45169

Accuracy of prediction using Reproducing Kernel Hilbert Spaces regression for carcass growth traits in Nellore cattle

Rafael Espigolan ; Daniel Gustavo Mansan Gordo ; Rafael Lara Tonussi  ; Gerardo Alves Fernandes Júnior ; William Bruno Fernandes de Andrade ; Luis Arthur Loyola Chardulo ; Fernando Baldi ; Lucia Galvão de Albuquerque
Download
Favorite this paper

The objective of this study was to compare the accuracy of genomic predictions obtained with different Reproducing Kernel Hilbert Spaces (RKHS) regressions for rib eye area (REA) and hot carcass weight (HCW) in Nellore cattle. Animals were genotyped using BovineHDBeadChip panel, with 777,962 SNP markers. After quality control, the final dataset included 2,710 animals and 453,431 SNPs for REA and 2,495 animals and 453,920 SNPs for HCW. Contemporary groups (CG) were defined as farm and year of birth, and yearling management group. REA and HCW were adjusted for fixed effects of CG and age of animal at slaughter, using the solutions from a pedigree-based animal model. So, the response variables were the phenotypes adjusted for fixed effects (Y*). Analyzes performed: Polygenic model (AM), Genomic BLUP (GM), RKHS regression with a grid for bandwidth parameter (h) equal to 0.1, 0.2, 0.5, and from 1 to 10 (KMh) and Kernel averaging, with h equal to 0.1 and 10 (KAM). A five-fold cross-validation with ten random replications was carried out and models were compared in terms of their predictive mean squared error (PMSE) and accuracy of prediction (ACC), defined as the correlation between the Y* observed and the Y* predicted divided by the square root of heritability of each trait. For REA and HCW, the KM Gaussian kernel approach with h equal 1 (KM1) showed the highest ACC, 0.39 and 0.34, and lowest PMSE, 45.14 and 260.22, respectively. The ACCs for AM, GM, and KAM were, in this order, 0.24, 0.39 and 0.39 for REA, and 0.16, 0.34 and 0.33 for HCW. The PMSE estimates were 46.97, 45.20 and 45.19 for REA, and 263.92, 260.44 and 260.25 for HCW. The models considering genomic information performed better than the polygenic model and, although the GM, KAM and KM1 approaches presented similar predictions, the KM1 showed the lowest PMSE estimate. Acknowledgements: Sao Paulo Research Foundation (FAPESP) grant #2009/16118-5, #2014/00779-0 and #2015/13084-3.