Optimising bias and accuracy in genomic prediction of breeding values

Abstract

Reference populations used for genomic selection (GS) usually involve highly selected genotyped individuals which may result in biased prediction of genomic estimated breeding values (GEBV). Bias and accuracy of GEBV in animal breeding programs was explored for various prediction methods. The data was simulated to compare Best Linear Unbiased Prediction of breeding values using pedigree based relationships (PBLUP), genomic relationships for genotyped animals only (GBLUP) and a Single Step approach (SSGBLUP), where information on genotyped individuals was used to infer realised relationship among all available genotyped and non-genotyped individuals that were linked through pedigree. In the SSGBLUP, varying weights (α=0.95, 0.50) for the genomic relationship matrix (G) relative to the A-matrix weights (1-α) were applied to construct an H matrix. Different selection and mating designs with various heritabilities (h²) and QTL models were tested to compare the methods. Results showed that the accuracy of the GEBV prediction increased linearly with an increase in the number of animals selected for genotyping in the reference data. For a random mating design with no selection (RR), all prediction methods were unbiased. Prediction bias was evident in GBLUP, when a smaller proportion was more intensely selected for genotyping but bias was smaller when the proportion of selectively genotyped animals was 20% or higher. The SSGBLUP (α=0.95) showed more accuracy compared to GBLUP and there was less bias with selective genotyping. However, PBLUP and SSGBLUP did show some bias with selection and assortative mating, probably due to not fully accounting for allele frequency changes due to selection of QTL with larger effects. This bias was larger in SSGBLUP than in PBLUP, likely due to the G- and A-matrices not being coherently scaled with allele frequency changes. SSGBLUP required lower values of α to decrease bias and increase accuracy of GEBV with selection and positive assortative mating. Models with a higher h² were more accurate and less biased in the prediction, compared to those with a lower h². Results suggest that selective genotyping in a breeding programme can lead to significant bias in prediction of GEBV when only evaluating genotyped individuals. The SSGBLUP method can provide more accurate and less biased estimates but more attention needs to be paid to appropriate scaling of A and G matrices in selected populations.