Identification of Retinol Binding Protein 1 Promoter Hypermethylation in Isocitrate Dehydrogenase 1 and 2 Mutant Gliomas

Patient Cohorts and Tumor Specimens

A total of 198 frozen or formalin-fixed paraffin embedded tissue specimens were obtained from the University of California, Los Angeles Brain Tumor Translational Resource (Los Angeles, CA). Remnant human brain tumor samples were collected from patients undergoing surgical resection and who provided written informed consent. The collection of human brain tumor samples was approved by the University of California, Los Angeles Institutional Review Board. Normal brain tissues were collected with University of California, Los Angeles Institutional Review Board approval from one patient undergoing non–tumor-related surgery and from four patients at the time of autopsy (the postmortem interval was less than 12 hours). IDH1 was sequenced on all samples, and IDH2 was sequenced on selected IDH1 WT samples. The clinical characteristics of the patient cohorts are listed in Tables 1 and 4. A diagram showing the composition of the cohorts is shown in Figure 1.

Cell Culture

The U87MG, U138MG, and U373MG glioma cell lines were obtained from Dr Paul Mischel (University of California, Los Angeles, Los Angeles, CA). Dr Carol Kruse (University of California, Los Angeles, CA) provided the D54MG cell line. Dr Glyn Dawson (University of Chicago, Chicago, IL) and Dr Anthony Campangoni (University of California, Los Angeles, CA) provided the HOG cell line. All glioma cell lines were cultured in Dulbecco’s modified Eagle medium/F12 medium (Invitrogen, Grand Island, NY) supplemented with 10% fetal bovine serum and 100U/mL penicillin/streptomycin. The human astrocytic progenitor cell line (APC) was obtained from Dr Ina Wanner (University of California, Los Angeles) and cultured in Delbecco’s modified Eagle medium/F12 medium with 10% fetal bovine serum (22).

Massively Parallel RRBS

RRBS was done using the protocol published by Meissner et al. (14,15). Details of the RRBS protocol, including generation of U87MG genomic DNA samples, RRBS quality control, and bioinformatics, are outlined in the Supplementary Methods and shown in Supplementary Figure 1 (available online). Briefly, DNA was isolated from U87MG glioma cells and frozen tumor tissues, digested with the restriction enzyme MspI to enrich for fragments containing CpG islands, and end-repaired using methylated cytosine. After adapter ligation, the DNA was size-fractionated by gel electrophoresis, and DNA fragments between 100 and 400 base pairs in length were isolated to minimize large fragments with poor sequencing coverage. Isolated DNA were then bisulfite treated, amplified, mixed with unmodified PhiX DNA (a bacterial genome inserted for quality control and to assess mapping), and sequenced on an Illumina Genome Analyzer IIx. The Novoalign software package (www.novocraft.com) was used to align the sequence data. Aligned sequence data were then sorted using the SAMTools software package (23) and stored in SAM format for further analysis.

Methylation status was determined at individual CpG sites, and the results were compiled to show the level of methylation at individual CpG islands. CpG islands were mapped by previously published definitions (24); and a strict quality control protocol (see Supplementary Methods, available online) was implemented to ensure the quality of the sequence base calls, alignment, and the final methylation data.

Methylation-Sensitive Restriction Enzyme Assay

MSRE, as described by Tran et al (25), was used to assess genome-wide methylation for 64 glioma samples. Briefly, genomic DNA was digested with the restriction enzyme BfaI and divided into two aliquots. One aliquot was digested with the methylation-sensitive restriction enzyme HpaII and labeled with Cy5 after polymerase chain reaction (PCR)-amplification (25). The second aliquot was digested with the methylation-insensitive enzyme MspI and labeled with Cy3 after PCR amplification. The PCR products were hybridized to an Agilent High-Density 2-color Human CpG Island Microarray (Agilent Technologies, Santa Clara, CA).

