Effect of Genetic Architecture and Partitioning of Training Population on GEBVs, SNP Effects and GWAS: A Simulation Study

Background/Objectives: Inconsistency of results in genome-wide association studies (GWAS) has been a challenge for animal breeders and geneticists. Understanding how different training subset configurations influence genomic estimated breeding values (GEBVs) and GWAS is essential for optimizing genomic evaluations. This study aimed to evaluate the impact of training population partitioning and QTL architecture on prediction accuracy, GEBV and SNP-effect correlations, and on the consistency of GWAS. Methods: A simulated population consisting of ten breeding generations was partitioned and evaluated on four training scenarios: animal ID, sex, generations, and generation correct.blocks. Moreover, four distinct genetic architectures were simulated, representing combinations of two QTL counts (100 and 1000) and two effect-size distributions (normal and gamma). Phenotypes were available for 10,000 individuals, which were genotyped for 50,000 SNP markers. Results: Across generation blocks, accuracy increased from earlier to more recent generations. GEBV correlations were consistently higher than SNP-effect correlations across scenarios. Adjacent generation blocks showed stronger correlations than distant blocks. Architectures with 1000 QTL yielded higher accuracy than 100 QTL architectures, while effect distribution had limited influence. Manhattan plots showed stable major QTL peaks across subsets. However, reduced peak magnitudes with more noise signals were observed in smaller training sets. Training population size and genetic distance strongly influenced genomic prediction performance. GEBVs were more stable than individual SNP-effect estimates across training configurations. Conclusions: These findings provide insights for interpreting why GWAS results fluctuate more than breeding values due to limited dimensionality of genomic information.