The challenges of genome-wide interaction studies: lessons to learn from the analysis of HDL blood levels

Abstract

Genome-wide association studies (GWAS) have revealed 74 single nucleotide polymorphisms (SNPs) associated with high-density lipoprotein cholesterol (HDL) blood levels. This study is, to our knowledge, the first genome-wide interaction study (GWIS) to identify SNP×SNP interactions associated with HDL levels. We performed a GWIS in the Rotterdam Study (RS) cohort I (RS-I) using the GLIDE tool which leverages the massively parallel computing power of Graphics Processing Units (GPUs) to perform linear regression on all genome-wide pairs of SNPs. By performing a meta-analysis together with Rotterdam Study cohorts II and III (RS-II and RS-III), we were able to filter 181 interaction terms with a p-value<1 · 10-8 that replicated in the two independent cohorts. We were not able to replicate any of these interaction term in the AGES, ARIC, CHS, ERF, FHS and NFBC-66 cohorts (Ntotal = 30,011) when adjusting for multiple testing. Our GWIS resulted in the consistent finding of a possible interaction between rs774801 in ARMC8 (ENSG00000114098) and rs12442098 in SPATA8 (ENSG00000185594) being associated with HDL levels. However, p-values do not reach the preset Bonferroni correction of the p-values. Our study suggest that even for highly genetically determined traits such as HDL the sample sizes needed to detect SNP×SNP interactions are large and the 2-step filtering approaches do not yield a solution. Here we present our analysis plan and our reservations concerning GWIS.

Figure 1. Flow diagram overview of the analysis plan.

Figure 2. The forest plots for β_int of the four most significant interaction terms after meta-analysis of the replication cohorts: rs2315598-rs2853228 (a), rs6848132-rs7863451 (b), rs3756856-rs11758333 (c) and rs4596126-rs11676467 (d).

Although the analysis in the discovery and the filtering was done with scaled phenotypes, for these forest plots, the HDL levels are not scaled in the Rotterdam Study cohorts.

Figure 3. The smallest detectable effect with the current sample size of 2,996 individuals at 80% (a), 50% (b) and 1% (c) power levels.

Figure 4. The overlap between the interaction terms with p-value<3.03 · 10⁻⁷ after a GWIS with GLIDE in RS-I only and after a GWIS with GLIDE in RS-I, RS-II and RS-III combined.