Extending the application of empirical methods for predicting the accuracy of genomic predictions

The accuracy of genomic predictions increases with the size of the reference population. Constructing reference data to underpin genomic selection can be expensive; predicting genomic accuracy is essential when designing effective reference data projects. Several theoretical predictions have been formulated to predict genomic accuracy. However, due to the theoretical effective number of chromosome segments (Me) term being underestimated, theoretical predictions of genomic accuracy may be inflated. Dekkers et al. (2021) proposed an empirical approach that has been demonstrated to be effective for estimating Me and genomic accuracy. Furthermore, when empirical Me estimates are used in theoretical accuracy equations, both empirical and theoretical predictions of genomic accuracy are comparable. However, the empirical method requires existing reference data; therefore, its application is limited when designing reference data projects, especially for new traits. This study aimed to develop a methodology that would enable empirical estimates of Me to inform predictions of genomic accuracy for a wide range of reference data scenarios, regardless of the availability of existing reference datasets. Pedigree, phenotypes and genotypes were obtained from the Australian Brahman BREEDPLAN genetic evaluation. The empirical method was applied using univariate models, with genotyped and phenotyped animals as the reference; reference sizes ranged from N = 1,493 to 15,965. Genomic accuracy was the same for the full empirical method and the theoretical calculations using empirical Me. Genomic accuracy from theoretical calculations with theoretical Me was, on average, 19% higher. Plotting empirical Me/N for each trait against N yielded a cubic curve, demonstrating that Me/N can be predicted for a given reference population size and is not dependent on trait-specific parameters (i.e., heritability). The ability to adequately describe this curve depends on having enough datasets to cover the entire range of reference sizes. To better define the cubic curve, the 400-day live weight reference population was sequentially reduced in size based on year of birth, and empirical Me was estimated and added to the plot. The power function to describe the curve was y=846.53x-0.777. Furthermore, the results showed that although differences in Me (and thus genomic accuracy) due to genomic diversity were observed, the cubic curve was similar across different breeds. Therefore, using another breed's curve may be an option for breeds that have not yet commenced genomic selection. A Brahman-specific plot of expected genomic accuracy versus reference size has been generated to assist in predicting genomic accuracy from reference populations of varying sizes. This paper has extended Dekkers' empirical method to allow for wider application, particularly in designing effective reference data projects, and produced a Brahman-specific plot of expected accuracy for reference size informed by their own breed's data.