Background: Epigenome-wide association studies are emerging in the field of cancer epidemiology with the rapid development of large-scale methylation array platforms. Until recently, these methods were only valid for DNA from flash frozen (FF) tissues. Novel techniques for repairing DNA from formalin-fixed paraffin-embedded (FFPE) tissues have emerged; however, a direct comparison of FFPE DNA repair methods before analysis on genome-wide methylation array to matched FF tissues has not been conducted.

Methods: We conducted a systematic performance comparison of two DNA repair methods (REPLI-g Ligase vs. Infinium HD Restore Kit) on FFPE-DNA compared with matched FF tissues on the Infinium 450K array. A threshold of discordant methylation between FF-FFPE pairs was set at Δβ > 0.3. The correlations of β-values from FF–FFPE pairs were compared across methods and experimental conditions.

Results: The Illumina Restore kit outperformed the REPLI-g ligation method with respect to reproducibility of replicates (R2 > 0.970), highly correlated β-values between FF-FFPE (R2 > 0.888), and fewest discordant loci between FF-FFPE (≤0.61%). The performance of the Restore kit was validated in an independent set of 121 FFPE tissues.

Conclusions: The Restore kit outperformed RELPI-g ligation in restoring FFPE-derived DNA before analysis on the Infinium 450K methylation array. Our findings provide critical guidance that may significantly enhance the breadth of diseases that can be studied by methylomic profiling.

Impact: Epigenomic studies using FFPE tissues should now be considered among cancers that have not been fully characterized from an epigenomic standpoint. These findings promote novel epigenome-wide studies focused on cancer etiology, identification of novel biomarkers, and developing targeted therapies.

See all the articles in this CEBP Focus section, “Biomarkers, Biospecimens, and New Technologies in Molecular Epidemiology.”

Cancer Epidemiol Biomarkers Prev; 23(12); 2622–31. ©2014 AACR.

Epigenome-wide association studies (EWAS; ref. 1) focused on disease risk (2, 3) and environmental exposures (4, 5) are emerging in the field of cancer epidemiology. DNA methylation is a stable epigenetic modification of DNA that occurs primarily at cytosine-guanine (CpG) dinucleotide pairs within CpG islands in close proximity to gene promoters (6). DNA methylation alterations are critical events in carcinogenesis that parallel genomic mutational alterations and often occur early in carcinogenesis (7–10). Genome-wide methylation profiling is technically feasible with the emergence of several large-scale methylation assays, including the Illumina Infinium 450K arrays (11, 12), methylated DNA immunoprecipitation (13), and reduced representation bisulfite sequencing (14). The Infinium 450K array is a high-throughput assay interrogating more than 450,000 CpG sites and is considered the platform best suited for large studies (11, 12). However, until recently, its use had been limited to high-quality high-molecular weight DNA (11) derived from flash-frozen (FF) tissues that are uniformly fragmented for unbiased whole genome amplification (WGA; ref. 15). Because of the formalin-induced damage (e.g., nonuniform fragmentation and cross-linking; ref. 16), DNA from formalin-fixed paraffin-embedded (FFPE) tissues has been generally unsuitable for WGA due to potential biased amplification (15). This limitation of the methylation array technologies has limited molecular epidemiologic studies focused on the epigenome of cancers for which only FFPE blocks are available.

Overcoming the technical limitations inherent to genome-wide methylation array of FFPE specimens is critical to leverage the power of epigenomic analyses for many cancers and to promote EWAS (1). To address the DNA requirements for the Infinium platform, Thirwell and colleagues proposed a novel method to repair random DNA fragmentation by random ligation (REPLI-g ligase protocol) before bisulfite modification. This method generated highly concordant methylation values (e.g., R2 = 0.97) from paired FF-FFPE tissues using the Illumina 27K array (17), which overall were replicated by Jasmine and colleagues (18). However, the performance of the Thirwell technique of FFPE-DNA (e.g., a comparison of paired FF-FFPE tissues) before analysis on the Infinium 450K array, which interrogates a larger portion of the genome, is unknown. With the development of the Infinium 450K methylation array, Illumina developed the Infinium HD FFPE Restore Kit (19) to repair bisulfite-modified FFPE DNA before WGA using a combination of DNA polymerases and ligases to restore DNA length. The Restore protocol differs from Thirwell protocol with respect to (i) the order of ligation and bisulfite modification, (ii) the ligase/polymerase methodology, and (iii) the minimum starting DNA requirements. Furthermore, the Thirwell REPLI-g ligase method is less expensive and requires less processing time; therefore, it may be an optimal choice for large-scale population studies if found to perform accurately on the Infinium 450K array. However, the performance of the two DNA repair methods (REPLI-g ligase and Illumina Restore Kit) have not been directly compared on DNA interrogated by the Illumina 450K array. Herein, we report an independent and systematic comparison of the REPLI-g ligase method and Ilumina Restore Kit for repairing FFPE-DNA before WGA and analysis on the Infinium 450K array. We examined several parameters that differed between the methods as well as several data normalization methods. The optimal method was successfully tested on 121 archived FFPE tissues from a completed national clinical trial, Radiation Therapy Oncology Group (RTOG) trial 98-11.

### Overview of experimental parameters

The REPLI-g ligase (LIG) and Illumina Restore (RES) methods for repairing DNA extracted from FFPE blocks were compared, as outlined in Fig. 1. The original Thirwell method (17) was tested using 500 ng of genomic DNA that underwent REPLI-g ligation and bisulfite modification with 4 μL (LIG1) or 8 μL (LIG2) of bisulfite-modified DNA used for WGA and Illumina 450K array. LIG1 and LIG2 were considered REPLI-g ligase technical replicates. Bisulfite modification degrades genomic DNA (20) and may counteract the function of the ligase when performed after ligation; therefore, bisulfite modification of 500 ng (LIG3) or 250 ng (LIG4) of genomic DNA was conducted before REPLI-g ligation. The minimum starting amount of DNA is 500 ng from FF and 250 ng (19) or 500 ng (17) from FFPE, depending on the repair method. DNA inputs of 500 ng (RES1) and 250 ng (RES2/RES3) were used for the Illumina Restore kit. RES2 and RES3 were technical replicates for the Illumina Restore method. DNA from RTOG tissues was processed according to RES2 parameters (Fig. 1).

