Background

Breast and prostate cancer are the most frequently diagnosed cancers in women and men respectively with over 200,000 cases each of new breast and prostate cancer estimated for 2006 in the United States [1]. Furthermore, prostate cancer is the second leading cause of cancer-related deaths in men and breast cancer is the second leading cause of cancer-related deaths in women. Family history is a well established risk factor for both breast and prostate cancer providing evidence for underlying genetic factors contributing to cancer occurrence. Accumulating research has identified a number of candidate genes and biologic pathways associated with increased susceptibility to cancer. However, even the most penetrant mutations, such as in BRCA1 and BRCA2, account for only 5–10% of cases and are present in <1% of the general population. Genome-wide association studies (GWAS) provide a comprehensive approach to identification of genetic variants associated with cancer risk unconstrained by existing knowledge and may permit detection of common genetic variants each with small associated cancer risk but great public health impact. Reports from two recent GWAS demonstrated the importance of this approach with the discovery of novel loci for breast cancer susceptibility [2, 3]. Four SNPs in the FGFR2 gene were strongly associated with breast cancer and the association was confirmed in a sample of cases and controls derived from three additional studies [3].

We used the Framingham Heart Study (FHS) Affymetrix 100K SNP genotyping resource for GWAS of breast and prostate cancer phenotypes. The FHS offers the advantage of a prospective longitudinal family-based community sample with participants who have been well-characterized throughout adulthood with respect to risk factors and diseases, including cancer. We report results of two complementary strategies to identify genome-wide associations with cancer phenotypes: 1) a simple low p-value SNP ranking strategy; and 2) 100K SNP associations within candidate genes and regions previously reported to be associated with these cancers in humans.

Methods

Study sample

The genotyped study sample comprised 1345 Original cohort (n = 258) and Offspring (n = 1087) participants from the 330 largest FHS families. The Overview [4] provides further details of this sample. There were 250 participants in the sample with cancer (excluding non-melanoma skin cancer) including 58 women with breast cancer, and 59 men with prostate cancer. The Boston University Medical Center Institutional Review Board approved the examination content of Original Cohort and Offspring examinations. All participants provided written informed consent including consent for genetic studies.

Cancer phenotype definitions and residual creation

The 5209 Original Cohort participants have been examined biennially since study inception in 1948 and the 5124 Offspring Cohort participants (children of the Original Cohort and spouses of the children) have been examined approximately every 4 years since enrollment in 1971. Cancer cases were identified at routine examinations or by health-history updates for participants who did not attend an examination. Medical records were reviewed by two independent reviewers (BEK, GLS). The vast majority of cancers were confirmed by pathology reports; <3.4% of cancer cases were based on death certificate or clinical diagnosis alone. The 1976 World Health Organization ICD-O coding was used to classify all primary cancers. Hence, topography, location (subdivision of site), histology or morphology (cell histopathology), behavior (degree of malignancy), and grade (histological grading & differentiation) were recorded along with date of diagnosis. Cancer cases reviewed through December 31, 2005 were included in this study. The proportion of women and men in the study sample with breast (8%) and prostate cancer (9%) respectively was similar to that in the full FHS sample.

Cox proportional hazards models were used to generate martingale residuals using the PHREG procedure in SAS to perform the regression analysis of time from study entry to cancer diagnosis or last contact free of cancer. Breast cancer was examined in women only and models were cohort-specific and adjusted for 1) age at entry and 2) age, parity, and body mass index at study entry. For prostate cancer, in men only, models were cohort-specific and adjusted for age at entry.

Genotyping

Affymetrix 100K SNP GeneChip genotyping and the Marshfield STR genotyping performed by the Mammalian Genotyping Service http://research.marshfieldclinic.org/genetics are described in the Overview [4]. SNPs were excluded if minor allele frequency <0.10 (n = 38062); genotypic call rate <0.80 (n = 2346); Hardy Weinberg equilibrium test p < 0.001 (n = 1595). There were 70,987 autosomal SNPs available for analysis after the exclusions.

