Sparse kernel models provide optimization of training set design for genomic prediction in multiyear wheat breeding...

Area of Science:

• Plant Breeding
• Quantitative Genetics
• Bioinformatics

Background:

• Genomic selection (GS) is crucial for accurate prediction in breeding schemes.
• Multigeneration data can enhance GS accuracy by increasing training set size.
• Genomic best linear unbiased prediction (GBLUP) utilizes kinship for predictions, but can be compromised by complex family structures in multigeneration data.

Purpose of the Study:

• To evaluate the impact of incorporating multigeneration data on genomic prediction accuracy.
• To assess the effectiveness of the sparse selection index (SSI) method, including a trimmed version, for optimizing training data in GS.
• To compare the prediction accuracy and stability of GBLUP and SSI using a large wheat dataset.

Main Methods:

• Utilized an 8-year multigeneration wheat grain yield dataset (n = 68,836).
• Implemented and compared GBLUP with traditional and trimmed sparse selection index (SSI) methods for training set optimization.
• Analyzed prediction accuracy gains by increasing historical data in the training set.

Main Results:

• Prediction accuracy increased with the inclusion of more historical data (e.g., ~0.05 improvement in GBLUP accuracy using 5 years vs. 1 year).
• The SSI method demonstrated a slight accuracy gain over GBLUP while significantly reducing the training set size.
• Trimmed SSI resulted in training sets with a balanced representation across generations, suggesting improved stability.

Conclusions:

• Incorporating multigeneration data enhances genomic prediction accuracy in wheat breeding.
• The trimmed SSI method offers an efficient approach to optimize training data for GS, improving prediction stability.
• SSI provides a more robust genotype ranking compared to GBLUP, especially with larger training sets.