Data Augmentation Enhances Plant-Genomic-Enabled Predictions.

Area of Science:

• Plant breeding
• Genomics
• Machine learning

Background:

• Genomic selection (GS) is a powerful tool in plant breeding but faces challenges in prediction accuracy.
• Factors influencing GS accuracy necessitate novel approaches for improvement.

Purpose of the Study:

• To investigate the efficacy of data augmentation (DA) in enhancing genomic selection prediction accuracy.
• To evaluate the impact of DA on top-performing plant lines within real-world breeding datasets.

Main Methods:

• Utilized deep neural networks with data augmentation (DA) to generate synthetic data for training.
• Applied DA techniques to 14 real-world plant breeding datasets for genomic selection.
• Compared prediction performance of DA-enhanced models against conventional methods.

Main Results:

• Data augmentation significantly improved prediction accuracy for the top 20% of lines across 14 datasets.
• Average gains in prediction performance for top lines were substantial (108.4% NRMSE, 107.4% MAAPE).
• Worse performance was observed when evaluating the entire testing set, indicating a focus on elite line prediction.

Conclusions:

• Data augmentation is a potent strategy for boosting prediction accuracy in genomic selection, particularly for identifying elite breeding lines.
• Further empirical validation is recommended to solidify the benefits and understand the nuances of DA in GS.
• DA shows promise for optimizing plant breeding programs by improving the identification of superior genotypes.