An exome-wide analysis of low frequency and rare variants in relation to risk of breast cancer in African American Women: the AMBER Consortium

Summary

We conducted the largest exome chip analysis to date for breast cancer in African ancestry women. Gene-based analyses identified associations between the PDE4D and FBXL22 genes and estrogen receptor negative breast cancer, based on extremely rare variants. Replication is needed.

Abstract

A large percentage of breast cancer heritability remains unaccounted for, and most of the known susceptibility loci have been established in European and Asian populations. Rare variants may contribute to the unexplained heritability of this disease, including in women of African ancestry (AA). We conducted an exome-wide analysis of rare variants in relation to risk of overall and subtype-specific breast cancer in the African American Breast Cancer Epidemiology and Risk (AMBER) Consortium, which includes data from four large studies of AA women. Genotyping on the Illumina Human Exome Beadchip yielded data for 170 812 SNPs and 8287 subjects: 3629 cases (1093 estrogen receptor negative (ER−), 1968 ER+, 568 ER unknown) and 4658 controls, the largest exome chip study to date for AA breast cancer. Pooled gene-based association analyses were performed using the unified optimal sequence kernel association test (SKAT-O) for variants with minor allele frequency (MAF) ≤ 5%. In addition, each variant with MAF >0.5% was tested for association using logistic regression. There were no significant associations with overall breast cancer. However, a novel gene, FBXL22 (P = 8.2×10^–6), and a gene previously identified in GWAS of European ancestry populations, PDE4D (P = 1.2×10^–6), were significantly associated with ER− breast cancer after correction for multiple testing. Cases with the associated rare variants were also negative for progesterone and human epidermal growth factor receptors—thus, triple-negative cancer. Replication is required to confirm these gene-level associations, which are based on very small counts at extremely rare SNPs.

Introduction

Genome-wide association studies (GWAS) have identified more than 90 genetic loci associated with breast cancer (1,2). Per-allele odds ratios have been modest (most <1.2), as is typical for GWAS findings. These low-penetrance loci, together with previously discovered high- and moderate-penetrance genes, fail to explain the majority of the genetic contribution to the disease (1–4). Most GWAS-based associations have been established in European or Asian populations, and the majority of these associations have failed to reach statistical significance in studies of African ancestry (AA) women (5–15). While larger AA sample sizes and an accounting for differences in linkage disequilibrium across ethnicities would likely result in more successful replications, the European-discovered risk variants may also explain less of the genetic contribution to breast cancer in AA women.

While some of the unexplained breast cancer heritability in AA women is surely due to unidentified common susceptibility SNPs in this population, another portion may be explained by less common (1–5%) and rare variants (< 1%). These lower frequency variants represent a large proportion of all human genetic variation but are poorly captured by most GWAS arrays (16). There are a growing number of examples of rare variants associated with complex disease, with findings for autism, schizophrenia, inflammatory bowel disease and diabetes (17). In addition to the already established high- and moderate-penetrance genes for breast cancer (3,4,18), novel low frequency risk variants for cancers including prostate (19) and ovarian (20,21) have also been reported. Still, it remains unclear how much rare variants contribute to the heritability of breast cancer and other complex diseases.

In recent years, the development of exome-wide arrays has allowed for the relatively inexpensive assessment of known rare exonic variants. Current exome arrays include >200 000 coding variants and were developed on the basis of whole exome sequencing data from ~12 000 individuals. Most of those sequenced were of European ancestry, but a small number of AAs and other ethnicities were included as well (16,22).

A case-control study nested in the Multiethnic Cohort (MEC) used the Illumina exome chip to investigate the role of rare exonic variation in the etiology of breast cancer (16). Single SNP analyses were conducted, as well as gene-based testing of the burden of rare alleles. Only one significant association was found, for splice-site SNP rs145889899 in the LDLRAD1 gene. This variant was only seen in AAs (with a frequency of 0.65% in AA controls) and had an odds ratio of 3.74. While no additional findings were significant, there was low power to detect genotype relative risks ≤ 2 in the AA participants due to the modest number of available AA cases (N = 591).

The present study combined the MEC exome chip data with exome chip data from three additional studies of breast cancer in AA women, forming the African American Breast Cancer Epidemiology and Risk (AMBER) Consortium, the largest exome wide analysis sample to date for AA breast cancer (3629 cases and 4658 controls). We primarily used gene-based methods for association analysis, given the relatively low power and high multiple testing burden for single SNP analyses of rare variants (22,23). Gene-based testing has the potential to increase power when multiple SNPs in a given gene are associated (22).

