Joint selection of SNPs for improving prediction in genome-wide association studies

It is of great interest to select single-nucleotide polymorphism (SNP) associated with diseases in genome-wide association studies (GWAS). Since genetic variants affect diseases in multiple ways, the joint analysis of SNPs is needed to understand the full effects of genetic variants. However, since the number of SNPs is large and there exists linkage disequilibrium (LD) among SNPs, it is not easy to identify the joint effects of SNPs on complex traits. Thus, the multi-step approach is commonly used for handling these problems. First, SNPs marginally associated with diseases are selected via single SNP analysis. Next, joint identification of putative SNPs via penalized regularization method is carried out for the preselected SNP set. Finally, SNPs from the joint identification step are ordered by a measure which is yielded from the joint analysis. Some current approaches have proposed scoring measures to select causal SNPs such as selection stabilities and effect sizes. In this paper, we discuss some pros and cons of these measures and propose new joint SNP selection measures based on re-sampling methods such as permutation and bootstrap. We illustrate the joint SNP selection based on our measure by using bipolar disorder data from Welcome Trust Case Control Consortium (WTCCC). We demonstrate that the proposed method substantially improves the prediction of disease status compared to other scoring measures.