Biomarkers for detection and prognosis of breast cancer identified by a functional hypermethylome screen

Abstract

Breast cancer (BC) is a disease with diverse tumor heterogeneity, which challenges conventional approaches to develop biomarkers for early detection and prognosis. To identify effective biomarkers, we performed a genome-wide screen for functional methylation changes in BC, i.e., genes silenced by promoter hypermethylation, using a functionally proven gene expression approach. A subset of candidate hypermethylated genes were validated in primary BCs and tested as markers for detection and prognosis prediction of BC. We identified 33 cancer specific methylated genes and, among these, two categories of genes: (1) highly frequent methylated genes that detect early stages of BC. Within that category, we have identified the combination of NDRG2 and HOXD1 as the most sensitive (94%) and specific (90%) gene combination for detection of BC; (2) genes that show stage dependent methylation frequency pattern, which are candidates to help delineate BC prognostic signatures. For this category, we found that methylation of CDO1, CKM, CRIP1, KL and TAC1 correlated with clinical prognostic variables and was a significant prognosticator for poor overall survival in BC patients. CKM [Hazard ratio (HR) = 2.68] and TAC1 (HR = 7.73) were the strongest single markers and the combination of both (TAC1 and CKM) was associated with poor overall survival independent of age and stage in our training (HR = 1.92) and validation cohort (HR = 2.87). Our study demonstrates an efficient method to utilize functional methylation changes in BC for the development of effective biomarkers for detection and prognosis prediction of BC.

Keywords: :

• biomarker
• breast cancer
• detection
• functional hypermethylome screen
• prognosis

Introduction

BC is the most frequently diagnosed female cancer. In the last year alone, 30% of newly diagnosed cancer cases in the US were BC.¹ Recent comprehensive sequencing efforts have identified several gene mutations in BC, but most occur at low frequency and their function in the etiology and development of BC is not yet understood.^2,3 Importantly, these studies suggest that each individual BC harbors only a few mutations. For example, BRCA1, BRCA2, ATM and TP53, occur in only 5–25% of cases and there is great heterogeneity among tumors.^4-8

In contrast, gene expression profiling has provided: (1) a molecular classification of BC into clinically relevant subtypes; (2) new tools to predict disease recurrence and response to different treatments and; (3) new insights into oncogenic pathways and the process of metastatic progression. Several expression profiling studies have shown that BC can be grouped into four categories with composite expression profiles:^9,10 the triple negative subtype, the ERBB2-like subtype and luminal-like subtypes A and B. Additionally, gene-expression profiling have provided prognostic signatures including the 70-gene signature (Mammaprint),¹¹ the 76-gene signature (Rotterdam signature),¹² the 21-gene recurrence score (OncotypeDx)¹³ and genomic grade (MapQuant Dx).¹⁴

One potential way to further explore expression differences among breast tumors is through an examination of epigenetic changes, such as DNA methylation. Accumulation of gene promoter DNA hypermethylation is an early event in tumorigenesis leading to silencing of cancer related genes.¹⁵ Epigenetic biomarkers have shown promise in other tumor types for correctly stratifying patients including prediction of response to Temozolomide by MGMT methylation in glioblastomas and prediction of recurrence in early stage lung cancer by methylation of CDKN2A/P16 and RASSF1.^16,17 TFPI2, a colorectal cancer (CRC) specific methylated gene discovered by recent efforts to characterize the CRC hypermethylome,¹⁸ shows promise as an early detection biomarker for CRC in stool DNA-based assays.¹⁹ However, to accomplish these objectives in BC, we need to identify more sensitive and specific markers for this disease. Traditional candidate gene approaches, which have yielded epigenetic biomarkers in other tumor types, have had minimal success in BC and genome wide epigenetic screens have not previously identified biomarkers that easily translated into early detection or prognosis prediction.^20-23

