Increasing Accuracy of Genomic Selection in Presence of High Density Marker Panels through the Prioritization of Relevant Polymorphisms

An increase in the density of marker panels did not result in a significant increase in the accuracy of genomic selection (GS) using either regression (RM) or variance component (VC) approaches. Increasing the number of variants using a RM approach will increase collinearity and reduce the effects of associated variants which will hamper the ability to prioritize relevant polymorphisms. Using the VC approach, increasing marker density, after a certain threshold, will not improve and could adversely affect the quality of (G) as the relative number of shared polymorphisms between any two individuals decreases. One way to increase the quality of G in the presence of HD panels is to prioritize the variants. The fixation index (F_ST), as a measure of allele frequency variation among sub-populations, could be used as a score to prioritize candidate SNPs. In this study we evaluated the impact of SNP prioritization using F_ST scores on the genetic similarity between individuals and on the accuracy of GS. A trait with heritability of 0.4 was simulated and the phenotypic distribution was divided into three subpopulations (bottom 5%, middle 90%, top 5%). Genomic data consisted of 400K SNP markers distributed on 10 chromosomes to mimic a 1.2 million SNPs marker panel in bovine. Using different quantiles of the F_ST distribution, 500 to 80,000 SNPs were selected. Similar numbers of SNPs were selected randomly for comparison. The matrix G was calculated for each set of selected SNPs based on their F_ST score and randomly selected counterpart. Using all 400K SNPs, 46% of the off-diagonal elements (ODE) were between -0.01 and 0.01, whereas only 16% of ODE were within that range when 20K SNPs were selected based on their F_ST scores for computing G. When 20K SNPs were selected at random, around 33% of the ODE fell within the same range. The number of ODE greater than 0.05 was significantly greater (27%) when G was constructed using F_ST-selected SNPs compared to 5, and 10% when G was calculated using all or randomly selected SNPs. Maximum accuracy (0.82) was achieved when 5 to 10K SNPs are selected based on F_ST scores compared to an accuracy of 0.71 and 0.54 using all or randomly selected SNPs. Accuracy could be improved by maximizing the genomic similarity between individuals by prioritizing relevant markers. A balance between the number of SNPs and the portion of the genetic variance explained is needed to achieve higher accuracy.