In order to have successful application of genomic selection, reference population and marker density should be chosen properly. This study purpose was to investigate the accuracy of genomic estimated breeding values in terms of low (5K), intermediate (50K) and high (777K) densities in the simulated populations, when different scenarios were applied about the reference populations selecting. After simulating the historical (undergoing drift and mutation) and recent (undergoing selection) population structures, 800 individuals were remained in reference population. Three scenarios were considered for reducing the reference population number including: 1) 400 individuals which had the highest relationships with the validation set, 2) 400 individuals which had the highest inbreeding, and 3) 400 selected individuals by random. The genomic breeding values were predicted for traits with two heritability levels (0.25 and 0.5) using best linear unbiased prediction (BLUP) with different markers and pedigree information combinations of included pedigree-based BLUP (ABLUP), which was used a numerator relationships matrix (A) only, genomic best linear unbiased prediction (GBLUP) which was used a genomic relationship matrix (G) only, and BLUP|GA, which combined both A and G by using a weight parameter (lambda). By increasing lambda, the prediction model was changed from GBLUP (lambda=0) to ABLUP (lambda=1). The results indicated that without considering the panel density effects, G matrix (lambda=0) and A matrix (lambda=1) usages had the highest and lowest prediction accuracy, respectively. Comparative analyses of different scenarios of reference population selection revealed that all individuals' inclusion in reference population yielded the highest estimation accuracy for breeding values (P<0.05). On the contrary, using reduced single nucleotide polymorphism (SNP) panels considerably decreased the accuracy of breeding value prediction. Individuals selecting in the reference set with a high genetic relationship to target animals, considerably improved the reduction in genomic prediction accuracy because of small reference population size.