In the current study, a comprehensive approach to define genome-wide functional methylation changes in BC (functional BC hypermethylome), by using whole transcriptome expression arrays on BC cell lines after pharmacological inhibition of DNA methylation, is presented.¹⁸ We find that approximately 100–500 genes may be hypermethylated and associated with transcriptional silencing in each primary BC. We have validated many of these genes in a series of primary BCs and find that methylation frequencies of these novel cancer genes show heterogeneity in primary tumors, with methylation frequencies ranging from 3% to 93%. We identified 33 cancer specific methylated genes and, among these, two categories of genes: (1) highly frequent methylated genes that detect early stages of BC and; (2) genes that show stage dependent methylation frequency pattern, i.e., lower methylation frequency in early stage BCs (stage 0 and 1) and a higher methylation frequency in advanced stage BCs (stage 3 and 4), which are candidates to help delineate BC prognostic signatures. Our efforts characterize the functional BC hypermethylome and provide novel sets of biomarkers for detection and prognosis prediction of BC.

Results

Gene discovery

We used our recently described genome-wide approach to identify functional methylation changes (functional cancer hypermethylome), i.e., genes silenced by promoter hypermethylation, in human cancer types for which representative cell lines are available.¹⁸ The functional BC hypermethylome, containing 2117 promoter-associated CpG island genes, was identified among four BC cell lines of varying metastatic potential [estrogen receptor (ER)/progesterone receptor (PR) positive MCF7 and T-47D and ER/PR negative MDA-MB-231 and MDA-MB-468] by our previously established criteria:^18,24 (1) no basal expression in untreated cells; (2) no re-expression with the histone deacetylase (HDAC) inhibitor Trichostatin A (TSA), and; (3) ≥ 0.5 log-fold re-expression after treatment with the demethylating drug 5-aza-2'-deoxycytidine (DAC). Two dimensional plotting of DAC vs. TSA induced gene expression changes, as analyzed on Agilent 4x44K microarrays, generated a characteristic spike of re-expressed genes for each of the BC cell lines (Fig. S1A). In order to enrich for functionally meaningful and frequently methylated candidate genes we determined the overlap between the four BC cell lines and identified genes that are methylated in more than one cell line. The overlap of the candidate hypermethylated genes is shown in Figure S1B.

To address the effectiveness of the pharmacological treatment and the sensitivity of the array approach, we examined seven guide genes (CDH1, CST6, HOXA9, PYCARD, SFRP1, TWIST1 and PLAU)—i.e., genes previously known to be hypermethylated and silenced in some of these BC cell lines (Table S1). We compared the re-expression status of these genes after DAC and TSA treatment by RT-PCR (Fig. S1C) with their location on the DAC vs. TSA scatter plots from the microarrays (Fig. S1A). All selected guide genes validated the microarray results by appearing within ≥ 0.5 log-fold DAC re-expression zone of the individual BC cell lines and by showing re-expression after DAC treatment in individual BC cell lines by RT-PCR.

Validation of cancer specific methylated genes

A hundred and 30 unique candidate genes, identified primarily from the overlap of the individual BC cell lines hypermethylomes (Fig. S1B), were selected for further downstream analysis. Table S2 shows the gene names and corresponding arrays in which these genes appeared.

We first determined the expression status of these 130 candidate genes in untreated and DAC treated BC cell lines by RT-PCR (Fig. S2A for validation schema and Fig. S2B for representative examples). Eighty-eight genes (67.7%, 88/130) showed lack of basal expression and DAC-induced re-expression in at least one of the corresponding cell lines from whose hypermethylome the gene was derived. The remaining 42 genes, either basally expressed or not re-expressed with DAC, were likely false positives and therefore given low priority. Based on these criteria, the expression validation rate for our microarray based strategy was 67.7%.