Materials and methods

Study population

This investigation was conducted using data from the AMBER Consortium, a collaboration of four of the largest studies of breast cancer in AA women. The AMBER Consortium has been described previously (24), and prior reports have detailed the individual studies: the Carolina Breast Cancer Study (CBCS) (25), the Women’s Circle of Health Study (WCHS) (26,27), the Black Women’s Health Study (BWHS) (28), and the Multiethnic Cohort (MEC) (29). Institutional Review Board approval was obtained for each study, and all participants provided written informed consent.

Briefly, the CBCS is a North Carolina population-based case-control study of women aged 20–74 years that began in 1993. The North Carolina Central Cancer Registry’s rapid case ascertainment system was used for case identification, and controls were selected through 2001 using Division of Motor Vehicles lists (age < 65 years) and Health Care Financing Administration lists (age ≥ 65). Interviewers collected questionnaire data and samples for DNA analysis in home visits.

The WCHS is a multi-site case-control study in New York City (NYC) (2002–2008) and New Jersey (NJ) (2006-present). Hospital-based ascertainment of cases aged 20–75 years was used in NYC, and controls were selected through random digit dialing. Cases in NJ are identified by the NJ State Cancer Registry using rapid case ascertainment, and controls are identified through random digit dialing and community-based efforts (27). Risk factor data and samples for DNA analysis are obtained during in-person interviews.

The BWHS is a prospective cohort study of 59 000 AA women from across the United States who enrolled by completing a postal health questionnaire in 1995. The age range at baseline was 21–69 years. Biennial follow-up questionnaires identify new cases of breast cancer, and these cases are confirmed by medical records or from state cancer registry data and the National Death Index. Nearly 27 000 BWHS participants provided saliva samples for DNA analysis.

The MEC is a prospective cohort study that began in 1993 with the enrollment of men and women aged 45–75 years from a range of ethnic groups in Hawaii and California. Data are collected by mailed questionnaire at 5-year intervals, and breast cancer cases are confirmed through the Hawaii and California state cancer registries and the National Death Index. Blood samples were collected from study participants for DNA analysis.

Eligible cases for the present analyses were AA women with incident invasive breast cancer or ductal carcinoma in situ (DCIS). For BWHS and MEC, controls were chosen from among women without breast cancer, and were frequency matched to cases on geographical region, sex, race and 5-year age group. ER status for cases was determined using pathology data from hospital records or cancer registry records.

Genotyping and QC

Genotyping of DNA from participants in the BWHS, CBCS and WCHS was performed by the Center for Inherited Disease Research (CIDR) using the Illumina Human Exome Beadchip v1.1. This array includes >200 000 coding variants, as well as tag SNPs for GWAS hits, a grid of common variants, and ancestry informative markers (AIMs). A description of the exome chip design is available from http://genome.sph.umich.edu/wiki/Exome_Chip_Design. CIDR used the GenTrain Version 1.0 calling algorithm in GenomeStudio version 2011.1, Genotyping Module 1.9.4. Manual review was conducted for all Y, XY pseudoautosomal and mitochondrial SNPs. Autosomal and X chromosome SNPs were also manually reviewed if a rare heterozygous cluster may have been missed by the GenCall algorithm and if the zCall algorithm (30) identified four or more possible new heterozygous points.

A total of 246 519 SNPs were genotyped, and 231 705 SNPs remained after excluding variants that failed technical filters imposed by CIDR, or QC filters recommended by the University of Washington. Briefly, genotypes with a GenCall (GC) score <0.15 were classified as missing, and SNPs were removed if they had poor cluster properties (ex. cluster separation <0.2 or <0.3 depending on allele frequency), call rates <0.98, Hardy–Weinberg Equilibrium P < 1×10^-4, >1 Mendelian error in trios from HapMap (31) or >2 discordant calls in duplicate samples. Mitochondrial and Y chromosome SNPs were also removed. Genotypes were attempted for 6936 participants from the BWHS, CBCS and WCHS, and were completed with call rate >98% for 6828 participants, which included 3130 cases (963 estrogen receptor negative (ER−), 1674 ER+, 493 ER unknown) and 3698 controls.

