We used the MIRA assay in combination with genome-wide CpG island arrays containing 27,800 loci to identify methylated CpG islands in 5 breast cancer cell lines (HCC1806, HCC1419, HCC1143, HCC1937, HCC1395) relative to normal human mammary epithelial cells (NHMEC) 7 8 . A primary selection based on probes showing > 5 fold enrichment of the methylated fraction in breast cancer cell lines versus NHMEC cell line DNA across 3 or more cell lines or a more stringent selection with probes scoring greater than 10 in 4 or more cell lines yielded a large number of loci showing methylation across breast cancer cell lines (see additional 1). These included genes/loci previously reported to be methylated in breast cancer (for example SIM1, PAX6, DLX4, RUNX3). In order to reduce the list to manageable numbers we carried out bioinformatic analyses to gather information on the expression of target genes in breast tissue, confirmation of the presence of a CpG island within the likely promoter region of target genes and functional annotation to determine likely roles in tumourigenesis, see materials and methods for more information. This led to a selection of 32 novel genes for further analysis (Figure 1, see additional file 2) in a stepwise fashion including the following stages:

1. Methylation analysis in breast cancer cell lines and normal control breast tissue DNA.

2. Methylation analysis in a panel of breast tumours (n = 40)

3. Gene expression analysis before and after 5-aza-dC treatment

4. Methylation analysis in an independent cohort of paired normal/tumour breast DNA panel (n = 20)

5. Expand methylation analysis to include lung, colorectal and prostate cancer, for the genes showing tumour-specific methylation in breast cancer.

We used COBRA digest to ascertain the methylation levels across a panel of 9 breast cancer cell lines (including the 5 lines used for MIRA) in CpG island regions upstream of the transcription start site. We confirmed frequent methylation (>30%) for 16 loci in breast cancer cell lines. These loci were unmethylated in DNA from normal/control breast. Sixteen loci either showed low or no methylation in the panel of breast cancer cell lines or were equally methylated in normal control breast DNA in the region analysed by our COBRA primers and were removed from further analysis (Figure 1). The remaining 16 positive genes demonstrating frequent methylation in breast cancer cell lines were analysed further by COBRA digest in a panel of primary breast tumours (n = 40). Nine (CIDE-A, COMP, DBC1, EMILIN2, EPSTI1, FBLN2, SALL1, SESN3, SIM2) of these genes (56%) demonstrated frequent methylation in primary breast tumours (frequency ranging from 26-63%) (Figure 1, 2, 3, 4, 5, 6)

The above nine frequently methylated genes were shown to be expressed in normal breast tissue (Figure 2, 3, 4, 5, 6). In order to determine the role of methylation in gene silencing, we treated breast cancer cell lines with the demethylating agent 5-aza-dC. Eight (CIDE-A, COMP, DBC1, EMILIN2, EPSTI1, FBLN2, SALL1, SIM2) of the above nine genes showed re-expression after treatment of the methylated cell lines with 5aza-dC (Figure 2, 3, 4, 5, 6). Hence methylation suppressed gene expression, whilst SESN3 showed no change in expression before and after treatment in both methylated and unmethylated cell lines. SESN3 was therefore not considered for further analysis.

To determine if methylation of the eight remaining genes was cancer-specific (i.e. tumour acquired), we analysed an independent set of paired normal/tumour breast DNA samples (n = 20). Five genes (62.5%) demonstrated tumour acquired methylation (EMILIN2, SALL1, DBC1, FBLN2 and CIDE-A) (Table 1, Figures 2, 3, 4, 5, 6), whilst the remaining 3 (COMP, EPSTI1, SIM2) showed equal methylation in the corresponding normal breast tissue DNA. The COBRA results for these 5 genes were confirmed by cloning and sequencing of bisulphite modified DNA in primary breast tumours (Figures 2, 3, 4, 5, 6). In majority of the breast tumour samples we also detected the unmethylated product. This is likely to be due to the contamination from normal cells, since none of the tumours were microdissected and due to the heterogeneous nature of the primary tumour samples.

To explore whether the above 5 genes showing tumour-specific methylation in breast cancer were methylated in other epithelial cancers besides breast, we analysed a panel of cancer cell lines corresponding to three common epithelial cancers (lung, colorectal and prostate). Lung tumour cell lines (NSCLC) demonstrated frequent methylation of CIDE-A, EMILN2, FBLN2 and SALL1. DBC1 methylation in lung cancer has been reported previously hence it was not analysed in this study 9 . Colorectal tumour lines showed frequent methylation of all 5 genes, whilst prostate tumour lines demonstrated frequent methylation of CIDE-A, DBC1, EMILIN2, SALL1 and less frequent methylation for FBLN2 (Table 2).

The above genes were further analyzed in primary tumours for the tumour types that demonstrated frequent methylation in the corresponding tumour cell lines. Where applicable N/T paired DNA was analysed (Table 2).

