Cow selective genotyping strategies for genomic selection programs in some simulated dairy cattle

Adding genotypic and phenotypic information from cows to the reference population has the potential to increase the accuracy and decrease selection bias of genomic estimated breeding values (GEBV) in dairy cattle populations (Mc-Hugh et al. 2011; Thomasen et al. 2014; Wiggans et al. 2010). However, the cost of genotyping limits its use to an informative subset of animals (Boligon et al. 2012) and hence looking for an optimal strategy is crucial herein. Now a day, selective genotyping strategies (Boligon et al. 2012; Jiménez-Montero et al. 2012) and imputation methods (Calus et al. 2014) are some suggestive tools to reduce genotyping costs. Therefore, the aim of this study was to assess the effect of inclusion of cow genomic data in the reference population for genomic prediction under scenarios encompassing different selection strategies, number of genotyped cows and heritability on accuracy and bias of genomic breeding values when compared to the use of only bulls in the reference population. Material and A dairy cattle population was simulated by mimicking real linkage disequilibrium (LD) extent and population structure using QMSim software (Sargolzaei and Schenkel, 2009). This population mimicked a 4-pathway dairy cattle selection program with the use of artificial insemination (AI) technology, followed by a population with the use of genomic selection (GS). Within the GS reference population, pseudo-phenotypes were simulated for bulls and cows by calculating different predefined accuracies based on de-regressed EBV of Holstein cattle in Canada. That reflected difference between accuracy of EBV from a bull and accuracy of EBV from a cow which is based on his daughters’ records and her own and daughters’ records, respectively. Parameters of population simulation in this study were based on Dehnavi et al (2018). Two traits with a heritability of 0.30 and 0.05 (representing production and functional traits, respectively) were simulated independently. The reference population consisted of 5000 top selected bulls plus either 2500, 5000 or 10000 cows. Cows were selected randomly (R), based on the highest breeding values (H), the most accurate breeding values from top distribution tail (AH), two tails of the distribution of breeding values (TT), or the most accurate breeding values from two distribution tails (ATT) (Fig 1). The simulation was repeated 20 times for each scenario. Genomic predictions were computed using snpBLUP method in SNP1101 software (Sargolzaei, 2014). Pearson correlation, mean square error (MSE) between true breeding value (TBV) and direct genomic value (DGV), and coefficient regression of TBV on DGV were calculated as measurement criteria for accuracy, error and bias of genomic predictions, respectively. The extent of linkage disequilibrium in the simulated AI population showed a similar pattern that observed in North American Holstein cattle (Fig 2). The ascertainment bias was introduced according to the minor allele frequency (MAF) distribution observed in North American Holstein data (Fig 3). The ATT scenario resulted in higher accuracy of GEBV (Fig 4), lower MSE (Fig 5) and bias (Table 1) compared to other scenarios. This scenario led to 0.123 to 0.215 gain in accuracy (∆ accuracy) compared to the use of bulls only for traits with heritability equal to 0.30 and 0.05, respectively (Fig 4). R and TT scenarios followed ATT. Gains in accuracy were 0.117 to 0.204 and 0.113 to 0.196 for scenarios R and TT, respectively (Fig 4). Regardless of EBV accuracy, selecting cows with high breeding values (H, AH) led to the lowest accuracy of GEBV (Fig 4) and the highest MSE (Fig 5) and bias (Table 1). Overall the observed gains in accuracy for the two traits varied from 0.062 to 0.149 and 0.009 to 0.07 for scenarios AH and H, respectively (Fig 4). An increase in the number of cows decreased both MSE (Fig 5) and bias (Table 1). The average regression coefficient of TBV on DGV across all scenarios varied from 1.12 for 2500 cows to 1.08 for 10000 cows for the trait with a heritability of 0.30 (Table 1). These coefficients ranged from 1.16 to 1.14 for 2500 to 10000 cows for the trait with heritability of 0.05 (Table 1). Similar to these results, using Guernsey breed Jenko et al (2017) showed when only half of population was genotyped, genotyping cows with phenotypes in extremes was superior up to 8-10 folds in accuracy for yield traits than genotyping cows at random or genotyping cows with upper tail phenotypes. Genotyping cows with tail phenotypes can cover on average 88% of the difference between the scenario where all the cows were genotyped or only half of them were genotyped at random (Jenko et al. 2017). The present study showed that sampling cows from the most accurate EBVs on extremes seemed to be more informative on all SNPs, both favorite and unfavorite alleles in the population. The inclusion of cows in the reference population increased the accuracy of the genomic predictions across all scenarios and decreased bias of them as well. However, sampling cows from the most accurate EBVs from two distribution tails seemed to be more useful to precisely and accurately predict genomic breeding values of young animals than the other sampling strategies that were investigated in this study.