Detecting PCOS susceptibility loci from genome-wide association studies via iterative trend correlation based feature screening

摘要

Ultrahigh dimensional data with binary response and categorical features has become increasingly prevalent in various fields. Applications using such data exist in genome-wide association studies (GWAS), medical imaging, finance, text mining, among others [ , ]. The most prevailing gene selection approaches used in genome-wide association studies consider an association of each of the genetic variant using univariate models (i.e, single-SNP models); however, they evaluate the association of each SNP in isolation from the others and hence ignore combined joint effects of multi-loci [ – ]. As a matter of fact, most complex diseases are reported to be mediated through multiple genetic variants, each conferring a small or moderate effect with low penetrance, which obscures the individual significance of each variant [ , ]. Furthermore, 70−−80 of genomes showing regions of high linkage disequilibrium (LD), which is the nonrandom association of alleles at nearby loci [ – ]. Malo et al. (2008) claimed that single-SNP approaches failed to differentiate truly influential SNPs from spurious SNPs that were merely in LD with the influential SNPs [ ]. Therefore, although widely used in GWAS data analyses for its simplicity, single-SNP models have limited power and yield both high false-positive and false-negative results [ – ].