Statistical Analysis

The statistical methods for genome-wide association analyses are described in detail in the Overview [4]. While there are various suggested methods for interpretation of genome-wide significance, we chose to use a conservative (p < 0.05/10-6 = 5 × 10-8) threshold to define genome-wide significance for this report.

Association

All cancer residual traits listed in Table 1 were computed using Cox proportional hazards models. The full set of FHS participants with the phenotype were used to create the residuals. The residuals were used to test for association between the genotyped subset of participants and the SNPs using family-based association test (FBAT) and generalized estimating equation (GEE) models. FBAT analyses were restricted to at least 10 informative families. The GEE tests tended to give an excess of very small p-values over what would be expected (see Overview [4]).

Table 1 Cancer Phenotypes for the Framingham Heart Study 100K Analyses
Full size table

SNP prioritization

We used several strategies to prioritize SNPs associated with cancer traits. First, we used an untargeted approach whereby SNP associations were ranked according to the strength of the p-value for each trait. Next we identified candidate genes reported to be associated with each cancer trait from review of the literature. Candidate genes were selected by searching PubMed (using susceptibility, gene, cancer and (breast or prostate) as keywords, last accessed 08-15-06), and the Entrez Gene and Online Mendelian Inheritance in Man resources, as well as recent text books. All available 100K SNPs in or near the a priori selected candidate genes were investigated for association with cancer traits. Finally for prostate cancer, we also examined SNP associations in the region on chromosome 8 (8q24) previously reported to be associated with prostate cancer in Icelandic families and confirmed in three case-control series [5] and African American men [6]. Further, for prostate cancer we examined the overlap in SNP associations in our study and the top 500 ranked SNPs from the Cancer Genetics Markers of Susceptibility (CGEMS) project sponsored by the National Cancer Institute. Because CGEMS used an Illumina platform for genotyping and the genotyping used in this study was performed with an Affymetrix platform, the gene_symbol from the UCSC annotation was used to link with CGEMS top 500 SNP list. Using this method, 1487 SNPs in FHS 100K (including SNPs with MAF < 0.1) correlated with the CGEMS top 500 SNPs related to a known gene.

SNPs were annotated with the UCSC genome browser tables using the May 2004 assembly http://genome.ucsc.edu/[7, 8]. All genes within 60 kb of the top ranked SNPs were identified. The physical location of the SNPs was based on Build 35 of Genome for this report; however, the 100K web browser was based on Build 36.

Results

The cancer phenotypes available in the FHS 100K SNP resource, including details of the sample size, number of cancer events, and covariate adjustment for each trait are listed in Table 1. In this report, we consider only two phenotypes: breast cancer in women (multivariable-adjusted) and prostate cancer in men. Among participants in the 100K sample the mean age at breast cancer diagnosis was 59 years (range 35 to 83 years) in Offspring Cohort women and 70 years (range 35 to 97 years) in Original Cohort women; the mean age at prostate cancer diagnosis was 66 years (range 43 to 85 years) in Offspring Cohort men and 76 years (range 53 to 95 years) in Original Cohort men.

For each of the cancer phenotypes, Table 2 provides the top 15 SNPs ranked in order by lowest p-value for the GEE models and for the FBAT models (all SNP associations can be viewed on the web) [9]. None of the SNP associations achieved genome-wide significance (p < 5 × 10-8) [4]. However, for prostate cancer, the top SNP in GEE models, rs9311171, is in CTDSPL (CTD {carboxy-terminal domain, RNA polymerase II, polypeptide A} small phosphatase-like), a gene that may play a role in tumor suppression [10].

