Performance of Amplified DNA in an Illumina GoldenGate BeadArray Assay

Advances in molecular biology techniques, in particular, high-throughput genomic analyses, have heralded a new beginning in studies of complex human disease. Gene association studies, aimed at elucidating the underlying genomic basis of a given phenotype, are enabled by technologies whereby thousands of variants may be studied in thousands of samples in a time- and cost-effective manner. These assays require high-quality DNA in sufficient amounts, and acquisition of such has proven to be problematic. Several whole genome amplification (WGA) methodologies have been developed, which allow the generation of microgram quantities of DNA from nanogram starting quantities.

One such method is multiple displacement amplification (MDA), which is an isothermal amplification using the highly processive bacteriophage Φ29 DNA polymerase (1). Φ29 DNA polymerase produces very long products through a hyperbranched strand displacement amplification. MDA WGA has been evaluated for fidelity in genotypes by a number of methods, including short tandem repeat polymorphisms (2-6), fluorescence correlation spectroscopy (7), TaqMan (4, 6, 8, 9), Illumina linkage panels (10, 11), Affymetrix linkage panels (12), Affymetrix genome wide association GeneChips (13, 14), and Illumina GoldenGate highly multiplexed assays (15-18). WGA DNA was also used in two genome-wide association studies on the Perlegen array (19, 20), but neither had included replicated genomic DNA samples. In a publication from the Breast Cancer Association Consortium (19), 5.4% of MDA WGA DNAs failed to reach a 90% call rate on the Perlegen array and 38,836 (1%) of single nucleotide polymorphisms (SNP) had call rates <95%.

Studies of the fidelity of MDA WGA DNA have most often involved a limited set of DNA samples, typically <20 samples (2, 4, 10, 12-14, 21, 22) or SNPs (7, 8). Few have compared samples and SNPs in significant numbers to provide a broad view of the performance of WGA DNA in the high-throughput genotyping technologies available today (19). The International Multiple Sclerosis Genetics Consortium included 456 MDA WGA DNAs in a linkage study of 3,417 samples using the Illumina BeadArray linkage mapping panel (GoldenGate; ref. 23); however, only 21% of the MDA WGA samples were successfully genotyped (23). Pask et al. (15) genotyped 86 paired genomic MDA WGA DNAs using an Illumina 384-plex BeadArray, with a failure rate of 10% for WGA versus 6% for the genomic DNA, with 98.8% concordance. Berthier-Schaad et al. (16) found 99.66% concordance between 48 paired WGA and genomic DNAs in a 1,536-plex BeadArray, with 88.95% of WGA samples yielding genotype data; however, SNP call rates were not specified.

In this report, we compare genotyping outcome using replicated MDA WGA and genomic samples as part of an ovarian cancer candidate gene association study of 1,536 SNPs and 2,368 samples using the Illumina GoldenGate assay (24). We assess consistency and quality of WGA:genomic DNA paired samples, genomic replicates, and replicates with independent and common WGA preparations. Nested within our endeavor to understand ovarian cancer genetics, we hope to inform on the appropriateness of MDA WGA preparations for use in future high-throughput studies.

DNA from 2,051 women participating in ongoing ovarian cancer case-control studies at the Mayo Clinic (n = 872) and Duke University (n = 1,179) was extracted from fresh peripheral blood using the Gentra AutoPure LS Puregene salting out methodology (Gentra, Inc.). For samples with limited DNA available (Duke University samples), MDA WGA was done using the REPLI-G protocol (Qiagen) and 200 ng genomic DNA as input; no additional “clean-up” protocols were used after amplification. The yield of amplified DNA ranged between 45 and 70 μg, with most (60%) yielding ∼50 μg. DNA concentrations were adjusted to 50 ng/μL, verified using PicoGreen dsDNA Quantitation kit (Molecular Probes, Inc.). The REPLI-G WGA DNA protocol produced long products, comparable with high molecular weight genomic DNA (Fig. 1). The study protocol was approved by the Institutional Review Boards of Mayo Clinic and Duke University, and study participants provided written informed consent.

A total of 2,368 samples were assayed on 96-well plates designed to assure quality of laboratory methods, facilitate laboratory workflow, and maximize statistical power to detect associations with ovarian cancer.