EMILIN2 was methylated in 33% colorectal carcinomas (CRC) and in 32% adenomas, whilst only 1.2% of corresponding normal tissues demonstrated any methylation. Methylation was confirmed using cloning and sequencing of bisulphite modified DNA (Figure 7). Whilst NSCLC and prostate tumours were either unmethylated or showed very low level frequency of methylation in both tumours and corresponding normals (Table 2).

SALL1 was methylated in 83% of CRC and in 89% of adenomas, whilst methylation frequency in corresponding normals was much less (38%). SALL1 demonstrated similar frequencies of methylation in tumour and matched normal samples for both NSCLC and prostate tumours (Table 2).

DBC1 was frequently methylated in CRC and prostate cancer cell lines (we did not analyse NSCLC since DBC1 is already known to be methylated in this cancer). In CRC and adenomas the methylation frequency was 100% but it was not cancer specific, all corresponding normals were also methylated. This could indicate that DBC1 undergoes tissue-specific methylation in certain tissues but not in others. Whilst weak methylation was seen in 1 out of 14 prostate tumours and 2 out of 4 normals (Table 2).

FBLN2 demonstrated tumour specific methylation in prostate tumours (36%) and was unmethylated in lung tumours. FBLN2 similar to DBC1 demonstrated very high frequency of methylation in colorectal tumours, adenomas and corresponding normal samples (Table 2).

CIDE-A was found to be methylated in 7/8 analysed primary CRC tumours and in all adenomas and corresponding normals. It has been recently demonstrated that CIDE-A expression is epigenetically controlled in certain tissue and cell types 10 , which when taken in consideration with our methylation analysis, suggests that normal colorectal tissue does not express CIDE-A. Whilst a similar percentage of methylation frequency was seen in prostate tumours and corresponding normals (21% and 25% respectively). CIDE-A was methylated in 1/16 NSCLC and 0/16 matched normals (Table 2).

The most interesting finding in the breast cancer cohort was that the methylation of EMILIN2 was associated with relapse (P = 0.031), presence of lymph node metastases (P = 0.047), with ER and PR positive tumours (P = 0.0009; 0.0082 respectively) and poor relapse free survival (P = 0.041) (Figure 8). CIDE-A methylation was associated with ER positive tumours (P = 0.016), whilst FBLN2 methylation was associated with PR positive tumours (P = 0.013).

EMILIN2 appears to be the most interesting gene due to the association between its methylation status and clinical-pathological features including survival. To confirm that the methylation of EMILIN2 is functionally relevant, we determined EMILIN2 expression in a subset of primary breast tumours (good quality RNA was available for these samples) using quantitative real-time RT-PCR. Expression analysis revealed that the breast tumours with EMILIN2 methylation showed lower levels of EMILIN2 expression compared to unmethylated tumours (Figure 9).

Five breast cancer cell lines (HCC1806, HCC1419, HCC1395, HCC1143 and HCC1937) and two normal human mammary epithelial cell lines (NHMEC lines 1585T and 3736T) were used for MIRA. CoBRA analysis used an additional four breast cancer cell lines (MCF7, T47D, MDA-MB-231 and HTB19) and one normal breast tissue DNA sample (AMS Biotechnology), replacing the NHMEC lines. Cell lines were grown as described previously 31 32 . Clinical and pathological features of the breast cancer cell lines are described in reference 31 and additional file 3. A set of 40 breast tumours (ductal carcinomas) and an independent set of further 20 breast tumours (ductal carcinomas) with corresponding normal breast tissue DNA were analysed. Additional file 4 gives the clinical-pathological information for the breast tumours used in this study. Ethical guidelines were followed for sample collection. A cell line panel consisting of nine colorectal, fifteen lung and five prostate cancer cell lines was used for initial analyses in these cancers. Further work was then carried out using colorectal adenomas and tumours, non small cell lung carcinomas and prostate tumours together with corresponding non cancerous tissue.

The MIRA assay was carried out on five breast cancer cell lines (HCC1806, HCC1419, HCC1395, HCC1143 and HCC1937) and two normal mammary epithelial cell lines (NHMEC lines 1585T and 3736T). The MIRA assay and chip hybridisation procedure was carried out as described previously in references 7 and 8. Human CpG island microarrays containing 27800 CpG islands covering 21 MB were purchased from Agilent Technologies.

Methylation enrichment factors of tumour to normal were calculated for all probes. This data was then analysed as follows:

• First selection, a single probe with a greater than 5-fold increase in MIRA-enriched fraction of tumour versus normal in 3 or more cell lines. This selected for 14,875 probes representing 6,912 individual hits.

• From these hits were removed any that did not represent a known, named gene. This selection removed 145 Open reading frames (Orfs), 3225 Chromosomal regions, 29 miRNAs, 143 Predicted genes (LOCs, MGCs, FLJs and KIAs) and 23 Agilent Plate related hits. This selection retained 3,347 genes (10,840 probes) which represent a low stringency candidate list.

• The low stringency candidate probe list was further refined by the removal of all probes not designated to be within promoter or divergent promoter regions. This selection retained 3,352 probes representing 1268 genes.