Figure 1.

Schematic of experimental parameters comparing FFPE DNA repair methods to FF DNA. DNA (500 ng) from FF tissue was processed according to Illumina Human Methylation instructions, including bisulfite modification followed by the standard Human Methylation processing protocol (Gold standard). Experimental conditions are separated for REPLI-g ligation (LIG) and Illumina Restore Kit (RES). LIG1: the original Thirwell method using 500 ng of genomic DNA processed by REPLI-g ligase and bisulfite modified (BS), and 4 μL of bisulfite-modified DNA used for the starting material for the Illumina Human Methylation array kit. LIG2: Thirwell method with output DNA increased to 8 μL of bisulfite-modified DNA, which is the same as used in the Restore Kit. LIG3 and LIG4: 500 ng and 250 ng of genomic DNA, respectively, were bisulfite modified, processed by REPLI-g ligase and 8 μL of material used for Illumina Methylation Array kit. RES1: the Illumina Restore protocol using 500 ng of genomic DNA, bisulfite modified, and processed per Restore Kit protocol (including 8 μL for Array steps). RES2 and RES3: technical replicates for Restore Kit using 250 ng of genomic DNA.

Figure 1.

Schematic of experimental parameters comparing FFPE DNA repair methods to FF DNA. DNA (500 ng) from FF tissue was processed according to Illumina Human Methylation instructions, including bisulfite modification followed by the standard Human Methylation processing protocol (Gold standard). Experimental conditions are separated for REPLI-g ligation (LIG) and Illumina Restore Kit (RES). LIG1: the original Thirwell method using 500 ng of genomic DNA processed by REPLI-g ligase and bisulfite modified (BS), and 4 μL of bisulfite-modified DNA used for the starting material for the Illumina Human Methylation array kit. LIG2: Thirwell method with output DNA increased to 8 μL of bisulfite-modified DNA, which is the same as used in the Restore Kit. LIG3 and LIG4: 500 ng and 250 ng of genomic DNA, respectively, were bisulfite modified, processed by REPLI-g ligase and 8 μL of material used for Illumina Methylation Array kit. RES1: the Illumina Restore protocol using 500 ng of genomic DNA, bisulfite modified, and processed per Restore Kit protocol (including 8 μL for Array steps). RES2 and RES3: technical replicates for Restore Kit using 250 ng of genomic DNA.

Close modal

### Tissues specimens

#### FF and FFPE specimens.

Matched FF colon adenocarcinoma tissue and corresponding archived FFPE blocks were obtained from three patients between 1999 and 2008 (2 males and 1 female; Supplementary Table S1). Histopathology was reviewed and confirmed by a dedicated gastrointestinal pathologist. Similar regions of invasive tumor within FF and FFPE specimens were marked and macrodissected. This study was reviewed by the University of South Florida Institutional review board and determined to be exempt.

#### Archived FFPE tissues from RTOG 98-11.

As a validation set, we utilized archived FFPE sections collected in the RTOG 98-11 trial for the treatment of squamous cell carcinoma of the anus (21). Sections from archived FFPE tissues (4 × 10 μm or 8 × 5 μm sections) were obtained from 186 cases, pathologically reviewed, and macrodissected (22).

### DNA extraction

DNA from FF tissue was isolated using the QIAamp DNA Blood Mini Kit (QIAGEN Inc.) following manufacturer's recommendations. DNA from sections of colon FFPE blocks (e.g., 10 μm × 8–10 slides) or RTOG tissues was isolated using the QIAamp DNA FFPE Tissue Kit (QIAGEN Inc.). DNA was evaluated using the NanoDrop1000 Spectrophotometer and double-stranded DNA concentration determined using the Qubit 2.0 Fluorometer.

### Quality control

Quality of FFPE DNA was tested in triplicate by real-time PCR using the Illumina FFPE QC kit (Illumina, Inc.). Amplification of the FFPE sample DNA was compared with the amplification of a Quality Control template (QCT). The real-time PCR threshold cycle (Ct), or the quantification value, was averaged across replicates and a ΔCt for each sample was calculated (CtFFPECtQCT). An FFPE DNA sample was deemed adequate for the Illumina 450K array if the ΔCt was <5.

### Bisulfite conversion

Genomic DNA was directly subjected to bisulfite modification for FF samples, LIG3, LIG4, and all RES samples; whereas, DNA for LIG1 and LIG2 was processed using the REPLI-g ligase before bisulfite modification (Fig. 1). Genomic DNA (500 or 250 ng) or REPLI-g ligated DNA in 45 μL volume was bisulfite modified using the EZ DNA Methylation Kit (cat#D5001, Zymo Research) following manufacturer's instructions. A volume of 9.5 μL M-elution buffer was used for all conditions except LIG3 and LIG4, which required 11.5 μL M-elution buffer.

### REPLI-g ligation of FFPE samples

The REPLI-g FFPE Kit (QIAGEN Inc.) was tested following Thirwell protocol (17) for LIG1 and LIG2. The sequence of bisulfite modification was altered for LIG3 and LIG4, with bisulfite-modified DNA treated with REPLI-g ligase according to published protocol followed by the Zymo Clean & Concentrator kit (Zymo Research Corporation) before processing for the Illumina 450K array (LIG 3 and LIG4).

