Genome-Wide, High-Resolution Detection of Copy Number, Loss of Heterozygosity, and Genotypes from Formalin-Fixed, Paraffin-Embedded Tumor Tissue Using Microarrays Free

Formalin-fixed, paraffin-embedded (FFPE) material tends to yield degraded DNA and is thus suboptimal for use in many downstream applications. We describe an integrated analysis of genotype, loss of heterozygosity (LOH), and copy number for DNA derived from FFPE tissues using oligonucleotide microarrays containing over 500K single nucleotide polymorphisms. A prequalifying PCR test predicted the performance of FFPE DNA on the microarrays better than age of FFPE sample. Although genotyping efficiency and reliability were reduced for FFPE DNA when compared with fresh samples, closer examination revealed methods to improve performance at the expense of variable reduction in resolution. Important steps were also identified that enable equivalent copy number and LOH profiles from paired FFPE and fresh frozen tumor samples. In conclusion, we have shown that the Mapping 500K arrays can be used with FFPE-derived samples to produce genotype, copy number, and LOH predictions, and we provide guidelines and suggestions for application of these samples to this integrated system. [Cancer Res 2007;67(6):2544–51]

The challenges associated with DNA derived from formalin-fixed, paraffin-embedded (FFPE) samples have prevented widespread application of FFPE DNA to many of the technologies available for high-quality DNA, although some options with lower genomic coverage are available (1–3). In this study, we show the feasibility and limitations of a genome-wide assessment of genotype, loss of heterozygosity (LOH), and copy number using FFPE DNA on the Affymetrix Mapping 500K array set, which includes the Mapping 250K Nsp Array and the Mapping 250K Sty Array (Santa Clara, CA). These arrays use a process termed whole-genome sampling analysis (WGSA; ref. 4), in which genomic DNA is digested and ligated to adaptors. A subset of digested fragments are then PCR amplified in a complexity reduction step before hybridization to the arrays. PCR proved to be the critical step when processing FFPE samples.

We compared several extraction methods to determine which protocol provides FFPE DNA most suitable for array analysis and found that a PCR-based assessment of DNA quality predicted the downstream performance of FFPE DNA samples better than age of FFPE sample. We identified a necessity for (a) in silico compensation against fragment size bias and (b) a fragment size filter during analysis of FFPE samples. We tested our new guidelines for FFPE DNA qualification and analysis on archival samples of various tissue types, storage times, and location sources. Quality of FFPE DNA varied but the methods outlined by this study enabled prediction of performance. These results show that FFPE DNA can be suitable for a combined study of genotype, LOH, and copy number on a whole-genome scale.

Sample selection and DNA extraction. Five primary endometrioid ovarian cancers were selected without screening for the initial portion of this study. For each sample set, normal lymphocytic DNA, fresh tumor tissue, and FFPE tissue were analyzed. Samples were collected between 1993 and 1999 as part of a larger study of ovarian cancer in women living in and around Southampton, United Kingdom (5). At the time of collection, DNA was extracted from blood samples and fresh tumor biopsies were snap frozen in liquid nitrogen. A portion of each frozen tumor biopsy was sectioned to assess the proportion of tumor. For samples 526T and 594T, microdissection was done (6) to obtain DNA with a >80% tumor component. DNA was extracted from the fresh frozen tissue using a salt chloroform method (7).

In 2002, a portion of each frozen tumor biopsy was formalin fixed and paraffin embedded as described previously (8), with all tumors fixed in 10% neutral buffered formalin for <24 h at room temperature. At the time of DNA extraction, the FFPE tumors had been embedded in paraffin blocks for 3 years. Five sections (10 μm) were deparaffinized twice in xylene (5 min) and rehydrated in 100%, 90%, and 70% ethanol (1 min each). The sections were stained with hematoxylin (4 min) and washed with water (1 min), acid alcohol (10 s), water (1 min), Scott's tap water (1 min), and water (1 min). The sections were then stained with eosin (3 min), rinsed with water (10 s), and dehydrated in 70%, 90%, and 100% ethanol (30 s each). Tumor cells were manually microdissected under a dissecting microscope as described previously (6) to obtain high-purity (>80%) tumor DNA. The tumor component for sample 594 was high enough that it was not stained or microdissected. DNA was extracted from the five endometrioid FFPE tissues using a modified Qiagen protocol (Valencia, CA; described below). Following DNA extraction from FFPE tissue, a salt precipitation DNA cleanup was done as described in the Affymetrix GeneChip Mapping Assay Manuals.

