Comparison of Genomic Prediction Models for General Combining Ability in Early Stages of Hybrid Breeding Programs
Parent Selection; Maize; Genomic Estimated Breeding Values
Genomic prediction studies in maize have primarily focus on prediction of hybrids in late stages of the breeding program and have largely ignored selection of inbred lines for use as parents in subsequent breeding cycles. This study evaluated the performance of genomic prediction models for selection of parents in subsequent breeding cycles using estimated general combining ability. Five genomic prediction models were evaluated under different SNP marker densities or with true QTL genotypes using stochastic simulations of entire maize breeding programs. The simulated maize breeding programs modelled over 20 years of breeding using AlphaSimR. The performance of the genomic prediction models was measured at the double haploid stage by tracking genetic gain for hybrids formed by crossing inbred lines from different heterotic pools. The results showed that the performance of the genomic prediction models depends on the marker density. Under low-density, the models that included pool-specific effects showed higher genetic gain than the models that have a common additive effect for both heterotic pools. In contrast, the performance of the models was similar when high-density markers were used. Using the true QTL genotypes showed the superiority of the models that included dominance effects. For heterosis, under low-density, the genomic prediction models showed a different performance. While using high-density markers and true QTL genotypes, the models that included dominance effects showed higher heterosis than the models that included only additive effects. In conclusion, the performance of the genomic prediction models is dependent on the marker density, genomic prediction accuracy increases with high-density, and the performance of models using SNP markers is different from what was expected using true QTL genotypes.