These 88 genes, which lacked basal expression and were re-expressed with DAC, were further tested for their methylation status in BC cell lines using MSP (Methylation Specific PCR). Eighty-one genes showed methylation in at least one cell line (Fig. S2C for representative examples of true positives and false positive). Since the ultimate goal of our approach was to identify genes that show cancer specific methylation, we further analyzed the methylation status of these 81 genes in a series of normal breast (NB) tissues (n = 6). Fourteen genes were methylated in more than one out of six NB tissues and were subsequently excluded from further analysis Fig. S2A). These stringent criteria yielded 67 cancer specific methylated genes that were silenced and methylated in BC cell lines and had no or low frequency methylation in NB tissues.

Forty genes out of 67 were randomly selected as a representative pool for gene methylation frequency analysis in a pilot cohort of 30 primary BCs (Hopkins training cohort). Table 1 summarizes the demographics and clinicopathological characteristics of patient samples. Out of the 40 genes, 33 genes [overall 42.5%, (33/40*67)/130] were methylated in primary BCs with frequencies ranging from 3% to 93% (Fig. 1A).

Table 1. Baseline characteristics of breast cancer patient cohorts

		Hopkins training cohort	Hopkins extended training cohort	Hopkins validation cohort	MDxHealth cohort
		n = 30	n = 97	n = 89	n = 132
Age, y	Median	53	54	54	57
	Interquartile range	45–61	46–63	44–63	48–66
Stage, no. (%)	DCIS	6 (20)	22 (21)	15 (17)	21 (16)
	Stage 1	6 (20)	31 (34)	24 (27)	19 (14)
	Stage 2	6 (20)	30 (31)	20 (23)	32 (24)
	Stage 3	6 (20)	6 (6)	24 (27)	51 (39)
	Stage 4	6 (20)	7 (7)	5 (6)	9 (7)
Grade, no. (%)	High	17 (57)	47 (49)	21 (24)
	Low to intermediate	13 (43)	42 (43)	32 (36)
	Unknown		8 (8)	36 (40)
Lymphovascular invasion, no. (%)	Present	9 (30)	20 (21)	27 (30)
	Absent	21 (70)	49 (50)	38 (43)
	Unknown		28 (29)	24 (27)
Perinodal invasion, no. (%)	Present	9 (30)	9 (9)	17 (19)
	Absent	21 (70)	25 (26)	8 (9)
	Unknown		63 (65)	64 (72)
Lymphnodes, no. (%)	Positive	17 (57)	24 (25)	41 (46)
	Negative	13 (43)	59 (60)	38 (43)
	Unknown		14 (15)	10 (11)
ER/PR receptor, no. (%)	Positive	21 (70)	45 (46)	46 (52)
	Negative	6 (20)	35 (36)	16 (18)
	Unknown	3 (10)	17 (18)	27 (30)
HER2 amplification, no. (%)	Positive	6 (20)	18 (19)	13 (15)
	Negative	18 (60)	50 (51)	55 (62)
	Unknown	6 (20)	29 (30)	21 (23)
Ki67 index, no. (%)	≥ 30%	11 (37)	33 (34)	34 (38)
	< 30%	10 (33)	34 (35)	30 (34)
	Unknown	9 (30)	30 (31)	25 (28)

Figure 1. Methylation frequencies of cancer specific methylated genes in primary BCs. Bar graph showing methylation frequencies (%) of 33 cancer specific methylated genes in the Hopkins training cohort (n = 30).

To summarize, starting with 130 selected candidates from the potential BC hypermethylome we verified 88 as basally silenced in untreated cells and re-expressed in DAC but not in TSA treated cells yielding in a microarray validation rate of 67.7%. Further downstream biological validation yielded a discovery rate of approximately two in three for identification of hypermethylated genes in cell lines (81/130) and approximately one in three for identification of cancer specific hypermethylated genes [42.5%, (33/40*67)/130]. Based on those discovery rates we hypothesize that the functional BC hypermethylome roughly consists of 100–500 genes [40% of the individual cell line hypermethylome sizes: T-47D: 94 (234); MDA-MB-468: 245 (612); MCF7: 318 (795); MDA-MB-231: 492 (1230)].

