Artur Oliveira Rocha
Genomic selection offers opportunities to accelerate genetic improvement in U.S. sheep populations, but its implementation depends on reliable phenotypes recorded in large scale, accurate and complete pedigrees, up to date (co)variance components and genetic parameters, representative genotyping, appropriate statistical models, and routine genomic validations. The overall objective of this dissertation was to evaluate genotyping strategies, genetic parameter estimation, statistical model refinement, genomic prediction strategies, and genomic validation methods to support modernization of the National Sheep Improvement Program (NSIP) genetic evaluations, using simulated U.S.-like sheep populations and industry-relevant NSIP datasets from Suffolk and Katahdin sheep representing terminal-sire and maternal breeding objectives. A review of U.S. sheep genetic evaluation identified major limitations for genomic selection, including small and dispersed flocks, incomplete pedigree, limited genotyping, outdated model assumptions, and the need for validation procedures that reflect real selection decisions. A simulated composite sheep population resembling U.S. Katahdin structure was then used to evaluate Single-step Genomic Best Linear Unbiased Prediction (ssGBLUP) under different genotyping and pedigree-error scenarios, including traits with heritabilities of 0.35 and 0.10, sex-specific genotyping proportions from 0 to 100%, pedigree error rates from 0 to 20%, and random, estimated breeding value (EBV)-based, or phenotype-based genotyping strategies. Random genotyping improved genomically enhanced estimated breeding value (GEBV) accuracy by up to 19% relative to selective strategies, whereas pedigree errors reduced accuracy and increased bias and dispersion, with missing pedigree having a larger effect on these validation metrics than misidentified sires. Genetic parameters and genetic trends were also estimated for U.S. Suffolk sheep, where positive genetic trends were observed for the main routinely evaluated traits, but declining maternal EBV trends for early body-weight traits indicated that maternal genetic effects should continue to be monitored, even in a terminal-sire breed. Updated genetic parameters and evaluation models for body weight, carcass quality, and reproduction traits were then obtained using a pedigree of 75,271 animals and 1,875 genotyped animals with 42,072 single nucleotide polymorphisms (SNP) after quality control. Direct heritabilities were moderate for growth traits (0.26 ± 0.019 – 0.44 ± 0.035) and carcass traits (0.35 ± 0.025 – 0.41 ± 0.026), but low for reproductive traits (0.05 ± 0.008 – 0.08 ± 0.008). The accuracy of prediction was slightly improved when incorporating genomic information using ssGBLUP, with small increases in additive relationships across-flocks. As more genomic information is collected in Suffolk sheep, those benefits will undoubtedly increase. Redefinition of the models fitted, coupled with the updated parameter estimates, had the greatest consequence on enhancing the genetic evaluation, and on rankings of EBV as compared to ssGBLUP. The Linear Regression (LR) method was evaluated across 6,144 simulated scenarios and in U.S. Katahdin data for weaning weight and post-weaning fecal egg count. LR accuracy was the most robust LR statistic, whereas LR bias and dispersion were more sensitive to selective genotyping and pedigree errors. Using whole-population additive genetic variance inflated LR accuracy by approximately 7% for weaning weight and 4% for fecal egg count relative to focal additive genetic variance, supporting the use of focal variance in LR validation. In this context, focal animals are a subset of animals with masked phenotypes. Finally, genomic relationship tuning and unknown parent groups (UPG) were evaluated in U.S. Katahdin sheep using a pedigree of 140,691 animals, 13,366 genotyped animals, and 29,877 SNP. Alternative Fixation Index-based tuning had minimal impact, whereas UPG definitions strongly affected validation metrics and genetic trends; pseudo-count-based joining was the most promising strategy, although no single UPG structure was optimal across traits. Overall, this dissertation showed that genomic selection combined with more refined models, genetic parameters, and validation strategies can improve NSIP genetic evaluations. The benefits are expected to be larger when implemented as part of a broader system that combines large-scale phenotypic records, representative genotyping of a large proportion of individuals, updated genetic models, appropriate treatment of incomplete pedigree, and rigorous genomic validations.