Fast Large-Scale Genomic Prediction Utilizing up to Full Sequence Genomic Data through SVD-Based Models

With genomic data increasing both in terms of individuals and loci, computational costs of genomic prediction are likely to increase substantially, e.g., through direct inversion of enormous genomic relationship matrices (GRM) for use in single-step evaluations, or analysis of sequence data using Bayesian variable selection models. To alleviate these problems, we propose singular value decomposition (SVD) based methods, performing reduced rank approximation of genomic data. SVD is computationally demanding, but can be performed on a reduced core sample of individuals and/or in parallel on different genome segments, making fast computation possible.

After SVD, genomic analysis can be performed either by principal component ridge regression (PCRR) or by a genomic animal model (GBLUP), with GRM defined by the chosen principal components. In the latter case, computation of the exact inverse GRM is fast, e.g., considerably faster than computing the original (non-inverted) GRM. Inverse GRM can now be computed for any number of genotyped individuals and loci, and the elements may be computed in parallel, individually or row-wise. In simulated sequence data, an SVD-based GBLUP model gave virtually identical GEBVs as the original GBLUP model (r = 1.000).

Furthermore, an SVD-based non-iterative BayesC model utilizing PCRR was developed. In simulated sequence data, the original and SVD-based BayesC models performed similarly in terms of accuracy and persistency of GEBVs over generations. The major difference was computing time, with the SVD-based model (after SVD) being equally fast as a simple SNP-BLUP model. The accuracy of both BayesC models declined less than SNP-BLUP/GBLUP over generations.