Revealing the interaction effect on the accuracy of genomic selection between marker density and genetic architecture

Genomic selection (GS) is a powerful tool for improving the performance of economic traits in the selection breeding of animals and plants. In order to drive down the costs of genotyping numerous individuals, many researches have focusing on using the low-density marker panel in the practice of GS, especially in the economic aquatic species with a small industrial scale and lagging selection breeding. The aim of this study was to investigate the interaction effect on GS accuracy between marker density and genetic architecture. Two available genotype data sets (large yellow croaker and common carp) were used to simulate the phenotypes that had varying complexity of genetic architecture. Two parameters, number of quantitative trait loci (nQTL) and heritability (h2) were used to control the complexity of genetic architecture in the process of simulation. Three different strategies were used to generate marker panels with different density. First was randomly selection (“random”) from full-size SNP marker panel and second was evenly selection (“even”). The third strategy (“gwas”) was that pre-selection of SNPs with smallest P-value that obtained through GWAS. In total, 52 scenarios were simulated in the present study. A strong interaction between marker density and genetic architecture was observed when using “gwas” strategy, which can be described as three different types of predictive curve. “Type I” indicated GS predictability first decreased rapidly and then decreased slowly with increasing marker density. “Type II” indicated an increase following by a decrease. “Type III” indicated a rapidly increase following by a slowly increase. As the complexity of the genetic architecture increases, the predictive curve gradually shifted from “Type I” to “Type II” and then to “Type III”. When using “random” or “even” strategy, the predictive curve is stable at “Type III” without variation at different scenarios of genetic architecture. These results indicated that SNPs with the largest informative value (effect size, P-value, et al.) related to target trait can be used to generate low-density marker panel to achieve the maximum GS predictability when the target trait have a simple genetic architecture. Additionally, a standard procedure was proposed to infer the complexity of genetic architecture for a target trait based on the shapes of predictive curve of GS. The result of our study will better guide the application of low-density marker panel in GS practice.