Currently Available Versions of Genome-Wide Association Studies Cannot Be Used to Query the Common Haptoglobin Copy Number Variant

We recently reported in the Journal that the common biallelic haptoglobin (Hp) copy number variant (CNV) (rs72294371) is predictive of coronary heart disease (CHD) among patients with elevated glycoslylated hemoglobin (HbA_1c) (1). In 2 independent populations, we observed that participants with both the Hp2-2 genotype (homozygous for presence of CNV) and an HbA_1c ≥6.5% had a 10-fold higher risk of CHD compared with those with at least 1 Hp1 allele and HbA_1c <6.5% (1). Our study used a candidate gene approach, and an accompanying editorial (2) suggested that the recently completed genome wide association (GWA) studies of millions of single nucleotide polymorphisms (SNPs) should identify the same genetic predictor of CHD, to confirm our findings. However, the Hp CNV is not a SNP; it is a 1.7 kb replication, and as detailed in the following text, our investigations and the work of 2 different groups (3,4) show that currently available GWA studies cannot tag Hp CNV or identify the association between this variant and risk of disease.

From our original published investigation of the Hp CNV and risk of CHD in the Nurses’ Health Study, we also had GWA data available on 726 women (~700,000 SNPs genotyped using the Affymetrix Genome-Wide Human 6.0 array, as previously reported [2]). In this sample, several SNPs were significantly associated with the Hp CNV when we conducted a GWA analysis of the Hp CNV as either an ordinal endpoint or coded dichotomously; the smallest p value was ~10⁻⁶⁰, and 48 SNPs were identified with p values of <10⁻¹⁰. We assessed the 3 × 3 genotype frequency tables for the Hp CNV and the SNPs with significant p values. Although none of the SNPs had alleles in perfect correspondence with the Hp CNV alleles, the contingency tables identified a few SNPs with similarities in genotype frequencies as the Hp CNV. For example, all Hp2-2s had the 0 genotype for rs17669033. However, not all participants with the 0 genotype for rs17669033 had the Hp2-2 genotype, and the overall correlation of rs17669033 and Hp CNV was r² = 0.01. Therefore, rs17669033 cannot be used as a surrogate for the Hp CNV’s genotype.

We used PLINK (5) and Haploview (6) software to calculate pairwise r², linkage disequilibrium (LD), and haplotypes between the Hp CNV and all SNPs 20 kb upstream and downstream of the Hp CNV on chromosome 16. We used both the genotyped GWA SNPs, and in addition, included all SNPs from the genome wide imputations to HapMap (7) phase II (MACH software version 1.0.16) and 1,000G. Many SNPs and 1 haplotype were significantly associated with the Hp polymorphism, but the r² values observed were all <0.10, and they could not predict the Hp CNV genotype. We ran stepwise regression models in SAS (version 9.2; SAS Institute, Cary, North Carolina) to investigate whether groups of SNPs could be used to predict the Hp CNV genotype. However, the highest model r² we could achieve was r² = 0.27, when SNPs from 6 different haplotypes were in the model, and this model did not predict the Hp CNV genotype. Furthermore, none of the GWA study groups of SNPs, individual SNPs, or haplotypes replicated our findings for the association between the Hp CNV and risk of CHD. Our data clearly show that currently available GWA studies have a blind spot for the Hp CNV polymorphism.

Rodriguez et al. (3) used several methods to test for the existence of tagging SNP(s) for the Hp CNV. Similar to our results, individual SNPs and haplotypes were in LD with the Hp CNV with D′ = 1, but none had r² values >0.16, and they could not tag the Hp CNV (3). A similar GWA analysis was also performed in the Diabetes Heart Study; when LD was assessed using both D′ and r² between SNPs and the Hp CNV, no SNPs could tag the Hp CNV (4). These studies published these results as secondary findings, which were not included in abstracts and titles that could surface in literature reviews, preventing the misconception that the association between the Hp CNV and CHD could be captured by GWA (8). It is important to understand the limitations of GWA studies with polymorphisms such as deletions and CNVs, and that a direct candidate-gene approach may be necessary for certain non-SNP polymorphisms. Difficulty in tagging non-SNP polymorphisms with GWA studies has been widely encountered in complex genomic regions.

In conclusion, we decisively demonstrate in a large population study of CHD that the Hp CNV polymorphism cannot be identified through the use of SNPs. Individual SNPs and haplotypes can be strongly associated with the Hp CNV without being of any use as a diagnostic predictor of the Hp CNV genotype. Therefore, one cannot use GWA studies to query the role of the Hp genotype in determining the risk of disease. To test whether the interaction we reported between the Hp CNV genotype and HbA_1c on risk of CHD (1) is replicable in other populations, a study design in which the Hp CNV is directly assessed in its association with incident disease is necessary.