119 Awardee Talk: Decoding genomic predictions and genome-wide association in large populations

Data recording is an integral part of breeding programs. Some breed associations and breeding companies keep historical data that may go back to the 1950s. While pruning old data can help decrease computing costs without sacrificing accuracy, this practice is avoided by most livestock industries. Besides phenotypes and pedigree, genomic information has been accumulating very fast, and dealing with that has become the biggest challenge in the genomic selection era. There are currently more than 8.5 million genotyped dairy cattle and 1.5 million Angus cattle in the US. We recently ran genomic evaluations for almost 4 million genotyped animals, 30 million animals in the pedigree, and 45 million phenotypes for five different breeds, and this talk will describe how these large-scale genomic evaluations are possible under single-step GBLUP (ssGBLUP). In ssGBLUP, a genomic relationship matrix (G) is needed, which may take around 116 TB to keep in memory (RAM) for 4 million genotyped animals. Although millions of animals can be genotyped, the dimensionality of the genomic information is limited, meaning there is a limited set of genotyped individuals representing the independent chromosome segments in the population (i.e., the non-redundant information). The dimensionality is around 6 k in pigs and chickens and 15 k in cattle, which are considered the “core” animals. This reduced rank is exploited to reduce computing costs by creating a generalized, sparse inverse of G with an approximately linear cost in computing and storage. In multi-breed evaluations, the dimensionality should be assessed within each breed to maximize the GEBV prediction accuracy. The model-based accuracy of GEBV is also an important piece of information to help selection but cannot be easily obtained in large, genotyped populations. Our recently developed algorithm also considers the dimensionality of genomic information and uses block sparse inversion to approximate prediction error (co)variance (PEC) for genotyped animals. With PEC for core animals, computing PEC for SNP is trivial, and the latter is used to obtain p-values for SNP effects in genome-wide association studies (GWAS). Using large amounts of information in GWAS may improve the resolution and discovery. Exploiting the limited dimensionality of genomic information helped lift computing limitations for ssGBLUP genomic evaluations and GWAS in large, genotyped populations. Because of that, ssGBLUP became a popular method for genomic selection in livestock. Although this method is quite efficient now, the ever-increasing number of genotyped animals will pose future challenges that will be addressed by constantly changing algorithms and evolving computing power.