Table 2 Cancer Phenotypes for FHS 100K Project: Results of Association Analyses*
Full size table

There were several additional associations not listed in Table 2 that were of interest. For prostate cancer, in GEE models rs906304 (rank 27, p = 0.000067), is in NCOR2 also known as SMRT. SMRT levels have been reported to be elevated in prostate cancer cells, and result in suppression of anti-proliferative target gene actions for the vitamin D receptor [11, 12]. In FBAT models, for prostate cancer SNP rs255561 (rank 17, p = 0.00039), is near XRCC4, a gene that plays a role in DNA repair and rs1897676 (rank 50, p = 0.0012), is in PTPRD. Protein tyrosine phosphatases are signaling molecules involved in the regulation of a variety of cellular processes including cell growth, differentiation, mitotic cycle, and oncogenic transformation [13]. For breast cancer in GEE models rs4146372 (rank 31, p = 0.00007), is near RAB21, SNP rs9307561 (rank 40, p = 0.0001), is near FAT4, and rs10512849 (rank 46, p = 0.00014), is in FGF10. These genes appear to play biologic roles in a variety of processes including tumor growth and suppression [13, 14]. In FBAT models, for breast cancer rs2836391 (rank 46, p = 0.0012), is in ERG, an oncogene important in the development of prostate cancer [15, 16].

Table 3 All SNP Associations within Selected Breast and Prostate Candidate Genes (up to 60 kb)
Full size table

Our second strategy was to identify from the literature candidate genes implicated in breast and prostate cancer susceptibility (see Additional data file 1). For prostate cancer, we identified 63 candidate genes. Twenty of these candidate genes had from 1 to 20 SNPs on the 100K chip whereas the remaining genes had no SNP coverage on the chip. For breast cancer, 75 potential candidate genes were identified, 28 of these genes had between 1 and 35 SNPs on the 100K chip and the remaining candidate genes were not covered on the chip. Two SNPs in MSR1 (rs9325782, GEE p = 0.008 and rs2410373, FBAT p = 0.021) were associated with prostate cancer and three SNPs in ERBB4 (rs905883 GEE p = 0.0002, rs7564590 GEE p = 0.003, rs7558615 GEE p = 0.0078) were associated with breast cancer (Table 3). For prostate cancer, a region on chromosome 8q24 was recently reported to be associated with prostate cancer risk in Icelandic men and confirmed in three case-control series of men of European ancestry and African American men [5, 6]. There were a total of 64 SNPs on the 100K chip in this 8q24 region (128 to 129.3 Mb interval). However, the reported risk SNP, rs1447295, was not included on the 100K chip and none of the 64 available SNPs were in linkage disequilibrium with the risk SNP. Five other SNPs in this region were associated with prostate cancer with a GEE or FBAT p-value < 0.01 (Table 4).

Table 4 SNPs in the Chromosome 8q24 region Associated with Prostate Cancer:GEE or FBAT p-value < 0.01
Full size table

The National Cancer Institute commenced the CGEMS [17] initiative to conduct genome-wide association studies to identify genetic factors related to prostate and breast cancer. We examined overlap between the top 500 ranked SNPs for prostate cancer in CGEMS phase 1a [18] and the results of the FHS 100K GWAS analysis for prostate cancer. The physical position of the SNP was used to detect overlapping associations and the results are shown in Table 5. Of note, many of the associations in Table 5 are in SNPs with very low minor allele frequencies and the results are presented according to minor allele frequency. WWOX gene, a tumor suppressor gene, that has been reported to play a role in prostate cancer [19], showed evidence of association (rs3751832, p = 0.0009) in our study sample.

Table 5 Prostate Cancer SNP Associations Common to Both CGEMS Top 500 Ranked SNPs and FHS 100K SNPs
Full size table

Discussion

Breast and prostate cancer are the two most frequently diagnosed cancers in the United States and result in substantial morbidity and mortality [1]. A number of breast and prostate cancer susceptibility genes and chromosomal regions have been identified [2, 3, 5, 20–34]. However, currently known genes account for only a fraction of the familial aggregation of breast cancer [25] and few prostate cancer susceptibility genes have even been identified. Risk for these cancers is likely mediated through variation in many genes, each conferring a relatively small risk for the disease. Genome-wide association studies provide an opportunity to discover novel genes and pathways that play a causal role in cancer occurrence and in turn may lead to new therapies for the prevention and treatment of cancer. Finding genetic associations with breast and prostate cancer risk that are robust across multiple studies may facilitate the identification of high risk individuals who can be targeted for early screening and preventive interventions.

