Analyses on the influence of the structure of the joint reference population in the German Warmblood horse using different cross-validation approaches for the trait withers height

The accuracy of genomic breeding value estimation depends strongly on the size and genetic composition of the reference population. For the German Warmblood horse, a sufficiently large reference population could be established through the collaboration of five German breeding association. The present study deals with the influence of the structure of the joint reference population on the accuracy of genomic prediction for the trait withers height. For the evaluation, 2,895 mares with genotype information for 60,801 single nucleotide polymorphisms (SNPs) were available. A two-step approach was used and three different (cross-validation) scenarios were tested and compared. For the first cross-validation approach, a classical 5-fold cross-validation with ten replicates was performed using the entire reference population. Since a previous analysis of genomic relatedness revealed substructures reflecting the different breeding foci of the participating breeding associations, in the second cross-validation approach, the reference population was divided by breeding orientation ('jumping' and 'dressage') and the prediction accuracy within the two subsets was investigated. In the third approach, prediction accuracy was estimated when horses of one breeding orientation were predicted using SNP effects estimated using horses of the other breeding orientation. Correlations between conventional breeding values and direct genomic breeding values ranged from 0.63 to 0.71 for the first cross-validation approach, from 0.60 to 0.69 for the second cross-validation approach, and from 0.12 to 0.36 for the third approach. The results show that the use of a reference population across breeds is currently the best strategy for the German warmblood horse. Furthermore, the findings indicate that sufficiently large kinship relations between calibration and validation are of elementary importance and that relatively small sample sizes can be effectively compensated for through ensuring as close genetic relationships between these two groups as possible.