The Delta Select channel catfish breeding program at the Warmwater Aquaculture Research Unit (WARU – USDA-ARS), in Stoneville, MS, started selecting animals to increase the harvest weight and carcass yield in 2006. Variance components and breeding value estimation were based on pedigree and phenotypes initially and Genomic selection was initiated in 2015. Thus, our objectives in this study were to estimate the variance components and breeding values for harvest weight and verify the accuracy, dispersion, and bias of breeding values from pedigree and genomic models.
Data were provided by WARU – USDA-ARS, with harvest weight records of 46,450 animals from 2008 to 2021, of which 12,279 were genotyped using 53,976 single nucleotide polymorphisms (SNPs). Variance components were estimated using pedigree REML and single-step genomic REML (ssGREML) algorithms as implemented in the BLUPF90+ program. To validate the estimated breeding values (EBV) and genomic estimated breeding values (GEBV) we estimated the dispersion, bias, and accuracy according to the Linear Regression (LR) method, where the focal group was composed of 1208 animals born in 2021 with genotypes, phenotypes, and complete parent information. In the LR method, (G)EBV of animals from the focal group was estimated based on a full and reduced dataset (i.e., no phenotypes for focal animals). Thus, the reduced dataset had phenotypes up to 2020 and the complete one had phenotypes up to 2021. The EBV and GEBV were predicted using REML-BLUP, REML-ssGBLUP, and ssGREML-ssGBLUP, for reduced and complete datasets using the same relationship matrices according to BLUP and ssGBLUP methodologies. Random effects included the additive direct genetic effect, the family environment effect and residual, and the contemporary group effects (year-pond-sex), and age nested within sex were fixed.
The heritability was lower in REML (0.118) than in ssGREML (0.298), also the additive genetic variance presented the same pattern (6,626 and 17,548 for REML and ssGREML). The residual and common family variances were smaller in ssGREML. REML estimates in ssGBLUP provided the most accurate and least biased GEBV (0.22 and 0.68). The accuracy using REML estimates in ssGBLUP increased by 89% and 24% compared with REML estimates in BLUP (0.36) and ssGREML estimates in ssGBLUP (0.55). The most biased predictions (0.72) and the biggest