JBLUP: The Joint Best Linear Unbiased Prediction Using BLUP and GBLUP Solutions

The approach of single-step best linear unbiased prediction utilizes phenotypic, pedigree and genomic information but is difficult to compute for large scale genetic evaluations. To achieve computing efficiency while using all available information for genetic evaluation, we developed the joint best linear unbiased prediction (JBLUP) by combining best linear unbiased prediction (BLUP) and genomic BLUP (GBLUP) as one measure of genetic prediction. For genotyped animals, JBLUP uses existing BLUP and GBLUP solutions. For non-genotyped animals, indirect GBLUP is estimated based on the GBLUP of genotyped animals with phenotypic observations through pedigree relationships and the inverse of the genomic relationship matrix without solving the joint system of equations for both BLUP and GBLUP. Three methods of weighting BLUP and GBLUP were proposed: observed accuracies of BLUP and GBLUP from k-fold validation studies, optimum weighting based on the best prediction accuracy from a range of validation studies, and animal-specific weighting using expected prediction accuracies of BLUP and GBLUP of each animal. The JBLUP with optimum weighting was compared with the single-step approach implemented by ssBLUP using 10-fold validation studies for four phenotypes on 2936 genotyped Duroc pigs with 41,108 SNPs and 14,126-32,033 non-genotyped Duroc pigs. Preliminary results showed that JBLUP had 98.5-101.0% of the ssBLUP accuracies for genotyped animals without phenotypic observations, and 107.4-109.0% of the ssBLUP accuracies for genotyped animals with phenotypic observations. These results indicated that JBLUP could be a computationally competitive alternative for implementing genetic evaluation utilizing all available pedigree, genomic and phenotypic information.