Identification of detection markers for bc using real-time beacon strategy

In order to develop the functional BC hypermethylome for translational purposes, we first tested the ability of using these novel genes for detection of BC. A subset of the cancer specific methylated genes were tested in a cohort (MDxHealth cohort) of BC of varying stages (n = 132; stages 0–4) along with a series of normal breast tissues (n = 105) in order to find a gene combination marker which could detect BC with high sensitivity and specificity using a highly sensitive real-time beacon-based strategy. Four genes including GREM1, HOXD1, NDRG2 and TAC1 were selected based on their high methylation frequency discovered in a test run. First, sensitivity for detection of BC was measured for each gene and each stage separately (Fig. S3). Then, combinations of up to three genes were analyzed and Fisher’s exact test was used to assess the significance of the resulting classification of cases and controls relative to the methylation status. Table 2 summarizes obtained values for single genes and the best gene combinations. The best result for a single marker was obtained with NDRG2 with a sensitivity of 88.9% and specificity of 93.9% for detection of BC. The highest sensitivity and specificity values were obtained by the combination of NDRG2 and HOXD1, with an overall sensitivity of 94.6% and a specificity of 89.8%. Addition of extra markers to the combination of NDRG2 and HOXD1 did not result in an improvement in sensitivity and specificity. This suggests that a relatively small number of highly sensitive and specific hypermethylated genes could result in optimal detection of breast cancer.

Table 2. Combinatorial beacon marker analysis for detection of breast cancers

(A) MDx Health (stages 0 to 4)
	#Cancer	#Normal	#Total	Sens*	Spec*
NDRG2	127	98	225	88.9	93.9
GREM1	127	99	226	69.3	91.9
TAC1	126	101	227	72.2	90.1
HOXD1	120	94	214	64.2	95.7
NDRG2 and GREM1	130	103	233	91.5	86.4
NDRG2 and TAC1	131	105	236	93.9	84.8
NDRG2 and HOXD1	129	98	227	94.6	89.8
GREM1 and TAC1	129	103	232	82.2	86.4
GREM1 and HOXD1	128	99	227	75.8	87.9
TAC1 and HOXD1	128	101	229	79.7	86.1
NDRG2 and GREM1 and HOXD1	130	103	233	94.6	82.5
NDRG2 and TAC1 and HOXD1	131	105	236	94.7	80.9
(B) MDx Health (DCIS/stage 0)
NDRG2 and HOXD1	21	100	121	85.7	89.9
(C) MDx Health (stages 0 to 1)
NDRG2 and HOXD1	40	100	140	90.0	89.9
(D) MDx Health (stages 0 to 2)
NDRG2 and HOXD1	70	100	170	92.9	89.9

* Maximum sensitivity/specificity trade offs were calculated using the fisher exact test

Detection strategies for breast cancer should be focused on the earliest stages of cancer to realize improvements in the management of this disease. Therefore, we tested the most promising gene combination on the earliest stages of breast cancer. The combination of NDRG2 and HOXD1 showed an overall sensitivity of 85.7% and a specificity of 89.9% for detection of DCIS (stage 0) and including stage 1 and stage 2 BCs into the assay, the sensitivity and specificity for detection of early stage BC (stages 0–2) improved to 90% and 92.86%, respectively (Table 2). Taken together, these data provide a new set of highly frequent cancer specific methylated genes that may be useful for the early diagnosis of BC.

Identification of novel prognostic markers for bc by correlating clinicopathologic patient data with gene methylation status

An urgent need in breast cancer treatment is the stratification of early stage BC patients into low and high risk recurrence patients for better therapy decision making. We utilized the Hopkins training cohort where we had well annotated clinical follow-up data for hypothesis-generation only. Using Pearson’s Chi-2 test, we tested for statistical association between common clinicopathologic patient characteristics and methylation status of the 33 cancer specific methylated genes evaluated in our Hopkins training cohort (n = 30). Eleven genes including CDO1, CKM, CRIP1, GREM1, KL, SFRP1, TAC1, TF, TNFRSF11B, TWIST1 and ZNF702 showed a significant association between gene methylation status and prognostic parameters associated with a worse survival outcome such as tumor stage, tumor grade, lymph node status, lymphovascular or perinodal invasion, estrogen/progesterone receptor status, and Ki67 proliferation index. Supplementary Table 3 summarizes the relationships between clinicopathological data and gene methylation in detail.

