Accuracy and bias of genomic prediction for second-generation candidates

Abstract

As a result of intensive genomic selection in Holstein breeds, the distance of selection candidate to genomic reference population has increased. Most genotyped candidates nowadays have no sire with daughters in milk yet at the time of being selected for breeding, they are referred to as second-generation candidates. Genomic model up to now has been optimized for genomic prediction of candidates with their sire or dam in reference population, referred to as first-generation candidates, and has not accounted for the breakdown of linkage disequilibrium from first- to second-generation candidates. To quantify the loss in accuracy and the bias of genomic prediction for the second-generation candidates, a special genomic validation was conducted, based on genotype and phenotype data from the December 2015 genomic evaluation for German Holstein. As a comparison, a regular genomic validation was done based on the same validation bulls by treating them as first-generation candidates. Accuracy of genomic prediction of direct genomic values, shown in observed R² values of Interbull GEBV Test, was significantly lower for second-generation than first-generation candidates, and the decrease in R² values from first- to second-generation ranged from 0.02 to 0.14 with a mean of 0.086 for 37 MACE traits. A similar drop in the R² value was found also in conventional pedigree index. Bias of genomic prediction, expressed as ratio of regression slopes between the two validation scenarios, deviated also from its expectation. Variance of direct genomic values of the second-generation candidates was too high, in relation to that of the first-generation candidates, with an average of the ratio being 0.95 across all the 37 traits. A shrinkage factor for SNP effect estimates was proposed for direct genomic values in order to reduce the over-prediction for the second-generation candidates. By doing so, the same set of SNP effect estimates can be used for differentiated prediction of genomic breeding values for both the first- and second-generation candidates. The genomic model for German Holstein has been optimized for properly predicting genomic breeding values of second-generation candidates and the optimized model was introduced in April 2016.