Genome-wide DNA methylation measurements in prostate tissues uncovers novel prostate cancer diagnostic biomarkers and transcription factor binding patterns

We collected the prostate cancer and benign-adjacent tissues used in this study at Stanford University Medical Center between 1999 and 2007 from patients undergoing radical prostatectomy with patient informed consent under an IRB-approved protocol. The percentage of prostate cancer epithelial cells in each sample was assessed by a pathologist specializing in genitourinary cancers on hematoxylin and eosin (H & E) stained frozen sections of the tissues from which the DNA was extracted. We selected those samples in which at least 90% of the epithelial cells were cancerous for nucleic acid extractions, and used the QIAGEN AllPrep DNA/RNA mini kit (QIAGEN) to extract DNA and RNA.

We assayed DNA methylation levels by using the Illumina Infinium HumanMethylation 450 K beadchip array (Illumina, San Diego, CA, USA) [ 18] and calculated the methylation beta score as: b = Intensity _Methylated/(Intensity _Methylated + Intensity _Unmethylated). We converted data points that were not significant above background intensity to NAs. We removed CpGs having greater than 10% missing values prior to normalization. Data was normalized with the ComBat R package [ 19]. Post-ComBat normalization, we observed that the Infinium I and II assays showed two distinct bimodal b-value distributions, so we developed a regression method to convert the type I and type II assays to a single bimodal b-distribution corresponding to Reduced Representation Bisulfite Sequencing (RRBS) b-values [ 20]. After the Methylation 450 K data was converted to RRBS b-values, any values less than zero were assigned zeros and values greater than one were assigned ones. The equations for correction are shown below:

Infinium I to RRBS:

Infinium II to RRBS:

Linear mixed model analysis of the methylation data was performed using the lme command in R, with patient as a random effect, and age and ethnicity as fixed effects. Logistic regression was performed using the glm command (family = binomial). The p-values were adjusted using the Benjamini and Hochberg method [ 21]. CpGs with a standard deviation of less than 1% across samples were removed prior to analysis.

We constructed RNA sequencing libraries using a transposase-mediated construction method described previously [ 22]. Four RNA-seq libraries were pooled into each lane and sequenced using Illumina HiSeq 2000 instruments to generate paired-end 50 sequencing reads (Illumina, San Diego, CA, USA). Read-pairs were aligned to Gencode (version 9.0) using TopHat (version 1.4.1), and the relative abundance of each transcript was quantified using Cufflinks (version 1.3.0) and BEDTools [ 23– 26]. Differential expression analysis was conducted based on tumor status using DESeq2 (version 1.8.1) with default settings in likelihood ratio test (LRT) mode. Transcripts from the X and Y-chromosomes were removed prior to differential expression analysis.

Chromosomal positions of significant CpGs were annotated using RefSeq (hg19 assembly) [ 27]. The Gene Set Enrichment Analysis (GSEA) tool was used to analyze enriched cellular pathways [ 28]. GSEA was run with Kegg and Reactome selected, and used an FDR-corrected q-value cutoff of 0.05.

Hierarchical clustering was performed using Cluster 3.0 [ 29]. Data was mean-centered and clustered by both gene and array using Euclidean distance with average linkage. Clusters were visualized using TreeView [ 30].

TCGA DNA methylation (Illumina Methylation 450 k) datasets and associated clinical data for prostate (PRAD_2013_09_07), lung (LUAD_2013_09_07), breast (BRCA_2013_09_07) and pancreatic (PAAD_2013_09_07) tissues were downloaded from the UCSC cancer genome browser at time of manuscript preparation. Datasets were normalized prior to validation analysis.

ENCODE transcription factor binding data was downloaded from http://​genome.​ucsc.​edu/​cgi-bin/​hgTrackUi?​db=​hg19&​g=​wgEncodeRegTfbsC​lusteredV3. We overlapped the CpGs found within gene regulatory regions (promoter, first exon or first intron) from the top 10,000 most significant CpGs from regression analysis with the ENCODE transcription factor binding sites, and used a Fisher’s exact test to determine transcription factor binding sites enriched for differential methylation over background. For EZH2 binding site overlap, we overlapped significant CpGs (FDR p-value < 0.05) with EZH2 binding data previously published [ 31]. For gene expression analysis, genes that were differentially expressed between tumor and normal (DESeq2-based FDR p-value < 0.05) were designated as overlapping a TF binding site if greater than 50% the binding site fell within the transcript promoter region. The promoter region was defined as 1000 bp upstream to 500 bp downstream of the transcription start site. Transcription factors with a Bonferroni-corrected p-value <0.05 were classified as significantly enriched.

