EPIC BeadChip array provided substantially more coverage of CpG-sites throughout the genome than in prior studies, thereby allowing identification and analysis of differentially-methylated CpG sites that had not been analyzed or revealed. This is further evidenced by the observation that only 15 of 91 “high-confidence,” and 2 of 23 “top-performing,” DMRs identified in our analysis were probed by the HumanMethylation27 BeadChip used in the 582 HGSOC cases included in TCGA. Nonetheless, analysis of the TCGA dataset showed general congruence with data obtained from the same probes using the EPIC BeadChip. The increased genomic coverage afforded by the EPIC BeadChip also provided improved resolution, facilitating bioinformatic identification of high-confidence DMRs containing multiple differentially-methylated probes, thereby improving the likelihood of successful development of sensitive and specific locus-specific assays. And while the number of loci that could be assessed for methylation by locus-specific assays was necessarily limited to a select few DMRs, the use of these assays afforded the necessary sensitivity to detect the low copy numbers of methylated epialleles found in microdissected epithelia that might have otherwise gone undetected using less-sensitive genome-wide analysis techniques. There are a number of limitations to our study that warrant consideration. Firstly, while we sought to utilize a balanced cohort of 96 gynecologic specimens that might provide insight into, not only methylomic differences between HGSOC and various normal tissues, but other gynecologic malignancies, as well, this naturally limited the overall number of HGSOC cases that could be included and somewhat lowered the statistical power for identifying HGSOC-specific DMRs. However, the use of the TCGA dataset and rigorous cross validation somewhat minimizes this concern. Secondly, our analytical validation by SMART-MSP used many samples from the original cohort. Ideally, a second, independent cohort would be used as a means of validating the utility of these DMRs as methylation biomarkers for potential clinical use ( 47 ). Thirdly, the number of STIC-lesions and healthy fallopian epithelia that could be analyzed is somewhat limited by inherent difficulties in their procurement. This is presently an all but ubiquitous limitation for STIC-lesion studies ( 16 , 17 ), and a concerted effort among the research community will likely be necessary to amass larger cohorts. Fourthly, the use of locus-specific SMART-MSP assays, while providing the requisite sensitivity for analyzing limited DNA yields from LCM specimens, allowed only a discrete number of loci to be assayed and precluded evaluation of intratumoral/intercellular heterogeneity within the samples. Nonetheless, the present study establishes an important proof-of-concept for the potential of epigenetic alterations to contribute to the development and progression of HGSOC and acts to provide the necessary impetus for future epigenome-wide and mechanistic studies.

Gynecologic cancers account for over 10% of all cancer-related deaths of women in the U.S. ( 1 ). Many of these deaths can likely be prevented if a sensitive and specific set of biomarkers for detecting the earliest stages of malignancy could be identified. This is particularly true in the case of high-grade serous ovarian carcinoma (HGSOC), the most lethal of all gynecologic cancers, where over 60% of women are not diagnosed until peritoneal dissemination has already occurred. While considerable effort has been made toward the identification of both protein and genetic biomarkers for HGSOC, such as CA-125 and TP53 mutations, these have yet to demonstrate sufficient utility to warrant recommendation for use in general screening ( 2 ). Furthermore, no novel biomarker has been approved for screening, diagnosis or monitoring of HGSOC or other gynecologic cancers in over two decades ( 3 ). There thus remains an unmet need for identification of biomarkers capable of sensitive and specific detection of these cancers, particularly at the earliest stages of disease. All forms of cancer, including gynecologic types, exhibit widespread epigenomic alterations when compared with normal tissues ( 4 , 5 ). Gene silencing by epigenetic alterations, including DNA methylation ( 6 ), is a well-known contributor to carcinogenesis and can occur during or even prior to the development of precursor lesions ( 7 – 9 ). A number of previous studies have likewise documented aberrant gene promoter methylation associated with both ovarian ( 10 ) and endometrial ( 11 , 12 ) cancers, as well as regions of differential methylation that occur in both cancer types ( 13 , 14 ). However, most studies aimed at identifying methylation biomarkers for the detection of gynecologic cancers have yet to achieve the requisite performance metrics for clinical implementation and there thus remains an unmet need for their discovery ( 3 ). A number of factors have likely impeded identification of clinically-useful methylation biomarkers for gynecologic cancers, particularly for early-stage HGSOC. Perhaps most notable is the recent discovery that the majority of HGSOCs originate, not within the ovary, but in precursor lesions of the fallopian tube, commonly referred to as serous tubal intraepithelial carcinoma (STIC) lesions ( 15 – 17 ). It is not until these lesions mature and seed into the ovary, a process that is estimated to occur over a period of seven years ( 16 ), where they rapidly progress toward malignancy. Thus, the etiological origin of HGSOC implies that the most useful biomarkers should be detectable in STIC lesions before they progress to the ovary and should also demonstrate specificity against noncancerous gynecologic tissues. However, analysis of STIC lesions remains inherently difficult as specimens are often scarce and extracting DNA of sufficient purity and yield for genome-wide epigenetic analysis is often not feasible. Not inconsequently, and to the best of our knowledge, STIC-specific epigenetic alterations have yet to be reported. A second issue hampering methylation biomarker discovery for HGSOC is the paucity of publicly-available methylation data with high genomic coverage. Unlike other cancer types, methylation data for HGSOC from The Cancer Genome Atlas (TCGA) were primarily acquired with first generation microarrays (Illumina HumanMethylation27) that provide less than 3% of CpG coverage compared with the current generation chip. The current methylation microarray from Illumina is the MethylationEPIC (EPIC) BeadChip, which contains over 850,000 CpG probes, over half of which are located within gene promoter regions ( 18 ). In the present work, we utilize an etiologically-informed approach to address some of the impediments that have previously hampered identification of robust methylation biomarkers for gynecologic cancers and early-stage HGSOC, in particular. We achieve this by first leveraging the current-generation, MethylationEPIC BeadChip to perform genome-wide methylomic analysis to screen for sites of differential methylation that occur in HGSOC in relation to other gynecologic malignancies and healthy gynecologic tissues. We then apply bioinformatic smoothing algorithms across adjacent CpG sites to identify the most robustly differentially-methylated regions (DMRs) that exhibit hypermethylation exclusively in HGSOC tissue when compared with healthy mucosae. We then develop and employ locus-specific methylation assays to analytically validate these DMRs and evaluate their performance as biomarkers for HGSOC. Lastly, we evaluate the potential utility of these biomarkers for detecting early-stage HGSOC by assessing the presence and specificity of hypermethylation within these DMRs in early precursor STIC lesions.