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).

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.

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).

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).

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).

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