Evaluation of Gene Effects from Genome-Wide Association Studies Using a Bayesian Semi-Parametric Model with Dirichlet Process Prior in an Indigenous Chinese Pig Breed

GWAS (genome-wide association studies) using SNP data often estimated gene effects based on single bi-allelic SNP loci, assuming that each effect size was either fixed or followed a normal distribution. In reality, however, a causative locus may consisted of multiple-alleles and its genetic variation may not be well captured by a single bi-allelic SNP which itself does not exactly represent the functional mutation but is only a linked marker. In the present study, causative gene effects were estimated using a Bayesian parametric random-effect model (BP-REM) and a Bayesian semi-parametric model with Dirichlet process prior (BSM-DPP), based on either single SNPs or haplotypes defined for a varying number of SNPs, in Ning Xiang pigs (a Chinese indigenous swine breed). The experiment population consisted of 302 Ning Xiang pigs, each genotyped by the GGP (GeneSeek Genomic Profile) PorcineHD chip (i.e., 68,516 SNPs). The mean (standard deviation) of minor allele frequencies of the 68, 516 SNPs were 0.1195 (0.1558). The 95% confidence interval of identity disequilibrium based on these SNPs was estimated to be between 0.0030 and 0.0164. The phenotypes were simulated to be genetically determined by five additive major genes located on distinct chromosomes, each of between 2 to 10 alleles. For simplicity, additive polygenic effects were assumed to have been taken accounted of and hence were not considered. The heritability due to the five major genes were set up to be 5-10%. The results showed that both model performed almost similarly well when estimating the effect of the bi-allele gene, whether fitted on single SNPs or haplotypes, but they both performed poorly with large deviations when estimating the effects of genes with > 2 alleles if fitting on single SNPs. With haplotypes, BSM-DPP performed slightly better than BPRM, particularly for genes with many (e.g., 10) alleles. Additional salient features of BSMD are that it relaxed the normality assumption for effect sizes of haplotypes and estimated their effects in clusters.