OvaPrint—A Cell-free DNA Methylation Liquid Biopsy for the Risk Assessment of High-grade Serous Ovarian Cancer

Epithelial ovarian cancer (EOC) is a malignancy associated with considerable morbidity and high rates of mortality, primarily due to its frequent presentation in an advanced stage (1). High-grade serous ovarian carcinoma (HGSOC), the most common histology of EOC, is the most lethal gynecologic malignancy, with a 5-year survival rate of 40% or less when diagnosed at late stage. Only about 15% of women with a suspicious mass will present with stage I cancer. The detection of earlier stage EOCs is associated with considerably better outcomes, with markedly improved 5-year survival rates for stage I and stage II tumors compared with stage III and IV disease (2). Detection of EOC at an early stage is difficult given early-stage EOC symptoms are either nonexistent or nonspecific and are often associated with more common benign gynecologic and gastrointestinal conditions (e.g., uterine fibroids, endometriosis, benign ovarian tumors, inflammatory bowel disease), and physiologic processes (e.g., menstruation, ovulation). In addition, about 20% of the 300,000 women undergoing surgery for a pelvic mass each year will have an EOC, with the remaining diagnosed with a benign tumor (3).

Currently, for women presenting with an adnexal mass and for whom a surgical intervention has been chosen, radiologic imaging, serum tumor markers, and the use a multivariate index assay (e.g., Ova1, Overa, ROMA,) may be used to determine the optimal surgical approach. Specifically, surgical outcome for women with EOC is improved when cytoreduction is performed by a gynecologic oncologist (4–9). However, in practice only 30%–50% of patients with an adnexal mass, that is subsequently determined to be EOC on final pathology, are referred to a gynecologic oncologist by these grading criteria (8). Currently, pathologic assessment of patient tissue is the only way to definitively diagnose EOC. Accordingly, developing an assay and algorithm that would preoperatively discriminate between benign and malignant pelvic masses would be of considerable clinical value and improve clinical management and referral of patients.

In women at increased genetic risk for developing EOC, risk-reducing surgery to remove the fallopian tubes and ovaries remains the only established prophylactic approach. Organizations like the Centers for Disease Control and the National Comprehensive Cancer Network have called for genetically high-risk women to undergo regular pelvic exams, pelvic ultrasounds, and serial measurements of serum levels of CA-125 prior to risk-reducing surgery. Many attempts have been made to improve the specificity of CA-125 for average risk women with a known pelvic mass. Approaches have included adding other serum proteins, such as beta2 microglobulin in the OVA1 and Overa tests (Aspira labs; ref. 10), HE4, or adding transvaginal ultrasonography for ovarian assessment based on the risk of malignancy index (11, 12). However, such screening has not been shown to provide any improvement in clinical outcomes should EOC develop and have significant limitations in their specificity which leads to an increase in the number of unwanted interventions (13). The negative outcome of the UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS) trial, which tested a two-step screening strategy using measurement of CA-125 and transvaginal pelvic ultrasound in over 200,000 women, affirms the need for a more specific biomarker panel/algorithm to identify women at increased risk for developing EOC or to detect malignancy in women already presenting with adnexal masses (13). It should be noted that the vast majority of women who go on to develop EOC present with no family history and have no germline pathogenic variants (14).

As pointed out in our recent review (15), developments in proteomics and genomics have produced novel, noninvasive methods for EOC detection and screening, but unfortunately most markers have been shown to only be useful for identifying late-stage disease (16). Liquid biopsies utilizing cell-free (cf)DNA are enabling noninvasive, dynamic assessment of disease status in patients with cancer with cfDNA methylation being particularly promising for early detection. DNA methylation is a fundamentally important modification for the maintenance of large genomes. There are several advantages to developing liquid biopsy approaches using aberrant DNA methylation over other molecular alterations such as point mutations or serum-based protein markers that have also been pointed out in our review (15–21). First, DNA methylation changes occur early in tumorigenesis and are highly stable, covalent chemical marks. Detection sensitivity of aberrantly methylated DNA is also enhanced by the increased frequency and distribution of DNA methylation changes compared with single-nucleotide variants or copy-number aberrations. In addition, DNA methylation measurements incorporate numerous regions, each with multiple CpG positions, allowing better limits of detection than for protein-based markers or DNA mutations. Furthermore, because aberrant CpG island hypermethylation rarely occurs in normal cells, DNA methylation can be detected with a notable degree of sensitivity, even in the presence of background methylation derived from normal cells. Finally, large-scale DNA methylation alterations are tissue- and cancer-type specific and therefore potentially have greater ability to detect and classify cancers in patients with early-stage disease.

