Towards a sweetpotato genomic-enabled breeding: optimizing two-stage analysis of multi-environment augmented trials

Area of Science:

• Plant breeding
• Quantitative genetics
• Genomic selection

Background:

• Single-stage (SS) models are efficient for multi-environment genomic selection but can be challenging to implement, particularly for unreplicated observational trials (OTs) common in early-stage breeding.
• Two-stage models offer a practical alternative when SS models are difficult to fit, providing a viable strategy for analyzing complex datasets.

Purpose of the Study:

• To compare the selection and prediction performance of various two-stage genomic selection models against a single-stage benchmark.
• To evaluate the suitability of deregressed best linear unbiased predictions (dBLUPs) and pedigree-based dBLUPs (dABLUPs) as inputs for second-stage models in unreplicated trials.
• To assess the benefit of pool-specific genomic prediction models versus models trained on the complete dataset.

Main Methods:

• Evaluated 1,138 sweetpotato clones across six OTs using partial diallel crosses.
• Compared different two-stage strategies (using BLUEs, dBLUPs, or dABLUPs as entries) against a single-stage model.
• Investigated the impact of diagonal versus full weight matrices in weighted second-stage models.
• Assessed pool-specific prediction models against models using the entire dataset.

Main Results:

• Differences in selection among second-stage models were minimal, with a slight advantage for models using dABLUPs and a full weight matrix.
• The choice of weighting scheme significantly impacted prediction performance, more so than the choice of entry.
• For pool-specific predictions, dABLUPs demonstrated superior performance compared to other entries.
• When using the complete dataset, differences between entries were marginal.

Conclusions:

• Two-stage models utilizing dABLUPs combined with a full weight matrix are recommended for analyzing augmented trials, offering a practical and effective approach.
• These methods provide genomic predictions and selections comparable to single-stage models, particularly valuable in scenarios with unreplicated trials.
• The study highlights the importance of appropriate weighting schemes and input data (dABLUPs) for optimizing two-stage genomic selection strategies in plant breeding.