### Restoration of FFPE samples

Infinium HD FFPE DNA Restore kit (Illumina, Inc.; ref. 19) protocol was carried out according to the manufacturer's instructions on 8 μL of bisulfite-treated FFPE DNA (RES1, RES2, and RES3). The DNA was eluted with DiH20 after a 5-minute incubation and stored at −20°C before the Infinium processing.

### Genome-wide methylation assay

The Infinium Human Methylation 450K Beadchip (Illumina, Inc.) measures DNA methylation at 482,421 CpG loci, which covers approximately 99% of RefSeq genes, 96% of UCSC CpG islands, and specific commonly methylated CpG sites in human cancers (www.ncbi.nlm.nih.gov/RefSeq, www.illumina.com). This protocol was carried out according to the manufacturer's recommendations. All samples utilized 8 μL starting material except for FF and LIG1, which used 4 μL of bisulfite-modified DNA as per Thirwell and colleagues (17). A Tecan Liquid Handling robot with the Te-Flow apparatus was used for the single base extension and staining, and the chips were scanned on a single HiScanSQ System (Illumina Inc.). Paired FF and FFPE DNA were processed on the same chip to reduce batch effects. RTOG DNA samples were run in 3 batches of 24 and included two interbatch technical replicates (Case ID 37 and 370).

### Data analysis

Three data normalization methods [minfi package-implemented Illumina (23), SWAN (24), and DASEN (25)] and raw β-values were examined. The β-value was calculated for the 485,512 CpG-loci from unmethylated (U) and methylated (M) signal [M/(U + M + 100)] and assigned a range between 0 and 1 (unmethylated to 100% methylated) within the Illumina Infinium software and provided within the raw idat file. The raw idat files were processed for all normalization packages in R/Bioconductor (26). The minfi Bioconductor package (ref. 23; version 1.8.7) generated three distinct MethylSet objects: (i) the preprocess-Raw method, (ii) preprocess-Illumina method with background correction and normalization via controls, and (iii) preprocess-SWAN method (24). β-Values for each MethylSet were extracted using minfi's getBeta method with the parameter type = “Illumina” to ensure that the β-values were calculated with an offset of 100. The DASEN normalization used the DASEN method in the WateRmelon package (ref. 27, version 1.2.2; ref. 25). All normalizations ranked the quality of experimental conditions identically; primary results are reported using the DASEN normalization method. Paired FF-FFPE methylation data is available as GEO series GSE59038 as both raw idat files and processed β-values (DASEN normalization).

Normalized β-values for CpG loci with a detection P value > 0.05 were removed. β-Values were analyzed as continuous variables. To quantitate the discordant β-values between FF DNA and FFPE-repaired DNA, we calculated a Δβ = β-valueFF − β-valueFFPE. The number of CpG loci with Δβ either above or below (e.g., absolute value of Δβ noted as |Δβ|) 0.3 was calculated. Correlation between FF and FFPE was determined using Pearson correlation coefficient. Significant differences in mean R2 or Δβ between REPLI-g ligase and Restore methods were determined using a paired t test (two-sided). All calculations and correlations were done in MATLAB. Principal component analysis (PCA) was performed using Evince v2.5.5 software (UmBio AB, www.umbio.com).

We examined two methods (REPLI-g ligase and Illumina Restore) for repairing DNA extracted from FFPE tissues and analyzed on the Infinium 450K array. Three paired sets of FF-FFPE colon tumors that varied by time since collection (4–13 years) and gender were used (Supplementary Table S1). The experimental design and standard nomenclature are presented in Fig. 1 and outlined in Materials and Methods.

### Patient variability drives the largest differences in β-values

The sources of variation in methylation β-values across the patient samples and all conditions were examined using PCA. As shown in Fig. 2A, the samples clustered primarily by patient (triangles, circles, and diamonds) in the first and second principal components (PC1 and PC2), which accounted for 36.6% and 27.8% of the variation in β-values, respectively. This indicates that patient-related differences in β-values are larger than differences by tissue storage type (FF vs. FFPE) or ligation method (LIG vs. RES). PC3 and PC4 (Fig. 2B) explain an additional 5.9% and 3.9% of the variation in β-values, respectively. There was a clear similarity among all Restore samples across patients (e.g., clustering of red, orange, and gray points) in PC3 and PC4 with little variability compared with the separation of the REPLI-g ligase-processed samples (dark blue, yellow, light blue, and purple points; Fig. 2B). The LIG3 samples (light blue points) were most similar to the Restore samples and clustered together across patients. In summary, interpatient variation in β-values was high; all Restore samples and LIG3 displayed the least variation in β-values between patients and were most similar to FF samples.

Figure 2.

PCA was used to identify the sources of variation in methylation β-values across the three patient samples and all conditions. Patient samples are denoted by shape (A, circle; B, triangle; and C, diamond). A, Scatter plot of samples in the first and second principal components (PC1 and PC2), which accounts for 36.6% and 27.8% of the variation. B, Scatter plot of PC3 and PC4, which explains 5.9% and 3.9% of the variation with separation by tissue storage type (FF vs. FFPE) as well as the ligation method (LIG vs. RES).

Figure 2.

PCA was used to identify the sources of variation in methylation β-values across the three patient samples and all conditions. Patient samples are denoted by shape (A, circle; B, triangle; and C, diamond). A, Scatter plot of samples in the first and second principal components (PC1 and PC2), which accounts for 36.6% and 27.8% of the variation. B, Scatter plot of PC3 and PC4, which explains 5.9% and 3.9% of the variation with separation by tissue storage type (FF vs. FFPE) as well as the ligation method (LIG vs. RES).