For the study of independent sample sets, DNA was extracted from FFPE tissue from 17 breast tumors and 8 colorectal tumors. FFPE blocks were collected from 11 pathology laboratories and ranged in age from 1 to 17 years. The formalin fixation and paraffin embedding protocols used for these tissues are not known but are likely to be quite varied. For breast tumors, 10 μm sections were deparaffinized, stained with H&E, and manually microdissected (described above). The colorectal tumors were not stained or microdissected due to their high tumor component. DNA was extracted from breast and colorectal tissues (described below), and as before, a salt precipitation DNA cleanup was done.

The collection and use of tissues for this study were approved by the appropriate institutional ethics committees.

Trial of DNA extraction methods for FFPE tissue. Five DNA extraction methods were trialed using whole 20 μm sections from three FFPE blocks. The methods that were compared were the MagneSil Genomic Fixed Tissue System (Promega,⁶

Madison, WI), ChargeSwitch gDNA Micro Tissue kit (Invitrogen,⁷ Carlsbad, CA), PureGene (Gentra Systems,⁸ Minneapolis, MN), DNeasy Tissue kit (Qiagen⁹), and a phenol/chloroform extraction. With the exception of the DNeasy Tissue kit and phenol/chloroform, the extractions were done according to the manufacturer's instructions. The extractions done with the DNeasy Tissue kit and with phenol/chloroform both were modified to include an initial incubation at 95°C for 15 min followed by 5 min at room temperature as described previously (9), before being digested with proteinase K for 3 days at 56°C in a rotating oven with periodic mixing and fresh enzyme added each 24 h. A salt precipitation was done on DNA from all five extraction methods.

DNA quality assessment and preparation. The extracted DNA was quantified using UV spectroscopy at 260 nm. Random amplified polymorphic DNA-PCR (RAPD-PCR; ref. 10) was done to assess the quality of DNA and maximum fragment lengths as described previously using 50, 5, or 0.5 ng DNA (11). Qiagen HotStarTaq was used, with 0.4 units per reactions (Qiagen⁹). Products were visualized with ethidium bromide on a 3% gel.

Preparation and application of DNA to the mapping arrays. Matched fresh and FFPE samples were analyzed on the Affymetrix GeneChip Human Mapping 10K v2 Xba Array and 50K Xba Array and prepared using the Mapping 10K v2 Assay kit and the Mapping 100K Assay kit (Affymetrix)¹⁰

The only exception to the manufacturer's protocol was that 10 cycles were added to the PCR cycling conditions for each FFPE sample.

Matched fresh tumor, FFPE tumor, and normal samples were assayed using the Mapping 250K Nsp Assay kit and the Mapping 250K Sty Assay kit¹⁰ and hybridized to the 250K arrays. The 500K assay was done according to the manufacturer's protocol, beginning with 250 ng DNA. Ninety micrograms of PCR product were fragmented and labeled, using additional PCRs when necessary for FFPE breast and colorectal samples.

Data analysis. Genotype calls were produced using the dynamic model algorithm (12) by the Affymetrix GeneChip Genotyping Analysis Software version 4.0. A stringent P value cutoff threshold of 0.26 was used. Concordance was determined by calculating the number of single nucleotide polymorphisms (SNP) that gave the same call in both fresh frozen and FFPE DNA from the same tumor and dividing this number by the total number of SNPs that were called in both samples.

LOH predictions were produced using dChipSNP software (dChip2005_f4 version¹¹

; ref. 13). LOH values were inferred using the Hidden Markov Model and restricting to SNPs on fragment sizes ≤700 bp.

