Comparison of Genomic-Enhanced EPD Systems Using an External Phenotypic Database

The American Angus Association (AAA) is currently evaluating two methods to incorporate genomic information into their genetic evaluation program: 1) multi-trait incorporation of an externally produced molecular breeding value as an indicator trait (MT) and 2) single-step evaluation with an unweighted G matrix (SS). Our objective was to quantify bias and accuracy of genomic predictions using these two approaches. Because phenotypic data was limited for actual carcass measures from genotyped bulls in the AAA database, we tested bias and accuracy using data from the germplasm evaluation program (GPE) at the US Meat Animal Research Center. Traits evaluated included birth, weaning, and yearling weight, maternal weaning weight, carcass weight, marbling score, ribeye area, and backfat thickness. The GPE has sampled 197 AAA bulls over 45 years and has typically characterized weight and carcass data on 8-15 progeny from each bull. Of these 197 bulls, 128 had been genotyped using high density arrays. From the AAA database, EPDs were derived using MT, SS, and non-genomic (NG) animal model methods. To detect differences in accuracy, these EPD were correlated to multi-breed EPD derived from GPE. These correlations were restricted to bulls that had high-density genotypes. In addition, to quantify bias, breed specific EPD regression coefficents were derived from GPE. Independent variables for regression analysis were derived by dropping the EPD from sampled bulls proportionally through the pedigree (1/2 reduction in each generation). These EPD regression coefficients were expected to be one if GPE conditions were similar to conditions in breed association databases. Regression coefficients were obtained for GPE progeny born after 1998 as bulls that produced these progeny were generally genotyped. The NG EPD were used as a benchmark for regression comparison as any bias observed without genomic data was expected to persist with the addition of genomic information. Resulting correlations were very similar for all three methods for weight traits as correlations among MT, SS, and NG EPD were greater than 0.99. For carcass traits, SS and NG EPD were correlated to each other at 0.97 or more while correlations of each with MT were less than 0.93. For all traits, SS correlations were highest with GPE EPD. In regard to bias, SS regressions were always similar to NG regression while MT regression were always lower. Data from outside sources such as GPE can be useful for evaluating alternative genetic prediction models.