Discordant Haplotype Sequencing Identifies Functional Variants at the 2q33 Breast Cancer Risk Locus

The large number of compelling disease loci identified from disease-association studies (e.g., 1–4) provides immense potential for novel insight into disease mechanisms and new opportunities for diagnosis, prevention, and treatment. To date, 2,511 genome-wide association studies (GWAS) of 1,353 traits have generated 8,345 hits at P ≤ 5 × 10⁻⁸ (5). However, the task of identifying the underlying functional sequence variants (FSV) has proven extremely challenging, and, thus far, very few FSVs have been identified for these types of low-risk disease loci (e.g., 6–8). A number of factors contribute to the challenge, including cost and an inadequate functional knowledge base. Performing DNA sequencing (DNAseq) of the disease loci would capture the FSVs, but this is often financially impractical for large case–control designs (GWAS for common traits can include 100,000s of cases and controls). Furthermore, empirical evidence suggests that the vast majority of FSVs for these loci will be common, noncoding variants influencing gene expression (9). Hence, the identification of the FSV via simple prioritization of the sequence data may be difficult because many variants will be statistically comparable, and our understanding of noncoding regulatory elements is incomplete.

Instead of complete DNAseq of a GWAS region, a standard approach to identify cFSVs is a more detailed statistical interrogation of the region (“fine-mapping”). Often these fine-mapping studies are undertaken in the same large sample sets as the initial analyses, they explore evidence for multiple signals and use statistical imputation to estimate the association evidence for other common and rare variants in the region that were not directly genotyped (10). A caveat of statistical imputation is that it relies on external imputation panels that are not only distinct from the target populations, but are often much smaller and sequenced only at very low depth (e.g., 4×), leading to potential inaccuracies, especially for rare variants. DNAseq of cases and controls would remove the need for imputation, but remains cost prohibitive for large case–control studies. Here, we propose a cost-efficient extreme discordant haplotype DNAseq design that is financially viable and can greatly aid in the identification of cFSVs.

At a predefined risk locus, we statistically phase the genotype data and perform haplotype mining to define risk haplotype/s that optimally distinguish cases from controls. Phasing does not require external data. Under the hypothesis that FSV/s lie on a risk haplotype backbone, selecting cases homozygous for risk haplotype and controls without risk haplotypes results in an extreme discordant design with good power to detect underlying FSVs with small sample sizes conducive to quick and affordable high-throughput sequencing (HTS). To illustrate the gain in power of the discordant design, consider a FSV with population risk allele frequency (RAF) = 0.32 and an allelic OR = 1.09 (i.e., RAF = 0.34 in cases). Even if the associated SNP is in perfect linkage disequilibrium (LD) with the FSV (r² = 1.0), greater than 85,000 samples are required to identify it with 80% power at a significance of α = 5 × 10⁻⁸. Now consider a haplotype discordant design: cases homozygous for the risk haplotype and controls with zero copies. If the risk haplotype were a perfect proxy, the FSV RAF would be enriched to 1.0 in cases and reduced to 0.0 in controls. Although enrichment will not be perfect, it will be greatly enhanced; thus, if RAFs become 0.9 (cases) and 0.1 (controls), a sample size of only 28 has 80% power at α = 5 × 10⁻⁸ (see Supplementary Methods for more details).

As proof-of-concept, we present a discordant haplotype HTS DNAseq study in 38 individuals for the 2q33 (CASP8/ALS2CR12) region, a breast cancer risk locus (11). In order to address the lack of knowledge surrounding gene regulation and better prioritize our sequence variants (SV), we supplemented publicly available annotations with RNAseq in normal and tumor breast tissue, and identified three FSVs affecting enhancer activity.

An overview of study design detailing the sample sets used for the DNAseq and RNAseq analyses can be found in the Supplementary Data.

