Methylation profiling identified novel differentially methylated markers including OPCML and FLRT2 in prostate cancer

ABSTRACT

To develop new methods to distinguish indolent from aggressive prostate cancers (PCa), we utilized comprehensive high-throughput array-based relative methylation (CHARM) assay to identify differentially methylated regions (DMRs) throughout the genome, including both CpG island (CGI) and non-CGI regions in PCa patients based on Gleason grade. Initially, 26 samples, including 8 each of low [Gleason score (GS) 6] and high (GS ≥7) grade PCa samples and 10 matched normal prostate tissues, were analyzed as a discovery cohort. We identified 3,567 DMRs between normal and cancer tissues, and 913 DMRs distinguishing low from high-grade cancers. Most of these DMRs were located at CGI shores. The top 5 candidate DMRs from the low vs. high Gleason comparison, including OPCML, ELAVL2, EXT1, IRX5, and FLRT2, were validated by pyrosequencing using the discovery cohort. OPCML and FLRT2 were further validated in an independent cohort consisting of 20 low-Gleason and 33 high-Gleason tissues. We then compared patients with biochemical recurrence (n=70) vs. those without (n=86) in a third cohort, and they showed no difference in methylation at these DMR loci. When GS 3+4 cases and GS 4+3 cases were compared, OPCML-DMR methylation showed a trend of lower methylation in the recurrence group (n=30) than in the no-recurrence (n=52) group. We conclude that whole-genome methylation profiling with CHARM revealed distinct patterns of differential DNA methylation between normal prostate and PCa tissues, as well as between different risk groups of PCa as defined by Gleason scores. A panel of selected DMRs may serve as novel surrogate biomarkers for Gleason score in PCa.

Introduction

Prostate cancer (PCa) is the most common cancer in men in the US. The natural history and clinical course of PCa is very heterogeneous: some are indolent with little effect on overall lifespan, whereas others can progress to lethal metastatic disease. The test commonly used to screen for PCa is the prostate specific antigen test (PSA); however, many other factors can affect PSA levels and result in high false-positive findings.¹ Furthermore, current therapeutic modalities have complications that may lead to significant morbidity.² Therefore, new methods are needed to distinguish indolent cancers from their aggressive counterparts.

One promising approach is DNA methylation profiling. Changes in DNA methylation, with accompanied epigenetic gene silencing, appear to be the earliest somatic genomic alterations recognized in human PCa and continue throughout disease progression.³ Hypermethylation of promoter CpG islands (CGIs) can silence tumor suppressor genes early during tumorigenesis, while significant global hypomethylation arises later in PCa progression.^4-6 It is likely that different patterns of DNA methylation may distinguish aggressive vs. indolent PCa and predict responses to specific treatments. Earlier studies have shown that promoter hypermethylation of specific genes are associated with PCa progression. The best characterized gene, glutathione S-transferase-π (GSTP1), shows promoter hypermethylation in more than 90% of PCas from numerous independent analyses.³ Other methylation-mediated gene silencing has been reported in more than 40 genes, including APC, RARβ, and RASSF1A, and the number continues to grow.^7-9 However, no consistent conclusions can be drawn regarding the predictive value of these methylation markers.

In addition, most studies that evaluate global gene methylation focus on the promoter CGI region and leave the regions outside CGIs largely unexplored. However, recent studies have demonstrated that methylation of CpG dinucleotides up to 2 kb away from CGI (the CGI “shores”) can better distinguish different tissues as well as cancer and normal tissues.^10,11 We applied the comprehensive high-throughput arrays for relative methylation method (CHARM), which is a custom-designed NimbleGen HD2 microarray containing approximately 4.6 million CpG sites across the genome, to thoroughly examine all CpG dinucleotides in the genome. Because CHARM includes all CpGs in the genome, it is a non-biased methylation array that allows the identification of differentially methylated regions (DMRs) located in both CGI and non-CGI regions. We have successfully identified a series of genome-wide DMRs that can stratify normal vs. PCa and low-grade vs. high-grade PCa cases based on the Gleason score.

Results

CHARM effectively identified methylation differences between normal vs. prostate cancer and low Gleason vs. high Gleason cases

The CHARM assay was performed using 10 normal prostate tissues, 8 low Gleason score (GS) and 8 high Gleason score PCa samples (Table 1). All probes were displayed in the scatter plot with a subset of probes grouped into significant DMRs (P<0.01, highlighted in red) (Fig. 1A), with 7% of the total probes constituting significant DMRs in the comparison of normal vs. cancer, and 2% in low Gleason vs. high Gleason. There were 2,101 hypermethylated and 1,456 hypomethylated DMRs in the PCa samples compared with normal samples (Fig. 1B, C). In addition, we observed a total of 913 DMRs showing differential methylation between low Gleason and high Gleason samples, with 613 hypermethylated DMRs and 300 hypomethylated DMRs in high Gleason samples (Fig. 1B, C). There were 666 regions overlapping the DMR groups of normal vs. cancer (18.6% of total DMRs) and low vs. high Gleason samples (72.9% of total DMRs) (Fig. 1D).