Close modal

### Reproducibility of FFPE-based genome-wide methylation is higher using Restore method

The reliability of each DNA repair method was assessed using β-values from the Infinium 450K array of technical replicates (RES2/RES3 and LIG1/LIG2). The correlation between Restore internal replicates (RES2 vs. RES3) ranged from 0.970 to 0.989 (Fig. 3A). The correlations between REPLI-g ligase replicates (LIG1 vs. LIG2) were lower (0.850–0.972, Fig. 3B). REPLI-g ligation resulted in more β-values off the diagonal (Fig. 3B) compared with the Restore replicates (Fig. 3A). There were 346 total CpG loci across patients with discordant β-values (|Δβ|) of >0.3 between RES2 and RES3 (Supplementary Table S2). In comparison, REPLI-g ligase replicates had a total of 12,718 loci across patients with |Δβ| > 0.3.

Figure 3.

Density correlation plot between internal replicate samples. Correlations are shown using (A) the Restore kit (RES2 vs. RES3) and (B) REPLI-g Ligase (LIG1 vs. LIG2) for each tumor. Colors range from blue (low point density) to red (high point density), with the highest density along the diagonal.

Figure 3.

Density correlation plot between internal replicate samples. Correlations are shown using (A) the Restore kit (RES2 vs. RES3) and (B) REPLI-g Ligase (LIG1 vs. LIG2) for each tumor. Colors range from blue (low point density) to red (high point density), with the highest density along the diagonal.

Close modal

### Highest correlations between FF-FFPE β-values are achieved using Restore

The goal of both methods is to repair DNA such that amplification bias during WGA due to potentially nonuniformly fragmented DNA from FFPE preserved tissues is reduced. To quantify whether the DNA repair methods reduce this bias and methylation values from FFPE-derived DNA are representative of FF-derived DNA, the β-values from FFPE-derived DNA were compared with β-values from matched FF-derived DNA. Overall, the average correlation between FF-FFPE pairs ranged from 0.804 to 0.938 (Table 1). For all samples, the correlations between FF and the Restore-treated FFPE DNA were significantly higher than the REPLI-g ligase-treated FFPE DNA (Mean R2: 0.91 vs. 0.86, respectively; P < 0.003). Correlations between FF-FFPE pairs using the Restore method were highest for RES1 (RES1 vs. FF, R2 = 0.888–0.938; Fig. 4A) but did not differ by DNA input amount (Table 1). Correlations using the REPLI-g ligase method were all lower than Restore. LIG3 had the highest correlations among the LIG samples (LIG3 vs. FF, R2 = 0.836–0.905; Fig. 4B). For Restore, β-values from 250 ng and 500 ng of FFPE DNA (Thirwell recommended starting amount) were highly correlated (RES1 vs. RES2, R2 = 0.974–0.989; Fig. 4C) and at a similar magnitude as seen between technical replicates (RES3 vs. RES2, R2 = 0.970–0.989; Fig. 3A).

Figure 4.

Representative density correlation plot between FF-FFPE pairs. Correlation of DASEN normalized β-values for (A) Restore (RES1 vs. FF) and (B) REPLI-g Ligase (LIG3 vs. FF) for each patient sample. C, correlation between FFPE Restore-processed samples by input DNA amount (500 ng RES1 vs. 250 ng RES2).

Figure 4.

Representative density correlation plot between FF-FFPE pairs. Correlation of DASEN normalized β-values for (A) Restore (RES1 vs. FF) and (B) REPLI-g Ligase (LIG3 vs. FF) for each patient sample. C, correlation between FFPE Restore-processed samples by input DNA amount (500 ng RES1 vs. 250 ng RES2).

Close modal
Table 1.

Summary of β-value correlation (R2) and discordance (|Δβ|) between FF-FFPE pairs for REPLI-g Ligation and the Restore methods

R2 of β-values FF-FFPE pairsaDiscordant loci between FF-FFPE pairsbOverlap of discordant locic
PatientOverallPatientOverallPairwise by patientOverall
ABCMean ± SDABCMean ± SD% CpG-LociABACBCABC
LIG1 0.857 0.812 0.900 0.856 ± 0.044 8,254 10,338 3,772 7,455 ± 3,355 1.54% 638 120 137 23
LIG2 0.849 0.810 0.901 0.853 ± 0.046 9,415 10,638 3,834 7,962 ± 3,627 1.64% 728 139 142 27
LIG3 0.905 0.836 0.902 0.881 ± 0.039 3,212 9,892 4,323 5,809 ± 3,579 1.20% 190 65 156
LIG4 0.860 0.804 0.898 0.854 ± 0.047 10,250 15,168 4,390 9,936 ± 5,396 2.05% 774 163 244 19
RES1 0.939 0.888 0.918 0.915 ± 0.026 1,124 3,402 3,206 2,577 ± 1,262 0.53% 111 51 65 19
RES2 0.941 0.875 0.919 0.912 ± 0.034 1,149 4,662 3,053 2,955 ± 1,759 0.61% 146 34 72 14
RES3 0.938 0.881 0.920 0.913 ± 0.029 1,214 3,969 3,263 2,815 ± 1,431 0.58% 136 35 65 15
R2 of β-values FF-FFPE pairsaDiscordant loci between FF-FFPE pairsbOverlap of discordant locic
PatientOverallPatientOverallPairwise by patientOverall
ABCMean ± SDABCMean ± SD% CpG-LociABACBCABC
LIG1 0.857 0.812 0.900 0.856 ± 0.044 8,254 10,338 3,772 7,455 ± 3,355 1.54% 638 120 137 23
LIG2 0.849 0.810 0.901 0.853 ± 0.046 9,415 10,638 3,834 7,962 ± 3,627 1.64% 728 139 142 27
LIG3 0.905 0.836 0.902 0.881 ± 0.039 3,212 9,892 4,323 5,809 ± 3,579 1.20% 190 65 156
LIG4 0.860 0.804 0.898 0.854 ± 0.047 10,250 15,168 4,390 9,936 ± 5,396 2.05% 774 163 244 19
RES1 0.939 0.888 0.918 0.915 ± 0.026 1,124 3,402 3,206 2,577 ± 1,262 0.53% 111 51 65 19
RES2 0.941 0.875 0.919 0.912 ± 0.034 1,149 4,662 3,053 2,955 ± 1,759 0.61% 146 34 72 14
RES3 0.938 0.881 0.920 0.913 ± 0.029 1,214 3,969 3,263 2,815 ± 1,431 0.58% 136 35 65 15