Haplotype mining was performed previously using hapConstructor (12) based on 45 tagging-SNPs across 220 kb spanning CFLAR, CASP10, CASP8, and ALS2CR12 in 3,888 individuals from the Utah and Sheffield Breast Cancer Studies (1,882 breast cancer cases and 1,896 controls). A six-SNP risk haplotype was defined (P < 5 × 10⁻⁶) that optimally extracted the association evidence across the region (13). Specifically, the alleles on the haplotype were ins-A-G-G-T-del across SNPs rs3834129-rs6723097-rs3817578-rs7571586-rs36043647-rs35010052. These six SNPs spanned CASP8 and the intergenic region between CASP8 and ALS2CR12. HapMC (14) was used to impute missing genotypes and infer phase on all 3,888 individuals. Female breast cancer cases were selected who were homozygous for the risk haplotype, and female cancer-free controls who had zero copies (with probability > 0.95). For these individuals, germ-line DNA (200 ng) from peripheral blood leukocytes was selected for the sequencing panel based on DNA quantity and quality requirements (3–5 µg of high quality, high molecular weight genomic DNA, free of RNA contamination, and at a minimum concentration of 100 ng/µL with A₂₆₀/A₂₈₀ and A₂₆₀/A₂₃₀ ratios ≥1.8 and ≥1.9, respectively; confirmed by Bioanalyzer). The final panel comprised 38 individuals (21 cases and 17 controls).

In order to capture any underlying FSVs tagged through long-range LD with the risk haplotype, we sequenced across a broader region. We thus defined the chromosome 2q33.1 low-risk breast cancer locus to be from 201,566,128 to 202,566,128 bp (hg19); the minimal region required to contain all variants in the HapMap or 1000G projects with a r² ≥ 0.1 with the previously reported variants associated with breast cancer (rs1045485 and rs10931936; refs. 15,16). The region is gene-rich with 18 genes: AOX2P, LOC100507140, BZW1, CLK1, PPIL3, NIF3L1, ORC2, FAM126B, NDUFB3, CFLAR, CASP10, CASP8, ALS2CR12, TRAK2, STRADB, ALS2CR11, TMEM237, and MPP4 (Fig. 1).

The SureSelect Target Enrichment system was used to generate libraries for next-generation sequencing based on all nonrepetitive genic and intergenic sequence in the 1 Mb region. Capture bait design was carried out using Agilent E-array and Windowsmasker (NCBI C++ Toolkit), resulting in 8,227 baits that mapped to 567,411 bp. Samples were multiplexed 12 per lane for ABI SOLiD sequencing to generate 50 bp single-end reads. Raw sequence reads were aligned using NovoalignCS (http://www.novocraft.com), developed specifically for color space SOLiD technology data. On average, 69.2% of reads accurately mapped to the 1 Mb region, and 609,165 bases had ≥10× coverage (mean, 76×). Alignment was followed by duplicate marking using Picard (http://broadinstitute.github.io/picard/) and GATK (17) best practice V2 for variant calling, including local realignment, recalibration, joint calling with UnifiedGenotyper, and variant filtration. A total of 1,197 positions were heterozygous in at least one individual in the sequencing panel, and passed quality control variant filtration with good variant quality score (Q > 30), 63 SVs were not acknowledged by the 1000G project and were removed from consideration, resulting in 1,134 SVs for comparisons between cases and controls.

In the homozygous discordant risk haplotype design, the FSV should reside on all case chromosomes and no control chromosomes. To prioritize variants based on this expectation, the allele frequency (AF) was calculated for both REF (the allele designated as the reference allele) and ALT (the alternate allele) alleles at each of the 1,134 SVs in both cases and controls. An SV was considered a cFSV if an allele had AF ≥ 0.9 in cases and AF ≤ 0.1 in controls. If so, that allele was considered the risk allele. Nine hundred and fifty-five SVs had an allele with AF ≥ 0.9 in cases, but only 18 of these also had AF ≤ 0.1 in controls. For these 18, the RAF varied from 0.905 to 1.0 in cases, and from 0.0 to 0.094 in controls. One variant aligned perfectly with expectations (rs3769821, RAFs of 1.0 and 0.0 in cases and controls, respectively). Chi-squared tests of allele counts by case/control group were calculated to help interpret these observed discrepant RAF frequencies. All were highly significant (6.9 × 10⁻¹³ ≤ P ≤ 2.8 × 10⁻¹⁸; Table 1). Background frequencies for the risk alleles for the CEU (Utah residents)/GBR (Great Britain) 1000G samples are shown in Table 1. 

