Skip to main content
ARS Home » Research » Publications at this Location » Publication #278337

Title: Iterative combination of national phenotype, genotype, pedigree, and foreign information

item Vanraden, Paul

Submitted to: Journal of Dairy Science
Publication Type: Abstract Only
Publication Acceptance Date: 3/5/2012
Publication Date: 6/28/2012
Citation: Van Raden, P.M. 2012. Iterative combination of national phenotype, genotype, pedigree, and foreign information. Journal of Dairy Science. 95(Suppl. 2):446(abstr. 447). 2012.

Interpretive Summary:

Technical Abstract: Single step methods can combine all sources of information into accurate rankings for animals with and without genotypes. Equations that require inverting the genomic relationship matrix G work well with limited numbers of animals, but equivalent models without inversion are needed as numbers increase. An equivalent model that includes extra equations to solve for the added contribution of genomic information was applied to national Jersey data. The extra equations solved for G g = u and A22 a = -u, where A22 contains pedigree relationships for genotyped animals and u contains genomic estimated breeding values (GEBV) from the previous iteration. Solutions for g and a were then added when solving for u. Multi-trait across country evaluations (MACE) were deregressed and inserted as extra records containing foreign information. The methods were tested on U.S. Jersey yield data containing 4.4 million lactation records, 4.1 million animals in the pedigree, 16,852 genotyped animals, and 7,072 bulls with foreign MACE records. Heritability was reduced from 0.35 in official evaluations to 0.23 to mimic the effect of cow adjustments. For genotyped young bulls, single-step evaluations were correlated by .966 to multi-step evaluations. Both had the same reliability when tested using 4 year truncated data to predict deregressed proofs from the last 4 years, but regressions for single-step evaluations were closer to expected values. The weight on ' was reduced to 0.8 in the single step method and polygenic variance was increased to 20% in the multi-step method, both to improve the regressions. Convergence was much slower when the same algorithm was applied to Holstein data, and correlations were poor even after thousands of iterations. The number of Holstein genotypes was 135,724, with 65 million lactation records and 50 million animals in the pedigree. Second order Jacobi iteration was used in this study, but preconditioned conjugate gradient algorithm should be faster. More efficient strategies are needed because algorithms that work well on small or medium sized datasets may not handle very large populations.