Comparison of restricted maximum likelihood and Bayesian approaches in estimation of genomic variance components of wool traits in Merino sheep

Accurate estimation of variance components using pedigree and genomic data plays a key role in prediction of breeding values. Since SNP markers in genomic selection are distributed across the genome, they may cover all quantitative traits loci and potentially explain all of genetic variation. In this study, genotype data from Merino sheep, genotyped by 50k Illumina SNP chip were used. Staple length and Fibre diameter traits were studied in this research. To study the association between allele frequency spectrum and captured additive genetic variance, all SNPs were partitioned in five MAF bins with the equal numbers of SNPs. Two statistical models including separate analysis for each category of MAF SNPs or joint analysis of all MAF groups were fitted. The analysis were performed using REML (parametric) and a Bayesian method implemented via Gibbs sampling and RKHS (semi-parametric) model. Using all common SNPs in REML approach, estimates of genomic heritability were 0.72 and 0.48 for Staple length and Fibre diameter, respectively. In Bayesian approach, genomic heritability for mentioned traits were 0.74 and 0.47 respectively. In the separate analysis, estimates of genomic heritability using REML and Bayesian approaches for each MAF class were similar, but in joint analysis estimates of two approaches were different. Overall, when the model is simple both approaches perform similarly while when model is complicated as joint analysis in present study, two approaches work different. Therefore, to determine which approach is more reliable, further research is required