Maintaining the accuracy of genomewide predictions when selection has occurred in the training population

Routine genomewide selection in maize (Zea mays L.) will lead to phenotyping only a subset of the lines in a biparental population between inbreds A and B. If the cross is used as part of the training population for predicting the performance of lines in a future cross, the training population would be a selected rather than a random subset of lines. Our objective was to determine if selection in the training population (i) reduces the response to selection and accuracy of genomewide selection in a biparental (A/B) population, and (ii) increases the genetic similarity of the best lines in the A/B population. A total of 969 biparental maize populations were evaluated at 4 to 12 environments from 2000 to 2008 for grain yield, moisture, and test weight. The parents of the 969 populations were genotyped with 2911 single nucleotide polymorphism (SNP) markers, and marker data were imputed from lower-density screening of the progeny in each biparental cross. Having phenotypic information on only a selected fraction (25%) of the lines significantly reduced the response to selection and predictive ability. However, augmenting the training set with the five poorest lines nearly restored the predictions to their original level of accuracy. Prior selection in the training population did not increase the genetic similarity (calculated from nonimputed SNP data) of the best lines in the A/B population. We concluded that including a small number of the poorest lines in a training population is a practical way to maintain the effectiveness of genomewide selection.