Table 1.

Clinical pathological characteristics of the patients of discovery and validation cohort.

	Discovery cohort		Validation cohort
	Low Gleason (GS=6)	High Gleason (GS≥7)	Low Gleason (GS=6)	High Gleason (GS≥7)
	N=8	N=8	N=20	N=33
Gleason Score
6 (3+3)	20	0	8	0
7 (3+4)	0	0	0	0
7 (4+3)	0	6	0	19
8-9	0	2	0	14
Age (y) at time of RRP*
mean ± SD	61±1	61±2	54±6	62±8
Range	59-62	59-63	44-70	45-76
Race
Caucasian	7	7	18	28
Asian	0	0	0	2
Black	0	1	0	1
Unknown	1	0	2	2
PSA at diagnosis, ng/ml
≤10	7	4	18	15
>10	1	4	2	18
mean ± SD	6.0±2.1	14.3±14.2	7.6±5.8	17.2±18.0
Pathological stage
T2	5	1	16	3
T3	3	7	4	30
Surgical margin status
Negative	4	4	16	16
Positive	3	4	4	17
Abuts	1	0	0	0
Extra capsular extension
Negative	5	2	2	10
Positive	3	6	18	23

^*The mean age of patients who provided adjacent normal samples (N=10) were 61±1 years (range, 59 to 63 years). Among these, 8 are Caucasians; one is black; and the race of the remaining one is unknown.

Figure 1.

Global views of CHARM assay results. (A) Scatter plots show methylation values of individual probes in comparisons of low Gleason vs. high Gleason samples (left panel) and normal vs. cancer samples (right panel). Each point represents a probe from the CHARM assay. Highlighted in red is the subset of probes with significant differential methylation values, in positions used to group them into DMRs. The rest of the probes are shown in blue with the color scale representing density. (B) Heatmaps of DMRs comparing low Gleason vs. high Gleason (left panel) and normal vs. cancer (right panel). Each row represents a DMR. The left panel presents a hierarchical clustering of genome methylation values in 8 high-Gleason and 8 low-Gleason patient samples. The 100 CHARM probes in 913 DMRs with the largest magnitude t-statistic between the 2 groups were used; each column denotes a patient. The right panel presents group comparison between 10 normal individual and the same 16 prostate cancer patients as shown in the left panel. The rows are in ascending order according to the mean methylation of the DMR in the cancer group. (C) Volcano plot for CHARM analysis data sets. Each dot represents an individual DMR. The x-axis represents the difference of maximum methylation value (maxidiff) of DMRs between the comparisons. A positive maxidiff value indicates more methylation and a negative value indicates less methylation in high Gleason (upper panel) and cancer samples (lower panel). The y-axis shows the P-value for the DMRs. The blue and red dots are the select top candidate DMRs according to the rank of both maxidiff values and nominal P-values subjected to further technical validation. (D) Venn diagram showing the number of significant DMRs identified in comparisons of low Gleason vs. high Gleason (blue) and normal vs. cancer samples (pink), and the number of overlapping DMR regions between the 2 comparisons.

To analyze the functional enrichment of the genes related to methylation changes, we performed integrative functional analysis using Ingenuity Pathway Analysis (IPA) (Table 3). In normal vs. PCa, the top network includes RNA posttranscriptional modification, cell cycle and cellular assembly, and organization networks. The top 3 bio-functions related to diseases and disorders are developmental disorder, skeletal and muscular disorders, and cancer. The top canonical pathway is the EIF2 signaling pathway. Regarding low vs. high Gleason cases, the top network involves embryonic development, nervous system development and function, and organ development. Similar to the results of normal vs. PCa, the top bio-function related to diseases is developmental disorder. The top canonical pathway is basal cell carcinoma signaling.

Table 3.

The most relevant functional networks and biological processes present in the gene lists.

