Adsa Logo White Adsa Title White

Accuracy of indirect predictions based on prediction error covariance from single-step genomic BLUP.

D. Lourenco

Events

06-24-2020

Abstract:

362
Accuracy of indirect predictions based on prediction error covariance from single-step genomic BLUP.
D. Lourenco*1, I. Aguilar2, A. Legarra3, A. Garcia1, Y. Masuda1, S. Tsuruta1, I. Misztal1. 1University of Georgia Athens, GA, 2INIA Las Brujas, Canelones, Uruguay, 3INRA Castanet Tolosan, France.

One of the ways to deal with the ever-increasing number of genotyped animals in single-step genomic BLUP (ssGBLUP) evaluations may be to use only genotyped animals with complete information in the official evaluation and compute indirect predictions (IP) for the remaining young genotyped animals. However, if IP are going to be published, there is a need for a measure of accuracy that reflects the standard error of IP. This measure should be similar to the accuracy of GEBV to validate the usefulness of IP. The objective of this study was to implement formulas to compute accuracy of IP based on the prediction error covariance matrix from ssGBLUP. Using field data, complete ssGBLUP evaluations were run with up to 60k genotyped animals. Reduced ssGBLUP evaluations considered genotypes for up to 55k animals. BLUPF90 was used to compute both complete and reduced evaluations. Accuracy of GEBV in the complete evaluation was computed based on PEV, whereas in the reduced evaluation the left-hand side of the mixed model equations was stored. Using POSTGSF90, the submatrix of prediction error covariance (PEC) for GEBV of genotyped animals was extracted and converted to PEC of SNP effects. Using the same software, GEBV were converted to SNP effects. PREDF90 was used to obtain IP and accuracy of IP for up to 5k young validation animals. Accuracy of IP was computed as a function of PEC for SNP effects and genotypes. Tuning and blending parameters were accounted for by the formulas implemented in POSTGSF90 and PREDF90. Correlations between accuracy of GEBV and IP for the 5k validation animals were greater than 0.98, as well as correlations between predictions. These results show that accuracy of IP can be calculated based on ssGBLUP evaluations without the need to run an extra SNP-BLUP evaluation to obtain PEC for SNP effects. Future steps will involve the expansion of this method to large-scale ssGBLUP evaluations, where the algorithm for proven and young is used, likely based on approximations.

Keywords: genomic predictions, reliability.