A panel of fresh-frozen breast tissue samples was assembled from 88 women who had surgery at the Huntsman Cancer Hospital from 2009 to 2012. This included paired tumor and normal (adjacent grossly uninvolved) tissues for 69 breast cancer patients and normal tissues from an additional 19 women undergoing breast reduction surgery. RNA was extracted from all tissue samples. For germ-line genotypes, DNA extracted from peripheral blood (N = 15), or from the normal tissue (N = 73) was used. For genotyping, we required 10 µL of DNA at a concentration of 50 ng/µL measured by PicoGreen. An Illumina BeadExpress assay was designed for the 18 cFSVs. One variant failed design (rs10635401) and another failed genotyping QC (rs3769820), leaving 16 cFSVs with good quality genotype data for differential expression analyses (carriage of risk allele vs. no copies). For RNAseq, we required 50 μL of RNA at a concentration of 25 ng/μL, and RIN ≥ 6.0. Bioanalyzer results indicated that quality was generally high (average RIN = 7.9), although one tumor sample yielded poor quality RNA (RIN=2.5) and was removed from consideration, resulting in a panel of 156 breast samples (88 normal, 68 tumor). RNAseq was carried out using Illumina TruSeq Stranded mRNA sample preparation with oligo dT selection, and samples were multiplexed 8 per lane on the Illumina HiSeq2000 to generate 50 bp single-end reads. Raw sequence reads were aligned using Novoalign (V2.08.01) against the reference genome plus extended splice junctions (hg19 annotations for gene model). We used USeq, an extensive package of open-source RNAseq workflow written and benchmarked by the University of Utah Bioinformatics Shared Resource (18), to quality control, wrap programs, manage, and prepare files. Two samples failed RNAseq QC and were removed (degraded signatures by Picard) leaving 154 samples (86 normal, 68 tumor), with an average of 19.1 M reads per sample for analysis. We performed differential expression analyses based on carriage of risk alleles at cFSVs using DESeq2 (19), which employs a negative binomial distribution to test for differences. The RNAseq panel provided 80% power to find ratios ≥1.19 or ≤0.84 (equivalent to fold differences in expression of ∼20%), at a nominal significance level (α = 0.05; ref. 20).

We selected CFLAR, CASP10, CASP8, and ALS2CR12, which span the central ∼0.25 Mb of the 1 Mb region, as potential target genes for our cFSVs. The transcriptome-normalized counts, averaged across all samples, were 79.9, 27.1, 38.4, and <1.0 per million mapped reads, respectively. This indicated that CFLAR, CASP10, and CASP8 were expressed sufficiently well in the breast to pursue expression-quantitative trait locus (eQTL) analyses, but that ALS2CR12 was not well expressed and was not considered further. Consistent with our data, Encyclopedia of DNA Elements (ENCODE) RNAseq data also indicated that CFLAR, CASP10, and CASP8 were expressed in breast tissues. All were expressed in the human mammary epithelial (HMEC) normal breast cell line, although the level of CASP10 expression was lower than the other genes; CASP8 and CFLAR were also expressed in the breast cancer cell line MCF-7. These three genes are also good functional candidates: CASP8 and CASP10 are proapoptotic, whereas CFLAR (caspase-8 and FADD-like apoptosis regulator) is antiapototic, structurally similar to caspase-8 but lacking the proapoptotic caspase death domains. Because it has been suggested that more than 40% of enhancers do not target the nearest gene, we tested for association between each cFSV and the all three genes (21).

Genome-wide germ-line SNP genotyping and whole-transcriptome RNAseq data for 97 normal (adjacent grossly uninvolved) tissues and 753 tumor breast tissue samples are available as part of The Cancer Genome Atlas (TCGA) Research Network (http://cancergenome.nih.gov/). Genotypes were available for only three SNPs from our list of 18 cFSVs. Transcriptome-normalized counts (with correction for differences in transcriptome length and specific gene annotations from hg18 to hg19) indicated expression levels of 149.1, 16.0, and 89.9 per million mapped reads for CFLAR, CASP10, CASP8, respectively, and differential analyses were performed on all three genes in both normal and tumor tissues. We note that the average standardized counts for CASP10 fell below 20 counts per million mapped reads, and more caution should be exercised in interpreting findings for that gene in these data.

The R package metaphor (http://cran.r-project.org/web/packages/metafor) was used to carry out a meta-analysis for the cFSVs, based on the ORs and 95% confidence intervals provided for 46,450 cases and 42,600 controls (Breast Cancer Association Consortium, BCAC) and 22,627 cases and 10,052 controls (9 GWAS) in Lin and colleagues (11).