The goal of the current study was to develop and evaluate OvaPrint, a cfDNA methylation liquid biopsy that can be used for the risk assessment of HGSOC. The current study differs in several critical ways from previous attempts to develop risk and screening markers for EOC. First, our study focused on HGSOC, avoiding a catch-all strategy that is fraught with statistical and biological concerns. Second, we heavily weighted stage I malignancies in the discovery of features and validation steps. Third, we used both normal fallopian tube and ovarian tissue as normal reference to account for the tubal origin EOC. Fourth, OvaPrint was designed for use in risk assessing both symptomatic and asymptomatic women with pelvic masses. Fifth, we made considerations for the heterogeneity within HGSOC. Finally, we conducted extensive validation of the assay. Taken together, OvaPrint is a highly sensitive and specific test that can be used for the early risk assessment and eventual screening of HGSOC.

A total of 403 tissue (n = 59) and plasma (n = 344) samples were obtained in this study (Fig. 1). A summary of sample demographics (age and race), diagnosis, and stage can be found in Supplementary Table S1. For feature discovery, we obtained 35 retrospectively collected HGSOC flash frozen tissue samples along with 15 normal fallopian tube and nine normal ovarian samples as controls (Supplementary Tables S1A and S2; Fig. 1). All tissue samples were histologically confirmed by a pathologist. Normal tissue samples were obtained from contralateral ovaries from 4 patients with EOC, 14 patients with uterine cancers, 2 patients with acute lymphoblastic leukemia, 2 patients with cervical cancer, and 2 patients with hemorrhagic cysts. For validation of selected markers in cfDNA, we obtained 128 retrospectively collected plasma samples with HGSOC (stage I N = 37, stage II N = 29, stage III N = 46, stage IV N = 16) and 100 plasma samples with benign ovarian masses (Supplementary Tables S1B and S3; Fig. 1) from patients with adnexal masses prior to surgery. All blood samples were collected in EDTA (purple top) tubes. Of these, 14 samples with poor library quality (see hybridization probe capture section below) and 16 samples pretreated with neoadjuvant chemotherapy prior to sample collection were excluded. Ultimately, a total of 372 samples were analyzed after sequencing. A total of 82 HGSOC plasma samples (27 stage I, 20 stage II, 35 stage III) and 100 plasma samples from benign ovarian tumors were included in the final model (Supplementary Table S1B). We also tested 16 stage IV HGSOC but these were excluded from the model and were analyzed separately. In addition, we included 13 cfDNA samples from patients with low-grade serous ovarian cancer (LGSOC), 22 samples with borderline serous tumors, 26 normal samples from healthy individuals, and 54 samples from patients with other histologic subtypes of EOC (Supplementary Table S1B; Fig. 1). These included a mix of high-grade, low-grade and borderline mucinous, endometroid, and clear cell carcinomas (Supplementary Table S1B; Fig. 1).

All samples were obtained under written informed consent and Institutional Review Board approval from the Gynecologic Tissue and Fluid Repository (GTFR, hereinafter referred to as “source 1”) at the University of Southern California and from seven additional commercial sources (“sources 2–8”). Sample source information can be found in Supplementary Table S2 [reduced representation bisulfite sequencing (RRBS) samples], Supplementary Table S3 (benign and normal plasma samples), and Supplementary Table S4 (all malignant plasma samples).

cfDNA was extracted using the MagMax Cell-Free DNA Isolation Kit (Thermo Fisher Scientific, RRID: SCR_008452) according to the manufacturer's protocol (22). Genomic (g)DNA was extracted from flash frozen tissues using AllPrep DNA/RNA Mini kit (Qiagen, RRID: SCR_008539) according to the manufacturer's recommendations. Tissues were homogenized using Bullet Blender homogenizer (Next Advance) for 5 minutes at full speed with a mixture of 0.9–2.0 mm RNase-free stainless-steel beads. Homogenates were passed through the QIAshredder (Qiagen) to remove any remaining particulate matter. Quantity and purity of the isolated DNA was determined using Tapestation D1000 High Sensitivity tapes (RRID: SCR_018435).

