Background

Breast cancer is the most common cancer in women, affecting at least 1 in 10 women in the western world. The potential to predict breast cancer and offer preventive measures is an effective intervention in women with an inherited predisposition to breast cancer due to mutations in BRCA1/2 genes [1]. However, these account for less than 10% of breast cancers [2]. While extensive genome-wide association studies have identified a number of single nucleotide polymorphisms (SNPs) associated with breast cancer risk [3], epidemiological models that include risk associated SNPs yield a receiver-operating-characteristic (ROC) area under the curve (AUC) of only 62%, a modest 4% improvement over the AUC of epidemiological models [4].

Predicting the likelihood of breast cancer development is therefore still challenging not only because the sensitivity of current strategies is low [4] but also because 11% to 52% of screen-detected breast cancers may be an over diagnosis of cancers which would have never become clinically evident [5–7]. Hence a biomarker that could predict the risk of developing breast cancer particularly in those with a poor prognosis and which is also independent of familial predisposition is urgently needed.

It is known that epigenetic variation contributes to inter-individual variation in gene expression and thus may contribute to variation in cancer susceptibility [8–10]. DNAme is the most studied mechanism of epigenetic gene regulation and represents a biologically and chemically stable signal. Aberrant DNA methylation is also a hallmark of cancer [9, 11], in particular increased promoter DNAme at stem cell differentiation genes (Polycomb-Repressive Complex 2 (PRC2) Group Target genes (PCGTs)) [12–19]. Initial evidence suggests that BRCA1 is a key negative modulator of PRC2 and that loss of BRCA1 inhibits stem cell differentiation and enhances an aggressive breast cancer phenotype by affecting PRC2 function [20]. Several proof of principle studies using a target gene approach or assessment of global DNA methylation analysing samples collected at the time of diagnosis provided the first evidence for the feasibility of breast cancer risk prediction using DNA methylation based markers [21–32]. It was also recently demonstrated that DNAme profiles in blood are able to predict cancer risk (on average 1.3 years in advance) within a group of women whose sisters had developed breast cancer [33].

Here we tested the hypothesis that women with an extremely high breast cancer risk (due to a BRCA1 mutation) carry a specific methylation signature in peripheral blood cells, which is also able to predict sporadic breast cancer incidence and death. We also tested whether this signature is tissue-specific.

Methods

Data from three different studies were used.

BRCA1 study

We analysed whole blood samples from two cohorts of BRCA1 mutation carriers and controls without a BRCA1 mutation (see Figure 1 and Additional file 1).

Figure 1
figure 1

Study design and identification/validation of the BRCA1 -mutation DNAme risk signature. AUC, receiver operating characteristics area under the curve; BC, breast cancer; FDR, false discovery rate; inv., invasive; WBC, white blood cells.

Full size image

MRC National Survey of Health and Development (NSHD)

We analysed both blood cells and buccal cells from a sample of women from the NSHD, a birth cohort study of men and women born in Britain in March 1946 [34–36]. A total of 152 (75 cancer cases and 77 controls) women were selected from those who provided both a peripheral blood and a buccal cell sample at the age of 53 years in 1999, who had not previously developed any cancer and who had complete information on epidemiological variables of interest and follow-up. We analysed >480,000 CpGs (using the Illumina 450 k array) in the 46 women who developed an invasive non-skin cancer (19 breast cancer, 5 reproductive tract and 22 other cancers; diagnosed 1 to 7 years after 53 years and an average of 4.75 years) and in the women (n = 77) who did not develop any cancer during the 12-year follow-up (for descriptive analysis see Additional file 2).

United Kingdom Collaborative Trial of Ovarian Cancer Screening (UKCTOCS)

We analysed serum DNA samples (which largely represent white blood cell DNA in this cohort - see Additional files 3 and 4) from postmenopausal women who developed breast cancer (n = 119) or remained cancer-free during the follow-up period (n = 122, maximum of 12 year follow-up (2001 to 2013)).