Variants rs3769823, rs3769821, and rs10197246 were chosen to investigate for ability to affect enhancer activity. To create the enhancer constructs, we performed PCR using genomic DNA from T-47D, which is homozygous for risk alleles at all three SNPs, and MCF10A, which is homozygous for neutral alleles, using the following primer pairs:

• rs3769821/rs3769823 forward (1,531 bp product), TACCTGAGCTCGCTAGCCGGATCAATGCTACAAAGACACG

• rs3769821/rs3769823 reverse, GGCCAGATCTTGATATCCCAGTCACCTCTGGAGGCATT

• rs10197246 forward (758 bp), TACCTGAGCTCGCTAGCCGCTGTTTAATTTCCATGCGTTT

• rs10197246 reverse, GGCCAGATCTTGATATCCTCTTTAGCAGTAGCACAACACAAA

Phusion HF master mix (NEB) and 10 ng genomic DNA were used for PCR, and PCR products were purified with AMPure XP beads (Beckman Coulter). We then digested pGL4.23 (Promega) with XhoI (NEB) and combined the digest with the purified PCR products to perform Gibson Assembly (NEB) according to the manufacturer's instructions. Clones were verified by Sanger sequencing, and no differences between the haplotypes, except for the expected alleles, were identified.

T-47D (ATCC HTB-133) and MCF10A (ATCC CRL-10317) were obtained from the ATCC, where short tandem repeat profiling analysis was used to authenticate the cell lines, and were grown within 6 months of resuscitation. Each line was cultured according to ATCC recommendations with the following modification: no insulin was added to the media for T-47D. Cells were plated at a density of 10,000 cells/well in 96-well plates, and after 24 hours, Lipofectamine 3000 (Life Technologies) was used to transfect the enhancer constructs as well as pGL4.23 as a negative control. Forty-eight hours after transfection, each well was assayed for luminescence using Steady-Glo Luciferase Assay System (Promega) and read on a GloMax luminometer (Promega). Luminescence was background subtracted using nontransfected wells and normalized by pGL4.23 levels. T tests were used to determine significance of expression differences between alleles.

The women whose samples were involved in the DNAseq and RNAseq experiments described here were enrolled in studies approved by the University of Utah Institutional Review Board or South Yorkshire Research Ethics Committee. Informed consent was obtained from all research participants.

Using HTS technology, we performed targeted DNAseq of all nonrepetitive sequence across 1 Mb at 2q33.1 (201.57–202.57 Mb, hg19) in 38 women selected for extreme discordance of a previously defined risk haplotype (21 breast cancer cases homozygous for the risk haplotype and 17 female, cancer-free controls with zero copies; ref. 13). We identified 1,134 high quality SVs that were also confirmed as SVs in the 1000 Genomes (1000G, http://www.1000genomes.org/) project data (Fig. 1).

In the optimal situation, the FSV would reside on all 42 case chromosomes and no control chromosomes. In selecting cFSVs, we allowed 10% discrepancy, such that an SV was considered a candidate if RAF ≥ 0.9 in cases and RAF ≤ 0.1 in controls. This resulted in 18 cFSVs (15 in CASP8 and 3 in the adjacent gene ALS2CR12): one was exonic (CASP8), 16 intronic, and one in the 3′ untranslated region of ALS2CR12 (Table 1). As expected, AF differences between the cases and controls were all highly significant due to the discordant design (all P < 10⁻¹²). Risk alleles were all very common in the general population (1000G RAFs = 0.27–0.40).

