Abstract
In genomic selection, the prediction accuracy is heavily influenced by the training set (TS) composition. Currently, two primary strategies for building TS are in use: one involves accumulating historical phenotypic records from multiple years, while the other is the “test-and-shelf” approach. Additionally, studies have suggested that optimizing TS composition using genetic algorithms can improve the accuracy of prediction models. Most breeders operate in open systems, introducing new genetic variability into their populations as needed. However, the impact of elite germplasm introduction in GS models remains unclear. Therefore, we conducted a case study in self-pollinated crops using stochastic simulations to understand the effects of elite germplasm introduction, TS composition, and its optimization in long-term breeding programs. Overall, introducing external elite germplasm reduces the prediction accuracy. In this context, Test and Shelf seem more stable regarding accuracy in dealing with introductions despite the origin and rate, being useful in programs where the introductions come from different sources over the years. Conversely, using historical data, if the introductions come from the same source over the cycles, this negative effect is reduced as long as the cycles and this approach become the best. Thus, it may support public breeding programs in establishing networks of collaborations, where the exchange of germplasm will occur at a pre-defined rate and flow. In either case, the use of algorithms of optimization to trim the genetic variability does not bring a substantial advantage in the medium to long term.