Low Gleason vs. high Gleason
Top networks	Score
Embryonic Development, Nervous System Development and Function, organ Development	50
Skeletal and Muscular System Development and Function, Connective Tissue Development and Function, Embryonic Development	35
Cellular Development, Gene Expression, Embryonic Development	33
Organismal Injury and Abnormalities, Skeletal and Muscular Disorders, Neurological Disease	33
Cellular Development, Nervous System Development and Function, Visual System Development and Function	29
Top bio functions related to diseases and disorders	P-value
Developmental Disorder	5.11E-09 – 6.14E-04
Gastrointestinal Disease	5.11E-09 – 6.14E-04
Neurological Disease	4.09E-08 – 6.32E-04
Psychological Disorders	4.09E-08 – 4.47E-04
Skeletal and Muscular Disorders	1.39E-07 – 6.14E-04
Top canonical pathways	P-value
Basal Cell Carcinoma Signaling	5.57E-05
Corticotropin Releasing Hormone Signaling	9.83E-05
Neuropathic Pain Signaling In Dorsal Horn Neurons	8.48E-04
nNOS Signaling in Neurons	1.02E-03
Factors Promoting Cardiogenesis in Vertebrates	1.42E-03
Normal vs. Cancer
Top networks	Score
RNA Post-Transcriptional Modification, Cell Cycle, Cellular Assembly and Organization	40
Nervous System Development and Function, Tissue Development, Tissue Morphology	38
Gene Expression, Cellular Development, Connective Tissue Development and Function	38
Cell Signaling, Carbohydrate Metabolism, Molecular Transport	38
Embryonic Development, Organismal Development, Skeletal and Muscular System Development and Function	36
Top bio functions related to diseases and disorders	P-value
Developmental Disorder	6.69E-18 – 3.69E-05
Skeletal and Muscular Disorders	3.60E-15 – 4.46E-05
Cancer	6.67E-15 – 4.25E-05
Gastrointestinal Disease	9.35E-14 – 1.69E-05
Neurological Disease	5.51E-11 – 4.46E-05
Top canonical pathways	P-value
EIF2 Signaling	1.16E-09
Transcriptional Regulatory Network in Embryonic Stem Cells	3.13E-08
Wnt/b-catenin Signaling	1.54E-05
Regulation of eIF4 and p70S6K Signaling	3.66E-05
Axonal Guidance Signaling	9.69E-05

Most DMRs are located at CGI shores

We next determined the distribution of the identified DMRs in the genome in relation to CGIs. If DMRs are distributed according to CpG density, the simulated frequency of DMR locations across the genome should follow the pattern shown as the white bars in Fig. 2 (the null hypothesis). However, our data demonstrated the distribution pattern of CHARM-identified DMRs to be as shown by the black bars. In both, the comparison of low vs. high Gleason grade (Fig. 2A) and the comparison of normal vs. PCa samples (Fig. 2B), significantly more DMRs were located within 2 kb of CGIs (i.e., on the “shore”) rather than within CGIs. In the comparison of normal vs. PCa, only 17.7% of the DMRs were in CGIs, compared to 68.2% in shores; an additional 14.1% of the DMRs were located greater than 2 kb away from CGIs. Similarly, in the comparison of low vs. high Gleason cases, the percentage of DMRs located in CGIs, shores and beyond was 21.1%, 59.5%, and 19.4%, respectively. The most enriched genomic regions for DMRs in both cases were 0-500 bp away from CGIs, with 31.3% in normal vs. PCa and 26.2% in low vs. high Gleason.

Figure 2.

Distribution of DMRs of low Gleason vs. high Gleason (A) and normal vs. cancer (B) in the genome in relation to CpG islands (CGIs). As denoted on the X axis, DMR positions are defined as “Island” (cover or overlap with more than 50% of a CGI), “Shore” (including overlap with 0.1–50% of a CGI, or located 0-500, 500-1000, or 1000-2000 bp from the nearest CGI), 2000-3000 bp from the nearest CGI, or more than 3000 bp from the nearest CGI. The y-axis represents percentage of each group for the DMRs of interest. The white bars represent the simulated DMR distribution (the null hypothesis, mean±2xSD). The black bars represent the locations of CHARM assay identified DMRs. * P<0.01 for simulation vs. DMR distribution.

