Deep learning methods improve genomic prediction of wheat breeding

Area of Science:

• Plant breeding
• Genomic prediction
• Machine learning
• Deep learning applications

Background:

• Genomic prediction (GP) models are crucial for evaluating unseen phenotypes in plant breeding.
• Deep learning (DL) shows promise for GP, but its application is often limited to small datasets.
• Previous DL models in plant breeding were primarily tested on image data and small-scale experiments.

Purpose of the Study:

• To evaluate the performance of a deep learning (DL) model against the best linear unbiased prediction (GBLUP) model using a moderately large dataset.
• To assess the genomic prediction accuracy of the DL model using five-fold cross-validation and a leave-one-environment-out (LOEO) strategy.
• To explore the potential of DL for predicting traits in plant breeding under more challenging, realistic scenarios.

Main Methods:

• A deep learning (DL) model, an extension of a multi-modal DL approach, was utilized.
• The DL model's performance was compared with the best linear unbiased prediction (GBLUP) model.
• Genomic prediction accuracy was assessed via a five-fold cross-validation and a leave-one-environment-out (LOEO) approach.

Main Results:

• The DL model outperformed the GBLUP model in genomic prediction accuracy for two out of five traits during five-fold cross-validation.
• Similar performance was observed for the remaining traits, indicating the DL model's broad applicability.
• The DL model demonstrated competitive performance when predicting complete environments using the LOEO method.

Conclusions:

• The deep learning model exhibits superior or competitive genomic prediction accuracy compared to GBLUP, particularly with moderately large datasets.
• This study validates the potential of DL models for enhancing trait prediction in plant breeding beyond small-scale or image-centric applications.
• The findings support the use of advanced DL techniques for more accurate and efficient crop improvement strategies.