Optimizing Fully-Efficient Two-Stage Models for Genomic Selection Using Open-Source Software

Genomic prediction requires a training set with phenotypic and genotypic data to calibrate a model, which then predicts genomic estimated breeding values (GEBVs) for non-phenotyped individuals. In plant breeding, these training sets often include large trials across multiple environments, making single-phase model fitting computationally expensive. An alternative is the two-phase approach: first, genotype means are adjusted for spatial variation; then, these adjusted means are used to predict GEBVs. However, unweighted regression in the second phase ignores correlations between estimation errors, while weighted regression, though more efficient for unbalanced designs, can be inadequate when the model oversimplifies reality. This occurs because its residuals are not free to absorb unmodeled effects. To address this, we incorporated estimation errors as a random effect, allowing residuals to absorb other sources of noise. This approach improved model performance and consistency, increasing genetic gain by up to 13.8% over five breeding cycles compared to the standard model.