In silico digestion of the whole human genome (HG18) using enzyme BfaI was used to predict sensitive (containing CCGG) and insensitive (not containing CCGG) fragments of DNA from tumor samples. Each BfaI fragment was assigned a methylation score based on the median Loess corrected log ratios of the Cy-5/Cy-3 signal for all probes mapping to that fragment. Methylation scores were then standardized on the basis of the distribution of values for all insensitive fragments.

Analysis of Methylation and Gene Expression in The Cancer Genome Atlas Dataset

Methylation data for 147 GBM samples were obtained from the TCGA (8). Methylation was measured at approximately 27 000 CpG dinucleotides for 62 samples, using the Illumina Infinium Human Methylation 27 assay (Illumina, San Diego, CA) and at 1500 CpG dinucleotides for 85 samples, using the Illumina Goldengate Methylation Cancer Panel assay. Level 3 data, including normalized methylation signal per gene per sample, were downloaded directly from the TCGA data portal (http://tcga-data.nci.nih.gov/tcga/). Gene expression data were available for 142 of 147 TCGA samples. Gene expression was measured using the Affymetrix HT Human Genome U133A microarray (Affymetrix, Santa Clara, CA). A TCGA gene-expression probe (203423_at, chr3:140718973-140741180) covers transcript variant 1 of the RBP1 gene.

Statistical Analysis

Differentially methylated CpG islands located in gene promoter regions were identified by performing the student t test comparing IDH1 MUT and WT samples. To minimize the number of statistically significant but biologically nonsignificant changes in methylation, arbitrary thresholds of 0.4, 0.4, and 3 (for RRBS, TCGA, and MSRE, respectively) were set as the minimum difference between the mean levels of methylation between IDH1 WT and MUT samples that may be biologically significant to warrant further evaluation. Different cutoffs were chosen for the three different datasets because of their technical differences. Data for the MSRE technique are expressed as a log ratio, whereas data for the RRBS technique are expressed as a direct percentage of methy lation. Also, to control the false discovery rate, a Q of less than or equal to 0.05 was set as the threshold for statistical significance, as described by Storey et al. (26), for the three datasets. A false discovery rate of 0.02% was estimated via permutation testing of the RRBS data [252 permutations (28), http://biosun1.harvard.edu/complab/dchip].

Correlation between IDH1 mutation and RBP1 methylation was determined using the Spearman correlation test. To examine the relationship between OS and RBP1 methylation, survival curves were estimated by the Kaplan–Meier method, and groups were compared using a log–rank test. Because of the tight correlation between IDH1/IDH2 mutations and RBP1 methylation, multivariable analysis with Cox proportional hazards model was done, which included variables including age (years, continuous variable), sex (male or female), performance status (Karnofsky performance score 80–100 vs ≤ 70), extent of resection (gross-total vs subtotal/biopsy), and IDH1/IDH2 mutation status or RBP1 methylation status separately. Statistical analyses by race/ethnic group were not done. The assumption of proportionality was verified by the statistical test of correlation between Schoenfeld residuals and ranked survival time. Statistical analyses were performed using the open-source R statistical analysis package (http://www.R-project.org). All tests were two-sided, and a P value of less than .05 was considered statistically significant.

Targeted Bisulfite Sequencing of Retinol Binding Protein 1

The methylation status of the RBP1 promoter CpG island was assessed by standard bisulfite sequencing utilizing a nested PCR protocol with the primer sets: stage 1 (forward, 5ʹ–TTTATTGGGTATTGGAAGATGTTG–3ʹ and reverse, 5ʹ–TCCAATCTACAACCTAAAAACTACC–3ʹ) and stage 2 (forward, 5ʹ–GGTATTGGAAGATGTTGGTTAA–3ʹ and same reverse primer as stage 1). The sequence of each sample was determined using Chromas Lite 2.33 (Technelysium Pty Ltd, South Brisbane, QLD, Australia). The level of methylation was semiquantitatively scored in quartiles by the relative heights of the methylated and unmethylated peaks. For RBP1 methylation determination by BiSEQ, the mean methylation levels (of the 21 CpG sites sequenced) above 50% were classified as methylated.