A panel of DMRs is validated using pyrosequencing

The top 10 DMRs ranked by nominal P-values based on the CHARM statistics in both comparisons are listed in Supplemental Table 2, with their location relative to the nearest genes and CGIs, and the methylation status. For DMRs located in untranslated regions (UTRs) or areas away from genes, the genes located nearest to the DMR were used for their names. Aiming to identify differential DNA methylation to distinguish high-risk from low-risk PCa, 6 top candidate DMRs selected from the low vs. high Gleason result list, which allowed for successful design and performance of the pyrosequencing assay, were tested using quantitative methylation pyrosequencing analysis to confirm the CHARM result. The panel of 6 DMRs included OPCML, ELAVL2, EXT1, IRX5, FLRT2, and MAB21L1. We analyzed 4 types of different comparisons for each DMR in the panel: 1) normal vs. low Gleason, 2) normal vs. high Gleason, 3) normal vs. PCa, and 4) low vs. high Gleason. All six DMRs were differentially methylated in normal vs. high Gleason (P<0.01) and normal vs. PCa (P≤0.01). For the comparison of normal vs. low Gleason, all except OPCML showed differential methylation patterns (P<0.05). For the comparison of low vs. high Gleason, OPCML, ELAVL2, EXT1, IRX5, and FLRT2 were also found to have significant differential methylation patterns (Fig. 3A-F). The directions of the methylation changes (hyper- or hypo-methylation) of these validated DMRs are all consistent with the CHARM results.

Figure 3.

Bisulfite pyrosequencing analysis of 9 select DMRs using samples from the discovery cohort. Each dot represents methylation rate at the indicated CpG site of individual samples. The green, blue, and pink dots represent the normal, low-Gleason, and high-Gleason samples, respectively. The lines connect mean value of each CpG site within each group. Gene names are specified in each panel. * P<0.01 and ** P<0.05 for low-Gleason vs. high-Gleason tumors. Please note that the scales differ between the individual graphs.

Among the previously reported genes differentially methylated based on Gleason score, APC, PDLIM4, SFN, and SERPINB5 have been studied extensively in PCa.¹² Therefore, even though they were not among the top DMRs identified in our CHARM study, we included them for pyrosequencing analysis of the discovery cohort. Our results showed that APC, PDLIM4, and SFN were significantly hypermethylated in high-Gleason samples compared with low-Gleason samples (Fig. 3G-I).

To further investigate the correlation between the novel DMRs and progression of PCa, an additional 53 independent PCa samples, including 20 low-Gleason and 33 high-Gleason grade tissues, were used as a validation cohort to test the 9 DMRs. The results demonstrated that the methylation levels of OPCML, FLRT2, SFN, and PDLIM4 differed significantly between low- and high-Gleason grade samples (Fig. 4A-D). ELAVL2 and EXT1 also showed a strong trend of differential methylation (Fig. 4E). Of the 4 DMRs with significant difference between low vs. high Gleason samples, OPCML and PDLIM4 were the most informative per P-value. The combination of these 2 DMRs might be an effective diagnostic panel for future validation: the AUC from the combined ROC curve was 0.91 (Fig. 4F).

Figure 4.

Bisulfite pyrosequencing analysis of 9 select DMRs using the validation cohort. Bisulfite pyrosequencing results are shown for (A) OPCML; (B) FLRT2; (C) SFN, and (D) PDLIM4. Each dot represents the percent of methylation of each sample at the indicated CpG site. The blue and pink dots represent the low Gleason and high Gleason samples, respectively. The lines connect mean value of each CpG site within each group. (E) Low-Gleason (blue) vs. high-Gleason (pink) bisulfite pyrosequencing results, shown as boxplots of mean values for ELAVL2, EXT1, IRX5, MAB21L1, and APC. Whiskers of the boxplots mark the 5^th and 95^th percentiles, the boxes mark the 25^th percentile, median, and 75^th percentile, while extreme values are shown as black dots. (F) Receiver operating characteristic (ROC) analysis of low-Gleason vs. high-Gleason using the validation cohort for prognostic potential of OPCML-DMR, PDLIM4-DMR, or the combination of the two. The area under the curve (AUC) is 0.86 for OPCML-DMR alone and 0.91 for the combination of OPCML-DMR and PDLIM4-DMR.

The impact of DMRs on biochemical recurrence

To further explore the prognostic value of the DMRs, we acquired a third cohort of patients with documented outcome data from the PCa repository (38 frozen tissues) and the Canary Foundation (133 FFPE tissues). Pyrosequencing of the 9 target DMRs was performed on all frozen and FFPE samples. Fifteen tumor specimens overlapped between the tissues acquired from the 2 specimen resources. The DMR methylation results for these 15 samples were compared. Fig. 5A shows OPCML DMR methylation in the 15 tumor samples demonstrating no significant difference between samples with different archiving method. We then combined the 2 sample sets as an outcome cohort. For the 15 overlapping samples, the pyrosequencing data from the Canary Foundation samples were used for analysis.

Figure 5.

