Issues in commercial application of single-step genomic BLUP for genetic evaluation in American Angus

Abstract Text: American Angus Association (AAA) has been using genomic information for genetic evaluations in a multistep approach since 2009. To improve accuracy while simplifying procedures, AAA is transitioning to single-step genomic BLUP (ssGBLUP) in the middle of 2016. Initial tests with ssGBLUP showed an increase in prediction accuracy of 25% for growth traits compared to traditional evaluations. Besides evaluation for growth traits, the goal of this study was to update the full pipeline for genetic evaluation with ssGBLUP methodology. The pipeline includes multi-trait models with linear and categorical traits, maternal effects, multibreed evaluations with external information, and a large number of genotyped animals but most of them with low EBV accuracy. Data included 9.7M animals in the pedigree, 184,354 genotyped animals, and at most 8.2M phenotypes for growth traits, calving ease (categorical), and carcass traits. The first issue during the implementation was the increasing number of genotyped animals. Single-step GBLUP requires the inverse of the genomic relationship matrix (GRM), which had a high computing cost and required around 1Tb of memory for this dataset. The algorithm for proven and young animals (APY) was used to approximate the inverse of the GRM. The number of core animals was set to 15,000, which was calculated as the number of eigenvalues of GRM explaining 99% of the variation. This algorithm reduced the memory usage to 40Gb and required 10% of the computing time while slightly improving the accuracy. Another issue was the increase in computing time for calving ease evaluation, which uses a threshold model, from 12 hours to 4.5 days. Resetting the preconditioned conjugate gradient iteration to solve the mixed model equations every 40 to 200 rounds helped decrease the time to 19 hours. The inclusion of external EBV for Red Angus was required for evaluation of growth traits. We developed software to support genomic and external information, and the implementation of a genomic multibreed model increased the computing time only by 2.5 hours. Current algorithm for approximation of accuracy of genomic EBV (GEBV) was too expensive for > 100,000 animals. A new algorithm was developed that does not require inverse of large GRM and accounts for multiple sources of information while avoiding double-counting. Correlations between accuracy from the new algorithm and true accuracy from PEV were higher than 0.85 for growth traits. Single-step GBLUP can be considered a mature methodology for commercial genomic selection in beef cattle.

Keywords: beef cattle, genomic selection