Partitioning of variance between multiple relationship matrices in BLUP analyses

Abstract

GWAS analyses have resulted in SNP sets that are more predictive for specific traits. Combining these SNPs in a genomic relationship matrix (GRM) with non-selected SNPs may dilute their predictive ability. Instead, predictive SNPs could be treated separately with their own variance. This study examines the partitioning of variance between multiple genetic effects defined by multiple relationship matrices. Univariate REML analyses were performed using GCTA for intramuscular fat (imf), carcase eye muscle depth (cemd), and carcase fat (ccfat) measured on approximately 9.5k genotyped sheep from multiple breeds. Genetic relationship matrices fitted included numerator relationship matrix (NRM) and two GRMs, one based on a standard SNP array (GRMC, 48.5k) and one based on SNPs selected from whole-genome sequence (GRMP, 2.7k). GRMs were constructed with either breed-specific allele frequencies, or population allele frequencies. Breed structure was accommodated by fitting random genetic groups. For GRMs constructed with population allele frequencies, the proportion of genetic variance attributed to GRMC was between 0.14 for ccfat and 0.38 for imf, while for GRMP it was between 0.36 for imf and 0.73 for cemd. The remaining genetic variance was explained by the NRM (range 0.02- 0.38). Similar proportions were observed for the multi-breed GRM. Proportions of genetic variances estimated for the NRM and GRMs can be used in singlestep models to increase prediction accuracy, but questions remain regarding the impact of co-linearity between effects. For example, using a breed-adjusted GRM resulted in an increase in genetic group variance relative to the other genetic effects.