*2.6. A Case Study of Genomic Prediction*

(Table S2).

*2.6. A Case Study of Genomic Prediction*  To assess GP prediction accuracy, a training-testing partition was generated with random assignment of breeding lines to either training or testing subsets. Considering the different improvement status of accessions in the population (cultivars, breeding lines, landraces or unknown types) and different levels of resistance, we randomly chose 20% of the 370 accessions in the population, that is, 93 accessions (52 cultivars, 21 breeding lines, 3 landraces and 17 unknown types) as validation dataset, that is, a five-fold random cross-validation set. To predict the PS of these 93 accessions, a RR-BLUP model using the SNP-500QTL set and the PS-mean of the remaining 277 accessions as TP set was built to predict PS. The predicted results are shown in Figure 5c and Table S2. The prediction accuracy was as high as 0.95 (*r* between observed and predicted PS). Similarly, a linear regression model of observed PS (*y*) to NPQTL (*x*) of the 277 accessions (the same TP as GP) produced *y* = -0.026*x* + 11.902 (Figure S2), which was similar to the regression equation previously obtained with the complete accession set (Figure 5b). Using this prediction model, predicted PS and intervals were calculated (Figure 5d, Table S2). The prediction accuracy of 0.92 for NPQTL was To assess GP prediction accuracy, a training-testing partition was generated with random assignment of breeding lines to either training or testing subsets. Considering the different improvement status of accessions in the population (cultivars, breeding lines, landraces or unknown types) and different levels of resistance, we randomly chose 20% of the 370 accessions in the population, that is, 93 accessions (52 cultivars, 21 breeding lines, 3 landraces and 17 unknown types) as validation dataset, that is, a five-fold random cross-validation set. To predict the PS of these 93 accessions, a RR-BLUP model using the SNP-500QTL set and the PS-mean of the remaining 277 accessions as TP set was built to predict PS. The predicted results are shown in Figure 5c and Table S2. The prediction accuracy was as high as 0.95 (*r* between observed and predicted PS). Similarly, a linear regression model of observed PS (*y*) to NPQTL (*x*) of the 277 accessions (the same TP as GP) produced *y* = −0.026*x* + 11.902 (Figure S2), which was similar to the regression equation previously obtained with the complete accession set (Figure 5b). Using this prediction model, predicted PS and intervals were calculated (Figure 5d, Table S2). The prediction accuracy of 0.92 for NPQTL was slightly inferior to that of the GP model. The observed PS values all fell within prediction intervals (Table S2).

slightly inferior to that of the GP model. The observed PS values all fell within prediction intervals