Copy number estimates for ovarian tumor samples using 500K data were determined by pairing tumor and matching normal samples in CNAG_v2.0.¹²

Nonpaired, nonmatching references were used for copy number prediction of 10K and 50K data. Log 2 ratios were imported into Spotfire DecisionSite (Spotfire,¹³ Somerville, MA) and the Affymetrix Integrated Genome Browser for visualization and comparison. Copy number estimates for breast and colon FFPE tumors were done using data from 48 HapMap samples (available online¹⁰) as a reference.

Estimated inter-SNP mean and median distances after exclusion of fragment sizes >700 bp were determined by first calculating the distance between all SNPs on each chromosome. Distances were then sorted per chromosome in descending order and the largest distances (representing centromeres) were removed for each chromosome, except for the acrocentric chromosomes 13 to 15 and 21 to 22.

Pearson (linear) correlations were calculated in Partek Genomics Suite (Partek,¹⁴

St. Louis, MO).

Microsatellite analysis. Nine microsatellite markers were used to assess LOH at three loci: chromosome 1q (D1S2816, D1S413, and D1S1726), chromosome 7p (D7S691, D7S670, and D7S2506), and chromosome 14q (D14S1011, D14S258, and D14S1002). Regions were selected where array-based LOH analysis showed discordant LOH results for fresh and FFPE-derived DNA. The forward primer was labeled with a 5′-fluorescent dye (FAM or HEX). The samples were analyzed using a 3130 Genetic Analyzer (Applied Biosystems,¹⁵

Foster City, CA) with POP7 polymer. An assessment of LOH was done using GeneMapper version 3.7. LOH was scored by calculation of the ratio of tumor DNA peaks (T1/T2) compared with that in the normal DNA to give a relative ratio (T1/T2)/(N1/N2). A ratio of 0 indicates complete allele loss and a ratio of 1 indicates no LOH. A ratio of <0.5 was scored as indicative of LOH.

DNA extraction from FFPE tissue. Five DNA extraction methods (phenol/chloroform, Qiagen DNeasy Tissue kit, Invitrogen ChargeSwitch, Promega MagneSil, and Gentra PureGene) were tested on consecutive sections from different FFPE ovarian tumor biopsies. Phenol/chloroform and modified Qiagen protocols (see Materials and Methods) provided the highest DNA yield as determined by UV spectroscopy; these yields were 2.2 times more than the average yield from any of the other three extraction protocols (Fig. 1A). RAPD-PCR, which uses primers of 10 bps to produce a ladder of amplicons, was also done to assess both amplification efficiency and maximum product size for each extraction protocol (11). Compared with DNA extracted from fresh lymphocytes, the FFPE-derived DNA from all extraction methods yielded consistently smaller PCR fragments, with a maximum reliable size of ∼800 bp (Fig. 1A). Phenol/chloroform and modified Qiagen extractions produced more intense and consistent PCR fragments across dilutions, suggesting that products were relatively free of contaminant inhibitors (Fig. 1A). DNA extracted with these two methods was processed through the PCR step of the Mapping 50K Xba Assay to further assess amplification efficiency. In this test, the modified Qiagen extraction provided a slightly higher PCR yield on average than the phenol/chloroform method (21.4 μg compared with 19.2 μg) and was therefore chosen for DNA extraction from FFPE tissues in this study.

Mapping 500K array performance. Five matched sets; each containing (a) nontumor, non-FFPE lymphocytic DNA, (b) fresh frozen ovarian tumor DNA, and (c) FFPE ovarian tumor DNA; were assessed for performance on the Mapping 500K arrays. All five FFPE samples had been stored for 3 years and provided average RAPD-PCR maximum amplicon sizes from 526 to 800 bp. During the PCR step of the Mapping assay, amplification products from all five FFPE tumors were concentrated <700 bp, a fragment size range that was reduced compared with non-FFPE samples (Fig. 1B). Decreased yield from the Mapping PCRs (Table 1) accompanied the decrease in amplicon size distributions. FFPE samples produced 63 to 83 μg PCR products for the Mapping 250K Nsp Array, whereas all non-FFPE samples produced >90 μg.