Genetic data from 499 cases (130 ER−, 294 ER+, 75 ER unknown) and 960 controls in the MEC were available from genotyping on a previous version of the exome chip (16) which contained >99% of the high quality variants from v1.1. Genotypes from MEC were combined with the data from the other AMBER studies into a data set containing 245 571 SNPs. Greater than 66 000 SNPs were monomorphic in the combined set and were omitted from analyses, as were SNPs with high quality data from only one of the two exome chips and SNPs with any discordant genotypes across the two chips for 30 MEC participants who were included on both. The final data set for analysis included 170 812 SNPs and 8287 participants: 3629 cases (1093 ER−, 1968 ER+, 568 ER unknown) and 4658 controls.

We used the CHARGE (Cohorts for Heart and Aging Research in Genomic Epidemiology) Consortium’s annotation of exome chip variants (version 6, 11/7/14) downloaded from http://www.chargeconsortium.com/main/exomechip (32). This annotation was performed with dbNSFP version 2.6 (33,34).

We used the smartpca program in the EIGENSOFT package (35) to conduct a principal components analysis (PCA) based on ~42 000 common SNPs, most of which were custom content additions to the exome chip for use in other AMBER projects. In a separate analysis, PLINK version 1.07 (36) was used to estimate identity by descent in participant pairs, and identified 130 sets of relatives across and within the individual studies, consisting of 270 individuals. These 270 individuals were flagged, as were 35 outlying individuals from the PCA, so that sensitivity analyses could be performed. Genotype principal components were tested for association with case status after controlling for the study covariates: study, DNA source (blood, saliva[Oragene], saliva[mouthwash]), and the matching variables. While no principal components were strongly associated in the multivariable model, we included terms for principal components with P < 0.1 in our analyses.

Association analysis

Gene-based association analyses for overall, ER+, and ER− breast cancer were conducted using the unified optimal sequence kernel association test (SKAT-O) (37), as implemented in the R package seqMeta (38). As a linear combination of the burden and SKAT (39) tests, SKAT-O achieves robust power whether a given gene has a high proportion of causal variants exerting effects in the same direction, or instead has many noncausal variants or variants exerting effects in opposite directions (22). We used the default SKAT-O option in the seqMeta package that considers rho = 1 (burden) and rho = 0 (SKAT) tests and selects the optimal of the two tests. Depending on which test is chosen, SKAT-O models the phenotype versus a weighted aggregation of either the variants (burden test) or the variant score test statistics (SKAT) to produce a gene-level P value that indicates the degree of enrichment of rare variant associations in that gene (37). We included variants with minor allele frequency (MAF) ≤ 5%, and used the beta distribution weights proposed by Wu et al. (39), which upweight rarer variants, for both tests. We used a Bonferroni correction based on the number of genes evaluated to assess the significance of the gene-based test results.

We performed separate gene-based analyses for three successively less stringent sets of exonic variants: (1) ‘NS_strict’ variants (based on Purcell et al. (40)): stopgain, stoploss, frameshift or predicted damaging by all five of the following algorithms: SIFT (41), mutationTaster category [A or D] (42), LRT (43), PolyPhen_HDIV (44) and PolyPhen_HVAR (44), (2) ‘NS_broad’ variants (Purcell et al. (40)): ‘NS_strict’ variants plus those variants that are predicted damaging by at least one of the five algorithms and (3) All nonsynonymous variants (‘NS_all’): ‘NS_broad’ variants plus all other missense and splice variants. Testing of these three sets of variants gave us more flexibility to find the best set of SNPs for gene-based analysis (ideally a set including most or all truly associated SNPs, but few, if any, unassociated SNPs).

Single SNP association analyses were conducted using logistic regression as implemented in PLINK version 1.07. These analyses were restricted to variants with MAF > 0.5% in order to avoid performing a large number of underpowered tests. We used a Bonferroni adjustment for the effective number of independent tests, applying the method of Gao et al. (45), to assess the significance of the single SNP results.

Both gene-based and single SNP analyses were adjusted for study, age, geographic region, DNA source and genotype principal components 5, 6 and 8 in a pooled analysis that combined individual level data across the four studies in AMBER. This approach was preferred over meta-analysis given prior evidence that pooled analysis is more powerful for gene-based testing of rare variants under conditions where pooling is appropriate (46).

