The potential of circulating tumor DNA methylation analysis for the early detection and management of ovarian cancer

Three-quarters of ovarian cancers (OC) are diagnosed when the tumor has spread into the abdomen and long-term survival rates of these women are low (10–30%) [1].

High-grade serous (HGS) OC accounts for 70–80% of OC deaths and the survival figures have not changed significantly over the last few decades [2]. Early diagnosis and personalized treatment still remain the biggest unmet needs in combating this devastating disease [2].

A number of OC biomarkers have been studied in the past. Among these, CA125, which was discovered more than 30 years ago [3], is still the “gold standard,” despite a modest positive predictive value when interpreted using a defined cut-off of 35 IU/mL [4], which has also been used as a reference standard in our work. Recently, the 35 most promising OC biomarkers were evaluated in the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial. The markers were tested in samples taken up to six months before OC diagnosis from 118 women and 951 age-matched controls and at a fixed specificity of 95%, CA125 sensitivity out performed all 35 markers [5]. However, the performance of CA125 dropped dramatically when samples taken > 6 months before diagnosis were evaluated [5]. Recently, we demonstrated that the performance of the Risk of Ovarian Cancer Algorithm (ROCA), based on CA125 serial profile, demonstrates superior performance characteristics during screening [6, 7]. CA125 kinetics are also increasingly being explored in women undergoing neoadjuvant chemotherapy (NACT) for predicting disease response and outcome [8–11]. Both require serial blood sampling, which, in the case of differential diagnosis, is never available in patients presenting clinically.

The vast majority of protein-based tumor markers are produced not only by cancerous but also non-neoplastic normal cells; CA125 is produced by mesothelial cells (i.e. peritoneum and pleura) and hence benign or inflammatory processes can result in aberrant elevations of serum CA125.

Recently, DNA-based markers, shed from tumor cells, have shown great promise in monitoring treatment response and predicting prognosis [12–16]. However, efforts to characterize the cancer genome have shown that only a few genes are frequently mutated in most cancers and that the location of the genetic mutation site differs across individuals with specific tumor types. Hence, the detection of somatic mutations is limited to patients that harbor a predefined set of mutations. The necessity of prior knowledge regarding specific genomic composition of an individual’s tumor tissue is one of the limiting factors when using these “liquid biopsy” approaches for early detection or differential diagnosis of a pelvic mass. Current sequencing technology allows for the detection of a mutant allele fraction of 0.1% (which is one mutant molecule in a background of 1000 wild-type molecules) [12, 17].

The development of a cell-free DNA based test for the early detection of cancer poses two major challenges: (1) low abundance of cancer-DNA in the blood; and (2) high levels of “background DNA” (shed from white blood cells [WBC] [18]) in serum samples that are separated from blood cells after significant time intervals. This is an issue with most currently available population cohort biobanks which could be used for the validation of potential screening markers using samples that have been banked years in advance of diagnosis.

Alteration of DNA methylation (DNAme) is: (1) an early event in cancer development [19–22]; (2) more frequently observed than somatic mutations; and (3) centered around specific regions, i.e. CpG islands [23]. Together with its chemical and biological stability, the detection of aberrant DNAme patterns in serum or plasma provide a novel strategy for cancer diagnosis as evidenced by several proof of principle studies [24–34]. DNAme analysis allows for the detection of specific patterns (i.e. full methylation of all linked 7–16 CpGs in a region of 120–150 bp) as opposed to single point mutations (e.g. in the TP53 gene) which is key to improving both the performance characteristics of the test and the detection limit of the assay. Plasma SEPT9 methylation analyses—currently the only cell-free DNA assay which is available for cancer screening in the clinical setting—demonstrates a specificity of 79% and a sensitivity of 68% for detection of colon cancers [31]. The clinical potential of serum/plasma-based cell-free DNA analysis is further exemplified by maternal plasma cell-free DNA testing for fetal trisomy which has already been integrated into clinical practice and demonstrates a higher sensitivity and a lower false-positive rate compared to imaging-based techniques [35].

We have employed two different epigenome-wide approaches to identify the most promising DNAme-based markers that discriminate OC vs benign pelvic conditions, developed serum tests using the discovered markers, and validated their performance in relation to the serum OC marker CA125.

The samples, techniques, and purpose of the three phases—marker discovery, assay development, and test validation—are summarized in Fig. 1.