Ethics

All studies were approved by the relevant research ethics committee or institutional review board. Informed consent was obtained by all volunteers and conforms with the Declaration of Helsinki. The BRCA1 study was approved by the ethics committee of the General University Hospital, Prague (No. 1199/07 S-IV). The NSHD epigenetics study was approved by the Central Manchester Research Ethics Committee (REC reference: 07/H1008/168). UKCTOCS was approved by the UK North West Multicentre Research Ethics Committees (North West MREC 00/8/34). Ethical approval for this nested case control study was obtained from the Joint UCL/UCLH Committees on the Ethics of Human Research (REC reference: 06/Q0505/102).

DNA methylation analysis

The DNA from whole blood and tissues was extracted at UCL [36] and at Gen-Probe [37]. Methylation analysis was performed using the validated Illumina Infinium Human Methylation27 BeadChip [16] or the Illumina Infinium Human Methylation450 BeadChip for NSHD samples. The methylation status of a specific CpG site was calculated from the intensity of the methylated (M) and unmethylated (U) alleles, as the ratio of fluorescent signals β = Max(M,0)/(Max(M,0) + Max(U,0) + 100). On this scale, 0 < β < 1, with β values close to 1 (0) indicating 100% methylation (no methylation) (see Additional file 4).

Data availability

Data from two of the studies in this manuscript have been deposited in the Gene Expression Omnibus repository under the accession numbers (GSE58119), (GSE57285), (GSE32396). The NSHD data are made available to researchers who submit data requests to mrclha.swiftinfo@ucl.ac.uk; see full policy documents at [38]. Managed access is in place for this 68-year-old study to ensure that use of the data is within the bounds of consent given previously by participants, and to safeguard any potential threat to anonymity since the participants are all born in the same week.

Statistical analyses

Differential methylation analysis

From the BRCA1 study, differentially methylated CpGs, with false discovery rate (FDR) corrected P values, between BRCA1 mutant carriers and BRCA1 wild type samples were identified via a multivariate logistic regression that was adjusted for age, batch and the presence of cancer.

Ensemble signature identification

The elastic net classification method was chosen for our study as it has been shown to be particularly effective when the number of predictors is far greater than the number of training points [39]. The elastic net method, as implemented in the glmnet R-package [40], identified a classifier comprising 1,829 CpGs with non-zero regression coefficients (see Additional file 4).

Validation

To evaluate its predictive accuracy, the identified classifier was tested on two independent datasets: (1) NSHD, and (2) UKCTOCS. For each individual, risk scores, based on their methylation profiles, were estimated and correlated to their disease status. An AUC value was then obtained via Somers’ Dxy rank correlation [41] (see Additional file 4).

Results

DNA methylation signature in white blood cells (WBC) associated with BRCA1 mutation status

We analysed DNAme of 27,578 CpGs in WBC samples from a total of 72 women with a known BRCA1 mutation and 72 women with no mutation in the BRCA1 or BRCA2 gene (Figure 1 and Additional file 1). The presence of a cancer has been shown to modulate the composition of WBCs and DNAme profiles in peripheral blood [42] and hence we used a mixture of women who did and who did not develop breast cancer in order to be able to adjust for this. Using a multivariate regression model that included age, cohort and cancer status as covariates we were able to rank CpGs according to the significance of the association between their DNAme profile and mutation status. On applying a relaxed threshold of FDR <0.3 we observed a total of 2,514 BRCA1-mutation associated CpGs, of which 1,422 (57%) were hypermethylated (hyperM) and 1,092 (43%) were hypomethylated (hypoM) in women who had a BRCA1 mutation (Figure 1, Additional file 5), representing a highly significant skew towards hypermethylated CpGs (Binomial test P < 1e-10). To arrive at a specific DNAme signature, which would allow classification of independent samples, we used the elastic net (ELNET) framework (see Additional file 4), which resulted in a signature consisting of 1,829 CpGs (Figure 2, Additional file 6).