Bisulfite pyrosequencing analysis of OPCML-DMR using the outcome cohort. (A) Bisulfite pyrosequencing results of OPCML-DMR are shown for FFPE tissues (white bars) and frozen tissues (black bars) derived from the same patient tumor, total 15 tumors. For (B), (D), (E), (F), and (G), bisulfite pyrosequencing results are of OPCML-DMR in tumor samples of the outcome cohort. Each dot represents the percent of methylation at the indicated CpG site for each sample. The blue and pink dots represent the no recurrence and recurrence samples, respectively. The lines connect the mean values of each CpG site within each group. (C) Kaplan-Meier plot showing the relationship of methylation of OPCML-DMR and time to biochemical recurrence. Patients with biochemical recurrence were divided into a high- (blue) or low- (pink) OPCML-DMR methylation groups. Log-rank test P=0.52. (H) Kaplan-Meier plot showing the time to biochemical recurrence for patients based on Gleason scores in the outcome cohort (Table 2).

We first examined the methylation status of the OPCML DMR, comparing patients with biochemical recurrence vs. no recurrence. No difference was evident between these 2 groups (Fig. 5B). In addition, within the recurrence group, the OPCML DMR methylation levels were not associated with a disease-free interval based on Kaplan-Maier estimation (Fig. 5C). We then examined the methylation status of the OPCML DMR comparing low-Gleason and high-Gleason score groups within the outcome cohort; again, no significant difference was found (Fig. 5D). It should be noted that the majority of the GS 7 samples of the outcome cohort were GS 3+4 (74 cases), which constituted 77% of all 96 GS 7 samples, whereas the GS 7 samples in the discovery and validation cohorts were all (100%) GS 4+3 samples. It has been well established that GS 3+4 tumors have a better prognosis than GS 4+3 tumors,^13,14 and that Gleason 3+4 tumors have been suggested to be treated similarly to Gleason 6 tumors clinically. Our data indeed showed that the GS 3+4 cases behaved similarly to GS 6 and were significantly different from GS 4+3 (Fig. 5E, P=0.004). The number of GS 3+4 cases was roughly balanced in the recurrence (40 cases) and no recurrence (34 cases) groups. The methylation status of the OPCML DMR was not able to distinguish the 2 groups (Fig. 5F). When GS 3+4 cases were excluded, comparison of OPCML DMR methylation of the outcome cohort showed a small difference between the recurrence and no recurrence groups (Fig. 5G, P=0.06). Of course, because this subgroup analysis of GS 3+4 vs. GS 4+3 is post-hoc, it carries a high risk of yielding false positive results. The Kaplan-Maier curve for biochemical recurrence based on the Gleason sore is shown in Fig. 5H.

The cancer genome atlas (TCGA) database

To compare results with published high-throughput methylation studies and to use an external cohort as a validation dataset, we sought to mine the methylation data from the TCGA prostate adenocarcinoma (PRAD) database, which were generated by HumanMethylation450 BeadChip array (450K; Illumina, San Diego, CA). Data from the 450K array were informative for all 9 DMRs that we investigated. However, of the 160 tumors with Gleason score information recorded, only 11 cases of GS 6 and GS 8 had biochemical recurrence. We were therefore not able to perform statistical analysis with enough power to compare low GS vs. high GS or recurrence vs. no recurrence. In the PRAD cohort, the total 116 cases of GS 7 are comprised of 72 GS 3+4 and 44 GS 4+3 cases, but no significant difference of OPCML DMR methylation was found between these 2 subcategories (data not shown). RNA-seq data were also available for the 160 tumors, but no correlation was found between expression and methylation levels for either OPCML or FLRT2 DMRs.

Integrative analysis of DMRs with copy number alterations

To evaluate whether the methylation of DMRs and the expression of DMR-related genes are correlated with genomic changes, we examined copy number alterations (CNAs) in all the discovery cohort samples except one low-GS case. Overall, larger numbers of CNAs were found in the high-Gleason group than in the low-Gleason group, including frequent loss of PTEN, TMPRSS2, and ERG, genes important in PCa.^15,16 There were also more DMRs and DMR-related genes affected by CNAs in the high-Gleason group (supplemental Fig. 2).

Discussion

With CHARM assay targeting genome-wide CpG sites to analyze differential DNA methylation comparing normal vs. PCa and low Gleason vs. high Gleason grade prostate tumors, we successfully identified series of DMRs in both comparisons. Most DMRs were found to be located in CGI shores instead of within CGIs, consistent with previous findings from other cancer types that show tissue-specific and tumor-specific DMRs are more concentrated in shore regions.^11,17 These results underscore the importance of methylation profiling through an unbiased, whole genome approach rather than the traditional approach that is focused on CGIs.