WGA DNA samples (n = 1,282) consisted of unique Duke University study participant DNA samples (n = 1,179), duplicate samples prepared in independent WGA preparations from a random 1.3% of Duke University participants (n = 15), and duplicate genotyping aliquots of a common WGA preparation from a random 7.5% of Duke University participants (n = 88).

Genomic DNA samples (n = 1,086) consisted of 1,021 study participant samples and 65 laboratory standard DNAs. Study participant samples included unique Mayo Clinic study participant samples (n = 872), duplicate samples from a random 2.9% of Mayo Clinic participants (n = 25), and duplicate samples from a 10.5% subset of Duke University study participants with adequate genomic DNA quantities (n = 124). Additional laboratory standard DNAs included replicates sets of a CEPH trio (Coriell Institute; father: NA10858, n = 8 replicates; mother: NA10859, n = 9 replicates; and daughter: NA11875, n = 9 replicates), replicate sets of 3 internally used laboratory standard DNAs (n = 12, n = 13, and n = 14 replicates, respectively).

We designed an Illumina GoldenGate assay for 1,536 candidate ovarian cancer SNPs (tagSNPs and putatively functional SNPs). Ninety three percent of the SNPs had Illumina SNP scores of >0.6. Samples (250 ng) were genotyped following the Illumina protocol (21). Genotype calls were made using the Genotyping module of the BeadStudio 2 software. WGA and genomic DNA were analyzed separately, as WGA DNA was expected to cluster differently from genomic DNA. Clusters were reviewed using the replicates and heritability information of the CEPH trio and laboratory standard DNA samples to refine clustering.

Initial laboratory quality assurance (QA) relied on the GenCall score, a quality metric indicating the reliability of called genotypes that is generated by the BeadStudio software. The GenCall_10 refers to the 10th percentile GenCall score in a particular distribution of GenCall scores. For loci, it represents the 10th percentile rank for all GenCall scores for that locus. For initial QA, samples and SNPs with GenCall_10 scores below 0.40 and/or call rates below 90% were failed for both genomic and WGA DNA. In addition, four samples (two WGA and two genomic) were found to be mislabeled; these were considered to have failed initial QA and were excluded from further analyses.

Data were summarized descriptively using frequencies and percents for categorical variables, and means and SDs for continuous variables. Sample-specific call rates were calculated as percent successfully genotyped SNPs among DNA samples (1,282, WGA; 1,021, genomic), including duplicates but excluding laboratory controls. Similarly, per-SNP call rates were calculated as percent successfully genotyped samples. Summary statistics were calculated among all SNPs and samples passing initial QA metrics, as well as using the more strict SNP and sample call rate thresholds of >95% and >99%. SNP concordance across four replicate types (WGA:genomic, independent WGA, common WGA, and genomic) was calculated as the number of concordant calls across all replicate pairs and all SNPs divided by the total number of nonmissing comparisons. Thus, samples replicated more than once contribute n*(n−1)/2 replicate pairs to the data where n is number of replicates. For each of the four replicate types, we also determined the number and percent of pairs completely concordant across all SNPs; this enabled us to assess consistency of errors across samples (i.e., whether discordant results were clustered within poorly performing samples).

We did a series of comparisons to assess whether SNP characteristics predicted a variety of per-SNP outcomes. SNP characteristics included Illumina-provided SNP scores (ranging from 0-1.1), Illumina-provided designability rank (which categorized SNP scores as 0, SNP score <0.4; 0.5, SNP score between 0.4 and 0.6; 1.0, SNP score >0.6), Illumina-provided validation class (1, nonvalidated; 2, two-hit or HapMap validated; 3, GoldenGate validated), proximity to nearest telomere in Build 36.2 (bp), GC content within 100 bp 5′ and 100 bp 3′ of each SNP (%GC), Hardy-Weinberg equilibrium (HWE) P value (χ² test), and minor allele frequency (MAF) among 941 Caucasian controls (women without ovarian cancer). Per-SNP outcomes included failure to pass initial QA in WGA samples, failure to pass initial QA in genomic samples, call rate in WGA samples, call rate in genomic samples, and concordance of the four replicate types (WGA:genomic, independent WGA, common WGA, genomic) at >90% per-SNP call rate. We assessed pair-wise associations of SNP characteristics with per-SNP outcomes using Spearman correlations for pairs of continuous variables, χ² tests for pairs of categorical variables, and Wilcoxon-Kruskal-Wallis tests for pairs composed of one continuous and one categorical variable. In addition, we assessed the correlation among SNP characteristics, and for correlated characteristics, we ran a series of analyses of covariance to evaluate the independence of the influence of SNP characteristics on genotyping outcome. All statistical tests were two-sided, and all analyses were carried out using the SAS software system (SAS Institute, Inc.).