Table 3. Statistical association between clinocopathological characteristics and methylation frequencies (%) of genes in Hopkins extended training cohort

Hopkins extended training cohort (n = 97)
Stage*		Stage 0,1 and 2	Stage 3 and 4	p value
	CDO1	50.00	100	< 0.001
	KL	5.00	23	0.018
	SFRP1	24.00	54	0.026
	TAC1	48.00	100	< 0.001
	ZNF702	12.00	38	0.015
Grade*		Grade 1 and 2	Grade 3	p value
	CKM	17.00	36	0.038
	GREM1	29.00	57	0.010
Lymphovascular or Perinodal Invasion*		Absent	Present	p value
	TAC1	49.00	80	0.011
	TF	32.00	64	0.009
ER/PR receptor*		Positive	Negative	p value
	CRIP1	7.00	29	0.007
	TNFRSF11B	9.00	26	0.048
Ki67 Proliferation Index*		Low (< 30%)	High (≥ 30%)	p value
	ZNF702	9.00	28	0.042

* p values were computed using pearson chi 2 test

Given the promising results from this training cohort, we extended our cohort sample size to 97 primary BCs (Hopkins extended training cohort). We found CDO1 (p < 0.001), KL (p = 0.018), SFRP1 (p = 0.026), TAC1 (p < 0.001) and ZNF702 (p = 0.015) as more frequently methylated in higher stage tumors, CKM (p = 0.038) and GREM1 (p = 0.010) in higher grade tumors, TAC1 (p = 0.011) and TF (p = 0.009) in lymphovascular and perinodal invasive tumors, CRIP1 (p = 0.007) and TNFRSF11B (p = 0.048) in ER/PR negative tumors, and ZNF702 (p = 0.042) in tumors with a high Ki67 proliferation index (Table 3).

Since 10 of these genes were consistently associated with adverse clinicopathological features, we further tested whether they were also associated with differences in overall survival of BC patients in two independent cohorts, the Hopkins extended training cohort (n = 97) (Fig. 2A and Table S4A) and the Hopkins validation cohort (n = 89) (Fig. 2B and Table S4B). Clinicopathological characteristics of both cohorts are summarized in Table 1. We performed univariate and multivariate Cox regression analysis and applied the Sidak adjustment method to control for multiple testing. DNA methylation of CKM (HR = 1.82, p = 0.126 in Hopkins extended training cohort; HR = 2.68, p = 0.029 in Hopkins validation cohort) and TAC1 (HR = 2.31, p = 0.045 in Hopkins extended training cohort; HR = 7.73, p = 0.046 in Hopkins validation cohort) predicted for poor overall survival. CDO1 (HR = 2.06, p = 0.072 in Hopkins extended training cohort; HR = 2.12, p = 0.112 in Hopkins validation cohort) showed a trend toward statistical significance to predict worse outcome in both cohorts. A statistically significant association with poor overall survival for these genes was lost when adjusted for age and stage, due to stage dependency of methylation.

Figure 2. Single gene and gene combinations for prognostication of mortality risk in BC patients. (A) Forest plots depicting univariate and multivariate HRs and corresponding 95% CIs for overall mortality risk associated with DNA methylation of single or combinations of genes in Hopkins extended training cohort (n = 97) and (B) Hopkins validation cohort (n = 89). Vertical line at 1 indicates HR = 1. Single gene marker and selected statistically significant gene combinations are shown. (C) Univariate and multivariate Cox proportional hazards regression based survival curves for prognostic gene marker combination TAC1 and CKM in Hopkins extended training cohort and Hopkins validation cohort.