RRBS was performed using the Ovation RRBS Methyl-Seq kit (Tecan Genomics, catalog no. 9522-A01, RRID: SCR_016771). The manufacturer's recommended protocol was followed. Briefly, 300 ng of gDNA was used as input into the reaction. DNA was digested with MspI, adapters ligated, and final repair performed. Libraries were bisulfite converted and PCR amplified.

Paired-end sequencing was performed on final library products on the NovaSeq 6000 platform (Illumina, RRID: SCR_016387). After demultiplexing, read quality was evaluated using FastQC (RRID: SCR_014583). FASTQ adapter trimming was performed by Trim Galore (RRID: SCR_011847). After trimming, paired-end reads were aligned to hg38 using Brabham Bioinformatics’ Bismark alignment software (RRID: SCR_005604); per-CpG methylation was also evaluated using Bismark. Bisulfite conversion rate was greater than 99% for all samples.

Differentially methylated regions (DMR) were called using metilene (23). Each HGSOC sample was compared with a pool of normal fallopian samples and normal ovary samples. The resulting DMR lists were filtered to contain only statistically significant DMRs (FDR P value < 0.05) with an |Δβ| > 0.2. From these data, we selected the top 1,677 DMRs for validation using hybridization probe capture.

For target validation, we used a hybridization probe capture method as described previously (22) and designed an assay for 1,677 DMRs identified by RRBS using the myBaits Custom Methyl-Seq (Daicel Arbor Biosciences). These targets spanned 532 kbp and contained 28,797 CpGs. The resulting probeset contained 115,739 oligonucleotide probes, each 80 bp in length.

Hybridization probe capture was performed on cfDNA from 82 patients with HGSOC, 100 patients with benign ovarian masses, 22 patients with borderline serous tumors, 13 patients with LGSOC, 11 patients with clear cell ovarian cancer, 12 patients with endometroid ovarian cancer, 22 patients with mucinous ovarian cancer, 9 patients with EOC with mixed histology, and 26 normal samples from healthy individuals (Supplementary Table S1B). A 5-point standard curve was generated by mixing 0% and 100%, commercially available methylation control DNA (Zymo Research, catalog no. D5014, RRID: SCR_008968) fragmented down to approximately 160 bp via sonication to simulate cfDNA. The resultant standards consisted of 0%, 7%, 25%, 50%, and 100% methylated DNA samples.

We used 10 ng of extracted cfDNA for bisulfite treatment using the EZ DNA Methylation-Gold Kit (Zymo Research, RRID: SCR_008968), followed by library preparation with the Accel-NGS Methyl-Seq DNA Library Kit (Swift Biosciences) and 11 cycles of indexing amplification using unique dual 8 bp indexing primers. Eight or more libraries were pooled for each enrichment reaction, with a total library mass of up to 1.6 μg insert-containing templates. Hybridization probe capture, sequencing, alignment, and methylation calling on all samples was performed as described previously (22), with the exception that hg38 genome build was used. For next-generation sequencing quality control (QC), alignment rate was calculated by Bismark (ref. 24; RRID: SCR_005604), bisulfite conversion rate was calculated using MethPipe (refs. 25, 26; RRID: SCR_005168), and cfDNA insert size distributions were calculated by qualimap bamQC (27); all QC metrics were aggregated using MultiQC (ref. 28; RRID: SCR_014982).

Beta values were calculated per CpG by the bsseq package and regions passing filtering (n = 1,441) were fed into a machine learning tool (29, 30). We then trained a binary classifier using a gradient boosted model (GBM; ref. 31) with grid search hyperparameter tuning to discriminate between HGSOC and benign samples. Rather than relying on a single training/testing split (32), we evaluated the performance using 10 repetitions of 10-fold cross-validation (Supplementary Fig. S1). Finally, we calibrated the GBM response value into a HGSOC probability score (hereinafter referred to as “OvaPrint score”) using a logistic regression model (32).

Beta values and metadata for both RRBS tissue, used for feature discovery, and hybridization probe capture cfDNA, used for model construction and evaluation, have been deposited on Zenodo (https://doi.org/10.5281/zenodo.8395817).