Clinical validation of the serum DNAme assay

We validated the combination of the three markers in Set 3 (Fig. 3a–c) alongside the CA125 serum marker (Fig. 3d). The average coverage (i.e. DNA strand reads by the sequencer for each sample and region) is > 500,000 (Additional file 1: Figure S7). Applying the above indicated cut-off thresholds for the three DNAme markers and 35 IU/mL for serum CA125 led to specificities of 90.7% and 87.1% and sensitivities of 41.4% and 82.8%, respectively (Table 1). Due to the fact that reaction #228 was only analyzed in Set 2, we combined Set 2 and Set 3 in order to redefine the thresholds. Whereas for #141 the threshold of 0.0008 remained unchanged, for #204 and #228 we further lowered the pattern frequency threshold to 0.00003 and 0.00001, respectively, leading to specificity and sensitivity of 91.8% and 58.3%, respectively (Table 1). Among these 48 HGS cancers (i.e. the most aggressive cancers), 6/11 (54.5%) stage I/II and 22/37 (59.5%) stage III/IV cancers were serum DNAme-positive. Importantly, there was no overlap between the DNAme-positive and CA125-false positive controls (Table 2).

		Specificity	Sensitivity
Set 3	CA125 (cut-off 35 IU/mL)	122/140 (87.1%; 95% CI = 80.1–92.0%)	24/29 (82.8%; 95% CI = 63.5–93.5%)
	3 DNAme-Marker Panel (thresholds based on Sets 1 and 2)	127/140 (90.7%; 95% CI = 84.3–94.8%)	12/29 (41.4%; 95% CI = 24.1–60.9%)
Set 2 & 3	3 DNAme-Marker Panel (new thresholds based on Sets 1, 2, and 3)	178/194 (91.8%, 95% CI = 86.7–95.1%)	28/48 (58.3%; 95% CI = 43.2–72.1%)

		H and BPM		HGS
		CA125-negative	CA125-positive	CA125-negative	CA125-positive
3 DNAme marker panel (new thresholds)	Negative	108	18	4	9
	Positive	14	0	1	15

Serum DNAme to predict response to platinum-based NACT

In order to further assess the cancer specificity and the dynamics of our three candidate markers in individual patients, we recruited 25 OC patients who received carboplatin-based NACT. Compared with the pre-treatment sample, all three DNAme markers decreased substantially and to a larger extent compared to CA125 after one and two cycles (Fig. 4a–d and Additional file 1: Figures S8–S10). Whereas CA125 dynamics were not a strong discriminator between chemotherapy-responders and non-responders (Table 3), serum DNAme dynamics (i.e. serum DNAme as defined in Sets 2 and 3, before chemotherapy compared to after two cycles) correctly identified 78% and 86% of responders and non-responders (Fisher’s exact test, p = 0.04) overall and 78% and 100% of responders and non-responders among those women who were left without residual disease after interval debulking surgery (Fisher’s exact test, p = 0.007) (Table 3).

Prediction chemosensitivity		Responder		Non-responder
		All	R0/1	All	R0/1
CA125	Positive → negative	2/10 (20%)	2/8 (25%)
	Positive → positive			6/8 (75%)	4/6 (66.7%)
DNAme	DNAme Positive → negative	7/9 (77.8%)	7/9 (77.8%)
	DNAme Positive or negative → positive			6/7 (85.7%)	6/6 (100%)

CA125 concentration of < 35IU/mL was deemed negative. Definitions of DNA methylation positivity are provided in Table 1. Note that amongst the 20 patients who had chemo-sensitivity data available, they were only included in the analysis if the pre- and/or post-treatment (after cycle 2) sample were positive (i.e. in 2 and 4 patients for CA125 and DNA methylation markers, respectively, neither sample was positive and hence response or lack of response could not be assessed)

Serum DNAme for early diagnosis of OC

In order to judge whether our marker panel is, in fact, capable of diagnosing OC early, samples predating OC diagnosis by up to two years (cases) and matched controls were used from the control (no screening) arm of the UKCTOCS cohort. The median time from venepuncture to serum separation was 21.97 h (interquartile range [IQR] = 19.91–24.34 h) for this sample set. As expected, both the amount of DNA/mL serum as well as the average DNA fragment size were substantially higher in UKCTOCS samples compared with the other samples used in this study (Fig. 5a and b). This is potentially due to DNA from WBCs leaking into the serum during the sample transport time, particularly during the warmer months of the year (Additional file 1: Figure S11). Nevertheless, a small proportion (on average, 19.9%) of the DNA consisted of smaller (50–250 bp) fragments representing DNA from apoptotic cells (including DNA from cancer cells) (Additional file 1: Figure S12). The “contaminating” majority of high-quality DNA not only dilutes the cancer signal but also skews the target sequence amplification towards WBC DNA. In order to adjust for these factors, we made an a priori decision to reduce the threshold for the three regions by a factor of 3 and split the analyses in samples above (high) and below (low) the median amount of DNA (Table 4). The three DNAme-marker panel was able to identify cases with a specificity of 88% and a sensitivity of 58%, when specifically assessed in samples with a DNA concentration lower than the median ng/uL value, and importantly which predated cancer diagnosis by up to two years (Table 4B and Additional file 3: Table S2). The sensitivity of the panel improved from 58% to 64% when exclusively assessing CA125-negative (<35 IU/mL) samples. As previously observed in the Set 3 analysis there is no overlap between CA125 and DNAme false-positive controls (Table 5). When directly comparing the performance of CA125 (applying a cut-off of 35 IU/mL) with the DNAme panel specifically in the “low” DNA samples, the DNAme panel had higher sensitivity (57.9% vs 42.1%) but lower specificity (88.1% vs 95.5%) compared to that of CA125 for the early detection of OC (Additional file 3: Table S3).

