Approximate single step genomic prediction for Norwegian Red cattle

Abstract

The exact single step Genomic Best Linear Unbiased Prediction (ssGBLUP) method has been used for breeding value estimation in the Geno breeding program since 2016. The number of animals with genotype information included in ssGBLUP has increased to over 210,000, making the exact inversion of the genomic relationship matrix computationally demanding. To address this, we tested two alternative approaches on ninety traits used for breeding value evaluation in the Norwegian Red cattle breed. The single step Algorithm for the Proven and Young Genomic Best Linear Unbiased Prediction (ssAPYGBLUP) approach consisted of a core dataset with 16,480 progeny-proven sires and sires of foreign origin, considering a 10% residual polygenic effect. The single step Singular Value Decomposition Genomic Best Linear Unbiased Prediction (ssSVDGBLUP) approach utilized genotypes from 5,186 progeny-proven sires, explaining 90% of genetic variation through chromosome-specific singular values. We compared estimates from these approximate methods to those from ssGBLUP for animals in the pedigree, and young genotyped animals for all the ninety traits. Correlations between ssGBLUP and ssAPYGBLUP estimates ranged from 0.976 to 1.000 for all the individuals in pedigree and from 0.940 to 0.995 for young genotyped individuals. For the ssSVDGBLUP and ssGBLUP approaches, correlations were between 0.971 and 1.000 for animals in the pedigree, and between 0.977 and 0.995 for young genotyped animals. When regressing ssGBLUP estimates to ssAPYGBLUP estimates, the linear regression coefficients were between 0.993 and 1.027 for all animals in the pedigree and between 1.005 and 1.061 for young genotyped animals. For the regression of ssGBLUP estimates to ssSVDGBLUP estimates, the linear regression coefficients were between 0.953 and 1.055 for all animals in the pedigree and between 0.866 and 0.949 for young genotyped animals. This means that predictions for young genotyped animals when using ssSVDGBLUP showed overestimation while predictions from ssAPYGBLUP were slightly underestimated.