*3.1. All Detected QTL Used as Markers in Genomic Prediction*

Three sets of QTL markers (SNP-500QTL, SNP-134QTL and SNP-67QTL) and a genome-wide SNP marker set (SNP-52347) were evaluated here. GP models built using SNP-500QTL consistently outperformed models derived with any of the other three marker sets (Table 3, Figure S1), lending credence to the robustness and reliability of the QTL identified using multiple single-locus and multi-locus GWAS statistical methods [4]. Most GWAS aim to detect large-effect QTL, such as the SNP-67QTL set. While potentially useful in MAS, these tend to explain a reduced portion of the phenotypic variation compared to more comprehensive models (Table 2). Consequently, the GP models built with such marker sets have lower GP accuracies. Therefore, using all potential QTL associated with the selective trait to build GP models is advantageous because it greatly improves prediction accuracy. Prediction accuracies of models obtained with SNP-134QTL and SNP-67QTL data sets were comparable (Table 3, Figure S1) and they explained a similar proportion of the phenotypic variation for PS (Table 2), confirming the redundancy or overlap between the two datasets. Removal of redundant QTL from SNP-134QTL to produce SNP-67QTL produced slightly higher accuracy models (Figure 3). Simplifying GP models by removal of redundant and unrelated markers will ease the practical implementation of GP in breeding programs.