aPearson correlation coefficient comparing FF versus FFPE β-values for each patient sample, by experimental condition. Mean R2 for LIG and RES samples differed significantly by Student t test (P = 0.0028).

bThe number of discordant CpG-loci defined as |Δβ| > 0.3 between FF and FFPE. Mean number of discordant CpG-loci differed significantly between LIG and RES samples (Student t test, P = 0.0012). Percent of CpC-loci with |Δβ| > 0.3 out of total CpG-loci evaluated (N = 485,512) for each condition.

cOverlap of individual discordant CpG-loci (|Δβ| > 0.3) across patient samples (A, B, and C), pairwise and overall.

### Restore yields fewer CpG loci with discordant FF-FFPE β-values

Discordant β-values (|Δβ|) between FF-FFPE tissues were examined for the different techniques by determining the average number of CpG loci above a |Δβ| threshold of 0.3. The mean number of loci with |Δβ| > 0.3 for RES samples ranged from 2,577 to 2,955 (0.53%–0.61% of all loci; Table 1). The mean number of loci with |Δβ| > 0.3 was 2- to 3-fold higher in the LIG samples than RES (5,809–9,936 loci, 1.2%–2.05%). Restore-processed DNA had significantly less CpG loci with |Δβ| > 0.3 between FF-FFPE compared with REPLI-g ligase-processed DNA (P < 0.002).

### FF-FFPE β-value discordance occurs at random CpG loci

Next, it was determined whether discordant β-values occurred in the same CpG loci (differential bias) across samples or whether they occurred at random (nondifferential bias) by examining the overlap in loci with |Δβ| > 0.3 across patient samples and various conditions. The number of loci that were consistently discordant between FF-FFPE across all three patient samples (A&B&C) ranged from 7 to 27 (Table 1). Of note, no CpG loci had a |Δβ| > 0.3 consistent across all seven experiments. These observations were consistent for Restore replicates (Supplementary Table S2), with no overlapping loci with |Δβ| > 0.3 across all three patients. These results suggest that loci with discordant β-values between matched FF-FFPE DNA samples occurred at random and were not due to specific poorly performing loci in FFPE samples.

### Impact of modulating experimental parameters

Experimental conditions that differed across the two methods were evaluated (Fig. 1). There were no differences for Restore-processed DNA by starting DNA (500 ng vs. 250 ng) for either the correlations between FF-FFPE or number of loci with |Δβ| > 0.3 (Fig. 4C and Table 1). For REPLI-g ligase, 500 ng starting DNA (LIG3) performed better than 250 ng (LIG4) for all metrics examined. LIG3 had the fewest number of loci with |Δβ| > 0.3 (mean loci = 5,809, 1.20% of all loci) among REPLI-g ligase experiments (Table 1) and clustered with RES samples by PCA (Fig. 2B). Bisulfite modification before REPLI-g ligation (LIG3/4) outperformed the reverse sequence (LIG1/2) with higher R2 (mean R2 = 0.867 and 0.855, respectively; Table 1). Finally, there was minimal impact of changing the volume of material taken forward for the Infinium assay (Mean |Δβ| > 0.3: 7,455 and 7,962 for LIG1 and LIG2, respectively; Table 1).

### Impact of data normalization

There is consensus that Infinium 450K β-values should be normalized (28, 29) and several methods have been proposed. We considered three normalization methods (minfi-Illumina, SWAN, and DASEN) in addition to using raw (un-normalized) β-values. Overall, all normalization techniques ranked the different experimental conditions identically on all performance metrics. Notably, there were differences observed in the magnitude and direction of the Δβ between FF-FFPE by normalization method. Compared with DASEN (Fig. 4), there was a tendency for minfi-Illumina to overestimate the methylation status of FFPE samples (vs. FF) with substantially more loci above the diagonal line (Supplementary Fig. S1). The Δβ between FF-FFPE samples were normally distributed with a mean Δβ of 0.002 using the DASEN method. There was a significant shift toward higher β-values within the FFPE samples (negative Δβ) when using minfi-Illumina (mean Δβ = −0.028, P = 6.03 × 10−8; Supplementary Fig. S2). Among all methods, DASEN also yielded significantly less loci with |Δβ| > 0.3 (mean |Δβ| > 0.3 = 7,831 across LIG samples and 2,794 for RES) while minfi-Illumina resulted in the highest number (mean |Δβ| > 0.3 = 18,082 across LIG samples and 5,528 for RES, P < 0.0001). The mean Δβ and number of discrepant loci (|Δβ| > 0.3) between FF-FFPE for all normalization methods are presented in Supplementary Tables S3 and S4, respectively.

