Impact of the structure of the genomic matrix on accuracy of evaluations for imputed genotypes

Abstract

Genotypes corresponding to 20 different low density chips currently imputed to the 50K illumina chip in the United Kingdom (UK) were created from bulls genotyped with the 50K chip. For each low chip type, 1200 bulls with genotypes were generated and then imputed to 50K. The accuracy of imputation was then computed as the correlation between the imputed genotypes and the original 50K genotypes within 4 classes of bulls: those with high, medium, low and no relationship with bulls in the reference (REF) population consisting of 14,280 bulls. Genomic relationship (G) matrices were computed using both the REF bulls and bulls with imputed genotypes. GBLUP was therefore undertaken with the bulls with imputed genotypes regarded as validation bulls for production traits and somatic cell counts. Four different types of G were used in the GBLUP: computed from genotypes but with 0.01 added to diagonal elements or (ii) with a weight of 0.05, 0.10 or 0.20 given to the A matrix. The impacts of these different G matrices on the genomic accuracies for the validation animals with imputed genotypes were evaluated. In general, the accuracy of imputation increased as validation animals have more relationship with animals in the REF population with accuracy varying from 0.87 to 0.999. Similarly, the accuracy of prediction increased as more polygenic effects were included in the construction of the G matrix for traits analysed. This increase was about 2 to 4 % for milk yield and 1 to 2% for SCC for most low density chips but was highest for chip 3. This could be attributed to more information coming from the A matrix resulting in increased accuracy.