IDH1 and IDH2 Sequencing

Genomic DNA was isolated from formalin-fixed paraffin-embedded or frozen tissue using the Recoverall Total Nucleic Acid Isolation Kit (Invitrogen, Grand Island, NY). Sequencing of IDH1 at residue 132 and IDH2 at residue 172 was determined by Sanger sequencing with the following primers: IDH1 (forward, 5ʹ–gcgtcaaatgtgccactatc–3ʹ and reverse, 5ʹ–gcaaaatcacattattgccaac–3ʹ) and IDH2 (forward, 5ʹ–CTCACAGAGTTCAAGCTGAAG–3ʹ and reverse, 5ʹ–CTGTGGCCTTGTACTGCAGAG–3ʹ). Purified PCR products were sequenced using the BigDye Terminator v1.1 and analyzed on a 3730 sequencer (both from Applied Biosystems).

Analysis by Quantitative Real-Time Reverse Transcriptase PCR

Total RNA was extracted from culture cells or tumor tissues using Trizol (Invitrogen, Grand Island, NY) according to the manufacturer’s instructions. Purity of the total RNA was then determined by the 260/280nm ratio, and the integrity was checked by electrophoresis on 1% agarose gel.

One microgram of RNA from each sample was reverse-transcribed to cDNA with the Reverse Transcription System (Promega, San Luis Obispo, CA) using oligo-dT primers. Normal brain cDNA isolated from one frozen surgical tissue, four frozen autopsy tissues, and two commercially available cDNA libraries (Biochain, Hayward, CA; Invitrogen, Grand Island, NY) were used as controls. Reverse transcriptase PCR (RT-PCR) was performed using Platinum DNA polymerase (Invitrogen, Grand Island, NY). The PCR product was separated on a 3% agarose gel. Quantitative real-time RT-PCR (qRT-PCR) using FastStart Universal SYBRR-Green Master (Roche, Mannheim, Germany) using a LightCycler® 480 System (Roche, Mannheim, Germany) was done. The primers were designed with Primer3 (27) (version 0.4.0, http://frodo.wi.mit.edu) software and were as follows: forward, 5ʹ–CAACTGGCTCCAGTCACTCC–3ʹ and reverse, 5ʹ–TGCACGATCTCTTTGTCTGG–3ʹ. The following conditions were used for amplification: 95°C for 3 minutes, 40 cycles at 95°C for 10 seconds, followed by 60°C for 30 seconds, and 72°C for 30 seconds. All samples were amplified in duplicate from the same RNA preparation, and the results are presented as the mean with standard error of the mean from three independent experiments. RBP1 mRNA levels were standardized to the levels of β-actin (internal control) and quantified by the relative Ct method (2^∆∆Ct).

Protein Expression by Western Blot

Total protein lysates were prepared using radio-immuno- precipitation assay buffer containing protease inhibitor to lyse the cells and tissues. The proteins (15 µg) were separated on a 4%–20% sodium dodecyl sulfate polyacrylamide gel and transferred to nitrocellulose membranes (0.45 µm, Bio-Rad, Hercules, CA). Western blot was performed with goat anti-CRBP1 (1:200) or rabbit anti-CRBP1 (1:200) polyclonal antibody (Sigma, St. Louis, MO), mouse anti-α-tubulin monoclonal antibody (1:4000, Sigma, St. Louis, MO), horse radish peroxidase-conjugated rabbit anti-goat (1:8000) or goat anti-rabbit (1:8000) IgG (Santa Cruz Biotech, Santa Cruz, CA), horse radish peroxidase-conjugated goat anti-mouse IgG (1:10 000, Jackson Immuno Research, West Grove, PA), and an enhanced chemiluminescence detection kit (Pierce, Rockford, IL). Densitometry was performed with Gel-Pro Analyzer 4.0 software (Media Cybernetics, Bethesda, MD).

