Multivariate Genomic Hybrid Prediction with Kernels and Parental Information

Area of Science:

• Plant breeding
• Quantitative genetics
• Genomics

Background:

• Genomic selection (GS) is crucial for predicting hybrid performance and optimizing breeding strategies.
• Leveraging genomic information enhances decision-making and success rates in hybrid breeding programs.
• Current methods can be improved by integrating parental information to boost genomic prediction accuracy.

Purpose of the Study:

• To explore the impact of incorporating parental phenotypic information as covariates within a multi-trait framework to improve genomic prediction of hybrid performance.
• To compare two approaches: direct use of parental phenotypic information (Pmean) versus using estimated breeding values (BV).

Main Methods:

• A multi-trait framework was employed to integrate parental information.
• Two approaches were tested: Pmean (direct parental phenotypes) and BV (parental breeding values).
• Prediction performance was evaluated using normalized root mean square error (NRMSE) and compared across approaches and kernel types (linear/nonlinear).

Main Results:

• Both Pmean and BV approaches improved prediction performance, reducing NRMSE by at least 4.24%.
• The Pmean approach, directly using parental phenotypic information, showed a slight but consistent advantage over the BV approach.
• No significant performance gain was observed when using linear versus nonlinear kernels.

Conclusions:

• Integrating parental phenotypic information significantly enhances the prediction performance of genomic selection for hybrids.
• The Pmean approach offers a slightly more effective strategy for improving hybrid prediction compared to using breeding values alone.
• These findings provide empirical evidence supporting the value of parental phenotypic data in optimizing hybrid breeding programs.