(A)	All samples			CA125-negative samples
	0–2 years	0–1 years	1–2 years	0–2 years	0–1 years	1–2 years
Specificity (CI%)	96.9% (90.8–98.6)	96.9% (90.8–98.6)	96.9% (90.8–98.6)	96.8% (91.5–99)	96.8% (91.5–99)	96.8% (91.5–99)
Sensitivity (CI%)	23.3% (12.3–39)	31.6% (13.6–56.5)	16.7% (5.5–38.2)	15.4% (5–35.7)	22.2% (3.9–59.8)	11.8% (2.1–37.8)
(B)	Low DNA samples			CA125-negative samples
	0–2 years	0–1 years	1–2 years	0–2 years	0–1 years	1–2 years
Specificity (CI%)	88.1% (77.3–94.3)	88.1% (77.3–94.3)	88.1% (77.3–94.3)	87.5% (76.3–94.1)	87.5% (76.3–94.1)	87.5% (76.3–94.1)
Sensitivity (CI%)	57.9% (34.0–78.9)	55.6% (22.7–84.7)	60.0% (27.4–86.3)	63.6% (31.6–87.6)	66.7% (24.1–94.0)	60% (17.0–92.7)
(C)	High DNA samples			CA125-negative samples
	0–2 years	0–1 years	1–2 years	0–2 years	0–1 years	1–2 years
Specificity (CI%)	90.3% (79.5–96)	90.3% (79.5–96)	90.3% (79.5–96)	90.2% (79.2–95.9)	90.2% (79.2–95.9)	90.2% (79.2–95.9)
Sensitivity (CI%)	12.5% (3.3–33.5)	30% (8.1–64.6)	0% (0–26.8)	0% (0–25.3)	0% (0–69)	0% (0–30.1)

		Controls (CA125)		Cases (CA125)
		Negative	Positive	Negative	Positive
3 DNAme marker panel	Negative	56	3	4	4
	Positive	8	0	7	4

We identified cancer-specific DNAme patterns in tissue and developed serum assays which we validated in various settings. Our findings suggested that cell-free DNA has the potential to detect ovarian cancer up to two years in advance of clinical diagnosis. Nevertheless, a cell-free DNA based test will have to overcome several hurdles before clinical implementation.

In healthy individuals, cell-free DNA is present at concentrations in the range of 0–100 ng/mL and an average of 30 ng/mL [45]. DNA derived from tumor cells is shorter than that from non-malignant cells in the plasma of cancer patients [46]. Our overarching goal was to develop DNAme-based markers for early OC detection. In order to realize this aim, samples must be derived from large population-based screening studies, such as the UKCTOCS trial, that have samples banked years ahead of diagnosis. Serum samples from ~ 100,000 women need to be collected to accrue sufficient OC numbers (i.e. a range of 40–50). Within the UKCTOCS, which involved collection and banking of serum samples from over 200,000 women, whole blood samples were couriered to the central laboratory within 2–48 h. Prospectively collected blood samples were spun down 2–12 hours after collection to mimic the UKCTOCS setting. The UKCTOCS, and to a lesser extent the other prospectively collected sets, contained higher than average amounts of cell-free DNA and fragments were longer on average. Both factors reflect the leakage of WBC DNA into serum. In order to compensate for this, we aimed for an extremely high coverage but noted that four of the UKCTOCS samples had at least one of the three markers with a coverage 100,000; we had made an upfront decision not to exclude any of the samples based on lower coverage but note that this is one of the limitations of this study. Despite these complicating factors the three-DNAme marker panel outperformed CA125 using a 35 IU/mL cut-off in detecting OC early in the group of women who had a DNA concentration lower than the median ng/uL value.

In order to provide further functional proof that the newly developed serum DNAme marker panel is cancer-specific and able to indicate the presence of active OC, despite competing with high levels of background WBC DNA inherent within the trial samples analyzed, we demonstrated that our serum DNAme-dynamics correctly identified 7/9 and 6/7 Carboplatin responders and non-responders, respectively.