For OvaPrint development, RRBS was first performed on flash frozen tissue from 35 HGSOC (25 stage I, eight stage III, two stage IV), 15 normal fallopian tube tissue, and nine normal ovarian samples (Figs. 1 and 2A; Supplementary Tables S1A and S2). The inclusion of both normal ovarian and fallopian tube tissue as well as utilizing a majority of stage I samples are unique design aspects of our study. Normal samples were histologically confirmed by a pathologist and obtained from contralateral adnexa from 4 patients with EOC, 14 patients with uterine cancers, 2 patients with acute lymphoblastic leukemia, 2 patients with cervical cancer, and 2 patients with hemorrhagic cysts. There were 1,677 statistically significant DMRs between HGSOC and normal samples. The median width of these DMRs was 242 bp [interquartile range (IQR) = 152–411 bp]; the median number of CpGs per region was 11 (IQR = 7–18). These regions strongly separated HGSOC from normal fallopian and ovary samples on a uniform manifold approximation and projection (UMAP) plot (Fig. 2B; ref. 33) and with hierarchical clustering (Fig. 2C). Hierarchal clustering also revealed two distinct HGSOC clusters, referred to as HGSOC 1 and HGSOC 2 with HGSOC 2 samples more closely resembling methylation patterns of normal tissue compared with HGSOC 1 (Fig. 2C); both subtypes were included in feature discovery.

To evaluate the generalizability of the HGSOC DMR signature, we accessed an external EPIC array dataset of HGSOC (n = 99) and normal fallopian (n = 12) tissues generated by Gonzalez-Bosquet and colleagues [ref. 34; Gene Expression Omnibus (GEO) Accession: GSE133556]. We compared methylation beta values between HGSOC and normal fallopian tissue at 772 (383 hypermethylated, 389 hypomethylated) of our 1,677 DMRs that overlapped at least one EPIC array CpG probe. We found that differential methylation between HGSOC and normal fallopian tissue in GSE133556 was consistent with our DMRs; mean Δβ of GSE133556 within hypermethylated/hypomethylated DMRs was 0.21 and −0.15, respectively (Fig. 2D; Wilcoxon P value < 0.0001 for both comparisons). Δβ values in GSE133556 between tumor and normal for each DMR strongly correlated with the Δβ we observed in our RRBS samples (Fig. 2E).

It is now widely accepted, based on genomic and histologic studies, that most cases of HGSOC arise in fallopian tubes (35). To assess whether HGSOC tumors were likely derived from fallopian tube or ovary, we performed RRBS analysis on a series of both normal fallopian (n = 15) or normal ovarian (n = 9) tissue. To evaluate the genome-wide similarity between HGSOC, normal fallopian, and normal ovary, we extracted beta values at every CpG locus with coverage ≥ 5X. We removed CpGs with very low beta value variance across the dataset (σ² < 0.001). This resulted in 2,231,688 CpGs that were used to evaluate genome-wide methylation profiles. Dimensionality reduction by UMAP showed HGSOC clustering more closely with normal fallopian tissue (Fig. 3A). HGSOC clustering was not affected by tumor stage (Fig. 3B). Furthermore, Euclidian distance was lower between HGSOC and normal fallopian tube tissue (distance = 198) compared with HGSOC and normal ovary (distance = 226). These findings provide epigenetic evidence as the basis for the fallopian tube to be the origin of HGSOC.

We designed a custom hybridization probe capture assay to selectively enrich DNA from 1,677 DMRs identified by RRBS as we have described previously (22). This hybridization probe, capture bait set constitutes the OvaPrint cfDNA methylation test. The targeted assay was used to profile plasma-derived cfDNA methylation in 182 samples from benign ovarian tumors (n = 100) and HGSOC (n = 82), including stage I (n = 27), stage II (n = 20), and stage III (n = 35). Of the 182 samples, 14 HGSOC had matching tissue previously sequenced by RRBS and the remaining samples comprised an independent set of samples. The benign cohort consisted of multiple, noncancerous ovarian masses including 48 benign serous cystadenomas, 12 fibromas, 11 benign mucinous cystadenomas, 10 endometriomas, and 5 benign teratomas, among several other less frequent benign masses (Supplementary Table S1B).

In addition, we also tested the generalizability of OvaPrint on 89 additional samples which consisted of borderline ovarian masses (n = 22), LGSOC (n = 13), clear cell ovarian cancer (n = 11), endometroid ovarian cancer (n = 12), mucinous ovarian cancer (n = 22), and EOC with mixed histology (n = 9). Finally, we included 26 normal samples from healthy individuals. In total, we captured and sequenced 313 plasma cfDNA samples (Supplementary Table S1B; Fig. 1) using OvaPrint.

