*3.2. Superior Performance of Genomic Prediction Combined with GWAS*

Surprisingly, the GP models built using SNP-52347 generated a lower prediction accuracy than the models with SNP-500QTL (Table 3, Figure S1), regardless of the statistical methods (Figure S1). Similarly, SNP-52347 explained a lower percentage of the phenotypic variation for PS than SNP-500QTL (Table 2). Besides interaction between SNPs, introduction of noise from genome-wide markers [30], the low prediction accuracy may also be owing to some of the erroneously called SNPs and imputation of missing SNP data. SNP-500QTL includes all or nearly all QTL potentially associated with PS; additional markers, not only failed to increase but actually reduced the prediction accuracy, further emphasizing the effectiveness of the QTL identification methodology adopted in our previously published GWAS study [4]. Similar findings were found for FHB in wheat where deoxynivalenol (DON) concentration QTL-linked markers significantly improve prediction accuracy compared to random genome-wide markers [30]. Markers linked to QTL underlying important traits are deemed more useful for prediction strategies because genome-wide markers may introduce noise, thereby reducing accuracy [30]. Using QTL for GP models may be beneficial to balance genetic backgrounds along with maximum gain of breeding value [31]. Genome-wide prediction models based on ~5000 SNPs from de novo GWAS for tropical rice improvement were as effective for prediction as the full marker set of 108,005 SNPs, indicating that the relationship between marker number and prediction accuracy is neither strict nor linear [32]. To sum up, combined applications of the QTL discovered via GWAS and the accelerated breeding cycles through GP facilitate the full use of genome-wide markers in crop disease resistance breeding [10,33]. Removal of redundant markers has the potential to alleviate the effect of the "large *p*, small *n*" issue.