To investigate differential DNA methylation associated with prostate cancer, we used the Illumina Infinium HumanMethylation450 BeadChip Methylation Assay, which covers more than 485,000 CpGs located throughout the human genome [ 18]. DNA methylation patterns were measured in 73 patient prostate cancer tissues and 63 benign-adjacent tissues, 52 of which are patient-matched (Table 1). Mixed model linear regression analysis identified (LME) 226,235 CpGs with significantly different methylation levels (LME, FDR-adjusted p-value <0.05) in cancer tissues compared to benign-adjacent prostate tissues. Of the 226,235 significant CpGs, ~67% had increased methylation and ~33% had decreased methylation in the cancer tissues compared to the benign-adjacent tissues (Fig. 1a). CpGs with higher methylation levels in tumor tissues were more likely to be within CpG islands (Fisher’s Exact Test, p-value 3.44e–154, OR = 1.18, 95% CI = 1.18–1.12), and statistically significant CpGs were also found in greater proportion in gene regulatory regions (promoter, first exon, or first intron) than in gene body regions (other exon, other intron, or 3′ proximal region), although this association did not reach statistical significance. (Table 2A, B).

To explore the genes and cellular pathways found in differentially methylated regions, we analyzed the top 10,000 most significant CpGs between the prostate cancer tissue and unaffected prostate tissue (LME, FDR p-value cutoff of <4.27e-13) (Fig. 1b). Of these, 75% had a higher methylation level in the cancer tissues. We divided the top 10,000 CpGs that were uniquely annotated to one gene by whether they resided in the gene regulatory region (promoter, first exon, and first intron) or the gene body (other exon, other intron, and 3 prime proximal region) and found that the CpGs with higher methylation in the cancer compared to benign tissue were statistically more likely to be associated with a gene regulatory region (Fisher’s Exact Test, p-value 0.015, OR = 1.10, 95% CI = 1.018–1.18) (Table 2C).

We used Gene Set Enrichment Analysis (GSEA) to determine which gene pathways are represented in the top 10,000 most significant CpGs [ 28]. We observed 3165 CpGs in the regulatory regions of 1589 genes with higher DNA methylation in the prostate cancer compared to benign tissue. GSEA analysis of those 1589 genes showed a strong signature for glycosaminoglycan metabolism, with five of the top 10 significantly enriched pathways associated with heparan sulfate metabolism and chondroitin sulfate metabolism (Additional file 1: Table S1). Other pathways included focal adhesion, pathways in cancer, Wnt signaling pathway, developmental biology and axon guidance. The enrichment for glycosaminoglycan metabolism pathways was specific to CpGs in regulatory regions with higher methylation in the cancer. Conversely, there were 776 CpGs located in gene regulatory regions of 621 genes with lower methylation in prostate cancer tissue. GSEA analysis of these genes showed enrichment for olfactory signaling, G-protein coupled receptor signaling, metabolism of carbohydrates, apoptosis, immune system, neuronal growth factor signaling pathway, and hemostasis (Additional file 2: Table S2).

We compared the DNA methylation data with transcription factor chromatin immunoprecipitation sequencing data (ChIP-seq) measured by the Encyclopedia of DNA Elements (ENCODE) Consortium to test whether there was an enrichment of transcription factor binding sites coinciding with the top 10,000 most differentially methylated CpGs between prostate cancer and benign-adjacent tissues. Enhancer of zeste homolog 2 (EZH2) was the most significantly enriched TF overlapping CpGs with higher methylation in the cancer tissues from our dataset (Fisher’s Exact Test, Bonferroni adj. p-value 7.54e-172, OR = 3.4, 95% CI = 3.14–3.68), and this observation was validated in The Cancer Genome Atlas (TCGA) prostate methylation dataset (Fisher’s Exact Test, Bonferroni adj. p-value 6.48e-120, OR = 2.48, 95% CI = 2.29–2.69) (Fig. 2a and Additional file 3: Table S3A and B). ENCODE TF binding data was generated from multiple types of cell lines. To determine whether EZH2 binding enrichment occurs in prostate cancer specifically, we compared the significant CpGs differentially methylated between prostate cancer and benign-adjacent tissue (FDR p-value cutoff of <0.05) with previously published EZH2 binding events from androgen-dependent (AD) and androgen-independent (AI) cell line models [ 31]. EZH2 binding events were significantly enriched in both the AD and AI contexts, although we observed a higher level of enrichment in the AI context (Fisher’s Exact Test, AD enrichment p-value =0.01, OR = 1.15, 95% CI = 1.01–1.30; AI enrichment p-value =0.00013, OR = 1.18, 95% CI = 1.08–1.29) (Additional file 3: Table S3C). Notably, in our tissue cohort, significant CpGs found in proximity to EZH2-bound sites are mostly hypermethylated (Fig. 2b). We also observed that a majority of transcripts that contain EZH2 binding sites in the promoter region that are differentially expressed between prostate cancer tissue and the benign-adjacent tissues have decreased expression in the prostate cancer tissue (Fig. 2b).

