Adsa Logo White Adsa Title White

Bias of dairy sheep evaluations using BLUP and single-step genomic BLUP with metafounders and unknown parent groups.

A. Legarra




Bias of dairy sheep evaluations using BLUP and single-step genomic BLUP with metafounders and unknown parent groups.
F. L. Macedo1,2, O. F. Christensen3, J. M. Astruc4, I. Aguilar5, Y. Masuda6, A. Legarra*1. 1INRA Toulouse, France, 2UdelaR Montevideo, Uruguay, 3Aarhus University Aarhus, Denmark, 4IDELE Toulouse, France, 5INIA Montevideo, Uruguay, 6University of Georgia Athens, GA.

Bias is a problem in pedigree-based and genomic-based predictions, and it hampers correct selection procedures. Assessing bias for small dairy cattle breeds, sheep, and goat is difficult. Also, there is a plethora of options to integrate Unknown Parent Groups in Single Step GBLUP. In this work we quantify possible biases in predictions for a dairy sheep breed (Manech Tete Rousse). This breed has a selection scheme for milk yield based on performance recording, progeny testing and Artificial Insemination (AI). The data comprises ~35 years, 1,842,295 performance records and 540,999 individuals in pedigree. In pedigree, there are 70% animals with sire and dam known, 15% with missing sire and 15% missing both. We defined 13 Unknown Parent Groups (or Metafounders). 3007 AI males were genotyped with 50k SNP chip. We tested models with and without genomic information (BLUP and SSGBLUP) and using 3 strategies to handle missing pedigree (Unknown Parent Groups (UPG), “Exact” UPG (EUPG), and Metafounders (MF). The Gamma relationship matrix across MF was estimated by GLS from genotypes of rams and showed mild correlation across MF. To quantify bias, we used method LR. We generated “partial” data deleting most recent records at every year from 2005 to 2014. Then we created “whole” data deleting records with cut-off years from 2007 to 2017. Then we compared (G)EBVs from “partial” and (G)EBVs of young rams from “whole” data across several pairs of cutoff dates, resulting in 65 comparisons. All models resulted in some overestimation of the genetic trend of 0.20 — 0.40 genetic standard deviations. As for the slope (over/underdispersion of (G)EBVs) BLUP_MF, BLUP_UPG, SSGBLUP_MF and SSGBLUP_UPG were unbiased (slopes near 1 with s.e. ~0.02 across comparisons) whereas SSGBLUP_EUPG was biased (slope 0.87 with s.e. 0.02). This is probably due to double counting. One particular truncation year (2008) showed bias for all methods (~0.70 for SSGBLUP_MF and ~0.90 for the other methods) and the likely reason was suboptimal collect of young males that particular year.

Keywords: genomic, bias, sheep.