Results

The present analyses included 3629 breast cancer cases (1093 ER−, 1968 ER+, 568 ER unknown) and 4658 controls. Table 1 shows the distribution of ER subtypes and age at diagnosis for the cases by study site.

Table 1.

Characteristics of participants in the AMBER Consortium by study site

	BWHS	CBCS	WCHS	MEC	ALL AMBER
Controls	2249	615	834	960	4658
Cases	901	1408	821	499	3629
ER+	498	741	435	294	1968
ER−	233	565	165	130	1093
ER unknown	170	102	221	75	568
Age at diagnosis
<40	47	204	85	0	336
40–49	262	459	215	9	945
50–59	302	381	292	108	1083
60–69	204	267	173	165	809
≥70	86	97	56	217	456

AMBER, African American Breast Cancer Epidemiology and Risk; BWHS, Black Women’s Health Study; CBCS, Carolina Breast Cancer Study; ER, estrogen receptor; MEC, Multi-Ethnic Cohort; WCHS, Women’s Circle of Health Study.

There were 184 100 annotation records for the 170 812 SNPs that passed QC filters: some SNPs mapped to more than one gene, and these multiple mappings were maintained for the gene-based analyses we performed. More than 80% of the SNP records were annotated as nonsynonymous, including missense, stopgain, stoploss, frameshift and splicing variants (see Supplementary Table 1, available at Carcinogenesis Online, for the full distribution of roles for the final SNP set). Over 80% of the SNPs had MAF <5% in AMBER, over 70% had MAF <1%, and nearly half of the SNPs had MAF <0.1%. QQ plots for the gene-based and single SNP association analyses we performed are shown in Supplementary Figure 1, available at Carcinogenesis Online. As is common for SKAT analyses of binary traits, there was inflation in the gene-based tests (47,48).

The number of gene-based tests conducted and the resulting alpha levels for significance are listed in Table 2 by outcome and SNP group. As the SNP functional group became more strict, fewer tests were conducted because fewer genes contained at least two SNPs in the given group. Fewer gene-based tests were conducted for the ER+ and ER− analyses compared to overall breast cancer because these subtype analyses had smaller sample sizes, which resulted in more monomorphic SNPs that were excluded.

Table 2.

Number of gene-based tests conducted and corresponding significance criteria

Analysis	Number of genes tested	Alpha level for significance
Overall breast cancer
‘NS_all’ SNPsa	14 652	3.4×10–6
‘NS_broad’ SNPsb	12 515	4.0×10–6
‘NS_strict’ SNPsc	3128	1.6×10–5
ER+ breast cancer
‘NS_all’ SNPs	14 515	3.4×10–6
‘NS_broad’ SNPs	12 316	4.1×10–6
‘NS_strict’ SNPs	2963	1.7×10–5
ER− breast cancer
‘NS_all’ SNPs	14 399	3.5×10–6
‘NS_broad’ SNPs	12 184	4.1×10–6
‘NS_strict’ SNPs	2865	1.8×10–5

^a‘NS_all’ SNPs: stopgain, stoploss, frameshift, missense or splicing.

^b‘NS_broad’ SNPs: stopgain, stoploss, frameshift, or predicted damaging by at least one of the following algorithms: SIFT, mutationTaster category [A or D], LRT, PolyPhen_HDIV or PolyPhen_HVAR.

^c‘NS_strict’ SNPs: stopgain, stoploss, frameshift or predicted damaging by all five algorithms.

Table 3 shows the five most significant genes for each SKAT-O run (see Supplementary Table 2, available at Carcinogenesis Online, for the top 50 genes for each set of variants). For overall and ER+ breast cancer, RTN4RL1 was the most significant gene for both the ‘NS_all’ and ‘NS_broad’ SNP sets, with nominal P values ranging from 1.8×10^–5 to 1.9×10^–4. These results were based on 6–10 SNPs that were used in burden tests (the SKAT-O method selected rho = 1 as optimal in these instances). For the ‘NS_strict’ variants, the most significant genes for overall and ER+ breast cancer were IQCA1 (P = 4.6×10^–4) and FSCN3 (P = 2.3×10^–4), respectively. None of the top results for overall or ER+ breast cancer survived a multiple testing correction based on the number of genes evaluated.

Table 3.