Next, we tested whether combinations of these genes would improve prognostic accuracy in our two cohorts. Methylation of TAC1 and CKM (HR = 2.55, p = 0.021 and HR = 3.15, p = 0.014), CDO1 and CKM (HR = 2.39, p = 0.026 and HR = 2.35, p = 0.070) and CDO1 and TAC1 (HR = 2.48, p = 0.017 and HR = 3.00, p = 0.032) was significantly associated with a worse outcome dependent on age and stage in both training and validation cohorts (Fig. 2A, B, Table S4A and B). Upon controlling for age and stage the marker combination TAC1 and CKM (HR = 1.92, p = 0.128 and HR = 2.87, p = 0.043) continued to predict for poor overall survival. The combination of three markers (CKM, CDO1 and TAC1) did not result in further prognostic improvement. Survival curves of the best prognostic gene marker combination TAC1 and CKM generated based on a Cox regression model are shown in Figure 2C.

Detection and prognostic biomarkers show distinct methylation frequency pattern

After we identified new potential markers for detection and prognostication of BC, we were interested in analyzing those marker genes in more detail to delineate a strategy for the prediction of a good detection vs. a good prognostic biomarker. For each marker gene we plotted the methylation frequency by tumor stage as obtained in our Hopkins training cohort (Fig. 3). We identified two groups of genes showing distinct methylation frequency patterns. One group of genes, including three (GREM1, HOXD1, and TAC1) of our four detection markers, showed a high methylation frequency (≥ 40%) in early stage BCs (stage 0 and 1). The other group of genes, including four (CDO1, CKM, KL, TAC1) of our five prognostic markers, showed a stage increasing methylation frequency pattern, i.e., lower methylation frequency in early stage BCs (stage 0 and 1) and a higher methylation frequency in advanced stage BCs (stage 3 and 4). Interestingly, TAC1, which is a detection and prognostic marker, fits the methylation frequency pattern of both groups, i.e., highly frequently (≥ 40%) methylated in early stage BCs and increasingly methylated from early to later stage BCs. NDRG2 and CRIP1 did not fit the methylation frequency pattern of their marker families. The early detection marker NDRG2 showed a 15% methylation frequency in early stage BCs using a highly stringent MSP assay, but a methylation frequency of about 80% by real-time beacon assay using methylation primers designed to be highly sensitive. This analysis underlines that potential marker for detection of BC can be identified by their highly frequent early stage methylation pattern while potential prognostic BC markers can be identified by a stage dependent methylation pattern and their association with adverse clinical parameters.

Figure 3. Distinct methylation frequency pattern for BC detection vs. BC prognostication marker. Methylation frequency of single marker genes (GREM1, HOXD1, NDRG2, TAC1, CDO1, CKM, CRIP1 and KL) was plotted against cancer stage as obtained in Hopkins training cohort (n = 30). Two groups, BC detection marker and BC prognostication marker, of genes showing distinct methylation frequency pattern were identified. Upper plot shows BC detection markers (GREM1, HOXD1, TAC1) that are characterized by methylation frequency ≥ 40% in early stage BC’s (DCIS and stage 1). Lower plot shows BC prognostication markers (CDO1, CKM, KL and TAC1) that are characterized by stage increasing methylation frequency. Note NDRG2, an early detection marker, did not reach the 40% methylation frequency cut off in early stage BCs by highly stringent MSP assay and CRIP1, a prognostic marker, did not follow the stage increasing methylation frequency pattern as seen with the other prognostic markers.

