Cross-validation of genetic and genomic predictions of temperament in Nellore–Angus crossbreds

Abstract

The objective of this study was to cross-validate predictions of genomic merit for overall temperament at weaning (TEMP) by alternative partitioning of training and validation sets in a Nellore–Angus crossbred population. Nellore–Angus F2 calves (embryo transfer and natural service reciprocal crosses), paternal half-siblings to the embryo transfer calves, and F3 calves had records for TEMP (subjectively scored on a 1 to 9 scale). Calves with DNA available were genotyped on the Illumina Bovine SNP50 version 1 and 2 assays, and, after quality control filtering, there were 34,913 SNP markers available for use. Calves with TEMP records and genotypes (n=769; mean±std of all TEMP records is 3.975±2.062) were used in this study. BayesB procedures with π̃=0.997 were used to obtain direct genomic values (DGV) with alternative partitioning of data into training and validation sets utilizing the family structure of this population. Training was conducted (scenarios 1–4) using the progeny of all but one of the four sire or grandsires, (5) using only F2 progeny, (6) by random assignment, and (7) using only embryo transfer F2 and natural service half-sibling cattle. In random assignment, the number of animals included in the validation population was the average of scenarios 1–5. The DGV generated in these training and validation sets were compared to traditional, pedigree-based breeding values from an animal model with equivalent training and validation sets. The training model included fixed effects of sex, birth year-season combination, and temperament scoring pen nested in birth year-season combination. Standardized accuracies were higher for DGV using BayesB procedures (0.226 on average) compared to EBV using a traditional pedigree-based animal model (0.122 average). Random allocation of individuals into training and validation groups resulted in the highest accuracies for DGV (0.503) and EBV (0.354) of the validation groups. Overall, accuracies of animals in validation groups were low (i.e., less than 0.35) using genomic predictions, which could be due to inadequate sample size, insufficient marker density, or limited relationship between individuals in the training and validation populations. The genomic method generated unique DGV among full- and half-siblings within families whereas pedigree-based prediction results in identical EBV among full-siblings. This is of particular interest for the cattle industry as genomic methods can provide more plausible estimates of genetic merit for crossbred or purebred cattle without records, even if the initial accuracy is low.