Table 1 describes overall genotyping performance on a per-SNP and per-sample level. We found that fewer MDA WGA samples were successfully typed compared with genomic samples: 1,197 of 1,282 WGA samples (93.4%) passed initial QA (GenCall_10 scores of ≥0.40 and call rates of ≥90%), whereas 1,014 (99.3%) of the 1,021 genomic samples passed (Table 1). We further observed that among samples successfully genotyped, fewer SNPs were called in the WGA samples compared with genomic samples: 1,432 of 1,536 SNPs (93.2%) passed initial QA (GenCall_10 scores of ≥0.4 and call rates of ≥90%) in the WGA DNAs (n = 1,197), whereas 1,480 SNPs (96.3%) passed QA in the genomic DNA (n = 1,014). All failed SNPs in genomic samples (n = 56; 3.6%) were intentionally uncalled in WGA samples; thus, the failed WGA SNPs (n = 104) include an additional 48 (3.1%), which passed initial QA in genomic samples. The reasons for not calling a particular SNP included poorly defined clusters, excessive replicate or Mendelian inheritance errors, excessive multiple clusters, observance of only a heterozygote cluster, and a high failure rate (>10%). These failed samples and SNPs were eliminated from all subsequent comparisons.

Among samples and SNPs passing initial QA, per-SNP and per-sample performance was quite high; mean call rates exceeded 99.5% for each. Although most SNPs and samples passing initial QA (both WGA and genomic) achieved per-SNP and per-sample call rates >95% (Table 1), differences between preparations were more apparent at higher thresholds. Among successful samples and SNPs, only 1,089 (91.0%) WGA samples achieved call rates >99%, whereas 956 (94.3%) genomic samples did. Similarly, only 1,324 (92.5%) SNPs were called in >99% of WGA samples, whereas 1,391 (94.0%) SNPs were called in >99% of genomic study participant samples. Therefore, our SNP failure rate of 7.6% using WGA lies within the lower end of the reported range of 4% and 18.5% using several genotyping platforms (3, 4, 11, and 14-16) when WGA DNA was used.

Several metrics are commonly used to predict success before genotyping and to eliminate poor-quality SNPs following genotyping. We assessed the relationship between these metrics and genotyping outcome defined by failure rates and call rates (Table 2). Illumina recommends selecting SNPs with a SNP score of >0.6 (designability rank, 1.0), as these have the highest probability of succeeding with GoldenGate chemistry, and 93% of the SNPs included in the project had scores within this range. Generally, we found that success of WGA DNA was predicted by these metrics. The Illumina-provided designability rank predicted failure in both WGA and genomic samples, whereas validation class did not correlate with failure. The Illumina-provided SNP score did not significantly differ in failed compared with typed SNPs in genomic samples, although lower SNP scores were observed in failed SNPs, and SNP score was a significant predictor of failure in the WGA samples. Call rates in both types of samples were also related to designability rank and SNP score, but validation was only relevant to WGA call rates (Table 2). This association with validation class and WGA call rate is difficult to interpret and suggests that two-hit or HapMap-validated SNPs have slightly lower call rates in WGA samples (99.52% versus 99.63% or 99.61%).