After alignment, samples had 28,971,805 mapped reads on average with a mean alignment rate of 82.95% (Supplementary Fig. S2A and S2B). We evaluated the percentage of reads mapped to target regions—a measure library of enrichment efficiency—and found 72.34% of reads were mapped within a target region (ref. 22; Supplementary Fig. S2C). The average bisulfite conversion rate was 99.3% (min = 98.8%, max = 99.5%; Supplementary Fig. S2D). Library fragment length distributions from all samples showed peaks at approximately 165 bp which is expected of cfDNA (Supplementary Fig. S2E).

We ran methylation standard curves to assess methylation linearity by comparing observed versus expected methylation values. In addition, two samples were used on two separate sequencing runs and served as technical and biological replicates. The control samples showed highly consistent beta values across all 1,677 target regions between runs (Supplementary Fig. S3A and S3B). We also constructed a 5-point standard curve from 0%, 7%, 25%, 50%, and 100% methylated DNA samples. The resulting standard curves from each target region for two sequencing runs were combined across runs and showed a strong linear relationship between observed and expected beta values (Supplementary Fig. S3C and S3D).

We used UMAP to further evaluate potential sample source, race, or age-specific effects. Beta values were extracted for all HGSOC and benign samples. We did not observe clustering based on sample source for benign samples (Supplementary Fig. S4A) or for HGSOC samples (Supplementary Fig. S4B). There were four sources for benign tumors and six sources for HGSOC (Supplementary Tables S3 and S4). Similarly, we did not observe samples clustering based on race (Supplementary Fig. S4C and S4D), suggesting that OvaPrint can be used across race. Finally, we examined the effect of patient age, as age-associated epigenetic changes are well documented (36); however, we did not observe any age-related clustering (Supplementary Fig. S4E and S4F). Taken together, these results show our targeted hybridization probe capture assay effectively enriches target regions, and is not influenced by sample source, race, or age.

Next, we wanted to determine whether the 1,677 DMRs comprising the OvaPrint targeted assay could discriminate between individuals with HGSOC and individuals with benign ovarian masses. After sequencing, mean beta values were calculated for each region in 82 HGSOC samples and 100 benign samples, which was used as input into machine learning. We used gradient boosting (31) to train a novel classifier that differentiates between HGSOC and benign ovarian tumors. A repeated cross-validation strategy was employed to obtain an unbiased and stable estimate of overall model performance (32) which constitutes the OvaPrint model. ROC analysis was used to evaluate model performance, and determined that OvaPrint has excellent power to discriminate between HGSOC and benign samples (Fig. 4A). The estimated area under the ROC curve (AUC) was 0.936 [95% confidence interval (CI): 0.90–0.97] indicating OvaPrint is highly sensitive and specific (Fig. 4A). Setting an OvaPrint score threshold of 0.5 on the calibrated scores resulted in an accuracy of 90.6%, with a sensitivity of 0.842, and a specificity of 0.960 [positive predictive value (PPV) = 0.95, negative predictive value (NPV) = 0.88; Fig. 4B]. To introduce greater nuance into the analysis, we also categorized samples into three distinct risk groups—namely, “high,” “moderate,” and “low”—rather than limiting the results to a simple binary outcome when assessing the likelihood of HGSOC. On the basis of the distribution of samples (Fig. 4B), we defined any sample with an OvaPrint score >0.65 as high risk, 0.65–0.35 as moderate risk, and <0.35 as low risk. Using this approach, 68 of 182 were high risk (HGSOC N = 66, benign N = 2), 10 of 182 were moderate risk (HGSOC N = 4, benign N = 6), and 104 of 182 were defined as low-risk samples (HGSOC N = 12, benign N = 92; Fig. 4B; Supplementary Table S5). We also tested 16 stage IV HGSOC cases. All 16 stage IV cfDNA samples were analyzed identically to the stage I–III HGSOC cases detailed above. We found OvaPrint predicted 13 of 16 stage IV samples as high risk, two as moderate risk, and one as low risk (Supplementary Fig. S5A; Supplementary Table S5). The overall accuracy was 94.8% with a ROC AUC of 0.960 (Supplementary Fig. S5B; sensitivity = 0.875, specificity = 0.960, PPV = 0.778, NPV = 0.980).

