Development of conditional cores via supersaturated designs for genetic diversity-based prediction models

Genetic diversity is fundamental to sustainable crop improvement. This study uses 1755 soybean [Glycine max (L.) Merr.] recombinant inbred lines from the soybean nested association mapping population, phenotyped for yield across nine environments to optimize the composition of the training set for a given testing set. Using supersaturated designs, we generated conditionally optimized training sets by iteratively selecting candidate genotypes according to a metric that quantifies diversity as the trace of the cross-product of the genomic matrix containing both a candidate training set and the fixed testing set. Two strategies were considered: (i) maximizing and (ii) minimizing genetic distances to the testing set. Our results show that maximizing genetic distance relative to the testing set improves prediction accuracy over random training sets under standard genomic best linear unbiased prediction. However, this advantage disappeared when models incorporated family structure or genotype-by-environment interactions, which consistently yielded superior overall accuracy regardless of training set selection. This result highlights the complexity of considering multi-omics integration for interpretability: determining whether an increase in accuracy comes from a better model specification or from capturing confounding factors. Finally, we also found that maximizing the genetic diversity of the training set with respect to the testing set preserves at least one individual from as many different families as possible. In contrast, minimizing diversity either excludes or overrepresents families. Further studies should consider a completely unbalanced dataset that mimics the real-life scenarios breeders face in their programs.