A practical approach for optimised partitioning of genomic relationship across chromosomes

Abstract

In genomic best linear unbiased prediction (GBLUP) models, genotyped markers are used to make a single genomic relationship matrix (GRM) and consequently each marker contributes similarly in explaining the genetic variance of traits. Some new methods incorporate markers effects in genomic prediction by applying different weights to markers in the GRM. These models often show small improvements in accuracy, but sometimes an increase in the bias of prediction. Alternatively, multiple GRMs made from markers located on each chromosome can be fitted in a GBLUP model. So, the chromosomes containing mutations with large effects on a trait can be used to explain more of the genetic variance. However, fitting many GRMs in a model is not always practical. In this study, for analysing final weight in Hereford cattle (2n=60), initially, we ran 30 models with 2 GRMs made from markers located on each chromosome (GRM_chr) and the markers from the remaining chromosomes (GRM_remaining-chrs). We found GRM_chr for chromosome 6 and 20 explained 20% and 23% of the total genetic variance, respectively, but the rest of GRM_chr failed to absorb any variance. Finally, the prediction model with 3 GRMs, GRM_chr for chromosome 6 and 20 and GRM_remaining-chrs, explained 22%, 26% and 52% of genetic variance, respectively, and compared to the model with a GRM made from all markers, log-likelihood was improved significantly (p<0.001). Although, our results show potential in improving the goodness-of-fit of genomic prediction model, further analyses are required to validate the improvement in accuracy of genomic prediction.