### Testing of Restore methodology in archived FFPE tissues

The optimized conditions of 250 ng FFPE-derived DNA using the Illumina Restore kit (RES2) were applied to a set of archived FFPE tissues collected between 1998 and 2005 as part of a cooperative group anal cancer clinical trial. Of the 186 cases, 121 (65%) had ≥250 ng DNA. In quality control testing, the ΔCt of DNA from RTOG FFPE tissues ranged from −1.03 to 4.73; all 121 samples passed quality control testing. Percent missing CpG loci were <3% for all but 1 sample (>5%), which was excluded. β-Values from two replicate samples repeated across batches (Case-ID 37 and 370) were highly correlated (R2 = 0.984 and 0.931, Fig. 5) and had few CpG loci with |Δβ| > 0.3 (29 and 1,483, respectively). The distribution of β-values for the 121 FFPE samples from RTOG exhibited a consistent pattern (Supplementary Fig. S3B) that was comparable with FF-FFPE pairs (Patient A, B, and C, Supplementary Fig. S3A). All samples, regardless of storage type, had β-value peaks at 0.2 and 0.8 and variability between β-values of 0.3 and 0.6. PCA for all samples showed no outliers or batch-effects (data not shown). Taken together, these data demonstrate that using the Illumina Restore Kit with 250 ng of genomic DNA from archived pretreatment biopsy specimens more than 10 years old can result in high quality epigenomic data.

Figure 5.

Density correlation plots for RTOG FFPE tissue internal replicates using Restore method. Correlations are shown using the Restore kit (RES2 condition) for archived FFPE tissues collected from Case-ID 37 (A) and 370 (B) within the RTOG 98-11 clinical trial. Colors range from blue (low-point density) to red (high-point density), with the highest density along the diagonal as expected.

Figure 5.

Density correlation plots for RTOG FFPE tissue internal replicates using Restore method. Correlations are shown using the Restore kit (RES2 condition) for archived FFPE tissues collected from Case-ID 37 (A) and 370 (B) within the RTOG 98-11 clinical trial. Colors range from blue (low-point density) to red (high-point density), with the highest density along the diagonal as expected.

Close modal

The Illumina Restore kit outperformed the RELPI-g ligation method for restoring FFPE-derived DNA before use on the Infinium 450K methylation array. The Restore method had the best overall performance regardless of starting amount of DNA tested and was consistent across several data normalization methods, although DASEN normalization performed best. The drawbacks of the Restore kit include added costs per sample (∼$80 vs.$24 in US\$, respectively) and additional processing time (∼4 hours per batch). The Thirwell method with modification may also be an acceptable option when sufficiently large effect sizes are expected (e.g., |Δβ| > 0.5). Our findings provide valuable guidance for selecting a DNA repair method for FFPE samples before analysis on the Infinium 450K array and highlight several important factors for consideration when designing epigenome-wide association studies, including cost, DNA requirements, and processing of resultant data.

Many cancers, such as anal, rectal, and esophageal, are treated with chemoradiation before surgery and only small pretreatment diagnostic FFPE biopsies are available for molecular analysis. The best method for repairing DNA from these tissue types for genome-wide methylation analysis (e.g., most representative of matched FF tissue) had not been determined. Our study provides evidence that “restored” FFPE-derived DNA generated β-values from the Illumina 450K array that were representative of FF tissues. We then validated these finding using DNA from pretreatment FFPE biopsies archived as part of the RTOG 98-11 anal cancer trial (21). DNA extracted from pretreatment FFPE biopsies met quality standards and yielded high-quality methylation data, as determined by a low percentage of undetectable probes, a consistent distribution of β-values across the array, and highly correlated β-values from sample replicates. Notably, the use of 250 ng starting DNA, instead of 500 ng as recommended by Thirwell and colleagues (17), increased our sample size by 43 cases (36% increase) that had DNA yields between 250 ng and 500 ng. The delineation of the optimal experimental conditions and restore methods should open the door for the characterization of genome-wide methylation in a number of cancers that remain heretofore inadequately studied from an epigenomic standpoint.

Using a comprehensive set of quality measures, we systematically compared two methods and three experimental parameters for repairing nonuniformly fragmented FFPE-derived DNA before WGA and evaluation on the Infinium 450K methylation array. To our knowledge, this is the first study to directly compare the Illumina Restore kit to the methods by Thirwell and colleagues (17) before analysis on the Infinium 450K array. The overall correlations of β-values between FF and FFPE tissues observed in this study was similar, if not greater, than those previously reported for the Infinium 27K array (17, 18). Lechner and colleagues demonstrated that FFPE DNA repaired using the REPLI-g ligase was sufficient for use on the Illumina 450K array when stringent quality control criteria were applied before data analysis and large effect sizes were observed [e.g., differential methylation between human papillomavirus (HPV) positive vs. HPV-negative head and neck cancers; ref. 30). In contrast to Jasmine and colleagues (18), our data suggest that there is minimal misclassification of β-values generated from FFPE DNA for both the REPLI-g ligase and the Restore; however, the Restore method outperformed the REPLI-g ligase method.

Overall, our findings that the Restore processed samples provided β-values representative of FF tissues are in agreement with those reported by Dumenil and colleagues (31), who analyzed 21 FF-FFPE colorectal cancer tissue pairs using the Illumina Restore and Illumina 450K array. They reported highly consistent β-values across FF-FFPE pairs, very low percentage of undetectable probes (<1%), and overlap in differentially methylated loci between FFPE and FF tissues (31). However, Dumenil and colleagues did not compare the Restore method to the less expensive REPLI-g ligation method nor did they consider more than one data normalization method. As both our study and Dumenil observed some differences in β-values between FF and restored FFPE-derived DNA (as expected), we strongly advocate that a minimum |Δβ|-threshold (e.g., |Δβ| > 0.3) be utilized, in addition to statistical significance, when identifying differentially methylated loci from Illumina 450K data.

