The identification of elite individuals is a critical component of most breeding programs. However, the achievement of this goal is limited by the high cost of phenotyping and experimental research. A significant benefit of genomic selection (GS) to plant breeding is the identification of elite individuals without the need for phenotyping. This study aimed to propose different calibration strategies using combinations between generations from different genetic backgrounds to improve the reliability of GS and to investigate the effects of LD in different types of mating systems: outcrossing (An) self-pollination (Sn) and hybridization (Hn). For this purpose, we simulated a genome with 10 linkage groups. In each group, two QTL were simulated. Subsequently, an F2 population was created, followed by four generations of inbreeding (S1 to S4, H1 to H 4, A1, to A4,). Quantitative traits were simulated in three scenarios considering three degrees of dominance (d/a = 0, 0.5 and 1) and two broad sense heritabilities (h2 = 0.30 and 0.70), totaling six genetic architectures. To evaluate prediction reliability, a model (RR-BLUP) was trained in one generation and used to predict the following generations of mating systems. For example, the marker effects estimated in the F2 population were used to estimate the expected genomic breeding value (GEBV) in populations S1 through A4. The squared correlation between the GEBV and the true genetic value were used to measure the reliability of the predictions. Independently of the population used to estimate the marker effect, reliability showed the lowest values in the scenario where d = 1. For any scenario, the use of the multigenerational prediction methodology improved the reliability of GS.

This paper aimed to evaluate the efficiency of subset selection of markers for genome-enabled prediction of genetic values using radial basis function neural networks (RBFNN). For this purpose, an F1 population from hybridization of divergent parents with 500 individuals geno-typed with 1,000 SNP-type markers was simulated. Phenotypic traits were determined by adopting three different gene action models – additive, additive-dominant, and epistasic , com-plying with two dominance situations: partial and complete with quantitative traits admitting heritability (h2) equal to 30 and 60%, each one controlled by 50 loci, considering two alleles per locus, totaling 12 different scenarios. To evaluate the predictive ability of RR_BLUP and the neural networks, a cross-validation procedure with five replicates were trained using 80% of the individuals of the population. Two methods were used: dimensionality reduction and stepwise regression. The square of the correlation between the predicted genomic estimated breeding val-ue (GEBV) and the phenotype value was used to measure predictive reliability. For h2 = 0.3 in the additive scenario, the R2 values were 59% for neural network (RBFNN) and 57% for RR-BLUP, and in the epistatic scenario, R2 values were 50% and 41%, respectively. Additionally, when analyzing the mean-squared error root, the difference in performance between the tech-niques is even greater. For the additive scenario, the estimates were 91 for RR-BLUP and 5 for neural networks and, in the most critical scenario, they were 427 for RR-BLUP and 20 for neu-ral network. The results showed that the use of neural networks and variable selection tech-niques allows capturing epistasis interactions, leading to an improvement in the accuracy of pre-diction of the genetic value and, mainly, to a large reduction of the mean square error, which indicates greater genomic value.