Additional biological features were also informative. The %GC content of the local sequence surrounding a SNP has been suggested to increase the likelihood of failure (16) or to increase likelihood of discordant genotype calls when MDA WGA DNA is used (3). Similarly, the distance to the telomere has been found to be a predictor for failure (16). We found that %GC was positively correlated with both WGA and genomic failure (P < 0.001), whereas distance to telomere predicted failure in WGA samples only. For WGA samples, the mean distance to telomere was 35.66 Mb for successful SNPs but 27.80 Mb for failed SNPs (P < 0.001). Among successful SNPs, call rates correlated with %GC and telomere distance for both WGA and genomic samples but seemed more strongly correlated in the WGA samples (Table 2). Figures 2 and 3 show SNP failure by categorized %GC and distance to telomere, respectively, and suggest that failure of WGA DNA genotyping is increased in regions with >50% GC content and within 50 Mb of a telomere. Because we observed that %GC and distance to telomere were themselves negatively correlated (r = −0.30; P < 0.001), additional analyses of covariance were done. We found that for both genomic and WGA, %GC remained significantly associated with SNP failure after adjustment for distance from telomere (P < 0.001), but that for distance from telomere, adjustment for %GC drastically attenuated the results (P value for association went from <0.001 to 0.58 for WGA DNA). These results suggest that distance from telomere was not independently associated with SNP failure rates after accounting for %GC.

HWE and MAF are routinely used to screen SNPs postgenotyping. We found that HWE did not relate to failure or call rates, but that MAF was inversely correlated with call rate in WGA and genomic samples (Table 2). Within the narrow call rate range, more common SNPs yielded lower call rates. We expect this to be due to easier ability to separate genotyping clusters when clusters differ greatly by size as occurs when low MAF SNPs are genotyped. With tighter clusters, samples that fall between two clusters are more likely to be uncalled (Fig. 4).

Consistency of genotype calls among replicate samples is critical for WGA evaluation. Generally, we found high genotype concordance (>99%) and, as expected, greater concordance with better-performing SNPs (higher call rates), as shown in Table 3. Concordance rates were virtually identical for SNPs with >90% call rate (data not shown) and SNPs with >95% call rate, although concordance improved for SNPs with >99% call rate (Table 3). Of 116 WGA:genomic pairs passing initial QA, there was excellent concordance overall. However, only one of the WGA:genomic sample pairs (0.9%) was completely concordant, whereas for the other replicate types (independent WGA, common WGA, and genomic), over 84% of pairs were completely concordant. This finding suggests that genotype error was consistent across WGA samples, and this only modestly improved with higher call rate thresholds. We observed fewer completely concordant pairs among samples amplified and genotyped twice (independent WGA replicates, 84.6%) than with samples amplified once but genotyped twice (common WGA replicates, 92.0%) or with genomic DNA replicates (95.8%), suggesting the introduction of errors in the amplification process. This implies that inclusion of replicates of WGA in a genotyping project may not provide as accurate estimates of concordance as including genomic replicates; although consistent genotypes may be seen across WGA replicates, they may not reflect the true genotype.

When we examined the nature of the discordance between WGA and genomic DNA samples, we found that 77.4% of the discordant calls at the >95% SNP call rate threshold and 81.6% of those at the >99% threshold were a reduction to homozygosity from a heterozygote state in the genomic samples. These likely reflect a bias in the WGA process and this has been reported by others using a variety of genotyping platforms (2, 4, 12, 25).

We assessed whether a variety of SNP metrics correlated with genotype concordance across the four replicate pairs using a 90% call rate threshold (Table 4). We found that the Illumina-provided designability rank and SNP score were highly correlated with concordance of WGA:genomic replicates (P < 0.001) but did not predict concordance rates among the WGA replicates and genomic replicates. This suggests that, when using WGA DNA, a higher SNP score threshold should be used to ensure genotyping accuracy. Surprisingly, we observed a correlation between HWE P value and concordance of genomic replicates only (P = 0.01); SNPs that deviated from HWE yielded lower concordance (Table 4).

Use of MDA WGA DNA in genotyping has received increasing attention in the past few years, promising to alleviate the barrier of limited participant DNA for conducting molecular epidemiologic studies in the era of ultra-high throughput genotyping. With few exceptions, reports of WGA DNA performance have been limited in scope in terms of number of DNA samples or number of variants tested. For two reports that used large numbers of samples in high-throughput genotyping applications, the results are not clear (15, 23), as the success of the WGA differed greatly. This report seeks to contribute to the growing interest in use of WGA, by assessing how well such DNAs did in a custom high-throughput genotyping assay. As errors are more likely to occur in samples with poor performance (that is, when call rates of SNPs decrease below a given threshold, the likelihood of genotype call errors increase for that sample), we selected two SNP call rate thresholds for samples to assess this performance. One (restricting to SNPs with >95% call rate) is commonly applied, whereas the other (>99%) is more stringent.