The assay was continued using 90 μg PCR product as the manual instructs or the total PCR yield when this was less than assay requirements. Importantly, the protocol was otherwise never modified. Normal and fresh tumor samples gave typical SNP call rates, with an average of 94.5% and 93.5%, respectively. These call rates are lowered due to application of a strict confidence score threshold (P ≤ 0.26; the default threshold is P ≤ 0.33). In contrast, FFPE samples achieved an overall average call rate of 79.84% and 75.17% for Nsp and Sty, respectively (Table 1). These decreased call rates are consistent with the poor amplification of larger fragments during PCR. Exclusion of SNPs on larger fragments significantly increased the call rates, such that incrementally more stringent fragment size restrictions incrementally increased call rates (Fig. 1C). In fact, stringent fragment size restrictions produced similar call rates between fresh frozen and FFPE samples, indicating that the Mapping 500K is well suited for FFPE DNA and identifying the limiting factor as the size of amplicons produced from the degraded DNA.

Concordance of genotype calls between paired FFPE and fresh frozen ovarian tumor DNA samples was examined to determine the reliability of genotypes from FFPE DNA. It is important to note that tumor heterogeneity lead to confirmed genuine differences in genomic content between matched FFPE and fresh frozen DNA, which would lower these concordance rates. Average overall concordance between FFPE and fresh frozen samples from the same tumor was 93.6%. Exclusion of the larger fragments increased concordance such that all SNPs located on fragment sizes ≤700 bp displayed an average of 97.4% concordance (Fig. 1D). Exclusion of several regions (chromosomes 1q, 7p, 15, and 16q) displaying heterogeneity between fresh frozen and paraffin sample 95 increased the concordance by >2% (Fig. 1D). These high rates of concordance, despite shown genetic differences between paired samples, underscore the reliability and reproducibility of genotype calls produced using FFPE-derived DNA samples with this platform. Importantly, it indicates the need to exclude SNPs on larger fragments for reliable genotype data. Because SNP fragment size is distributed randomly across the genome, the general effect of excluding larger fragment sizes is to reduce the overall resolution without preferentially losing extensive coverage in specific regions (see Supplementary Fig. S1). The effect of fragment size on concordance was specific to FFPE samples and is not observed in comparisons between frozen samples (data not shown).

LOH and copy number assessment. The reliability of genotype assignments using paraffin samples suggests their suitability for LOH predictions. In fact, FFPE and fresh tumor pairs produced similar LOH profiles when including SNPs on fragments sizes ≤700 bp (Fig. 2A). Regions of inconsistent LOH predictions between paired samples (for example, see Fig. 2A , boxes) were predicted independently by both Nsp and Sty arrays and appeared along concentrated regions, rather than being sporadically distributed across the genome, suggesting that they reflected true biological differences between the samples. We assessed several discordant regions of LOH using conventional microsatellite marker analysis and in all cases, the microsatellite analysis confirmed that the array predictions were genuine (data not shown).

The ability to associate copy number estimates with SNP genotypes relies on quantitation of SNP probe intensities (14). Because larger fragment SNPs were inadequately amplified during WGSA, these SNPs were noninformative for copy number analysis of FFPE samples (Supplementary Fig. S2A). Exclusion of these large fragment SNPs significantly increased the amplitude (signal) of copy number shifts and at the same time reduced the SD (noise) associated with the copy number estimates for all FFPE samples but not the fresh frozen samples (Supplementary Fig. S2B). This increase in signal to noise ratio justifies the use of such a filter, which maintained 308,788 SNPs for FFPE copy number analysis (Table 2). Probe intensities from the remaining smaller fragment SNPs predicted copy number profiles for FFPE samples consistent with those from matching fresh frozen material (Fig. 2B). Equivalent copy number changes were predicted between FFPE and fresh frozen pairs both across different chromosomes and different sample sets (Fig. 2C).

