Designing Genomic Prediction to Increase Plant Breeding Efficiency

Successful plant breeding continues to rely on effective tools. Data mining techniques will provide a great boost to the plant breeding process. Here we report a study of implementing representative data set selection to design genomic prediction. First, we introduced three methods in representative data set selection: PAM from cluster analysis, DivRank from ranking in network, and MaxCD from genetic mating scheme. By treating representative data sets as training sets, we conducted genomic prediction with a maize diallel comprised of 276 hybrids from 24 maize diverse founders of NAM population. Our results indicated that all these three methods outperformed the method of using random samples as training sets with higher and more stable prediction accuracy. We further tested these methods in genomic prediction with a factorial mating scheme and found that accuracies by PAM, DivRank, and MaxCD were higher than random sampling. Analysis using a large wheat population with 2,556 hybrids presented more supporting evidence. The proposed representative data set selection methods provided new insights into design optimization for resource allocation. We expect that research in data mining would generate additional helpful guidelines for plant breeding.