The most significant gene-based test results for each analysis

Gene	Nominal P value	Corrected P valuea	rhob	Cumulative MAF (%)	Number of SNPs included in the test
Overall breast cancer, ‘NS_all’ SNPs
RTN4RL1	1.3×10–4	1	1	4.31	10
TPCN1	2.0×10–4	1	0	0.32	9
RARA	2.3×10–4	1	1	2.42	2
KIF3C	3.0×10–4	1	1	2.62	6
OBSCN	4.4×10–4	1	0	75.88	146
Overall breast cancer, ‘NS_broad’ SNPs
RTN4RL1	1.9×10–4	1	1	1.96	6
RARA	2.3×10–4	1	1	2.42	2
GPRASP1	3.0×10–4	1	0	5.35	5
NCAPG2	3.3×10–4	1	0	2.91	9
TMEM130	3.6×10–4	1	0	0.11	2
Overall breast cancer, ‘NS_strict’ SNPs
IQCA1	4.6×10–4	1	0	3.09	3
PDE4A	1.2×10–3	1	1	0.53	2
ECT2L	2.0×10–3	1	0	1.33	2
GCKR	2.5×10–3	1	1	0.21	4
ACSF3	2.6×10–3	1	1	0.17	3
ER+ breast cancer, ‘NS_all’ SNPs
RTN4RL1	1.8×10–5	0.25	1	4.37	10
OR2W5	7.3×10–5	1	1	16.16	11
CYSRT1	1.0×10–4	1	1	0.23	2
RNF130	1.3×10–4	1	0	0.69	3
GABPA	1.3×10–4	1	1	0.02	2
ER+ breast cancer, ‘NS_broad’ SNPs
RTN4RL1	3.2×10–5	0.39	1	2.00	6
RNF130	1.3×10–4	1	0	0.69	3
GABPA	1.3×10–4	1	1	0.02	2
OR5H15	3.5×10–4	1	1	0.05	2
TCHP	3.9×10–4	1	1	4.35	11
ER+ breast cancer, ‘NS_strict’ SNPs
FSCN3	2.3×10–4	0.67	1	0.06	3
GUF1	7.0×10–4	1	0	0.22	3
ZIM3	7.8×10–4	1	1	0.87	2
TSC2	8.0×10–4	1	0	0.61	5
TBPL2	9.1×10–4	1	0	0.09	2
ER− breast cancer, ‘NS_all’ SNPs
PDE4D	1.2×10–6	0.017	1	0.02	2
PLEKHG5	2.2×10–5	0.31	1	9.36	18
CCNDBP1	4.4×10–5	0.64	0	0.06	3
DIMT1	5.8×10–5	0.84	0	0.03	2
TEX12	1.6×10–4	1	0	0.20	2
ER− breast cancer, ‘NS_broad’ SNPs
PDE4D	1.2×10–6	0.015	1	0.02	2
LRRC8D	1.8×10–5	0.22	0	0.26	3
CCNDBP1	4.4×10–5	0.54	0	0.06	3
MRPS31	9.8×10–5	1	1	1.12	2
NCR1	1.1×10–4	1	0	0.50	2
ER− breast cancer, ‘NS_strict’ SNPs
FBXL22	8.2×10–6	0.023	1	0.02	2
CCNDBP1	3.8×10–5	0.11	0	0.05	2
SCARB1	9.6×10–5	0.28	1	0.02	2
QRSL1	9.8×10–5	0.28	0	0.07	2
MFGE8	1.1×10–4	0.31	0	0.05	3

MAF, minor allele frequency.^aBonferroni correction for the number of genes tested.

^bThe rho parameter indicates whether the SKAT test (rho = 0) or burden test (rho = 1) gave the smallest P value.