In addition to limiting fragment size, compensation against fragment size bias was necessary to produce reliable copy number predictions. Although bias due to amplicon size can be negligible when using high-quality DNA, it becomes exaggerated when the DNA sample is degraded (Fig. 3, top). For FFPE samples, the mean copy number was grossly affected by the size of the amplicon carrying the SNP, such that smaller amplicons SNPs predicted gains and larger amplicons SNPs predicted losses in copy number. Quadratic regression helped to neutralize this fluctuation in mean copy number (Fig. 3, middle). Exclusion of SNPs on amplicons >700 bp before regression effectively removed the fragment size bias from copy number detection (Fig. 3, bottom). Copy number analysis of FFPE samples was done using the freely available CNAG_v2.0 software¹² (15), which automatically uses compensation against fragment size bias and includes an option to exclude SNPs based on fragment size. Alternate software tools that lack this compensation produced copy number estimates from FFPE samples that were noisier even with exclusion of large fragment sizes (data not shown).

Comparison of Mapping 10K, 100K, and 500K array performance. Although the various Mapping arrays all use the same technology and similar assays for genotype and copy number analysis, they each have differences that may influence their compatibility with FFPE samples. Particularly, the Mapping 500K and 10K arrays share the same amplicon distribution during the PCR step of WGSA, but the Mapping 100K assay relies on a wider amplicon size distribution (250–2,000 bp). Consequently, Mapping 100K data are more significantly affected by DNA degradation; for example, there are only 59 SNPs on fragment sizes <500 bp on the Mapping 50K Hind array. Previously, we showed the application of FFPE DNA to the 10K arrays (3) although without the analytic tools applied here. Now, we compared performance of FFPE samples on all Mapping arrays. As expected, call rates and concordances were poor when FFPE DNA was applied to the Mapping 100K assay, whereas performance was similar for the Mapping 500K and 10K arrays (Table 1; Supplementary Fig. S3). Furthermore, both the Mapping 500K and the 10K arrays, but not the Mapping 100K arrays, provided correct copy number predictions from FFPE DNA, whereas the Mapping 500K arrays best accommodated SNP filters to retain high genomic resolution (Supplementary Fig. S3).

Prediction of mapping array performance for a range of FFPE samples. DNA from FFPE samples can vary in quality as a result of the fixation protocol, years of storage, the extraction protocol, tissue source, and several other uncontrollable and controllable variables. To both identify a method for qualifying FFPE DNA samples for array analysis and test our guidelines for FFPE DNA extraction and data analysis, we measured the performance of an additional 25 FFPE tissue sources processed at separate institutes and stored for 1 to 17 years (Supplementary Table S1). These samples were not prescreened nor selected based on expected performance. Experiments were done without matched fresh frozen or nontumor samples. In a small test set, we found that application of 90 μg PCR product from FFPE samples increased call rates by several percentage points (data not shown); therefore, we assayed these samples using 90 μg whenever possible, even if this required pooling extra PCRs.

For each sample, we noted the largest amplicon size produced during RAPD-PCR as well as the size range of PCR products during the Mapping assay. Call rates were calculated for SNPs on fragment size ≤200 bp, 250 bp, 300 bp, and so on to determine the size at which call rates dropped <90%. This call rate drop-off value was used to indicate genotyping efficiency and reliability because fragment sizes with high call rates provided high concordance as well. Call rate drop-off values ranged from 250 to 750 bp compared with 700 to 850 bp for the five FFPE ovarian tumors. Therefore, most of these samples would provide reduced resolution for genotype and LOH. Copy number detection was more robust than genotype, and those cutoffs ranged from 300 bp up to no filter requirement at all. Plots of copy number versus fragment size were evaluated to determine the optimal fragment size filter for copy number analysis. These plots can be viewed in CNAG_v2.0, and various fragment size filters can be applied until the mean copy number for the SNPs retained in analysis are consistent across fragment size (Fig. 4C,, left). An example of this entire workflow is shown in Fig. 4A, to C and results are listed in Supplementary Table S1. As shown for a 733-kb hemizygous loss highlighted in this example, the fragment size filter suggested by this process was able to increase the signal to noise ratio by preferentially removing the noisy SNPs instead of the informative SNPs and at the same time was also able to retain higher resolution by not overfiltering (Fig. 4C , right).

