The accuracy of genomic prediction by singular value decomposition of the genotype matrix

Abstract

Increasing the order of the genotype matrix by augmenting the number of animals or genetic markers represents a computational challenge for genomic selection. The objective of this study is to evaluate the effect of the number of components in singular value decomposition (SVD) of the genotype matrix on the accuracy of genomic prediction. SVD is a data reduction method that could be useful to save computational time in the prediction of breeding values for genomic selection. Our study combined simulated phenotypes and genetic architectures with real genotypic data from chromosome 1 of Atlantic Salmon. QTL effects were sampled from random normal distributions. We employed principal component ridge regression (PCRR) and principal component-based algorithm for inverting the genomic relationship matrix (PCIG) to obtain predicted breeding values. We used three different heritabilities for the simulated traits. The analysis showed that accuracy increases steeply for the first few principal components until it stabilizes at around 100–200 components and no meaningful gain is obtained with additional components. Accuracies obtained with PCRR using only a fraction of the total number of components were higher than when using genotypes of all SNPs. The peak accuracy varied across replicates when QTL were sampled from a normal distribution. We conclude that SVD, as used in PCRR and PCIG, is a useful tool for genomic selection involving very large genotype matrices because it can save computational effort without compromising the accuracy of genomic prediction.