The array-based DNA methylation data normalization field is still nascent with new normalization techniques being proposed and methods continuing to be debated (see review by Wilhelm–Benartzi and colleagues, ref. 28). This is in contrast to the more mature array-based methods such as gene expression. An important observation of this study is that the selection of the best experimental protocol was independent of the four different normalization methods investigated. Nonetheless, the DASEN (25) technique performed best among the methods tested. This could be explained by the fact that DASEN is a global sample-to-sample normalization technique while minfi-Illumina and SWAN normalize each sample independently.

The validity of our findings is strengthened by our experimental rigor, including the inclusion of all paired samples on the same chip to reduce chip-to-chip variation. Although our experimental study consisted of only three paired cases, these specimens are representative of typical FFPE tissues that would be included in larger studies with variation in specimen age and patient gender. This was also evident by the variability observed in the methylation data. The experimental findings were successfully tested on 121 anal cancer FFPE tissues collected within a clinical trial, thus demonstrating the applicability of this method to clinical specimens from several pathology laboratories. We included two types of cancer in this study, colorectal and anal to represent (i) tumors with sufficient material to obtain FF and FFPE matched samples that provided >2 μg of DNA and (ii) small pretreatment biopsies with minimal amounts of DNA, respectively. This study did not include normal tissues, as examined by Jasmine and colleagues (18). However, as the goal was to identify a method that generated methylation results within FFPE tissues as similar as possible to matched FF tissue, the lack of normal tissues or use of different tumor types did not impact our conclusions.

This study demonstrated that FFPE-derived DNA processed using the Illumina Restore provides robust genome-wide methylation results that are similar to those from optimally stored matched FF tissues. This DNA repair method is recommended above the REPLI-g ligation method for future epigenomic studies. Aberrant methylation occurs during critical processes of aging, development, and carcinogenesis (2, 3); as such, these recommendations will have widespread implications and should greatly increase the breadth of diseases that can undergo methylomic profiling. This may, in turn, lead to further elucidation of disease pathogenesis, identification of novel biomarkers, and the development of targeted therapies.

No potential conflicts of interest were disclosed.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the H. Lee Moffitt Cancer Center and Research Institute.

Conception and design: E.M. Siegel, D. Shibata

Development of methodology: E.M. Siegel, A.E. Berglund, B.M. Riggs, S.A. Eschrich, A.O. Ajidahun, D. Shibata

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): B.M. Riggs, A.O. Ajidahun, D. Coppola

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E.M. Siegel, A.E. Berglund, S.A. Eschrich, R.M. Putney, D. Coppola, D. Shibata

Writing, review, and/or revision of the manuscript: E.M. Siegel, A.E. Berglund, B.M. Riggs, S.A. Eschrich, A.O. Ajidahun, D. Coppola, D. Shibata

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): E.M. Siegel, A.E. Berglund, B.M. Riggs, S.A. Eschrich, A.O. Ajidahun

Study supervision: E.M. Siegel, D. Shibata

The authors thank Drs. Thomas Sellers and Alvaro Monteiro for their critical review of this article and Drs. Bill Grady and Andrew Kaz for critical review of our study design. The authors are also grateful for their ongoing collaboration the RTOG 98-11 Investigators (Drs. Jaffer Ajani and Chandan Guha) and Statisticians (Ms. Kathryn Winters and Jen Moughan) in the use of RTOG 98-11 tissue and data.