Notably, a small subset of HGSOC samples (n = 12) were consistently misclassified with low OvaPrint scores (<0.35) across multiple cross-validation repeats, which we termed “misclassifiers” (MC). These MC samples were investigated for demographic, clinical, or technical factors that were unifying. First, we evaluated whether race or age were associated with MC samples as both factors are known to affect DNA methylation (36–38). There was no apparent effect of age or race on whether a sample misclassified or not, albeit there was some missing data for age and race (Supplementary Table S1B). Stage was also not associated with whether a HGSOC misclassified or not. Next, we investigated whether MC samples had abnormal methylation patterns in tissue. We cross-referenced samples that were analyzed by RRBS and OvaPrint. UMAP analysis and hierarchical clustering of beta values across all 1,677 target regions, and all HGSOC samples, was used to determine whether methylation signatures of MC samples differed in HGSOC tissue. Of the 35 HGSOC samples run on RRBS, 14 tissue samples had patient-matched cfDNA that was analyzed using OvaPrint. Of these, two of 14 were MC samples. We found that the corresponding tissue samples of MC tumors did not separate from other HGSOC samples by UMAP, and that they clustered with samples that had high OvaPrint scores and classified correctly (Supplementary Fig. S6A and S6B). We also did not observe anomalous hierarchical clustering of MC samples, nor did they cluster with normal fallopian or ovarian tissue (Supplementary Fig. S6C and S6D).

We also examined whether sample source for plasma samples was associated with sample misclassification (Supplementary Fig. S4). Notably, source 4, whose samples came from three specific hospitals (sites A, B, and D), had an association with misclassified samples. Specifically, source 4 “site D” contributed four of 12 of the MC samples (representing 50% of the samples obtained from that source), disproportionately more than any of the other sample sources. This source was unable to provide additional information or verify the samples as they no longer receive samples from site D. It is important to note that no benign samples were obtained from this source. Without the unverifiable samples from source 4/site D, the sensitivity and specificity would be 0.88 and 0.96, respectively with an AUC of 0.947 (PPV = 0.941, NPV = 0.914). While this does not explain all MC, the sensitivity of our assay was most impacted by a potential source effect and points to the importance of sample-dependent variables in liquid biopsy efforts.

To investigate the potential clinical utility of OvaPrint in screening asymptomatic women, we obtained cfDNA from healthy donors without known pelvic masses. For this, we analyzed 26 normal cfDNA samples using OvaPrint. UMAP analysis of cfDNA methylation demonstrated that normal samples clustered with benign tumors (Fig. 5A), indicating that the same features used for risk assessment in symptomatic women can also potentially be used in asymptomatic women. Next, adding the 26 normal samples to the benign group resulted in an equally high AUC (0.928, 95% CI = 0.887–0.968). OvaPrint classified 22 of the 26 normal samples as low risk, 0 samples as moderate risk, and four samples as high risk (Fig. 5B; Supplementary Table S5). Given the specificity of the assay when considering benign versus HGSOC, the four normal samples classifying as high risk could be suspicious of malignancy but without follow-up data this is speculative. Taken together, these results show OvaPrint has potential screening utility; however, additional normal samples and prospective studies are required to fully investigate the screening utility of OvaPrint.

For this study, we focused on HGSOC because it is the most common and most aggressive EOC subtype and our results show that OvaPrint is an excellent biomarker for HGSOC. However, EOC also includes clear cell, endometrioid, and mucinous histologic subtypes. Low-grade and borderline EOC were also included in this analysis. A sample summary is provided in Supplementary Tables S1 and S5. Although these other histologic subtypes represent less than 20% of all EOC, it would still be clinically useful to know how OvaPrint performs on other EOC histologic subtypes and tumors of lower malignant potential. To this end, we tested OvaPrint on plasma cfDNA from patients with mucinous ovarian carcinoma (N = 22; 11 low-grade, 11 borderline), endometrioid adenocarcinoma (N = 12; 2 high-grade, 9 low-grade, 1 borderline), clear cell carcinoma (N = 11; all high-grade), and EOC with mixed histology (N = 9; 8 high-grade, 1 low-grade). For mixed histology samples, five of nine had a serous component. In total, there were 21 high-grade, 21 low-grade, and 12 borderline cases for other histologic subtypes (Supplementary Tables S1B and S5).

OvaPrint classified two of 11 low-grade mucinous samples as high risk, one low-grade as moderate risk, and eight low-grade mucinous as low risk (Fig. 6A and B; Supplementary Table S5). For borderline mucinous samples, five of 11 were classified as high risk and six as low risk. For endometrioid adenocarcinoma, four of nine low-grade tumors were classified as high risk and five as low risk. Two high-grade and one borderline tumor were classified as high risk. For clear cell carcinoma, four of 11 samples classified as high risk; the remaining seven samples were classified as low risk. Finally, we observed five high risk, one moderate risk, and three classes classifying as low risk in samples with mixed EOC. Overall, OvaPrint classified 11 of 21 non-serous high-grade EOC as high risk (ROC AUC = 0.818, sensitivity = 0.57), one as moderate risk, and five as low risk.