Figure 2
figure 2

CpGs (n = 1829), which are differentially methylated in WBCs between BRCA1 mutation carriers and wild type controls and which comprise the ‘ BRCA1 -mutation DNA methylation signature’. Heatmap of normalised methylation values (blue = relative high methylation, yellow = relative low methylation) of CpGs comprising the BRCA1-mutation DNAme signature. The first colour bar at the top denotes the two main clusters where ‘red’ reflects the samples with a BRCA1 mutation whereas ‘green’ reflects samples without a mutation in BRCA1 or BRCA2 gene. The distribution of cancer cases is given in the second colour bar indicating women who had developed a breast cancer in purple. Right panel shows the enrichment of the top components of the gene set enrichment analysis in the hyper- and hypomethylated subset of CpGs; PCGT; Polycomb repressor complex 2 Group Target. Dashed line separates hypermethylated from hypomethylated CpGs.

Full size image

Given that PCGT methylation is a hallmark of almost all cancers and that a BRCA1 defect in normal non-neoplastic cells is likely to silence PCGTs and compromise cell differentiation [20], we posited that our BRCA1 DNAme signature may be able to predict sporadic breast cancer. Interestingly, Gene Set Enrichment Analysis (GSEA) [43, 44] on the 1,074 hypermethylated (Additional file 7) and 755 hypomethylated (Additional file 8) CpGs of the BRCA1- mutation signature demonstrated the association of BRCA1 mutation with promoter hypermethylation of PCGTs. Indeed, the top categories of genes, associated with the hypermethylated CpGs in BRCA1 mutation carriers, were significantly (P <10-10) enriched for stem cell PCGTs irrespective of the definition used (Figure 2, Additional file 7). In contrast, none of the gene categories associated with those CpGs which are hypomethylated in BRCA1 mutation carriers reached significance based on adjusted P values (Additional file 8). Even the GSEA on the 105 CpGs with a more stringent FDR (<=0.05) associated with BRCA1 mutation in white blood cells demonstrated the enrichment of PCGTs (P < =0.02) (Additional file 9).

BRCA1-mutation DNAme signature and breast cancer risk in peripheral blood cells in the NSHD

In order to test whether the BRCA1-mutation DNAme signature is able to identify women who will develop breast cancer we analysed one of the best available characterised longitudinal cohorts (Additional file 2). Applying the BRCA1-mutation DNAme signature (out of the 1,829 BRCA1 CpGs, 1,722 were present on the 450 k Illumina methylation array), yielded a breast cancer risk AUC = 0.65 (0.51 to 0.78, P = 0.02) (Figure 3A). Interestingly, the BRCA1 signature also significantly predicted the future development of invasive non-breast cancers (AUC = 0.62; 0.50 to 0.74; P = 0.04) (Additional file 10A).

Figure 3
figure 3

Validation of the BRCA1 -mutation DNAme signature in two independent prospective cohorts. ROC curves and AUC statistics to predict future breast cancer (BC) incidence applying the BRCA1-mutation DNAme signature in white blood cells (WBCs) (A) and in buccal (BUCC) cells (B) of the NSHD cohort and in serum DNA of the UKCTOCS cohort (C). Overlap of the top CpGs differently methylated in WBC between BRCA1 mutant and wild type (BRCA1 study) and the top CpGs differently methylated in serum DNA between women who have developed oestrogen receptor positive BCs and women who remained cancer-free (D). ROC curve and AUC statistics to predict deadly BCs applying the BRCA1-mutation DNAme signature in serum DNA in the UKCTOCS cohort (E) and Kaplan Meier curve (and hazard ratio (HR)) of future breast cancer patients with a high and low BRCA1-mutation DNAme score in serum DNA (F).