We report GWAS results for breast cancer and prostate cancer phenotypes in a community-based sample of adults from two generations of the same families. Although none of the SNP associations achieved genome-wide significance in GEE or FBAT models, this resource has the potential to detect novel cancer susceptibility genes and to explore the relevance of promising candidate gene associations to human cancer. Our results can be compared to those from other genome-wide association studies such as the National Cancer Institute's CGEMS [17]. Although the two studies used different genotyping platforms limiting overlap in the SNPs examined, we were able to determine the physical position of the SNPs. Using this strategy, SNPs in the ERRB4 gene (CGEMS DSSNP_ID rs2371438 and FHS 100K SNP rs10497958) were associated with prostate cancer. ErbB proteins are widely expressed in prostate cells [35] and may play a role in tumor development, growth and progression in human prostate cancer [36, 37]. We also examined the 8q24 region previously associated with prostate cancer risk. CGEMS investigators recently reported a second independent risk SNP (rs6983267) within the 8q24 strongly associated with prostate cancer [34]. The Affymetrix 100K GeneChip did not include either of the previously reported risk SNPs; however, we did identify five other SNPs in this region associated with prostate cancer. The underlying biologic mechanism mediating prostate cancer risk associated with the SNPs and chromosomal region remains unknown. A two-stage approach, genome-wide association followed by selective genotyping of SNPs with suggestive evidence of association, may provide an efficient strategy for pursuing initial genome-wide results [2, 38, 39].

Several important limitations merit comment. First, this study used cancer cases identified through surveillance of a multigenerational community-based sample. The enrollment and examination of Original Cohort and Offspring Cohort participants began years before DNA collection occurred. Thus, a survival bias may have been introduced. Our cases may be comprised of early-staged and less lethal cancers. To address this potential bias, we adjusted for covariates using the full Framingham sample, and used the residual traits for the subset of individuals genotyped using the 100K Affymetrix GeneChip to test for association with the SNPs in linear regression models. Residual traits from Cox models typically are not ideally distributed for linear regression models, but our adjustment method using the full Framingham sample precludes the testing of SNP associations with cancer traits using Cox models. Second, we had a small number of cancer events (250 all cancer cases, 58 breast cancer cases and 59 prostate cancer cases) limiting our ability to detect SNP associations. In a recent small GWAS of age-related macular degeneration that included 96 cases and 50 controls, an association with the CFH gene was identified [40] and confirmed in larger studies [41–43]. However, in that report, individuals homozygous for the CFH risk allele had a sevenfold increased likelihood of age-related macular degeneration [40]. It is very unlikely that common genetic variants for cancer phenotypes will confer a risk for cancer susceptibility of that magnitude. For example, the odds ratio associated with the risk marker identified for prostate cancer in region 8q24 was 1.72 in the combined Icelandic sample [5]. Furthermore, the associations between prostate cancer and the SNPs with low minor allele frequency (Table 5) are likely to be false positive associations given the small number of prostate cancer cases in our sample. Third, the 100K Affymetrix GeneChip provides limited coverage of the genome; many of our a priori candidate genes did not have any SNP coverage on the chip and coverage of some candidate genes that were present on the chip was suboptimal. Importantly, the replicated risk SNP, rs1447295, for prostate cancer [5] was not included on the chip. NHLBI has committed funds for a 550 K genome-wide scan on all FHS participants. This will enable us to confirm our initial 100K SNP associations in a larger sample with a greater number of cancer cases and with denser coverage of the genome. We did not examine epistasis or gene-environment interactions which may modify the associations noted in this study. Lastly, most of our associations are likely to be due to chance. Replication studies are needed to determine if any of the results we report are indicative of true associations. It is important that our data be used in conjunction with data from other samples given the high probability of false positive associations.

Conclusion

In summary, the untargeted genome-wide approach to detect genetic associations for cancer traits provides an opportunity to identify novel biologic pathways related to cancer occurrence and to direct future study of candidate genes that hold the most promise for relevance to cancer risk in humans. Enhancing our understanding of the mechanisms responsible for cancer susceptibility may in turn identify novel strategies for early detection, prevention, and treatment of breast and prostate cancers. These data serve as a resource for replication in other population-based samples.