Optimization of the Number of Marker for Genomic Selection in a Historical Data Set of Southeastern US Winter Wheat

Genomic Selection (GS) is a method for obtaining genomic estimated breeding values (GEBV) of individuals on the basis of a genome-wide marker data. Genotyping by sequencing (GBS) generates high-density genome coverage with thousands of markers at low cost, making the use of this tool attractive for plant breeders. One question that arises is how many markers are necessary to generate accurate predictions of performance of individuals. In order to test the effect of marker numbers on predictive ability, one can evaluate accuracies of different marker data subsets.  The objective of this work was determine how filtering criteria and overall marker numbers effect predictive ability for conducting genomic selection in wheat.  A large GBS marker data set on 334 soft red winter wheat lines from different southeastern USA breeding programs was partitioned based on different filter criteria, including read depth, percentage of missing data and frequency of heterozygous individuals. The effect of the marker subsets on prediction accuracy determined using phenotypic data from the Gulf-Atlantic Wheat Nursery over seven years. Results from 16 different filters yield a different number of markers (1 to 70,317) that is a function of the stringency of the kind of filter used. Stringent filters yield a small number of markers resulting in models with low predictive ability, while reduction in filter stringency increase the predictive ability of the model until reach a plateau around 4,000 markers. That information suggest that breeders can allocate resources related with marker information in different ways.