Two-stage joint selection method to identify candidate markers from genome-wide association studies

Abstract

The interaction among multiple genes and environmental factors can affect an individual's susceptibility to disease. Some genes may not show strong marginal associations when they affect disease risk through interactions with other genes. As a result, these genes may not be identified by single-marker methods that are widely used in genome-wide association studies. To explore this possibility in real data, we carried out a two-stage model selection procedure of joint single-nucleotide polymorphism (SNP) analysis to detect genes associated with rheumatoid arthritis (RA) using Genetic Analysis Workshop 16 genome-wide association study data. In the first stage, the genetic markers were screened through an exhaustive two-dimensional search, through which promising SNP and SNP pairs were identified. Then, LASSO was used to choose putative SNPs from the candidates identified in the first stage. We then use the RA data collected by the Wellcome Trust Case Control Consortium to validate the putative genetic factors. Balancing computational load and statistical power, this method detects joint effects that may fail to emerge from single-marker analysis. Based on our proposed approach, we not only replicated the identification of important RA risk genes, but also found novel genes and their epistatic effects on RA. To our knowledge, this is the first two-dimensional scan based analysis for a real genome-wide association study.

Multistage analysis strategies for genome‐wide association studies: summary of group 3 contributions to Genetic Analysis Workshop 16

“…The roles of variants in the PTPN22 gene on chromosome 1 were also confirmed by several of these studies. The three NARAC studies of Cho et al [2009], Niu et al [2009], and Wu et al [2009] used Stage 1 regression-based analysis (e.g., log-linear or logistic-regression models) to reduce the noise of non-significant SNPs, thereby limiting the number of SNPs analyzed in Stage 2. The results from these three Stage 1 studies contained statistically significant HLA-DRB1 variants, or chromosome 6 variants in regions that are in high LD with HLA-DRB1 variants.…”

“…The 26 windows were on chromosome 6 and were in high LD with HLA-DRB1 . Wu et al [2009] found 395 SNPs with significant marginal effect with Bonferroni p -value < 0.1, most of which were also on chromosome 6p21.…”

“…While the first stage in four of the studies [Cho et al, 2009; Niu et al, 2009; Wang et al, 2009a; Wu et al, 2009] had the same goal, namely to eliminate the noise of non-significant SNPs in preparation for Stage 2, this first step was performed in four different ways. Cho et al [2009] and Niu et al [2009] performed, respectively, a series of single-marker logistic regression and log-linear analyses to screen each SNP separately.…”

“…Wu et al [2009] also used a sequence of three logistic regression models to screen the SNPs: a simple model regressing only on each SNP and, after removing the top 200 SNPs from the single-SNP analysis, two exhaustive searches using all pairs of SNPs, first without and then with a SNP-SNP interaction term. (The top 200 SNPs were removed to prevent a SNP with an extremely high marginal association from forcing significance on the two-locus model.)…”

“…We note that neither LASSO nor the elastic net allow missing data. Accordingly, Cho et al [2009] only used individuals with complete genotypes; Sung et al [2009] imputed the partially missing genotypes using Merlin [Chen and Abecasis, 2007], and Wu et al [2009] imputed missing genotypes using Beagle [Browning and Browning, 2007]. …”