Model comparison on genomic predictions using high-density markers for different groups of bulls in the Nordic Holstein population

作者：Gao H; Su G^*; Janss L; Zhang Y; Lund M S

来源：Journal of Dairy Science, 2013, 96(7): 4678-4687.

DOI：10.3168/jds.2012-6406

• genomic prediction
• reliability
• high-density marker
• genetic relationship

This study compared genomic predictions based on imputed high-density markers (similar to 777,000) in the Nordic Holstein population using a genomic BLUP (GBLUP) model, 4 Bayesian exponential power models with different shape parameters (0.3, 0.5, 0.8, and 1.0) for the exponential power distribution, and a Bayesian mixture model (a mixture of 4 normal distributions). Direct genomic values (DGV) were estimated for milk yield, fat yield, protein yield, fertility, and mastitis, using deregressed proofs (DRP) as response variable. The validation animals were split into 4 groups according to their genetic relationship with the training population. Group(smgs) had both the sire and the maternal grandsire (MGS), Group(sire) only had the sire, Group(mgs) only had the MGS, and Group(non) had neither the sire nor the MGS in the training population. Reliability of DGV was measured as the squared correlation between DGV and DRP divided by the reliability of DRP for the bulls in validation data set. Unbiasedness of DGV was measured as the regression of DRP on DGV. The results showed that DGV were more accurate and less biased for animals that were more related to the training population. In general, the Bayesian mixture model and the exponential power model with shape parameter of 0.30 led to higher reliability of DGV than did the other models. The differences between reliabilities of DGV from the Bayesian models and the GBLUP model were statistically significant for some traits. We observed a tendency that the superiority of the Bayesian models over the GBLUP model was more profound for the groups having weaker relationships with training population. Averaged over the 5 traits, the Bayesian mixture model improved the reliability of DGV by 2.0 percentage points for Group(smgs), 2.7 percentage points for Group(sire), 3.3 percentage points for Group(mgs), and 4.3 percentage points for Group(non) compared with GBLUP. The results indicated that a Bayesian model with intense shrinkage of the explanatory variable, such as the Bayesian mixture model and the Bayesian exponential power model. with shape parameter of 0.30, can improve genomic predictions using high-density markers.

• 出版日期2013-7
• 单位中国农业大学