For CpGs with lower methylation in the prostate cancer tissues in comparison with the adjacent-unaffected tissue, there were two TFs with significant overlap that were bound in these regions: FOXA2 and SETDB1 (Additional file 3: Table S3D). However, we were unable to validate the enrichment we observed for these TFs in the TCGA prostate methylation dataset (Additional file 3: Table S3E).

To discover DNA methylation patterns that best distinguish prostate cancer tissue from benign-adjacent tissue, we performed logistic regression on the 100 most statistically significant CpGs from the linear mixed model regression (FDR p-value cutoff of <8.22e-15). We tested each combination of three CpGs within the top 100 most significant CpGs, as models containing three CpGs resulted in the smallest Akaike Information Criterion (AIC) value. We calculated the Area Under the Curve (AUC) for each Receiver Operating Characteristic (ROC) curve to identify the model with a maximal AUC. The top DNA methylation diagnostic model based on AUC from our analysis consists of cg00054525, cg16794576, and cg24581650 (Fig. 3, Additional file 4: Table S4). This DNA methylation model produces a ROC curve with an AUC of 0.97 in our cohort of prostate tissues and has a specificity of 98.4% and a sensitivity of 87.5%, indicating that DNA methylation status at these three genomic positions has very high predictive power for distinguishing malignant tissue from benign tissue (Fig. 4a). The corresponding waterfall plot demonstrates the high accuracy our top DNA model performs in classifying the prostate tissues (Fig. 4a). Based on analysis of methylation data from other tissue types, these methylation differences are unique to prostate cancer cells, as DNA methylation levels at these sites performed poorly in distinguishing lung, pancreatic, and breast cancer tissue from benign tissue (Additional file 5: Figure S1). The top three diagnostic CpGs are in close proximity to four total transcripts based on annotation, including CYBA, ERGIC1, HLA-J, and NCRNA00171. Out of these four transcripts, ERGIC1 has a statistically significant difference in mRNA expression level between prostate malignant tissue and benign tissue (DESeq2, adj. p-value 6.4e-06) (Additional file 6: Figure S2).

We utilized prostate data from TCGA as a validation cohort for our DNA methylation signature (Table 1). The TCGA methylation data was also measured using the Human Methylation 450 BeadChip and included 213 prostate cancer tissues and 49 normal tissues. Our model, based on 3 CpGs, could distinguish normal from malignant prostate tissues with a sensitivity of 84.5% and a specificity of 91.7% in the TCGA dataset, resulting in a ROC curve with an AUC of 0.920 (Fig. 4b, Additional file 4: Table S4). To determine how our top diagnostic model performs in the context of benign prostate hyperplasia (BPH), we used a previously published cohort (GEO accession: GSE55599) to see if our top DNA methylation model could distinguish prostate cancer tissue from prostate tissue obtained from patients with benign-hyperplasia and found that our model could perfectly discriminate these two types of tissues (Additional file 7: Figure S3) [ 32].

Additionally, we investigated prostate diagnostic markers from significant CpGs from the linear mixed model analysis that exclusively demonstrated an increase in methylation in cancers, as biomarkers that are hypermethylated in the cancer tissues are potentially more easily translatable to the clinic. In this context, the top model consists of cg15338327, cg00054525, and cg14781281 (Additional file 8: Figure S4), resulting in a ROC curve with an AUC of 0.97 in our dataset and an AUC of 0.92 in the TCGA prostate validation dataset (Additional file 9: Figure S5A-B). This hypermethylated diagnostic model also performed well at distinguishing benign-hyperplasia prostate tissue from prostate cancer tissue, with an AUC of 0.85 (Additional file 9: Figure S5C-D).