The research was supported in part by the American Society of Colon and Rectal Surgeons Foundation, under grant number LPG-097 (to E.M. Siegel and D. Shibata). This work has been supported in part by the Cancer Informatics and Molecular Genomics Core Facility at the H. Lee Moffitt Cancer Center and Research Institute, an NCI-designated Comprehensive Cancer Center, under grant number P30-CA76292.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Michels
KB
,
Binder
AM
,
Dedeurwaerder
S
,
Epstein
CB
,
Greally
JM
,
Gut
I
, et al
Recommendations for the design and analysis of epigenome-wide association studies
.
Nat Methods
2013
;
10
:
949
55
.
2.
Rakyan
VK
,
Down
TA
,
Balding
DJ
,
Beck
S
.
Epigenome-wide association studies for common human diseases
.
Nat Rev Genet
2011
;
12
:
529
41
.
3.
Feinberg
AP
.
Genome-scale approaches to the epigenetics of common human disease
.
Virchows Arch
2010
;
456
:
13
21
.
4.
Feinberg
AP
.
Phenotypic plasticity and the epigenetics of human disease
.
Nature
2007
;
447
:
433
40
.
5.
Besingi
W
,
Johansson
A
.
Smoke-related DNA methylation changes in the etiology of human disease
.
Hum Mol Genet
2014
;
23
:
2290
7
.
6.
Jones
PA
.
Functions of DNA methylation: islands, start sites, gene bodies and beyond
.
Nat Rev Genet
2012
;
13
:
484
92
.
7.
Timp
W
,
Feinberg
AP
.
Cancer as a dysregulated epigenome allowing cellular growth advantage at the expense of the host
.
Nat Rev Cancer
2013
;
13
:
497
510
.
8.
Brait
M
,
Sidransky
D
.
Cancer epigenetics: above and beyond
.
Toxicol Mech Methods
2011
;
21
:
275
88
.
9.
Taby
R
,
Issa
JP
.
Cancer epigenetics
.
CA Cancer J Clin
2010
;
60
:
376
92
.
10.
Sjoblom
T
,
Jones
S
,
Wood
LD
,
Parsons
DW
,
Lin
J
,
Barber
TD
, et al
The consensus coding sequences of human breast and colorectal cancers
.
Science
2006
;
314
:
268
74
.
11.
Bibikova
M
,
Le
J
,
Barnes
B
,
Saedinia-Melnyk
S
,
Zhou
L
,
Shen
R
, et al
Genome-wide DNA methylation profiling using Infinium(R) assay
.
Epigenomics
2009
;
1
:
177
200
.
12.
Dedeurwaerder
S
,
Defrance
M
,
Calonne
E
,
Denis
H
,
Sotiriou
C
,
Fuks
F
.
Evaluation of the Infinium Methylation 450K technology
.
Epigenomics
2011
;
3
:
771
84
.
13.
Mohn
F
,
Weber
M
,
Schubeler
D
,
Roloff
TC
.
Methylated DNA immunoprecipitation (MeDIP)
.
Methods Mol Biol
2009
;
507
:
55
64
.
14.
Gu
H
,
Smith
ZD
,
Bock
C
,
Boyle
P
,
Gnirke
A
,
Meissner
A
.
Preparation of reduced representation bisulfite sequencing libraries for genome-scale DNA methylation profiling
.
Nat Protoc
2011
;
6
:
468
81
.
15.
Bosso
M
,
Al-Mulla
F
.
Whole genome amplification of DNA extracted from FFPE tissues
.
Methods Mol Biol
2011
;
724
:
161
80
.
16.
Gilbert
MT
,
Haselkorn
T
,
Bunce
M
,
Sanchez
JJ
,
Lucas
SB
,
Jewell
LD
, et al
The isolation of nucleic acids from fixed, paraffin-embedded tissues-which methods are useful when?
PLoS ONE
2007
;
2
:
e537
.
17.
Thirwell
C
,
Eymard
M
,
Feber
A
,
Teschendorff
A
,
Pearce
K
,
Lechner
M
, et al
Genome-wide DNA methylation analysis of archival formalin-fixed paraffin-embedded tissue using the Illumina Infinium HumanMethylation27 BeadChip
.
Methods
2010
;
52
:
248
54
.
18.
Jasmine
F
,
Rahaman
R
,
Roy
S
,
Raza
M
,
Paul
R
,
Rakibuz-Zaman
M
, et al
Interpretation of genome-wide infinium methylation data from ligated DNA in formalin-fixed, paraffin-embedded paired tumor and normal tissue
.
BMC Res Notes
2012
;
5
:
117
.
19.
Illumina
.
Infinium HD FFPE Restore Protocol, Catalog # WG-901-2004
.
2011
[cited 2012 06/20/2012]; Available from
: http://supportres.illumina.com/documents/myillumina/5c3d90a3-793c-4a8b-932b-0434590f98ef/infinium_ffpe_sample_restore_booklet_15014614_c.pdf
20.
Grunau
C
,
Clark
SJ
,
Rosenthal
A
.
Bisulfite genomic sequencing: systematic investigation of critical experimental parameters
.
Nucleic Acids Res
2001
;
29
:
E65
5
.
21.
Ajani
JA
,
Winter
KA
,
Gunderson
LL
,
Pedersen
J
,
Benson
AB
III
,
Thomas
CR
Jr
, et al
Fluorouracil, mitomycin, and radiotherapy vs fluorouracil, cisplatin, and radiotherapy for carcinoma of the anal canal: a randomized controlled trial
.
JAMA
2008
;
299
:
1914
21
.
22.
Siegel
EM
,
Eschrich
SA
,
Winter
KA
,
Riggs
B
,
Berglund
A
,
Ajidahun
A
, et al
Epigenomic characterization of locally advanced anal cancer: an RTOG 98-11 specimen study
.
Dis Colon Rectum
2014
;
57
:
941
57
.
23.
Hansen
K
,
Aryee
M
.
Minfi: Analyze Illumina 450k Methylation Arrays
.
R package version 12
;
2012
.
24.
Maksimovic
J
,
Gordon
L
,
Oshlack
A
.
SWAN: Subset-quantile within array normalization for illumina infinium HumanMethylation450 BeadChips
.
Genome Biol
2012
;
13
:
R44
.
25.
Pidsley
R
,
Wong
CC
,
Volta
M
,
Lunnon
K
,
Mill
J
,
Schalkwyk
LC
.
A data-driven approach to preprocessing Illumina 450K methylation array data
.
BMC Genomics
2013
;
14
:
293
.
26.
Gentleman
RC
,
Carey
VJ
,
Bates
DM
,
B
,
Dettling
M
,
Dudoit
S
, et al
Bioconductor: open software development for computational biology and bioinformatics
.
Genome Biol
2004
;
5
:
R80
.
27.
Schalkwyk
LC
,
Pidsley
R
,
Wong
CC
,
wfcbN
T
,
Defrance
M
,
Teschendorff
AE
, et al
wateRmelon: Illumina 450 methylation array normalization and metrics
.
R package. 1.4.0. ed
;
2013
.
28.
Wilhelm-Benartzi
CS
,
Koestler
DC
,
Karagas
MR
,
Flanagan
JM
,
Christensen
BC
,
Kelsey
KT
, et al
Review of processing and analysis methods for DNA methylation array data
.
Br J Cancer
2013
;
109
:
1394
402
.
29.
Wu
MC
,
Joubert
BR
,
Kuan
PF
,
Haberg
SE
,
W
,
SD
, et al
A systematic assessment of normalization approaches for the Infinium 450K methylation platform
.
Epigenetics
2014
;
9
:
318
29
.
30.
Lechner
M
,
Fenton
T
,
West
J
,
Wilson
G
,
Feber
A
,
Henderson
S
, et al
Identification and functional validation of HPV-mediated hypermethylation in head and neck squamous cell carcinoma
.
Genome Med
2013
;
5
:
15
.
31.
Dumenil
TD
,
Wockner
LF
,
Bettington
M
,
McKeone
DM
,
Klein
K
,
Bowdler
LM
, et al
Genome-wide DNA methylation analysis of formalin-fixed paraffin embedded colorectal cancer tissue
.
Genes Chromosomes Cancer
2014
;
53
:
537
48
.