The performance of OvaPrint was also assessed in 13 LGSOC and 22 borderline serous tumors. We found that OvaPrint classified seven of 13 LGSOC samples as low risk, three as moderate risk, and three as high risk (Fig. 6A and B; Supplementary Table S5). For borderline serous samples, 21 of 22 were classified as low risk; one sample was classified as high risk. Out of 38 total low-grade EOC samples, OvaPrint classified 24 as low risk, five as moderate risk, and nine as high risk. A breakdown of tumor stage can be found in Supplementary Table S1. Notably, 63 of 89 total samples were stage I or II (Supplementary Table S1B).

Taken together, these results indicate that while OvaPrint has a lower sensitivity in non-HGSOC EOC histotypes, its performance in non-serous EOC is quite reasonable given the lack of other reliable tests. Still, it will be prudent to improve the sensitivity of OvaPrint for non-serous tumors.

In this study, we developed OvaPrint, a cfDNA methylation liquid biopsy designed to discriminate benign pelvic masses from HGSOC preoperatively that has also shown its potential clinical utility in screening by examining cfDNA from healthy individuals. OvaPrint achieved an overall accuracy of 91%, a PPV of 0.95, and a NPV of 0.88 (0.91 if site D samples are removed), surpassing other commercial tests such as OVA1 and ROMA (10, 39–41) which are primarily driven by CA-125. While CA-125–based tests are sensitive to EOC (ROMA = 94% sensitivity at 75% specificity; OVA1 = 92% sensitivity at 54% specificity) they are nonspecific, leading to a low PPV and high false positive rate (ROMA PPV = 60%, OVA1 PPV = 31%). A high PPV indicates that a positive test result will accurately identify patients with HGSOC, minimizing false positives which would lead to unwanted surgical interventions by gynecologic oncologists. The high NPV demonstrates that OvaPrint can accurately identify patients who are truly negative, minimizing false negatives which would leave lethal gynecologic malignancies undetected. For patients with early-stage undiagnosed EOC presenting with an adnexal mass and normal tumor markers, OvaPrint may encourage referral for appropriate surgical intervention and clinical staging over surveillance. Such interventions may even lead to a shift in stage at time of diagnosis.

EOC is an umbrella term for what is a heterogenous collection of malignant neoplasms. On the basis of histopathology and molecular genetic alterations, EOC are divided into five main types: HGSOC (70%), endometrioid (10%), clear cell (10%), mucinous (3%), and LGSOC (<5%) that account for over 90% of cases. These histotypes are characterized by differences in epidemiologic and genetic risk factors, precursor lesions, patterns of spread, and molecular events during oncogenesis, response to chemotherapy, and prognosis (22, 42). Therefore, not surprisingly, OvaPrint has a lower performance in other forms of EOC but may still have some clinical utility in its current form because other tools are not currently available. Still, studies are ongoing to identify and add additional features that are more specific to other subtypes of EOC.

Our results suggest that OvaPrint has some utility in identifying non-HGSOC histotypes of EOC as well as borderline tumors, although the sensitivity was lower in these subtypes. A high score is highly suggestive of a malignancy, and the clinical implications are 2-fold: (i) tumor rupture should be avoided (this is particularly relevant if ovarian cystectomy is being considered as the rate of rupture is higher as compared with oophorectomy, especially if a minimally invasive approach is utilized) and (ii) patients with a high score should be referred to a gynecologic oncologist who is trained to perform an adequate staging, or if needed, debulking surgery. Whether a high score translates to more aggressive biologic behavior is an area of future investigation.