As we did not observe an overlap between false-positive CA125 and false-positive DNAme samples, it is probable that DNAme false positivity is largely triggered by technical artefacts as a result of extremely low thresholds down to a pattern frequency of 0.000003 (i.e. three cancer patterns in the background of 1,000,000 DNA fragments with a non-cancer pattern). Of note is that for the serum sets, which have been prospectively collected within EpiFemCare, there was a substantial age difference between women who presented with benign pelvic masses and women who presented with OC. This age skew was completely intentional as our main purpose was to benchmark DNAme markers against CA125 false-positive controls and to assess whether CA125 false-positive controls are also DNAme-false positive. The main sources of false positivity are endometriosis, pelvic inflammatory disease, and fibroids—all conditions which are substantially more prevalent (or occur exclusively) in premenopausal (i.e. younger women), whereas OC is far more prevalent in older women. False CA125-positivity can usually be explained by a CA125-producing benign condition [47].

At the UKCTOCS prevalence screen [39], the ROCA identified elevated/intermediate risk in 0.93% of women, of whom 0.9% (42/4642) were diagnosed after repeat CA125 testing, ultrasound, additional imaging, and clinical assessment. Applying the three-marker DNAme test, with a conservative (i.e. excessive background DNA will not be an issue in prospective samples) estimate of specificity and sensitivity of 90% and 60%, respectively, as a second line test to ROCA-elevated women at risk could substantially decrease the time to diagnosis in at least half the women with OC.

OC is a low prevalence disease (i.e. lifetime risk in the general female population is 1–2 per 100 women [48]). The consequence of a positive screening test is an operation under general anesthesia (i.e. laparoscopic or open) to remove one or both ovaries/Fallopian Tubes. Hence, a high specificity of the screening test is of the utmost importance because the positive predictive value strongly depends on the prevalence of the disease and the specificity of the test [34]. Using a highly sensitive marker panel, able to detect > 80% of stage I/II OCS (i.e. a combination of CA125 and HE4 [49]), to pre-screen the entire population in order to narrow down the group of women who have a high likelihood of OC followed by the cell-free DNA test is a highly promising strategy to achieve a stage shift with at least 50% of cancers (instead of 25% in the absence of screening) diagnosed in stage I/II.

In addition to the use of serum and high levels of contaminating normal DNA from blood cells, the current work has some further limitations. First, the number of samples (specifically when considering only the low-DNA samples) in the UKCTOCS cohort was limited. Second, we were unable to assess whether the panel is specific for OC or whether it may additionally detect other cancer types. In serum set 3, we also analyzed serum samples from 37 patients with non-OCs (ten, 20, and seven with cervical, endometrial, and colorectal cancers, respectively). Two of ten (20%), 5/20 (25%), and 1/7 (14%) were deemed positive based on the final three-marker panels. This may indicate that our panel also detects other cancers arising from the Mullerian tract (i.e. cervical and endometrial cancers). In order to further elaborate on the aforementioned, we assessed the TCGA data. Whereas there was no CpG site on the 450 k methylation array for region #141, for regions #204 and #228 there were two (cg15015892 and cg05021743) and one (cg22344703) CpGs, respectively, represented on the Illumina array. Also, other cancers could potentially be identified using these markers (Additional file 1: Figure S13). Third, we did not directly compare the methylation levels in the primary tumor and the matched serum samples. As we have shown (Additional file 1: Figure S6), methylation levels in the primary cancers are relatively homogenous across different stages of OCs. Hence, any differences in methylation levels detected in the serum reflect conditions such as cancer cell turnover, release of cell-free DNA via lymph-vessels into the bloodstream, and half-life in the circulation—all factors which cannot be assessed by directly measuring DNAme in the cancer.

Our method of defining tumor-specific methylation patterns and quantifying the molecules exhibiting such patterns, instead of determining methylation levels, shows promising results regarding its applicability in liquid biopsy testing. While, in this study, the most promising tumor markers were all fully methylated, the method, per se, is not biased towards hyper- or uniformly methylated patterns and, as such, is also applicable to situations where the markers of interest show hypo- or heterogenous methylation, respectively.

Overall, our study provides a proof of principle that serum DNAme markers have the potential to detect OC within two years in advance of diagnosis and may therefore be able to guide personalized OC treatment. The recent advance of purpose-made blood collection tubes that stabilize cell-free DNA and prevent leakage of DNA from blood cells [50] will facilitate clinical implementation of DNAme pattern detection in cell-free DNA as a clinical tool in cancer medicine. In addition, recent evidence demonstrates that using DNAme patterns will allow for tissue-of-origin mapping in circulating cell-free DNA [51, 52] which supports the view that a DNAme marker panel is likely to cover a number of tumor-entities.

Overall, and for the first time, our study suggests that serum DNAme markers have the potential to diagnose OCs up to two years in advance of current diagnosis and may potentially enable individualized OC treatment. The recent advance of purposed blood tubes will facilitate clinical implementation of DNAme pattern detection of cell-free DNA as a clinical tool in cancer medicine.