Discussion

In the current study, we used a comprehensive approach to define genome-wide functional methylation changes in BC by using whole transcriptome expression arrays on BC cell lines after pharmacological inhibition of DNA methylation. We have recently described this approach in CRC, where the CRC hypermethylome consists of 300–500 hypermethylated genes per tumor, including multiple novel candidate tumor suppressor genes.^18,19,25 We now expand our efforts into BC, where it appears that a typical functional BC hypermethylome contains approximately 100–500 methylated genes associated with transcriptional silencing per tumor. Initially, our expression microarrays revealed approximately 2000 candidate hypermethylated genes in BC cell lines. By validating a subset of 130 candidate hypermethylated genes for expression, using stringent criteria of lack of basal expression and re-expression with DAC but not with TSA, we gained a microarray validation rate of 67.7%. On further validation for methylation in cell lines and primary BCs, but lack of methylation in normal breast tissue, the discovery rate for identification of BC specific methylated genes was refined to 40%. Based on this discovery rate, we calculated that each breast primary tumor has roughly 100–500 methylated genes (40% of the individual cell line hypermethylome size). This wide range as compared with CRC (300–500) may reflect the greater heterogeneity of BC.

In the process of array validation, we have more completely characterized 33 new hypermethylated genes in BC with methylation frequencies ranging from 3% to 93%. We demonstrate two potential ways in which these hypermethylated genes in BC can be used. First, we identified genes including GREM1, HOXD1, NDRG2 and TAC1 that are hypermethylated at high frequency in early stage BCs. With these genes examined using a sensitive real-time beacon assay, these new targets could be important for detection of BC. Indeed, in this study we find that the combination of NDRG2 and HOXD1 detects BC with sensitivity up to 94% and specificity of 90%. Detection of DCIS and early stage BCs (stage 1 and 2) by NDRG2 and HOXD1 remained highly sensitive (90%) and specific (93%). To evaluate a clinical potential of these markers for early detection of BC, we suggest their validation in body fluids of early stage BC patients. Second, we identified another category of gene including CDO1, CKM, CRIP1, KL and TAC1, whose methylation frequency increases with tumor stage and whose methylation status correlates with adverse clinical prognostic indicators. These prognostic biomarkers may prove useful for therapeutic decision-making. In particular, the combination TAC1 and CKM appears promising, since simultaneous methylation of both of these genes is associated with a 3-fold increased risk of worse patient outcome. Again, a potential use of these markers in a clinical setting to stratify early stage BC patients into groups of low and high risk recurrence patients needs to be validated in a cohort of solely early stage BCs. The feasibility of clinically useful epigenetic markers in BC is becoming more evident, not only by this study, but also by a recently published study be Hill et al.,²⁶ which supports this paradigm as well.

To summarize, we have characterized the functional BC hypermethylome and estimate the size of the hypermethylome for an individual BC to 100–500 genes. We have also demonstrated the way in which these novel findings can be used in the management of this disease. First, through validation of a subset of highly frequent methylated genes in early stage BCs using the sensitive real-time beacon strategy, we were able to define the gene combination marker NDRG2 and HOXD1 as one of the most sensitive and specific methylation signature for detection of BC. Second, we utilized a subset of genes, which are less frequently methylated at early stage but higher at advanced stage, to develop prognostic biomarkers. While additional prospective studies will be needed for further validation of these genes in larger populations and body fluids of BC patients, this study not only stresses the importance and extends the knowledge of functional methylation changes in BC from that in other genome-wide epigenetic screens,^21-23,27,28 but also demonstrates an efficient and effective method for utilizing functional methylation changes in BC patients.

Materials and Methods

Cell culture, treatment and microarray analysis

BC cell lines (MDA-MB-231, MDA-MB-468, MCF7 and T-47D) were purchased from American Type Culture Collection (ATCC), cultivated and maintained in appropriate media containing 10% serum according to ATCC’s recommendations. The HCT116 derivative cell line lacking the major DNA methyltransferases DNMT1 and 3b (DNMT1^−/− and DNMT3b^−/−; Double Knockout or DKO) cells were maintained as previously described.²⁹ Pharmacological treatments of BC cells, RNA extraction, RNA probe preparation and hybridization on 4x44K Agilent arrays, data analysis and processing as well as promoter CpG island determination analysis using inclusive Gardiner-Garder and Frommer parameters³⁰ were performed as previously described.¹⁸

