Potential of marker selection to increase prediction accuracy of genomic selection in soybean (Glycine max L.)

Abstract

Genomic selection is a promising molecular breeding strategy enhancing genetic gain per unit time. The objectives of our study were to (1) explore the prediction accuracy of genomic selection for plant height and yield per plant in soybean [Glycine max (L.) Merr.], (2) discuss the relationship between prediction accuracy and numbers of markers, and (3) evaluate the effect of marker preselection based on different methods on the prediction accuracy. Our study is based on a population of 235 soybean varieties which were evaluated for plant height and yield per plant at multiple locations and genotyped by 5361 single nucleotide polymorphism markers. We applied ridge regression best linear unbiased prediction coupled with fivefold cross-validations and evaluated three strategies of marker preselection. For plant height, marker density and marker preselection procedure impacted prediction accuracy only marginally. In contrast, for grain yield, prediction accuracy based on markers selected with a haplotype block analyses-based approach increased by approximately 4 % compared with random or equidistant marker sampling. Thus, applying marker preselection based on haplotype blocks is an interesting option for a cost-efficient implementation of genomic selection for grain yield in soybean breeding.

Keywords: Genomic selection; Glycine max; Prediction accuracy; Sampling method.

Fig. 1

Decay of linkage disequilibrium (r ²) with physical map distances between markers. The curve was fitted using locally weighted polynomial regression

Fig. 2

Distributions of haplotype block SNPs and SNPs for the 20 soybean chromosomes

Fig. 3

a Histogram of minor allele frequency and b polymorphism information content of 5275 SNPs

Fig. 4

Scatter plots of the first two principal components (PC) for 235 soybean varieties clustered into North Spring soybean (NSs) and Huanghuai Summer soybean (HHSs) subpopulations

Fig. 5

Box-Whisker plots of cross-validated prediction accuracies of plant height and yield per plant, with the method of ridge regression best linear unbiased prediction

Fig. 6

Cross-validated prediction accuracies of ridge regression best linear unbiased prediction based on three marker sampling strategies for plant height (a) and yield per plant (b). Marker subsets were selected using a random sampling (RSM), a haplotype block-based sampling strategy (HBA), and evenly sampling method (ESM)