The most significant genes for ER− breast cancer were PDE4D (P = 1.2×10^–6 using either the ‘NS_all’ or ‘NS_broad’ SNP sets) and FBXL22 (P = 8.2×10^–6 using the ‘NS_strict’ SNP set), and these survived a correction for multiple testing. The PDE4D and FBXL22 results were each based on burden testing (rho = 1) of two SNPs with a cumulative MAF ~0.02%. Details of the four SNPs contributing to these significant test results are shown in Table 4. All of the contributing SNPs are nonsynonymous coding SNPs for multiple isoforms of PDE4D or FBXL22. These SNPs had good genotyping cluster properties (Supplementary Figure 2, available at Carcinogenesis Online) and 100% genotyping pass rates in the present study. Each SNP had an MAF ~0.01% in the AMBER analysis of ER− cases and controls, due to the presence of one rare allele in one invasive ER− case. The rare allele carriers were four independent participants (one for each SNP) with varying ages at diagnosis and percentages of AA (Table 4). All four of these women had triple-negative breast cancer (tumors negative for estrogen receptors, progesterone receptors and human epidermal growth factor receptor 2). Among all 1093 ER− cases, 599 had been classified as triple negative based on available data on all three molecular markers. The four SNPs of interest are monomorphic in AAs from the 1000 Genomes Project (49) Phase 3 and the NHLBI ESP (National Heart, Lung, and Blood Institute Exome Sequencing Project) (50). These same projects report very low allele frequencies (≤0.2%) for these SNPs in European ancestry populations (Table 4).

Table 4.

SNPs contributing to significant gene-based tests for ER− breast cancer

SNP	rsID	Function	Fail ratea	MAF: analysis sample (%)b	MAFs: 1000G Phase 3	MAFs: NHLBI ESP	The 1 AMBER subject with one rare allele
							Disease status	Study	Age at diagnosis	% African ancestryc
Gene PDE4D
exm456537	rs201360779	Missense	0	0.01	Monomorphic except in FIN (1%)	Monomorphic in AAs; 0.05% in EAs	Invasive, ER−, PR−, HER2−	BWHS	45	68.6
exm456565	rs200725508	Missense	0	0.01	No data	Monomorphic in AAs; 0.03% in EAs	Invasive, ER−, PR−, HER2−	WCHS	61	99.6
Gene FBXL22
exm1168070	rs201654150	Missense	0	0.01	No data	Monomorphic in AAs; 0.02% in EAs	Invasive, ER−, PR−, HER2−	BWHS	48	82.6
exm1168081	rs149590841	Missense	0	0.01	Monomorphic except in ITU (0.5%)	Monomorphic in AAs; 0.10% in EAs	Invasive, ER−, PR−, HER2−	BWHS	35	61.8

AAs, African Americans; EAs, European Americans; ER−, estrogen receptor negative; 1000G, 1000 Genomes; HER2−, human epidermal growth factor receptor 2 negative; FIN, Finnish; ITU, Indian Telugu in the UK; NHLBI ESP, National Heart, Lung, and Blood Institute Exome Sequencing Project; PR−, progesterone receptor negative.

^aGenotyping fail rate in this study.

^bMinor allele frequency in the AMBER analytic sample of ER− cases and controls. All four SNPs have an MAF ~0.01%, which corresponds to the presence of one rare allele in one subject.

^cPercent African versus European ancestry as estimated by ADMIXMAP (51) using 2624 AIMs (ancestry informative markers).

Sensitivity analyses were run for the PDE4D and FBXL22 genes versus ER− breast cancer, excluding first-degree and second-degree relatives identified via the genotypes, as well as PCA outliers who clustered with HapMap 3 Europeans, Mexicans, or Asians. Results became more significant with these exclusions (PDE4D P = 9.9×10^–7; FBXL22 P = 7.4×10^–6).

Single SNP association analyses were performed for the 58 776 SNPs with MAF >0.5%. The correlation among these SNPs yielded the equivalent of 50 245 independent tests (45); therefore, the threshold for significance was set at 9.95×10^–7. SNP rs8100241, a previously reported risk marker at the ER−/ triple-negative GWAS locus on chromosome 19p13.11 (12,52–56), met this study-wide threshold for ER− disease (P = 1.7×10^–7). The A allele at rs8100241 had a frequency of 40% in the present study and was associated with a decreased risk of ER− breast cancer (OR 0.75, 95% CI 0.68, 0.84). No other individual SNPs reached statistical significance (Supplementary Table 3, available at Carcinogenesis Online), including the LDLRAD1 SNP rs145889899 (P = 0.17), for which an association had been reported in the MEC exome chip study (16).

Discussion

In these analyses, we observed significant associations between the PDE4D and FBXL22 genes and ER− breast cancer in a relatively large sample of AA women, using gene-based testing of rare exonic variants. Two nonsynonymous coding SNPs in each of these genes were responsible for their significant associations. The minor allele at each of these four SNPs was present in one invasive ER− case each (a different case subject for each of the four SNPs). These four cases were all triple-negative breast cancers. The four SNPs of interest from these two genes were absent in the AMBER controls. This is consistent with their reported monomorphism in AA samples from the 1000 Genomes Project and the NHLBI ESP.

