Invited review: efficient computation strategies in genomic selection

The purpose of this study is review and evaluation of computing methods used in genomic selection for animal breeding. Commonly used models include SNP BLUP with extensions (BayesA, etc), genomic BLUP (GBLUP) and single-step GBLUP (ssGBLUP). These models are applied for genomewide association studies (GWAS), genomic prediction and parameter estimation. Solving methods include finite Cholesky decomposition possibly with a sparse implementation, and iterative Gauss-Seidel (GS) or preconditioned conjugate gradient (PCG), the last two methods possibly with iteration on data. Details are provided that can drastically decrease some computations. For SNP BLUP especially with sampling and large number of SNP, the only choice is GS with iteration on data and adjustment of residuals. If only solutions are required, PCG by iteration on data is a clear choice. A genomic relationship matrix (GRM) has limited dimensionality due to small effective population size, resulting in infinite number of generalized inverses of GRM for large genotyped populations. A specific inverse called APY requires only a small fraction of GRM, is sparse and can be computed and stored at a low cost for millions of animals. With APY inverse and PCG iteration, GBLUP and ssGBLUP can be applied to any population. Both tools can be applied to GWAS. When the system of equations is sparse but contains dense blocks, a recently developed package for sparse Cholesky decomposition and sparse inversion called YAMS has greatly improved performance over packages where such blocks were treated as sparse. With YAMS, GREML and possibly single-step GREML can be applied to populations with > 50 000 genotyped animals. From a computational perspective, genomic selection is becoming a mature methodology.