To assess the cFSV for potential functional consequences, we performed eQTL analysis using high quality RNAseq data for CASP8, CASP10, and CFLAR in 86 normal tissue and 68 tumor samples, examining expression differences based on germ-line DNA genotype data for 16 of 18 cFSVs (Table 1). The expression of CASP8 was significantly decreased for carriers of 12 of 16 cFSVs within normal breast tissue (Table 2), most significant for a group of five cFSVs at the 3′ ends of CASP8 and ALS2CR12, between positions 202,143,928 and 202,153,920 bp (expression ratio = 0.86, P = 3.6 × 10⁻⁴; Figs. 2 and 3). Reduction in expression with carriage of these cFSVs was consistent in the subsets of adjacent grossly uninvolved and breast reductions (data not shown). No significant differences based on carriage of risk alleles were found for CASP8 in tumor (although ratios were in the same direction as for the normal tissue, but less extreme). No differential expression was apparent based on cFSVs in CASP10 or CFLAR in normal or tumor tissues. We replicated eQTL results using data from TCGA Research Network (http://cancergenome.nih.gov/) project, where RNAseq data were available for 97 normal breast and 753 tumor tissue samples, and germ-line genotypes for 3 of 18 cFSVs. Similarly, decreased CASP8 expression was found for risk alleles at three cFSVs in normal tissue, with the most significant association for rs6743068 (ratio = 0.86, P = 2.6 × 10⁻⁴), which resides in the cluster of five most significant cFSVs. As before, there was no evidence for CASP8 expression differences based on cFSVs in tumor tissues, or for CFLAR or CASP10 expressions in either tissue type.

This small but highly focused haplotype discordant study demonstrates that cFSVs can be identified that are associated with gene expression. However, equally important is how this information aligns and adds value to results from a traditional “fine-mapping” study. To achieve this, we aligned our cFSVs with those suggested by the fine-mapping study by the BCAC that used 89,050 samples and 1,733 imputation fine-mapping SNPs across the same 1 Mb region (11). The BCAC study identified one region that achieved genome-wide significance (referred to as iCHAV1, P = 1.1 × 10⁻⁹; ref. 11). The BCAC iCHAV1 SNP set corresponds closely with the set of cFSVs identified here. Nine SVs were present in both sets, seven present only in the BCAC set, and nine present only in the discordant haplotype set (Table 3). The seven SVs present only in the BCAC results were in genomic regions 1000G designated as inaccessible to HTS technologies and were not captured by our sequencing baits. This highlights a limitation of HTS technology, but also suggests caution with imputation in this region, which is based on low coverage sequence data (6 of the 7 missed SNPs were imputed). We genotyped the remaining SV, rs10197246, which indicated it would have passed our cFSV criteria (Table 3). Of the nine cFSVs present only in our discordant haplotype study, several only narrowly missed BCAC inclusion. For example, our set included both rs6735656 and rs10635401, but only the latter was captured in the BCAC set due to slight superiority in significance in that data. Interestingly, in a combined BCAC + 9 GWAS meta-analysis, the order of significance is reversed for these two SNPs (Table 3). Hence, the discordant haplotype DNAseq study not only aligned with, but also added value to, the very large BCAC fine-mapping study.

Based on the combined 25 cFSVs, we used publicly available annotations and RNAseq eQTL results to prioritize for functional follow-up (Table 3). The three most compelling cFSVs are rs3769823, rs3769821, and rs10197246. All three yield highly significant results in both the discordant haplotype and traditional fine-mapping approaches and are associated with CASP8 expression differences in normal breast tissue. The first two reside 435 bp apart in a promoter proximal region that exhibits DNAse I hypersensitivity in mammary epithelial and breast cancer cell lines, and binds multiple transcription factors (STAT3, MYC, Pol2, and CTCF) based on ENCODE data (22). Also notable are results from an eQTL study in blood (23), indicating highly significant decreased CASP8 expression associated with these cFSVs (Z = −5.44, P = 5.3 × 10⁻⁸) and suggesting the expression signature is also readily observed in blood (Table 3). The latter cFSV, rs10197246, resides in a potential distal regulatory element, 80 kb from a CASP8 promoter, that is bound by MYC in the breast cancer cell line MCF-7. Hence, MYC-triggered apoptosis may be one possible mechanism for the association, worthy of investigation.

We tested for functional differences between allelic variants of rs3769823, rs3769821, and rs10197246, by performing enhancer assays based on a luciferase reporter gene in cell lines derived from MCF10A (normal breast tissue) and T-47D (breast tumor). Figure 4 shows the enhancer activity of the risk and neutral alleles in each cell line. We found that the region containing the risk alleles of rs3769821 and rs3769823 drove significantly lower gene expression in both the normal and tumor cell lines compared with the same region harboring the neutral alleles. The region surrounding the risk allele of rs10197246 produced approximately 3-fold lower expression compared with the neutral allele in the normal breast cell line, but there was no difference between alleles in tumor cells. Overall, the functional results are consistent with our eQTL study and suggest that these cFSVs cause reproducible reduction of CASP8 expression in normal breast tissue.