Years of storage and overall call rates displayed some correlation to copy number and call rate drop-off values, but PCR-based analyses had higher predictive power for these performance metrics (Fig. 4D). The Pearson's correlation of median RAPD-PCR values to copy number drop-off was 0.93, indicating high predictive power. Comparison of array performance to PCR-based DNA quality tests gave R² values above 0.8. In contrast, R² values were <0.7 when comparing performance with years of storage or comparing copy number drop-off with overall call rate. These results indicate that a PCR-based test of DNA quality is a reasonable method for predicting whether a FFPE DNA sample will be amenable to array analysis.

Six of the 25 samples (two breast and four colorectal) were not applied to the arrays because no RAPD-PCR products were produced. Sample 0588 also failed RAPD-PCR, but it was still applied to the array. Consistent with the RAPD-PCR prediction, this sample was the only example, in which call rates broken up by fragment size never exceeded 90%, and data from even the smallest fragment SNPs were too noisy for copy number analysis.

There exists a large and growing deposit of archived clinical tissues, yet DNA extracted from these samples is usually degraded, contaminated, and of general low quality. This study expands the usefulness of the Mapping 500K arrays to DNA derived from FFPE samples, showing that the limiting factor for FFPE application is the size distribution of PCR amplicons during WGSA. The maximum amplifiable fragment size, which is correlated to array performance, varied between samples and may be influenced by both extent of DNA degradation and modification as well as the amount of inhibitors remaining in the sample. Use of a suitable DNA extraction protocol, such as the DNeasy Tissue kit, is important for obtaining DNA amenable to the assay, but other factors, such as years of storage and fixation process, will be harder to control. This underscores the necessity for a pre-WGSA quality control step that includes PCR of larger fragment sizes, such as RAPD-PCR or multiplex PCR (16). This study attempts to outline guidelines for qualifying FFPE DNA samples and analyzing qualified samples, but not all FFPE blocks will yield DNA suitable for the Mapping arrays.

FFPE DNA that is applied to the arrays may still vary in quality and therefore require more or less stringent fragment size filters. Despite reduction in coverage to accommodate loss of larger fragments, high resolution for genotype, LOH, and copy number assessment can still be maintained (Table 2; Supplementary Fig. S1). This is true because of the large number of SNPs on small fragments and because fragment size seems to be the only limiting factor. For example, with exclusion of SNPs on amplicons >700 bp, as was required for the first set of five FFPE samples, 308,788 SNPs were retained for analysis, providing a median and mean inter-SNP distance of 4.3 or 9.5 kb, respectively. Although the 10K array is also suitable for analysis of degraded DNA (3), the large SNP coverage and the small fragment emphasis of the Mapping 500K arrays make it ideal for FFPE sample analysis.

The percentage of FFPE samples archived in banks that could be applied to the arrays with limited loss in genomic resolution would be influenced by the methods of fixation and extraction used at various institutes. Importantly, all samples stored for 6 years or fewer provided copy number data for a minimum of 234K SNPs in this study. Some of the samples applied to the arrays required extremely stringent filters against fragment size, resulting in significantly decreased resolution of genomic data. Potentially, researchers may choose only to analyze DNA samples of such low quality when the FFPE sample is considered to be particularly precious. Importantly, RAPD-PCR results predicted that these samples would display decreased performance on the array and a PCR screen could be applied to avoid application of poorly doing samples. With the advent of more standardized protocols for sample processing in the future and with advances in DNA extraction, a higher proportion of FFPE samples may be applicable to the arrays.

Despite the large banks of FFPE samples available for retrospective studies that include follow-up analysis of patient outcome, most of these studies currently focus on frozen samples because of the limited options available for paraffin samples. Additionally, FFPE processing holds advantages for tissue storage during prospective studies, in which many biopsies are collected but only a fraction of them are applied to downstream assays with selection based on clinical outcome. These results outline guidelines for the application of FFPE samples to the same genome-wide platform already available to high-quality DNA samples, thus enabling widespread retrospective and prospective analysis of tumor samples in their most common form of storage.

Grant support: National Breast Cancer Foundation postgraduate research scholarship (E.R. Thompson).

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.

We would like to thank Giulia Kennedy, Manqiu Cao, Yaron Turpaz, and Guoliang Xing for technical input and discussions, Michael Shapero for his helpful suggestions and critical reading of the manuscript, and Dr. Alex Dobrovic for his help with DNA extraction.