Of the DMRs identified from the comparison of low- vs. high-Gleason grade, 72.9% overlapped with 18.6% of total DMRs from the normal vs. cancer comparison. This suggests that, to a large extent, the same genomic regions acquired epigenetic changes in the transition from benign to cancerous prostate tissue as in disease progression from low- to high-Gleason grade. Our results also demonstrated higher frequency of hypermethylation in PCa in comparison with normal samples, supporting previous study results.¹⁸ Furthermore, we found more frequent events of hypermethylation in high-Gleason tumors than in low-Gleason ones.

We identified novel candidate DMRs as candidate biomarkers to distinguish PCa vs. normal tissues, as well as low- vs. high-Gleason grade tumors. OPCML (opioid-binding cell adhesion molecule) is one of the top ranking DMRs that showed significant differential methylation patterns between high Gleason grade vs. low Gleason grade (high- vs. low-risk) groups in PCa. Interestingly, OPCML was originally reported to be epigenetically inactivated and have tumor-suppressor functions in epithelial ovarian cancer.¹⁹ OPCML acts as a broad tumor suppressor for multiple carcinomas including PCa with frequent epigenetic inactivation.²⁰ In epithelial ovarian cancer, OPCML has been shown to negatively regulate a specific receptor tyrosine kinase (RTK) repertoire consisting of EPHA2, FGFR1, FGFR3, HER2, and HER4 receptors. OPCML functions through directly binding the extracellular domains of RTKs, shifting their trafficking pathways and downregulating RTK levels via polyubiquitination-associated proteasome degradation, eventually leading to signaling and growth inhibition.²¹ The epigenetically inactivated region of OPCML reported by previous studies is located in the CGIs within the promoter region. In contrast, the OPCML DMR identified in our study is located in the CGI shore region. Considering that prostate and ovarian cancer are both sex-hormone regulated tumors and that OPCML exhibits interesting features of a tumor suppressor in multiple cancer types including prostate and ovarian, we hypothesize that the OPCML-DMR we identified may serve as an effective stratification marker for high- vs. low-risk PCa. Our finding of a predictive value of 0.89 based on AUC in distinguishing high- vs. low-Gleason tumors using OPCML-DMR alone (Fig. 4F) supports this hypothesis. However, our data has not shown prognostic value regarding biochemical recurrence using OPCML-DMR.

The other DMR validated in both the discovery and validation cohort is FLRT2. Together with 2 other factors, FLRT1 and FLRT3, the fibronectin leucine rich transmembrane proteins (FLRT) make a novel extracellular matrix protein family.²² FLRT2 has been shown to function in cell adhesion and/or receptor signaling. Kunkel's group reported XFLRT3 and XFLRT2 as novel transmembrane modulators of fibroblast growth factor (FGF)–MAP kinase signaling.²³ Furthermore, FLRT2 was demonstrated to interact with fibroblast growth factor receptor 2 (FGFR2) in mouse embryonic craniofacial tissue lysates. Stable knockdown or overexpression of FLRT2 in the chondrogenic cell line ATDC5 results in a corresponding decrease and increase of FGFR2 mRNA and protein expression, as well as downstream ERK phosphorylation levels.²⁴ FGF signaling regulates many important biological processes including cell proliferation, differentiation, and migration during development.²⁵ The human FGFR2 gene encodes FGFR2b and FGFR2c isoforms,^26-29 The hallmark of tumor invasion and metastasis, epithelial-to-mesenchymal transition (EMT) is accompanied by a switch from exclusive expression of FGFR2b to FGFR2c in the rat PCa model.³⁰ Moreover, restoration of FGFR2b led to suppression of PC-3 cell growth in vitro as well as reduced tumor formation in vivo.³¹ Therefore, FGFR2b is considered a tumor-suppressor in PCa. To date, there has been no report indicating FLRT2 is directly related to PCa; however, several papers implicate that FGFR2 might bridge the gap. Therefore, given the fact that human and mouse FLRT2 share 97% homology at protein level,³² and FGFR2b is interacting with and regulated by FLRT2 in mice and could potentially be anti-oncogenic in PCa, we hypothesize that the differential methylation of the region encompassing the FLRT2 transcription start site (TSS) might play a novel role in regulating the expression and interaction of FLRT2 and FGFR2, thereby modulating PCa progression.