In our proof-of-principle analysis of the 2q33 locus for breast cancer, we considered our focused extreme discordant haplotype study together with the much larger BCAC traditional fine-mapping study, in addition to tissue-specific RNAseq and public annotations, to select rs3769823, rs3769821, and rs10197246 as the three most compelling cFSVs for breast cancer risk. The first two SVs are 435 bp apart in CASP8, whereas rs10197246 resides 81 kb telomeric within the adjacent ALS2CR12 gene. Despite the 81 kb physical separation, all three cFSVs target CASP8, based on the eQTL data. Although all three lie on the risk haplotype, LD between rs3769823/rs3769821 and rs10197246 is not particularly strong (r² = 0.66), suggesting that there may be two functional hits on the haplotype. This hypothesis is consistent with the luciferase enhancer activity results that illustrate that risk alleles at both rs3769823/rs3769821 and rs10197246 independently and significantly decreased enhancer activity. Of particular interest was that CASP8 expression differences were consistently stronger in normal breast tissue compared with breast tumor material, both in the frozen tissue samples (RNAseq data) and the representative cell lines (enhancer assays). These observations support the hypothesis that these three variants are functionally relevant breast cancer risk variants that create an environment more conducive to tumorigenesis by reducing CASP8 gene expression (and hence levels of apoptosis) in the normal breast. Moving forward, MYC-triggered apoptosis may be one possible mechanism worthy of investigation, given rs3769823 and rs10197246 appear good functional candidates and both reside in regions bound by MYC in the breast cancer cell line MCF-7. The importance of CASP8 in breast cancer tumorigenesis was also recently underscored by its identification as one of 32 genes significantly somatically mutated in breast cancer tumors (24), and one of only 22 genes identified as significantly mutated in at least three different common tumor types (25), suggesting that other common cancers may follow the pattern for susceptibility that we see here.

Our interrogation of the 2q33 breast cancer risk locus implicates extremely common FSVs (RAF∼0.28) with small effect sizes (∼17% decreased gene expression), consistent with the common disease common variant model. The largely noncoding candidates identified and the consistent stronger significance in normal tissues suggest that greater availability of resources for gene expression and other epigenetic data in normal tissues may prove critical in the prioritization and identification of common, low-risk variants. Initiatives, such as the Roadmap Epigenomics Project and ENCODE, will certainly help in this regard as they come to fruition (26). Here, we were successful in illustrating differences in 88 local samples that could be replicated based on 97 samples from TCGA. However, in general, much larger sample sizes may be needed to generate robust findings.

Our small, focused, discordant haplotype DNAseq study identified 18 cFSVs for the CASP8/ALS2CR12 region. These results aligned with and extended the 16 cFSV suggested from a traditional fine-mapping association study almost four orders of magnitude larger, leading to a combined set of 25 cFSVs, and resulting in three variants (rs3769823, rs3769821, and rs10197246) implicated as functionally relevant breast cancer risk alleles influencing CASP8 expression. The discordant haplotype sequencing and traditional fine-mapping approaches have distinctly different strengths and weaknesses: statistically, technologically, and economically. Large population-based association studies attain good power to identify common, low-risk loci with accurate risk estimates, but fine-mapping efforts remain largely constrained to imputation because sequencing costs are prohibitive. The discordant haplotype design attains excellent power specifically for the locus for which they are optimized and is therefore cost effective for DNAseq, but needs to be carried out separately for each association signal if there is more than one in a region. A noteworthy limitation of both approaches is the challenge posed by repetitive DNA sequences, with around 50% of the human genome consisting of repetitive elements (27). For targeted capture, it is not possible to design uniquely mapping capture baits for these regions, and for any HTS technology, alignment is difficult across repetitive elements. In our discordant haplotype study, for example, we were able to capture only 61% of the region at a read depth of ≥10× (average 76× depth across the full capture). In particular, we did not identify 7 SVs that were available in the BCAC imputation study. Highly repetitive sequence affects imputation studies differently. Repetitive regions are likely to harbor increased rates of misalignment, leading to nonrandom genotype errors in the reference panel, from which imputations are made. Quality scores may not easily identify such errors. An example of this in the exome design is the appearance of highly variable “promiscuous” genes within gene families with a high degree of identity at the nucleotide level. In addition, the 1000G phase III data, frequently used as a whole-genome imputation reference panel, have only an approximately 4 to 6× average read depth across the majority of the noncoding regions of the genome, leading to low quality genotypes and an overabundance of homozygotes. To investigate this phenomenon, we downloaded all Caucasian 1000G whole genome .bam files that were indicated to have passed quality control. We performed a best practice haplotypeCaller GATK variant calling pipeline on these 191 genomes and inspected the data for the 7 SVs imputed in the BCAC study but not captured by baits in the current DNAseq study. As expected, the 1000G sequencing coverage for these 7 SVs was low (median DP = 5–7×). Call rates ranged from 83% to 99%. Of those called, individual-level genotype qualities were relatively poor (median GQ = 15–24; GQ < 20 often considered unreliable, GQ ≥ 30 often used as a quality filter). Hence, both designs are affected by repetitive genome sequences, albeit in slightly different ways, adding further value to a joint-interpretation approach. We conclude that these two approaches are extremely complementary and suggest that DNAseq in a nested discordant haplotype design within larger case–control studies could play an important role in identifying comprehensive short lists for functional studies.

