Aquaculture Reports (Apr 2022)
The impact of genetic relationship between training and validation populations on genomic prediction accuracy in Atlantic salmon
Abstract
The potential of genomic selection (GS) to improve production traits has been widely demonstrated in many aquaculture species. Atlantic salmon breeding programmes typically consist of sibling testing schemes, where traits that cannot be measured on the selection candidates are measured on the candidates’ siblings. While annual testing on close relatives is effective, it is expensive due to high genotyping and phenotyping costs. Accurate prediction of breeding values in distant relatives could significantly reduce the cost of GS. This study aimed to evaluate the impact of decreasing the genomic relationship between the training and validation populations on the accuracy of genomic prediction for two key traits; body weight and resistance to sea lice; and to assess the interaction of genetic relationship with SNP density. Phenotype and genotype data from two year classes of a commercial breeding population of Atlantic salmon were used. The accuracy of genomic predictions were close to zero when the prediction was performed across year class, albeit this may reflect a lack of genetic correlation between the same traits measured in the different year classes. Within a year class, systematically reducing the relatedness between the training and validation populations resulted in decreasing accuracy of genomic prediction; when the training and validation populations were set up to contain no relatives with genomic relationships > 0.3, the accuracies decreased by 44% for sea lice count and by 53% for body weight. Less related training and validation populations also tended to result in highly biased predictions. No clear interaction between decreasing SNP density and relatedness between training and validation population was found. These results confirm the importance of close genetic relationships between training and selection populations in salmon breeding programmes, and suggests that prediction across generations using existing approaches would severely compromise the efficacy of GS.
