Artificial intelligence meets genomic selection: comparing deep learning and GBLUP across diverse plant datasets

Area of Science:

• Plant genetics
• Computational biology
• Agricultural science

Background:

• Genomic selection (GS) is crucial for accelerating plant breeding.
• Genomic best linear unbiased predictor (GBLUP) is a standard method for GS.
• Deep learning (DL) offers potential for capturing complex genetic architectures.

Purpose of the Study:

• To compare the predictive performance of DL models and GBLUP in plant breeding.
• To assess the impact of hyperparameter tuning on model accuracy.
• To provide practical guidelines for selecting genomic prediction models.

Main Methods:

• Comparative analysis of DL models and GBLUP across 14 real-world plant breeding datasets.
• Rigorous hyperparameter tuning for each model and dataset.
• Evaluation of predictive accuracy and reliability.

Main Results:

• DL models frequently outperformed GBLUP, particularly with smaller datasets, by capturing non-linear genetic patterns.
• Neither DL nor GBLUP consistently surpassed the other across all traits and scenarios.
• Successful DL model implementation is highly dependent on precise parameter optimization.

Conclusions:

• DL and GBLUP methods are complementary in genomic prediction.
• Model selection should align with specific trait characteristics and breeding program priorities.
• Optimized genomic prediction models are essential for robust plant breeding outcomes.