Other novel DMRs identified in the discovery cohort, namely ELAVL2, EXT1, IRX5, and MAB21L1, all showed elevated methylation in high Gleason score tumors. ELAVL2 (embryonic lethal abnormal vision-like 2) encodes a highly conserved, neural-specific RNA-binding protein.³³ EXT1 (Exostoses-1) is a reported tumor suppressor in mice. Hypermethylation of the EXT1 CGI promoter led to transcriptional silencing of the EXT1 gene and subsequent loss of heparin sulfate.³⁴ EXT1 expression was significantly lower in benign prostatic hyperplasia (BPH) and PCa in comparison with normal prostate tissue.³⁵ IRX5 has been linked to human PCa through downregulation via vitamin D3 [1,25(OH)2D3] in LNCaP cells; VitD3 is a potent inhibitor of the proliferation of many different cancer cell types. Knockdown of IRX5 resulted in an increase in p21 protein expression, G2-M arrest, and apoptosis, partially mediated by p53.³⁶ The MAB21L1 gene has a highly polymorphic tandem CAG trinucleotide repeat in the 5′ UTR potentially associated with neurologic and psychiatric disorders.^37-39 Further studies are needed to explore the biological functions of the methylation status of these DMRs in PCa. The changes in the expression of these genes in prostate cancer, both the prevalence and extent, remain to be ascertained.

In this study we also included 3 well-studied DMRs: APC, PDLIM4, and SFN. These were all reported to be associated with the Gleason score in PCa;¹² however, we were not able to validate them as consistent DMRs in all of our cohorts. Our results showed significant differential methylation of APC in the discovery cohort but not in the validation cohort and, alternatively, differential methylation of PDLIM4 and SFN in the validation cohort but not in the discovery cohort. None of these 3 genes showed differential methylation patterns in the comparison of tumor samples with or without biochemical recurrence. Assays based on patient tissues are limited by the impact of infection/inflammation, which may partially explain the lack of significance of these markers in our cohort. For example, APC methylation is associated with inflammation.⁴⁰ We do not have information regarding inflammation in these samples. Age may also complicate the results as DNA methylation changes with aging.

Even though CHARM is a robust and non-biased method to assess the methylation of CpG sites throughout the whole genome, it requires large amount of input DNA (5 µg). This limits the sample availability for the CHARM assay in this study. Another limitation of the study is the sample size of our validation and outcome cohorts. Additional samples with complete clinical information are critical to validate our novel DMRs in order to determine the clinical utility of these markers. A recent European study on over 400 total samples reported 6 novel candidate DNA methylation markers for PCa; 3 of these markers, C1orf114, AOX1, and HAPLN3, were independent predictors of time to biochemical recurrence after radical prostatectomy.⁴¹ Additional emerging candidate markers for prognosis were also published.^42,43 Cross validation of these potential prognostic markers in various cohorts will be important to confirm truly useful methylation markers for clinical practice.

In summary, whole-genome methylation profiling with CHARM revealed distinct patterns of differential DNA methylation between normal prostate and PCa tissues, as well as between different risk groups of PCas defined by Gleason scores. We identified several novel DMRs including OPCML, ELAVL2, EXT1, FLRT2, and IRX5 to effectively distinguish low- from high-Gleason grade PCas. In particular, OPCML and FLRT2 were further validated in independent cohorts. However, these DMRs showed no significant difference of methylation in patients with or without biochemical recurrence. Further investigation is needed to validate the prognostic value of the novel DMRs and to explore the biological functions of their differentially methylated status in PCa.

Patients and materials/methods

This study was approved by the Institutional Review Boards (IRB) of the Fred Hutchinson Cancer Research Center and the University of Washington.

Study subjects

All tissue specimens in this study were obtained from the PCa repository, University of Washington (PI: Robert Vessella) and Canary Foundation (PtdIns: Pete Nelson). The tumor cell content of all the cancer specimens was 75% or higher. The discovery cohort included fresh frozen PCa tissues (n=16) and matched normal prostate tissues (n=10) from a total of 16 age-matched PCa patients. The validation cohort included 53 fresh frozen PCa tissues. The clinical pathological characteristics of the patients in the discovery and validation cohorts are shown in Table 1. The outcome cohort initially included 38 frozen tissues (from PCa repository, UW) and 133 FFPE tissues (from Canary Foundation). Fifteen tumor samples overlapped between the tissues acquired from the 2 specimen resources. Based on comparable results, we combined the 2 sample sets to form the outcome cohort and used the pyrosequencing results from the FFPE tissue for further analysis of the overlapped samples. Therefore, the outcome cohort included a total of 156 samples. Of these, 86 had no biochemical recurrence and 70 did. The clinical pathological characteristics of the patients of the outcome cohort are shown in Table 2. Gleason scores were previously reviewed by a pathologist.

Table 2.

Clinical pathological characteristics of the patients of outcome cohort.

