Haplotype-based Genomic Prediction of Breeds Not in Training

In this study we characterized breed differences and compared genomic prediction among multi-breed beef cattle populations using haplotype alleles constructed from phased genome-wide SNP genotypes. The Bovine 50K SNP genotypes of 1,872 Angus, 1,341 Charolais, 1,229 Gelbvieh, 3,221 Hereford and 3,252 Simmental animals were phased and imputed within-breed to the density of the 700K Bovine HD panel. Haplotype alleles were constructed within consecutive non-overlapping genomic windows of 1 million-bp (Mbp) for real 50K and imputed HD genotypes, and 100 kilo-bp (Kbp) for imputed HD genotypes. Alleles with a frequency within breed >1% were defined as common.  The 5 breeds shared < 1% of 1Mbp common alleles, but > 9% of 100Kbp common alleles. The first three principal components of haplotypes characterized breed differences better than SNP genotypes. Genomic prediction was performed for birth weight, weaning weight, yearling weight, back fat thickness and ribeye area using deregressed proofs with parental averages added back. Method BayesC was used to estimate SNP effects (SNP model) or effects of common haplotype alleles (haplotype model). Prediction accuracies of the 100Kbp haplotype model were similar or slightly higher than the SNP model when training populations included animals from the same breed as prediction candidates, but were significantly higher for certain traits and breeds when training populations excluded the breed of the prediction candidates. These results were supported by the observation that linkage disequilibrium (LD) was much higher or complete between multi-allelic haplotypes and bi-allelic SNPs across breeds, compared to pair-wise LD between single SNPs. Furthermore, the haplotype model was computationally more efficient than the SNP model because the number of common 100Kbp haplotype alleles was less than half the number of HD SNPs. In conclusion, the haplotype model is a more accurate and efficient alternative to the SNP model for genomic prediction when animals of the same breed as prediction candidates are not available for training.