| dc.description.abstract | Several of the major pig breeding companies have implemented genomic selection during the last years, but still there is put much effort in optimizing the use of genomic information. The Norwegian pig breeding company, Norsvin has implemented GS by using adjusted single step method (ssGBLUP), which utilize genomic relationship coefficient in addition to pedigree relationship. There has not been done any change of the variance components, which is calculated based on pedigree relationship.
The purpose of this paper is to look into consequences of implementation of GS by comparing methods with pedigree- (A), genomic- (G) and combined (H) relationship matrices for estimation of variance components and quality of the methods. There were used three different methods, best linear unbiased prediction (BLUP) which uses A-matrix as variance structure, genomic best linear unbiased prediction (GBLUP) using G-matrix and adjusted single-step (ssGBLUP) using H-matrix.
The methods were tested on four traits: growth (days from 40 to 120kg live weight), feed consumption (kg feed from 40kg to 120 kg live weight), lean meat percentage (percentage meat of carcass) and total born (still born plus live born). The first three characters were recorded at Delta, Norsvins test station for boars, for individuals born between 2011 and 2014. All individuals had both genotype and phenotype. The dataset included 4578, 4635 and 4606 individuals in the phenotype file, 6686, 6829 and 6788 in the genotype file and 12118, 12263 and 12214 in the pedigree file for growth, feed consumption and lean meat percentage, respectively. The pedigree files were seven generations deep for all four traits. Total born were recorded on sows for each farrowing in the period from January 2010 to March 2015. In total there were 129186 records registered for 62106 sows. Number of genotyped sows were only 3030, a marginal portion of the phenotyped animals. The traits were tested by univariate linear animal models.
Results from variance analysis showed no significant variation for the variance components dependent on relationship matrix. Comparison of log likelihood between methods using A- and H-matrices showed marginal better likelihood for the method using H-matrix for all traits except total born, where they were the same. The methods using genomic relationship, GBLUP and ssGBLUP, obtained similar but higher results for predictive ability than the BLUP method (e.g. correlation between predicted and observed phenotype, corr(y ̂,y) for total born, which showed the least difference between BLUP and ssGBLUP, was 0.28 and 0.30 for BLUP and ssGBLUP respectively). Regression coefficients deviated from one for all methods, indicating that the methods were biased.
To conclude, the non-significant results for variance components indicate that there is unnecessary to estimate variance components based on genomic relationship when implementing GS. The choice of method had larger effect when estimating breeding values than estimating variance components as the GBLUP and ssGBLUP obtained better likelihood and predictive ability than BLUP. Regression coefficients showed that all methods were biased. Improved predictive ability for total born implied that adjusted single step method is a convenient way to implement GS, obtaining better predictions of breeding values for both genotyped and non-genotyped animals. | nb_NO |