Although the association we report for the PDE4D gene is with ER− (and triple-negative) breast cancer, one of the two rare SNPs contributing to this association (rs201360779) was also seen in one invasive ER+ case. Thus, this gene may affect the risk of both ER subtypes. The two contributing PDE4D SNPs in our study were predicted to be damaging by mutationTaster (42), although SIFT (41) predicted that these mutations would be tolerated. The PolyPhen HDIV and HVAR models (44) predicted damaging results from rs200725508, but these algorithms predicted benign results for rs201360779 (with the exception of the HDIV prediction of ‘possibly damaging’ for this SNP for one PDE4D isoform).

The 2013 GWAS meta-analysis by Michailidou and colleagues (1) reported a breast cancer association with SNP rs1353747, which is located in an intron of PDE4D. In that study, the G allele at this common SNP showed weak protective associations for both ER+ and ER− disease. In this study, rs1353747 was not associated with either disease subtype.

PDE4D is located on chromosome 5q11.2–12.1 and encodes phosphodiesterase subtype 4D, a member of the PDE4 family of phosphodiesterases, which multiple tumor cell types express as major regulators of cAMP degradation (57). PDE4D may function as a tumor-promoting factor by causing lower cAMP concentrations, which have been linked to increased survival and proliferation of cancer cells. This oncogenic role is supported by experiments showing that inhibition of PDE4D causes apoptosis and growth retardation in multiple types of cancer cells, including breast, but not in nonmalignant epithelial cells (57).

Lin et al. (57) reported PDE4D homozygous deletions in 198 of 5569 (3.6%) primary tumors from The Cancer Genome Atlas (TCGA) projects and TumorScape (58), with most being internal microdeletions. They also found microdeletions in established cancer cell lines including breast. These microdeletions were associated with increased expression of the protein, and they affected upstream conserved regions 1 (UCR1) and 2 (UCR2) of the gene. UCR1 and UCR2 inhibit PDE4D activity, likely by forming complexes with the PDE4D catalytic domain before cAMP enters the site. Lin et al. showed that a short isoform of PDE4D with no functional UCR1 or UCR2 promoted cancer cell growth, while a long isoform that contained both UCR1 and UCR2 did not. In the present study, the two rare missense mutations contributing to the PDE4D gene-level association were located upstream of UCR1 and UCR2 and were risk variants (not protective). It could be hypothesized that these variants act by inducing a change in protein structure that disrupts the interaction of the UCRs with the catalytic domain of PDE4D, thereby increasing protein activity.

FBXL22 has not previously been associated with breast cancer. This gene is located on chromosome 15q22.31 and encodes F-box and leucine-rich repeat protein 22. This F-box protein, a ubiquitin ligase component, has been shown to promote the degradation of sarcomeric proteins, and is critical for maintaining cardiac contractility in vivo (59). It is unclear what biological mechanism might link FBXL22 to breast cancer development. Nevertheless, the two rare SNPs contributing to the FBXL22/ER− association in this study met strict criteria for variant functionality: these nonsynonymous SNPs were predicted to be damaging by five different algorithms (41–44).

Single SNP analyses confirmed an ER− association for the GWAS locus on chromosome 19p13.11 (12,52–56). The associated SNP in the present study was the common missense variant rs8100241 in the ANKLE1 gene. This SNP has shown significant associations with overall (52), ER− (55) and triple-negative (53) breast cancer in prior studies of mostly European ancestry subjects. These studies reported odds ratios ranging from 0.84 to 0.88 for the A allele at this SNP, as compared with the odds ratio of 0.75 reported in the present analysis for ER− breast cancer. It should be noted, however, that the large GAME-ON meta-analysis (http://gameon.dfci.harvard.edu) reported weaker effect estimates for rs8100241: the odds ratio reported for overall breast cancer was 0.95 (95% CI 0.92, 0.99; P = 0.017), and the odds ratio for ER− breast cancer was 0.94 (95% CI 0.83, 1.07; P = 0.36).

Although the present study sample is the largest exome wide analysis sample to date for AA breast cancer, this analysis was underpowered to detect per-allele odds ratios <1.5, except when cumulative risk allele frequencies per gene approached or exceeded 5%. Further power limitations existed for analyses by ER status.