• Finally, a stricter selection criteria of a single probe with a greater than 10-fold increase in 4 or more cell lines was applied. This selection retained 895 probes representing 276 genes which represent a high stringency candidate list (see additional file 1)

A large number of target genes were identified using both approaches, particularly the low stringency approach. The low stringency candidate list was compared to a publically available expression array data for an experiment demonstrating the re-expression of genes in the MCF7 breast cancer cell line after treatment with a de-methylating agent (GEO data base accession ID GSE5816). Genes in both the low stringency candidate list and the re-expressed list were then considered. Functional annotation of target genes from both lists was carried out using the functional annotation table function in DAVID http://david.abcc.ncifcrf.gov 33 34 . Additional file 2 shows the results from this analysis for the 32 target genes chosen to be analysed in this study. We also carried out literature searches using Pubmed http://www.ncbi.nlm.nih.gov/pubmed/ for the keywords (breast cancer, Cancer & Methylation) and searched SAGE expression data(GeneCards-http://www.genecards.org/cgi-bin/cardsearch) and Oncomine https://www.oncomine.org/. UCSC genome browser http://genome.ucsc.edu/ was used to confirm the presence of CpG islands within the likely promoter regions of genes.

Combined bisulphite restriction analysis (CoBRA) was used to analyse all genes in this study. Bisulphite modification of DNA was carried out as described previously 35 . Semi-nested CoBRA primers were designed to amplify specific regions within the CpG island close to the start of transcription for all 32 genes (primer sequences are shown in additional file 5 for the final 5 positive genes). Digestion of PCR products with BstUI (CGCG) or TaqαI (TCGA) was carried out overnight at 37°C or 65°C respectively prior to visualisation on a 2% agarose gel. In vitro methylated DNA was used as a positive control. Confirmation of CoBRA results was obtained by bisulphite sequencing. PCR products from selected samples were cloned into the pGEM-T easy vector (Promega) according to manufacturers' instructions. Single colony PCR was carried out on up to 12 colonies for each cloned sample and methylation index (MI) for each sample calculated as a percentage; number of CpG dinucleotides methylated/total number of CpG dinucleotides sequenced × 100.

Cell lines were maintained in DMEM media (Sigma Aldrich) supplemented with 10% FCS and 2 mM glutamine at 37°, 5% CO₂. Breast cancer cell lines HCC1806, HCC1419, HCC1937, HCC1395, MCF7, HTB19, T47D and MDA-MB-231 were treated with 5 μM 5-aza-2'-deoxycitidine (5-aza-dC). Treatment was carried out over 5 days with daily media changes and addition of fresh 5-aza-dC. RNA bee (AMS biotechnology) was used to extract RNA from 5-aza-dC treated and non-treated cells. Control breast total RNA was bought from AMS Biotechnology. cDNA was synthesised from 1 μg total RNA using SuperScript III (Invitrogen) and random hexamer primers. Touchdown PCR was used for expression analysis (primer sequences are shown in additional file 6 for the final 5 positive genes). GAPDH analysis was carried out concurrently for each gene using primers and PCR conditions described previously 35 . For each gene we used 2 methylated and one unmethylated cell line for expression analysis.

Quantitative real-time RT-PCR was performed as described previously 36 . Briefly, cDNA was made from total RNA extracted from frozen normal and tumour breast tissues. We quantified transcripts of the TBP (TATA box-binding protein) gene as the endogenous RNA control. Each sample was normalized on the basis of TBP content. Results, expressed as N-fold differences in target gene expression relative to the TBP gene (termed Ntarget), were determined by the following formula: Ntarget = 2Δ^Ctsample, where the ΔCt value of the sample was determined by subtracting the average Ct value of the target gene from the average Ct value of the TBP gene. The Ntarget values of the samples were subsequently normalized such that the mean ratio of the normal breast samples would equal a value of 1. The nucleotide sequences of the primers used for PCR amplification were as follows: (a) EMILIN2-U (5'-GCTTGAAGAAAGCCACAGATAATGAA-3') and EMILIN2-L (5'-CTGAGGCCCCTCTTTTGGATCTA-3') with a EMILIN2-specific product size of 116 bp, (b) TBP-U (5'-TGCACAGGAGCCAAGAGTGAA-3') and TBP-L (5'-CACATCACAGCTCCCCACCA-3') with a TBP-specific product size of 132 bp. PCR was performed using the SYBR Green PCR Core Reagents kit (Perkin-Elmer Applied Biosystems). The thermal cycling conditions comprised an initial denaturation step at 95°C for 10 min and 50 cycles at 95°C for 15 s and 65°C for 1 min. Experiments were performed with duplicates for each data point.

Methylation results were summarised using cross-tabulations. Comparisons of methylation rates of genes within the same sample were analysed using McNemar's tests, and exact P-values were obtained. Comparisons of methylation rates between normal and tumour cells were analysed using chi-squared tests. Associations between methylation status and clinicopathological characteristics were assessed using a two-sample t-test for the continuous variable age and chi-squared tests for categorical variables. All analyses were conducted using Stata Version 10.1 (Stata Corporation: College Station, TX, USA).