Accuracy and computational efficiency of genomic selection with high-density SNP and whole-genome sequence data.

Abstract The prediction of complex or quantitative traits from single nucleotide polymorphism (SNP) genotypes has transformed livestock and plant breeding, and is also playing an increasingly important role in prediction of human disease. Genomic predictions are made using a prediction equation derived from regressing the phenotypes of the individuals in a reference population on all available SNP simultaneously. Genomic selection is then selection of animals or plants for breeding based on these genomic predictions. As the rate of genetic gain that can be achieved with genomic selection is proportional to the accuracy of the genomic predictions, a key focus is now to increase the accuracy of genomic predictions. This can be achieved by increasing the size of the reference set, using denser markers, and using appropriate genomic prediction models. A wide range of genomic prediction models have been proposed, some of which use marker selection and either linear or non-linear Bayesian models for regression. The non-linear Bayesian models give higher accuracy of genomic prediction for some traits, particularly as marker density increases, but at the cost of high computational burden. Strategies to improve computational efficiency of the non-linear Bayesian methods are becoming more important, as the ultimate marker density is whole-genome sequence, and this is increasingly affordable in many species. In this paper, we review the performance of alternative models for genomic prediction. Strategies that have been proposed to improve the computational efficiency of implementing these models are evaluated. Finally, we outline what is required to enable genomic prediction from whole-genome sequence data.