Full size image

Consistent with the view that DNAme is tissue-specific, our DNAme signature - derived from peripheral blood cells from women with known BRCA1 status - was not able to predict invasive breast cancer (Figure 3B) or invasive non-breast cancer (Additional file 10B) in the buccal cell DNAme profiles obtained at the same time from the same women who provided blood DNA.

BRCA1- mutation DNA methylation signature and breast cancer risk in serum DNA in the UKCTOCS cohort

Less than 10% of invasive breast cancers are due to a BRCA1 mutation [45] and therefore it is unlikely that the predictive capacity of the BRCA1-mutation DNAme signature in the NSHD cohort was due to the correct identification of BRCA1 mutation carriers. Nevertheless in order to further substantiate that the BRCA1-mutation DNAme signature identifies sporadic cancers, we performed a nested case–control study within the UKCTOCS cohort (a 202,638 postmenopausal women cohort, who based on their family history were not at an increased risk of ovarian or breast cancer - see Additional files 3 and 4). As BRCA1-associated cancers are far more likely (75%) to be oestrogen receptor (ER) negative [46], we solely focused our analysis on women who provided a blood sample between 0.42 and 4.18 years (average 2 years) before they developed an ER positive invasive breast cancer (n = 119) and matched (on age at blood donation and recruitment centre) them to 122 women who did not develop a breast cancer during the follow-up period (5.61 to 12 years, average follow-up 11.92 years). As there was no whole blood DNA samples available from the women in UKCTOCS, we used serum-free DNA as a source of material for this analysis. Since >95% of blood samples were only spun down 24 to 48 h after the blood draw, it was important for us to identify the likely source of DNA in the serum samples. Although we were not able to definitely identify the source, the evidence clearly pointed towards an enriched for WBC DNA (see Additional file 11). The BRCA1-mutation DNAme signature predicted the development of an ER positive breast cancer with an AUC = 0.57 (0.50 to 0.64; P = 0.03) (Figure 3C) independent of whether the sample was taken less or more than 2 years prior to diagnosis (see Additional file 12). Importantly, the BRCA1- mutation DNAme signature also substantially overlapped with an ER + breast cancer specific risk signature (Additional file 13), which we derived de novo in the UKCTOCS cohort (P <2 x 10-33, Figure 3D). Of note, in the breast cancer specific risk signature, we also observed enrichment of biological terms, all crucially involved in stem cell differentiation and biology (Additional file 14). Again, these stem cell gene categories were only enriched among CpGs hypermethylated in cases, but not among CpGs hypomethylated in cases (Additional file 15). This observation is particularly pertinent given that NIPP1, PRC2, MSX1 and NANOG all suppress differentiation through occupation and suppression of specific gene sets.

BRCA1-mutation DNAme signature identifies women years in advance of fatal breast cancer diagnosis

In order to test whether the BRCA1-mutation DNAme signature is able to predict not only incidence but also breast cancer mortality we performed ROC statistics in the UKCTOCS set comparing women who died from breast cancer (n = 10) during the follow-up period with women who did not develop breast cancer (Figure 3E) and found an AUC = 0.67 (0.51 to 0.83; P = 0.02). In line with these findings women with a higher than average BRCA1-mutation DNAme signature score were 8.46 (95% CI 1.06 to 67.69) -fold more likely to die from breast cancer (P = 0.04) than those with lower than average scores (Figure 3F). Interestingly, apart from the number of nodes, none of the other clinico-pathological features or treatment modalities was associated with the BRCA1-mutation DNAme signature in these ER positive breast cancers (Additional file 16).

BRCA1-mutation DNAme signature and association with epidemiological and hormonal risk markers

Next, we were interested whether our DNAme signature could be explained by any of the breast cancer risk factors we had available for the UKCTOCS cohort. Interestingly, neither any of the epidemiological breast cancer risk factors nor any of the hormones (Tables 1, 2 and 3) we have analysed in the same serum samples was associated with our BRCA1-mutation DNAme signature. Interestingly, when we analysed women with and without a family history [47] separately, both BC incidence and death was predicted by our BRCA1-DNAme signature only in the group without family history (Additional file 17), but not in the (obviously very small) group of women with any family history (Additional file 18).