The significant gene-level findings reported here are based on four SNP variants that appear only once each in the AMBER sample of ER− cases and controls. Given these very small counts and the inflation seen in the gene-based test statistics, our results should be interpreted with caution. A simple Fisher’s exact test for the presence of a rare allele in PDE4D versus ER− case/control status yields a P-value of 0.036, as does the same test for FBXL22. Fisher’s exact test is conservative and does not upweight rarer variants or account for covariates; however, the modest P value from this test emphasizes the need for replication to verify associations between these genes and ER− breast cancer.

This is not to say that the rare variant calls for the four SNPs of interest are questionable. There is good reason to believe that these calls were accurate in our study. These SNPs are known rare variants that the exome chip was designed to capture. Each SNP was seen in two or more studies that contributed sequence data for development of the exome chip: the minor allele at rs201360779 was seen 29 times across the ~12 000 sequenced individuals, the minor allele at rs149590841 was seen seven times, and the minor alleles for the other two SNPs were each seen three times (ftp://share.sph.umich.edu/exomeChip/ProposedContent/codingContent/). In AMBER, all four SNPs showed high quality genotype clusters and clear separation of the heterozygous calls from the remainder of the genotypes. In addition, SNP rs201360779 was not a true singleton in AMBER, having also been seen in one ER+ case.

The exome chip used here has inherent limitations. First and most obvious, this array-based method includes only selected rare variants and is therefore not as exhaustive as exome sequencing in capturing rare exonic variants. Second, this chip does not attempt to assay rare variation in noncoding regions. Third, the chip was designed using exome sequencing data from mostly European samples. Therefore, rare variants in non-Europeans are not as well captured, and our data set may have lacked information on some important rare SNPs in AA populations (22).

Another potential limitation of our study is a current limitation of the field: the use of traditional methods such as PCA (or linear mixed models) to adjust for population stratification in rare variant association studies. These methods may not adequately control for population structure in some rare variant analysis settings; thus, the development of new methods for this purpose is an area of active research (23).

We should also acknowledge that while our multiple testing correction adjusted for the number of genes analyzed, there were additional levels of testing that were not included in this correction. Multiple outcomes (overall, ER+, ER− breast cancer) and SNP functional groups (NS_all, NS_broad, NS_strict) were also analyzed. We did not correct for the multiple breast cancer outcomes because there was considerable overlap among the groups of patients analyzed for overall, ER+ and ER− breast cancer, and we considered these to be tests of related hypotheses. There was also a high amount of interdependence among the three SNP functional groups, which would render a Bonferroni correction overly conservative. Nevertheless, implementing an adjustment for all of the ER− gene-based tests conducted across the three SNP functional groups results in a corrected P value of 0.035 for PDE4D, although the corrected P value for FBXL22 becomes non-significant (0.240).

In summary, an exome-wide gene-based analysis of rare variants found significant associations between the PDE4D and FBXL22 genes and ER− breast cancer in a collaborative study of AA women. The previous GWAS finding of a breast cancer risk marker in the PDE4D gene supports the idea that rare variants in this region in particular might affect breast cancer risk. Replication is needed to confirm the gene-level associations reported here, which are based on very small counts at extremely rare variants.

Supplementary material

Supplementary Tables 1–3 and Figures 1 and 2 can be found at http://carcin.oxfordjournals.org/.

Funding

This work was supported by the National Institutes of Health (P01 CA151135 to C.B.A., J.R.P., and A.F.O., R01 CA058420 to L.R., UM1 CA164974 to L.R. and J.R.P., R01 CA098663 to J.R.P., R01 CA100598 to C.B.A., UM1 CA164973 to L.L.M., L.K.M., and C.A.H., R01 CA54281 to L.N.K., R01 CA063464 to B.H., P50 CA58223 to M.A.T. and A.F.O., U01 CA179715 to M.A.T. and A.F.O.); the Department of Defense Breast Cancer Research Program, Era of Hope Scholar Award Program (W81XWH-08-1-0383 to C.A.H.); the Susan G. Komen for the Cure Foundation (to M.A.T. and A.F.O.); the Breast Cancer Research Foundation (to C.B.A.); and the University Cancer Research Fund of North Carolina (to M.A.T. and A.F.O.).

Glossary

Abbreviations

AA

African ancestry

GWAS

Genome-wide association studies

MAF

minor allele frequency