Development of RRBS for Methylome Profiling in Glioma

To examine potential tumor suppressor genes that may be hypermethylated in IDH1 MUT tumors, we applied the RRBS protocol (14,15) to profile the glioma methylome. As a part of the protocol validation, we initially performed RRBS on DNA isolated from U87MG cells, a widely used GBM cell line. U87MG DNA that had been treated with the CpG methyltransferase enzyme SssI was used as positive control. A negative control was obtained from PCR-amplified genomic DNA for which CpG methylation was not amplified. Data from U87MG DNA demonstrated no major gaps in our sequencing coverage (Supplementary Figure 2, A, available online) and showed the expected methylation patterns in the positive and negative controls (Table 2 and Supplementary Figure 2, B–D, available online). For reference, the high-resolution methylome of U87MG derived from RRBS is listed in Supplementary Table 1 (available online).

Methylated Genes in IDH1 Wild-Type and Mutant Gliomas

RRBS was then applied to characterize the methylome of five pairs of pathologically matched IDH1 WT and MUT frozen glioma samples. The clinical characteristics of the ten tumors are listed in Table 1. Similar to results with U87MG DNA, approximately 19 million aligned reads were generated per sample, yielding methylation information on approximately 2 million CpG sites within 23 900 CpG islands, with an average of 59% coverage per island (Table 2). Approximately 14 000 covered CpG islands were associated with unique promoters.

The mean methylation levels of each CpG island were compared between the five IDH1 WT and five IDH1 MUT tumors. Using an absolute difference in methylation greater than 40% and a cutoff of P value of less than or equal to .05, we found 346 CpG islands that were statistically significantly differentially methylated between IDH1 WT and MUT tumors; 125 of the CpG islands were associated with promoters (Supplementary Table 2, available online). Given the small sample size, only 81 CpG islands, including RBP1, passed multiple testing correction with Q less than or equal to 0.05. The false discovery rate in this cohort, using the above statistical significance criteria with correction, was estimated at less than 0.02% by permutation testing.

Whole genome characterization revealed that methylation of IDH1 MUT tumors was statistically significantly increased compared with IDH1 WT tumors (Table 2 and Figure 2 A). This finding is consistent with a glioma CpG island methylator phenotype observed by Noushmehr et al. and Christensen et al. (8,9). Compared with publicly available data, 106 of the 346 genes identified in our study were among the 1550 hypermethylated genes found by Noushmehr et al. (8). Annotation analysis of hypermethylated genes using the software package DAVID (28) revealed enrichment of several annotation terms including transcriptional regulation and apoptosis (Figure 2 B).

RBP1 Hypermethylation in IDH1 and IDH2 Mutant Gliomas

To screen for candidate tumor suppressor genes that were hypermethylated in IDH1 MUT tumors, we employed the criteria that the hypermethylated CpG island had to be in the 5ʹ promoter region, show statistically significant hypermethylation within all three screening datasets (RRBS, MSRE, and TCGA), and correlate with decreased gene expression from the TCGA dataset. Several genes met these screening criteria, and a portion of the complete list (346 hypermethylated CpG islands, Supplementary Table 2, available online) is provided in Figure 2 C. One of the identified genes was RBP1 (Figure 3 A and Table 3). Analysis of data obtained from the TCGA database showed increased RBP1 methylation in 23 IDH1 MUT tumors compared with 124 WT tumors (Figure 3 B, and Table 3), and RBP1 promoter methylation was associated with decreased expression of CRBP1 in 141 GBM samples from the TCGA (R ² = .625, P < .001) (Figure 3 B).

Given that RBP1 was found to be hypermethylated in the three datasets, correlated with decreased gene expression, and has been reported in the literature to be involved in the synthesis of the transcription regulator ATRA (17,29,30), we sought to evaluate RBP1 methylation by gene-specific BiSEQ in the Total Cohort of 198 glioma patients with available frozen or paraffin-embedded tissues. All 198 patients in the Total Cohort had RBP1 methylation determined by BiSEQ and included the 41 patients in the initial methylation screen (10 by RRBS and 31 by MSRE). RBP1 was hypermethylated in 76 of 79 IDH1 MUT tumors, 3 of 3 IDH2 MUT tumors, and 0 of 116 IDH1/IDH2 WT tumors (Table 4). The RBP1 promoter was found by BiSEQ to be either essentially fully methylated or fully unmethylated across the 21 CpG sites evaluated (Supplementary Figure 3, available online). Also, in samples with BiSEQ and RRBS/MSRE data, the results from the techniques were concordant, supporting the overall accuracy of our RRBS/MSRE dataset (Supplementary Figure 4, available online). These data show that both IDH1 and IDH2 mutations are highly correlated with RBP1 hypermethylation in the 198 gliomas we analyzed (Spearman R ² = 0.94, 95% confidence interval = 0.92 to 0.95, P < .001).

