Optimal Design of Genomic Prediction in Maize Hybrid Breeding

Given a set of n parental lines, generating and phenotyping all n²-n reciprocal hybrids are formidable challenges, when n > 100. Although genomic prediction is expected to identify promising hybrids with phenotyping some of hybrids in the yield trails, optimal designs are urgently needed to allocate available costs and resources efficiently in commercial breeding. Here we propose an optimal design of genomic prediction based on two criteria: connectedness between training and test populations, and diversity within training population. We test this new strategy in a partial diallel design, comprised of 276 crosses from 24 maize diverse founders in nested association mapping. Our results suggest that this design outperforms other existing designs in terms of prediction accuracy and reliability. We also show how these two criteria affect prediction results separately. Our results prove the advantage of integrating these two criteria in genomic prediction. Furthermore, this research provide a general guideline for the design of genomic prediction project in maize hybrid breeding.