Comparative analysis of different approaches for dealing with candidate regions in the context of a genome-wide association study

Abstract

Genome-wide association studies (GWAS) test hundreds of thousands of single-nucleotide polymorphisms (SNPs) for association to a trait, treating each marker equally and ignoring prior evidence of association to specific regions. Typically, promising regions are selected for further investigation based on p-values obtained from simple tests of association. However, loci that exert only a weak, low-penetrant role on the trait, producing modest evidence of association, are not detectable in the context of a GWAS. Implementing prior knowledge of association in GWAS could increase power, help distinguish between false and true positives, and identify better sets of SNPs for follow-up studies.Here we performed a GWAS on rheumatoid arthritis (RA) patients and controls (Problem 1, Genetic Analysis Workshop 16). In order to include prior information in the analysis, we applied four methods that distinctively deal with markers in candidate genes in the context of GWAS. SNPs were divided into a random and a candidate subset, then we applied empirical correction by permutation, false-discovery rate, false-positive report probability, and posterior odds of association using different prior probabilities. We repeated the same analyses on two different sets of candidate markers defined on the basis of previously reported association to RA following two different approaches. The four methods showed similar relative behavior when applied to the two sets, with the proportion of candidate SNPs ranked among the top 2,000 varying from 0 to 100%. The use of different prior probabilities changed the stringency of the methods, but not their relative performance.

Figure 1

Top ranked SNPs detected by each method. Venn diagram with the number of candidate SNPs ranked among the top 2,000 markers by each method in Candidate Sets 1 (A) and 2 (B). GWA, standard allelic test; FDR, PSA followed by FDR [8]; EMP, PSA followed by empirical correction; PO1 and PO2, posterior odds [6] with prior probabilities for candidate SNPs of 0.1 and 0.01; FPRP1 and FPRP2, false-positive report probability [7] with prior probabilities for candidate SNPs of 0.1 and 0.01.