RBP1 Hypermethylation and CRBP1 Regulation

To assess if RBP1 hypermethylation is associated with decreased CRBP1 expression, mRNA and protein expression were measured in various cell lines and glioma tumor samples by quantitative RT-PCR and western blot. Nontransformed APC cells were unmethylated for RBP1, and CRBP1 expression in APC cells was used as the reference. All of the tested glioma cell lines demonstrated RBP1 promoter methylation and decreased CRBP1 mRNA, despite being IDH1 wild-type (Figure 4 A). Western blot results were consistent with decreased CRBP1 protein expression in glioma cell lines (Figure 4 C). Also, we performed quantitative RT-PCR and western blot on 12 IDH1 WT, 43 IDH1 MUT, and 1 IDH2 MUT glioma samples using available frozen tissue. In all 56 samples, RBP1 was unmethylated in IDH1/IDH2 WT and methylated in IDH1/IDH2 MUT (data not shown). RBP1 mRNA levels from seven normal brain samples were used as controls. Figure 4 B shows that IDH1/IDH2 MUT tumors (n = 44) had statistically significantly decreased mRNA levels compared with WT tumors (n = 12) (IDH1/IDH2 MUT tumors vs WT tumors: mean = 17.93, SD =39.3 vs mean = 208.9, SD =337.5, respectively; P < .001). Western blot showed high CRBP1 protein expression in WT tumors (n = 12), but essentially undetectable levels in mutant tumors (n = 44) (Figure 4 C), which was confirmed by densitometric analysis (relative CRBP1 in IDH1/IDH2WT vs MUT tumors: mean = 193.3, SD = 224.2 vs mean = 0.02, SD = 0.15, respectively; P < .001) (data not shown).

Relationship Between Overall Survival and RBP1 Methylation Status

We identified 124 primary GBM patients in our cohort who had treatment-naïve tumor samples and who had received standard chemoradiation postoperatively (21). In our cohort, patients with WT GBM tumors (n = 101) had a statistically significant decrease in median OS compared with patients with IDH1 or IDH2 MUT GBM tumors (n = 23) (20.3 months vs 36.9 months, respectively; hazard ratio of death = 2.91, 95% confidence interval = 1.50 to 5.66, P = .002) (Figure 5 A). As expected from the close association between RBP1 promoter methylation and IDH1 or IDH2 mutation, RBP1-unmethylated patients (n = 102) had a statistically significantly decreased median OS of 20.3 months vs 36.8 months for RBP1-methylated patients (n = 22; hazard ratio of death = 2.48, 95% confidence interval= 1.30 to 4.75, P = .006) (Figure 5 B). Multivariable analysis showed that either IDH1/IDH2 mutation or RBP1 methylation status is an independent predictor of OS when IDH1/IDH2 mutation or RBP1 methylation status was analyzed separately with the clinical variables age, sex, performance status, and extent of resection (Supplementary Tables 3 and 4, available online). When patients were stratified by whether or not they received adjuvant isotretinoin with temozolomide, RBP1-unmethylated patients who were treated (n = 19) and those who were not (n = 83) had a similar median OS (22.0 months vs 19.8 months, respectively; P = .25) (Supplementary Figure 5, A, available online). The median OS for RBP1-methylated patients who had received isotretinion (n = 7) and untreated patients (n = 15) was also similar (2.5 months vs 36.7 months, respectively; P = .53) (Supplementary Figure 5, B, available online).