Optimizing training sets to identify superior genotypes in hybrid populations.

Area of Science:

• Plant breeding and genetics
• Genomic selection (GS)
• Quantitative genetics

Background:

• Identifying superior hybrids is crucial for commercial plant breeding and large-scale cultivation.
• Genomic selection (GS) uses marker data to predict breeding values, but training set optimization is key for accuracy.
• Existing methods for training set selection require evaluation for hybrid populations.

Purpose of the Study:

• To evaluate and extend training set optimization methods for constructing predictive models in hybrid populations.
• To assess the performance of different methods in identifying top-performing genotypes with high true breeding values (TBVs).
• To provide a flexible framework for optimizing training sets in practical breeding programs.

Main Methods:

• Investigated three training set optimization methods: MSPE_Ridge(v2), CD_mean(v2), and GV_average.
• Developed three evaluation metrics to assess predictive performance for identifying genotypes with the highest TBVs.
• Conducted simulation experiments using real genotype data from wheat, maize, and rice.

Main Results:

• GV_average demonstrated high computational efficiency and generated informative training sets across various sizes.
• GV_average occasionally compromised genomic diversity in small training sets.
• CD_mean(v2) proved a more reliable alternative for maintaining genomic diversity in small training sets.

Conclusions:

• The GV_average method offers computational advantages for training set optimization in genomic selection.
• CD_mean(v2) is recommended when genomic diversity is critical, especially for smaller training sets.
• The proposed framework enhances the accuracy of genomic prediction in hybrid breeding programs.