Table 1 Characteristics of the samples used from the UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS)
Full size table
Table 2 Additional characteristics of the samples used from the UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS)
Full size table
Table 3 Characteristics of the samples used from the UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS)
Full size table

Discussion

Here we have provided several novel lines of evidence indicating that DNAme profiles obtained in cells from women with a BRCA1 mutation have the potential to indicate future breast cancer development (and death) many years in advance of diagnosis. Our findings also show that genes encoding developmental transcription factors integral for stem cell differentiation and biology are hypermethylated in women predisposed to breast cancer.

Our data suggest that the BRCA1-associated DNAme signature is a risk predicting signature rather than an early detection signature, because: (1) the DNAme signature was derived from WBCs in women with a known BRCA1 status and was adjusted for cancer status (analysis included BRCA1 carriers without cancer at the time of sample draw); (2) the time from sample draw to diagnosis had no dramatic impact on the strength of association between DNAme and potential for breast cancer development; (3) the signature was validated in two independent cohorts; (4) we observed a very strong overlap of CpGs associated with BRCA1 mutation (BRCA1 study) and CpGs indicating future breast cancer risk (UKCTOCS); and finally (5) the signature was also associated with invasive non-breast cancers.

The observation that the top ranked hypermethylated BRCA1-mutation associated CpGs are highly enriched for PCGTs which we and others have previously shown to be an epigenetic hallmark of cancer tissue [12–18] and which are among the earliest, if not the earliest, molecular changes in human carcinogenesis [18] was an exciting finding because it fully supports recent data demonstrating that a BRCA1 defect leads to retargeting of the PRC2 and reduces cell differentiation.

Two key issues remain unclear. First, which factors lead to a BRCA1-mutation DNAme pattern in the absence of a BRCA1 mutation? It is likely that a combination of risk factors or factors which we have not captured (for example, early life events, transgenerational inheritance, and so on) contribute to epigenetic modifications which are in common to those associated with BRCA1 mutation. Second, is the BRCA1-mutation DNAme signature in WBCs functionally relevant or just simply an indicator of breast cancer risk? The fact that the signature is indicative of breast cancer mortality would support the view that subtle epigenetic mis-programming of immune cells may lead to general immune defects which in turn supports the development and proliferation of cancers. However, all these suggestions are highly speculative and need validation in further independent cohorts using well-defined subsets of blood cells or epithelial cells.

There are limitations to this study. First, we analysed whole blood DNA or serum DNA representing whole blood DNA and not a specific subset of peripheral blood cells. Second, although we found some good preliminary evidence that DNAme profiles in buccal cells are better at predicting future breast cancer risk (data not shown), we did not analyse buccal cells from BRCA1 mutation carriers, nor did we have access to independent prospective buccal cell data. Third, we used the 27 k array, instead of the 450 k array, to generate the BRCA1-mutation DNAme signature.

In summary, our data highlight DNAme analysis as a promising tool to predict future breast cancer development. Future epigenome-wide studies should focus on using epithelial cells like buccal - or epithelial cells from the uterine cervix which are hormone sensitive and more likely to capture an ‘epigenetic record’ of breast cancer risk factors. Such studies are more likely to provide the level of specificity and sensitivity which is required for a clinically useful risk prediction tool.

Conclusions

In summary, our DNAme signature derived from blood cells from BRCA1 carriers is able to predict breast cancer risk and death years in advance of diagnosis albeit with a modest AUC. Our data further support the notion that DNAme modification at stem-cell differentiation genes, even in unrelated tissues, is an early event associated with carcinogenesis.