This study provides broad-based corroboration of the use of WGA DNA in high-throughput genotyping based on findings from a large number of paired WGA:genomic DNAs as well as the large number of WGA samples included in the study. Examining the performance and concordance using several call rate thresholds also provides guidance. Although the number of SNPs failing to achieve a >95% call rate is greater when WGA DNA is used, successful genotyping for 93.4% of samples and 93.2% of SNPs attempted was achieved. In this report, <1% overall discordance was achieved at 3 thresholds of SNP call rates used. Other studies documenting the frequency of mismatch between genomic and WGA have reported 0.05% to 2.9% discordance from 4 to 83 DNAs using Affymetrix 10K or Illumina GoldenGate genotyping arrays (10, 12, 15, 16). We used a relatively high amount of DNA as input into the WGA, and this may account for the difference in success between this and other studies using Illumina BeadArrays (16, 23) where the amount of DNA varied and was below 20 ng. Bergen et al. (3) reported improved genotyping performance as the amount of input DNA increases. In addition, the lack of correlation between HWE P values and concordance of the genomic:WGA DNA likely results not only from the quantity input, but also the use of high quality DNA for the WGA, corroborating the findings of Liang et al. (9). When DNA quality is poorer, deviation from HWE is more frequently observed.

We observed several SNP characteristics related to genotyping failure, call rate, and performance. The Illumina-provided designability rank and SNP score were critical, in particular, to failure of WGA and accuracy of genotypes in WGA DNA compared with genomic DNA. These measures were also predictive of call rates for both WGA and genomic samples. The Illumina-provided validation class was related to WGA call rate, such that two-hit or HapMap-validated SNPs had slightly lower success. A previous report using an Illumina GoldenGate platform showed that likelihood of failure was found to correlate with distance to telomere (16). Although we saw no relationship between concordance and telomere distance, we observed that SNPs farther from the telomere were more likely to succeed in WGA and to yield higher calls rates in both types of DNA. Lage et al. (26) reported that loci near the ends of yeast chromosomes were underrepresented in array-CGH in MDA-amplified DNA. We also found that SNPs with lower %GC were more likely to yield genotypes and that %GC was more predictive of failure than distance to telomere. Thus, when using WGA DNA, SNP selection should take these characteristics of SNPs into account, as Bertier-Schaad et al. discussed (16). Although MAF correlated with call rates, we recommend that concerns over statistical power take precedence over variation within the narrow call rate range we observed.

We used the REPLI-G MDA WGA method for this study; however, other methods are available, such as GenomePlex (Sigma Aldrich), which may more reliably amplify highly degraded samples. This method, previously known as OmniPlex, typically yields shorter products, which might affect the performance of some genotyping assays. Bergen et al. (3) found MDA WGA DNA to be superior to GenomePlex WGA DNA with a significantly greater proportion of double-stranded WGA DNA, a greater short tandem repeat polymorphism and SNP genotype completion rate, and reduced sensitivity to %GC content. Hansen et al. (27) reported overall good performance of OmniPlex WGA DNA in an Illumina GoldenGate assay but did not consider characteristics of the SNPs in assessing performance.

In conclusion, we report that use of MDA WGA DNA is feasible and informative and that genotyping performance and concordance are quite good, although not as optimal as genomic DNA. Regarding genotype error, we note that, in the case-control setting with case and control samples similarly handled, association testing would be affected only to the extent of any nondifferentially misclassified exposure with a possible bias to the null. As recommended by Illumina Corporation, we observed that it is critical to cluster WGA and genomic samples separately. Others have further reported that the combination of replicate WGA was able to increase both the performance and accuracy of genotyping (2). We here suggest additional considerations regarding SNP selection before genotyping and screening before analysis. In particular, we recommend that, in addition to Illumina-provided design metrics, additional biological SNP characteristics, particularly distance to telomere and %GC (<50% GC preferred), be considered in the creation of GoldenGate BeadArray assays using MDA WGA DNA.

No potential conflicts of interest were disclosed.

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 thank Stephanie Anderson for data management, David Rider for SNP annotation, and Laura Huennekens for manuscript preparation.