Gene expression and methylation analysis

Total RNA was isolated¹⁸ or purchased from Stratagene (normal human breast). cDNA synthesis and RT-PCR were performed as published.¹⁸ Primer sequences for RT-PCR were designed using Primer3. For MSP analysis, DNA was extracted following standard phenol-chloroform extraction and bisulfite modified using the EZ DNA methylation Kit^TM (Zymo Research). Primer sequences were designed using MSPPrimer³¹ and Methylation-specific PCR was performed as described by Herman, et al.³² RT-PCR and MSP primer sequences as well as PCR conditions are available upon request.

Real time beacon methylation analysis

Analyte quantifications were performed using real-time MSP assays, which consisted of parallel amplification/quantification processes with specific primers and probes for each analyte and Molecular Beacon^® assay formats on an ABI Prism^® 7900HT instrument (Applied Biosystems). The MSP results were generated using the SDS 2.2 software (Applied Biosystems), exported as Ct (cycle threshold) values, and then used to calculate copy numbers based on a linear regression of the values plotted on a standard curve of 8 - 0.8 x 10⁶ gene copy equivalents, using plasmid DNA containing the bisulfite modified sequence of interest.

Human tissue samples

Formalin-fixed, paraffin-embedded tissue from normal breast and primary BCs were obtained from the pathology archives of Johns Hopkins Hospital, USA in accordance with all rules and regulations of Institutional Review Boards and HIPAA compliance. All paraffin-embedded samples underwent independent review to confirm diagnosis by a pathologist (EG). DNA of normal human breast from women with no history of BC was purchased from Stratagene (n = 6). For real-time beacon methylation analysis an additional cohort (MDxHealth cohort) of 132 primary BCs and 105 normal breast tissues from non-cancerous patients were obtained from the pathology archives of the Johns Hopkins Hospital, USA, Université de Liège, Belgium, Biona, Belgium and Utrecht Medical Center, the Netherlands.

Statistical analysis

Clinicopathological characteristics and gene methylation status were correlated using Pearson’s Chi 2 test. Hazard ratios (HRs), confidence intervals (CIs) and survival curves were calculated using univariate and multivariate Cox regression (adjustment for age and stage). To control for multiple testing, Sidak adjustment method was used to correct for family-wise error rate. For BC detection markers, significance of classification 2X2 tables, i.e., cases with unmethylated genes or gene combinations vs. controls with unmethylated genes or gene combinations, was assessed by Fisher’s exact test. p values of less than 0.05 were considered significant. All statistical analysis was performed using the STATA 9.2 software package.

Abbreviations:
BC	=	breast cancer
HR	=	hazard ratio
ER	=	estrogen receptor
PR	=	progesterone receptor
HER2	=	human epidermal growth factor receptor 2
CRC	=	colorectal cancer
DAC	=	5-aza-2’-deoxycytidine
HDAC	=	histone deacetylase
TSA	=	trichostatin A
RT-PCR	=	Reverse Transcriptase PCR
MSP	=	methylation-specific PCR
IVD	=	in vitro methylated DNA
NB	=	normal breast

Acknowledgments

We thank Sharon Metzger and Theresa Sanlorenzo-Caswell from the Johns Hopkins Tumor Registry for their help with clinicopathological information on patients as well as Kathy Bender for administrative support.

Disclosure

James G. Herman and Stephen B. Baylin receive research funding and are consultants to MDxHealth. Leander Van Neste, Valérie Deregowski, Ilse Vlassenbroeck and Wim Van Criekinge are employees of MDxHealth.

Financial Support

This work was supported by the Susan G. Komen Foundation, Mary Kay Ash Foundation, American College of Surgeons-Society of University Surgeons fellowship, German Academic Exchange Service (DAAD), and Dr. Jost Henkel Stiftung.