Unveiling the interactions between design optimization, modeling strategies, and spatial analysis in genomic selection.

key messagePartially-replicated designs are more efficient than fully-replicated ones. They can be further enhanced through optimization and present positive synergy with fully-efficient models and 2D P-splines. In plant breeding, traditional field trials rely on balanced designs with high replication to ensure robust breeding value predictions. However, genomic technologies leveraging genomic relationship matrices have reduced the need for extensive replication, favoring partially-replicated (p-rep) designs. These designs, while effective, pose statistical challenges due to low replication and imbalance. To address these, we conducted simulations using datasets from maize, sunflower, and oat, evaluating interactions between experimental designs and modeling strategies across varying field sizes. Classical and p-rep designs, generated via randomization or optimization, were compared, using a novel, faster implementation of the CDmean optimization criterion developed to overcome computational limits. Modeling strategies included single-stage models, unweighted and fully efficient two-stage analysis, and spatial analysis with blocking structures or high-resolution 2D P-splines. Our findings showed p-rep designs consistently outperformed classical designs in predicting unobserved lines, achieving 1-25% higher accuracies depending on the scenario. Within observed training sets, p-rep designs exhibited slightly lower accuracy than classical designs, but it is compensated with increased selection intensity. Design optimization enhanced accuracy by  1% over randomized p-rep designs, while 2D P-splines improved accuracy by 1-2% when paired with single-stage or fully efficient models.