	No recurrence	Recurrence
	N=86	N=70
Gleason Score
6 (3+3)	48	7
7 (3+4)	34	40
7 (4+3)	2	20
8-9	2	3
Age (y) at time of RRP
mean ± SD	59±7	61±7
Range	47-78	43-76
Race
Caucasian	69	56
Asian	2	0
Black	1	2
Hispanic	1	0
Indian (Native American)	1	0
Multi-Racial	12	11
Unknown	0	1
PSA at diagnosis, ng/ml
≤10	71	52
>10	14	18
Unknown	1	0
mean ± SD	6.9±3.7	10.4±11.3
Pathological stage
T2	72	47
T3	14	23
Surgical margin status
Negative	73	35
Positive	13	33
Abuts	0	2
Perforated Capsule
Negative	75	46
Positive	11	24
Post RRP recur
Detectable Mets	0	10
Disease Free Interval	N/A	44±36

DNA extractions

Genomic DNA of frozen tissue was isolated using Gentra Puregene Kit (Qiagen, cat. no. 158667). Genomic DNA of FFPE tissue was isolated using QiAamp DSP DNA FFPE tissue kit (Qiagen, cat. no. 60404) per user manual for CytoScan DNA purification from FFPE tissue (Affymetrix, cat. no. 901835).

CHARM microarray

CHARM microarrays were performed as previously described.⁴⁴ Briefly, 5 µg of DNA from each specimen was randomly fractionated with HydroShear, divided into 2 equal portions to be treated with and without McrBC, a restriction enzyme that cleaves DNA containing 5-methylcytosine preceded by a purine nucleotide, then size fractionated by agarose gel electrophoresis, purified and subjected to whole-genome amplification prior to hybridization with the CHARM array (customized NimbleGen HD2 array, Roche, cat. no. B7074-00-01). The raw data were analyzed with the R/Bioconductor CHARM package (Analysis of DNA methylation data from CHARM microarrays, http://bioconductor.org/packages/release/bioc/html/charm.html). Differential methylation was quantified for each pairwise tissue comparison by the difference of averaged and normalized methylation values (ΔM). Z scores were calculated using ΔM and standard errors (s.e.m) {z = [ΔM /s.e.m. (ΔM)]}. Probes carrying z scores with a false discovery rate (FDR) ≤5 % and contiguous in positions were grouped into DMRs. DMRs with P-values lower than 0.01 in permutation test were considered statistically significant.

Bisulfite conversion and pyrosequencing

DNA (100 ng) was used for bisulfite conversion, which was performed using EpiTect Bisulfite Kit (Qiagen, cat. no. 59104) or EZ-96 DNA Methylation-Direc Kit (Zymo Research, cat. no. D5022), according to the manufacturers' protocols. Primer sets with one biotin-labeled primer for the amplification of the bisulfite-converted DNA were either from a pre-designed PyroMark CpG assay (Qiagen, 978746) or custom designed using PyroMark Assay Design software version 2.0.1.15 (Qiagen, cat. no. 9019077). PCR reactions were performed using PyroMark PCR kit (Qiagen, cat. no. 978703) according to the manufacturer's protocol. The pyrosequencing analysis was carried out using PyroMark Q24 or MD96 Systems (Qiagen, cat. no. 9001514) to assess the quantitative methylation of the target genomic regions (Supplemental Table 1).

Integrative functional analysis

The analysis of the most relevant functional networks and biological processes present in the gene lists was generated through the use of IPA (Ingenuity Systems).

Chromosomal genomic array testing (CGAT)

Genomic DNA was tested on either CytoScanHD or SNP6.0 (Affymetrix, cat. no. 901835 and 901153, respectively), as per manufacturer's protocols. Copy number alteration and loss of heterozygosity were assessed using an algorithm established at the Seattle Cancer Care Alliance based on the Hidden Markov model. Data analysis and visualization was performed with Nexus software (BioDiscovery).

Statistical analysis

The significance of differences between the null hypothesis and DMR distribution in Fig. 2 was determined by the Χ² test. For analysis of pyrosequencing results, since methylation levels exhibited parallel profiles across CpG sites within a given DMR, statistical comparisons of methylation were based on the average methylation level across CpG sites. Comparisons between groups were by 2-sample t-test. Linear regression was used to adjust for age and other clinical factors. Logistic regression was used to evaluate area under the receiver operating characteristic (ROC) curve, to assess the ability of single or multiple DMRs to distinguish PCa with high vs. low Gleason grade.

Disclosure of potential conflicts of interest

No potential conflicts of interest were disclosed.

Funding

This study was supported by P01 Pacific NW SPORE and SUB CA085859 awarded by the National Cancer Institute.