Framework for crossing selection in a two-part strategy breeding program

Genomic selection has revolutionized breeding programs by enabling shorter cycles. Among the architectures that exploit this potential, the two-part strategy system stands out. In this system, a population improvement module applies rapid recurrent genomic selection to raise the genetic mean, while a product development module converts that gain into commercial lines. Recent trials and simulations have demonstrated that these strategies can increase the annual gain by up to six-fold compared to conventional phenotypic selection. However, their success depends on maintaining predictive accuracy and preserving genetic variance across cycles. Genetic drift, recombination, and the decay of linkage disequilibrium gradually separate the selection population from the training set, reducing the correlation between predicted and realized genomic values. At the same time, intense selection accelerates allele fixation, decreasing variance and increasing inbreeding. Another challenge lies in additive complementarity between parents: crosses involving genotypes that individually carry few favorable alleles may still produce superior offspring if those alleles combine appropriately. To reconcile immediate gain with the maintenance of diversity, this study proposes a framework that performs crossing selection to explicitly plan matings up to a target generation, from which the lines will be extracted, rather than optimizing only the next cycle. A crossing tree is generated by recombining all individuals once in every generation. Then, an exploration-exploitation method is applied to guide new simulations toward the most promising branches of the tree. Simulations conducted in R and AlphaSimR compared the proposed framework with a two-part program that directs matings solely based on the highest estimated breeding values within each full-sib family, with SNP effects estimated via RR-BLUP. Seven cycles of the breeding programs were simulated, with 30 repetitions. The multi-generational targeting approach resulted in greater cumulative genetic gain (average of 2% in the end of the 7^th cycle), lower loss of variability (average of 5% in the end of the 7^th cycle), and faster stacking of favorable alleles while keeping the number of crosses manageable. Although the method requires higher computational cost, it can reduce operational expenses by decreasing the number of necessary crosses, saving greenhouse space and genotyping budget, while providing flexibility to balance short- and medium-term genetic gains.