The focus on advanced stage and grouping all EOC together in prior screening and detection studies has been a major limitation of past studies (15). These limitations reduced power and made it more difficult to discriminate what are, biologically, very different types of tumors and could explain why ovarian cancer screening tests have largely been unsuccessful. In the current study, these caveats were avoided by focusing on the most common and aggressive form, HGSOC. In addition, samples from stage I HGSOC formed the majority of samples analyzed during feature discovery and validation. While many prior studies have relied on The Cancer Genome Atlas data for feature discovery, here we performed a de novo analysis using RRBS analysis on a cohort of HGSOC and normal controls. Finally, by utilizing both normal ovary and normal fallopian tube tissue as reference for feature discovery, we factored in the tubal origin of HGSOC, which has a basis in DNA methylation patterns associated with fallopian tube tissue. Also including normal ovarian tissue as a control enhanced the discrimination between tumors and benign neoplasms that arise from ovary. OvaPrint thus mitigated the problems that other screening studies likely had. We also demonstrated that HGSOC is heterogenous, as evidenced by DNA methylation subgroups after RRBS analysis (Fig. 2). In addition, there was a group of tumors that we refer to as “misclassifiers”, because they consistently failed to classify as HGSOC across multiple cross-validation repeats during machine learning and all had low OvaPrint scores. Clinical demographics such as age, race and stage were not underlying reasons. We cannot make a determination about the role of family history or survival as that information was not available for most patients.

Despite the valuable findings of the work presented in this article, it is important to acknowledge some limitations and ways to address them moving forward. First, the sample size for non-serous EOC tumors, as well as plasma from healthy individuals was small but we are planning prospective validation studies and plan to add additional subtype-specific features to future versions of OvaPrint, which is expected to increase sensitivity and specificity in other EOC subtypes. Second, the clinical significance of a high OvaPrint score in LGSOC and borderline cases remains unexplored. However, we plan to investigate whether a high/moderate risk test result in these cases is associated with worse outcomes in an ongoing prospective study. Finally, we were unable to identify a unifying factor between MC samples. We believe that misclassification of these samples was in part related to a sample-dependent cause due to sample source effect but other technical or biological factors still cannot be ruled out. Additional studies comparing patient tissues from MC samples with correctly classifying samples will help shed light on the causes and will help increase the overall sensitivity of OvaPrint.

In summary, EOC and specifically HGSOC remain devoid of clinical assays that either preoperatively confirm malignancy or that can be used for community-based or high-risk screening, especially for early-stage tumors. In this study, we report on the development of OvaPrint, a cfDNA methylation liquid biopsy for the risk assessment of HGSOC in women with adnexal masses. The focus on HGSOC highlighted the importance of not grouping all EOC into one category for biomarker development, a strategy that has likely hampered previous efforts. The current data are especially compelling for several reasons, highlighted by the feature discovery in tissue that was validated in an independent cohort of nearly 200 plasma samples which demonstrated excellent NPV and PPV. The development of OvaPrint, which is an accurate, noninvasive preoperative test for ovarian cancer has the potential to improve diagnosis and cancer care for the almost 20,000 women diagnosed yearly. Prospective studies are ongoing to further validate OvaPrint for risk assessment of adnexal masses of HGSOC and other EOC in symptomatic and asymptomatic women.

M. Spillman reports grants from TGEN during the conduct of the study, as well as other support from University of Arkansas for Medical Sciences and Texas Oncology outside the submitted work. L. Roman reports grants from Wright Family Foundation during the conduct of the study, as well as personal fees and other support from Global Coalition for Adaptive Research outside the submitted work. B. Salhia reports other support from CpG Diagnostics during the conduct of the study, as well as personal fees from AstraZeneca outside the submitted work; in addition, B. Salhia has a patent for Cell-Free DNA Methylation Test #18/546,472 pending and licensed to CpG Diagnostics Inc. No disclosures were reported by the other authors.

D.N. Buckley: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. J.P. Lewinger: Software, formal analysis, validation, investigation, methodology, writing–review and editing. G. Gooden: Resources, formal analysis, validation, investigation, methodology, project administration, writing–review and editing. M. Spillman: Conceptualization, investigation, writing–review and editing. M. Neuman: Investigation, writing–review and editing. X.M. Guo: Investigation, writing–review and editing. B.Y. Tew: Data curation, validation, investigation, methodology, writing–review and editing. H. Miller: Investigation, writing–review and editing. V.U. Khetan: Investigation, writing–review and editing. L.P. Shulman: Writing–review and editing. L. Roman: Conceptualization, resources, funding acquisition, writing–review and editing. B. Salhia: Conceptualization, resources, formal analysis, supervision, funding acquisition, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing.

We would like to thank the Keck Genomics Platform (KGP) at the University of Southern California for performing the sequencing detailed in this article.

This project was funded by a grant from the Wright Foundation (B. Salhia, J.P. Lewinger, H. Miller) and by the Swing Against Cancer Fund at the Norris Comprehensive Cancer Center, University of Southern California.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.