The objectives of this study were to describe, using the goat SNP50 BeadChip (Illumina Inc., San Diego, CA), molecular data for the French dairy goat population and compare the effect of using genomic information on breeding value accuracy in different reference populations. Several multi-breed (Alpine and Saanen) reference population sizes, including or excluding female genotypes (from 67 males to 677 males, and 1,985 females), were used. Genomic evaluations were performed using genomic best linear unbiased predictor for milk production traits, somatic cell score, and some udder type traits. At a marker distance of 50kb, the average r(2) (squared correlation coefficient) value of linkage disequilibrium was 0.14, and persistence of linkage disequilibrium as correlation of r-values among Saanen and Alpine breeds was 0.56. Genomic evaluation accuracies obtained from cross validation ranged from 36 to 53%. Biases of these estimations assessed by regression coefficients (from 0.73 to 0.98) of phenotypes on genomic breeding values were higher for traits such as protein yield than for udder type traits. Using the reference population that included all males and females, accuracies of genomic breeding values derived from prediction error variances (model accuracy) obtained for young buck candidates without phenotypes ranged from 52 to 56%. This was lower than the average pedigree-derived breeding value accuracies obtained at birth for these males from the official genetic evaluation (62%). Adding females to the reference population of 677 males improved accuracy by 5 to 9% depending on the trait considered. Gains in model accuracies of genomic breeding values ranged from 1 to 7%, lower than reported in other studies. The gains in breeding value accuracy obtained using genomic information were not as good as expected because of the limited size (at most 677 males and 1,985 females) and the structure of the reference population.

BackgroundAll progeny-tested bucks from the two main French dairy goat breeds (Alpine and Saanen) were genotyped with the Illumina goat SNP50 BeadChip. The reference population consisted of 677 bucks and 148 selection candidates. With the two-step approach based on genomic best linear unbiased prediction (GBLUP), prediction accuracy of candidates did not outperform that of the parental average. We investigated a GBLUP method based on a single-step approach, with or without blending of the two breeds in the reference population.MethodsThree models were used: (1) a multi-breed model, in which Alpine and Saanen breeds were considered as a single breed; (2) a within-breed model, with separate genomic evaluation per breed; and (3) a multiple-trait model, in which a trait in the Alpine was assumed to be correlated to the same trait in the Saanen breed, using three levels of between-breed genetic correlations (ρ): ρ = 0, ρ = 0.99, or estimated ρ. Quality of genomic predictions was assessed on progeny-tested bucks, by cross-validation of the Pearson correlation coefficients for validation accuracy and the regression coefficients of daughter yield deviations (DYD) on genomic breeding values (GEBV). Model-based estimates of average accuracy were calculated on the 148 candidates.ResultsThe genetic correlations between Alpine and Saanen breeds were highest for udder type traits, ranging from 0.45 to 0.76. Pearson correlations with the single-step approach were higher than previously reported with a two-step approach. Correlations between GEBV and DYD were similar for the three models (within-breed, multi-breed and multiple traits). Regression coefficients of DYD on GEBV were greater with the within-breed model and multiple-trait model with ρ = 0.99 than with the other models. The single-step approach improved prediction accuracy of candidates from 22 to 37% for both breeds compared to the two-step method.ConclusionsUsing a single-step approach with GBLUP, prediction accuracy of candidates was greater than that based on parent average of official evaluations and accuracies obtained with a two-step approach. Except for regression coefficients of DYD on GEBV, there were no significant differences between the three models.