L. Cannon-Albright has expert testimony in Myriad Genetics. No potential conflicts of interest were disclosed by the other authors.

Conception and design: N.J. Camp, G.J. Burghel, S. Knight, J. Gertz, A. Cox

Development of methodology: N.J. Camp, S.H. Rigas, G. Wang, S. Knight, R. Abo

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): N.J. Camp, M.A. Parry, S.H. Rigas, P.-Y. Tai, K. Berrett, I.W. Brock, B. Jones, P.S. Bernard, L. Cannon-Albright, M.W.R. Reed, J. Gertz, A. Cox

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): N.J. Camp, W.-Y. Lin, G.J. Burghel, T.L. Mosbruger, M.A. Parry, R.G. Waller, V. Rajamanickam, R. Cosby, B. Jones, S. Knight, J. Gertz

Writing, review, and/or revision of the manuscript: N.J. Camp, G.J. Burghel, R.G. Waller, R. Cosby, R.E. Factor, P.S. Bernard, L. Cannon-Albright, S. Knight, R. Abo, T.L. Werner, J. Gertz, A. Cox

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): N.J. Camp, A. Bigelow, R.G. Waller, I.W. Brock, B. Jones, D. Connley, R. Sargent, L. Cannon-Albright

Study supervision: N.J. Camp, B. Jones

The authors gratefully acknowledge funding support. They thank Neil Hall, Lisa Olohan, and Xuan Liu of the Liverpool MRC Genomics Hub for DNA sequencing. At the University of Utah, they thank Derek Warner and Michael Kline of the Genomics Core for genotyping; Brian Dalley of the High Throughput Genomics Core for RNA sequencing; David Nix and Brett Milash of the Bioinformatics Shared Resource; the Biorepository and Molecular Pathology core; Roger Edwards and the Research Informatics Core; and the Women's Cancer Disease-Oriented Team at the Huntsman Cancer Institute. At the University of Sheffield, they acknowledge Helen Cramp, Sabapathy Balasubramanian, Simon S. Cross, and Sue Higham for subject recruitment and data collection. The results described are in part based upon data generated by the TCGA Research Network: http://cancergenome.nih.gov/. Finally, they thank all the participants in this multicenter study for making this research possible.

This work was supported by The Avon Foundation (02-2009-080 to N.J. Camp); the NCI (NCI R01 CA163353 to N.J. Camp), and a pilot grant to N.J. Camp/J. Gertz from the Cancer Control and Population Sciences and Nuclear Control Programs of the NCI Huntsman Cancer Center (P30 CA42014); The Susan G. Komen Foundation (BCTR0706911 to N.J. Camp); Yorkshire Cancer Research (YCR S305PA, S295 to A. Cox and SPP060 to S.H. Rigas); and Cancer Research UK, CRUK (C9528/A11292 to A. Cox). Ascertainment in Utah was made possible in part by the Utah Cancer Registry (UCR) and the Utah Population Database (UPDB). The UCR is funded by Contract No. HHSN261201300017I from the NCI, with additional support from the Utah State Department of Health and the University of Utah. Partial support for all datasets within the UPDB is provided by the HCI and its NCI Cancer Center Support grant, P30 CA42014.

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.