Exploring Deep Learning for Complex Trait Genomic Prediction in Polyploid Outcrossing Species.

Area of Science:

• Plant breeding
• Genomics
• Machine learning

Background:

• Genomic prediction (GP) uses genome-wide markers to estimate genetic merit, but applications in polyploids are limited.
• Deep learning (DL) offers potential for GP by capturing complex genetic interactions like epistasis, crucial in polyploids.

Purpose of the Study:

• To evaluate the predictive accuracy of linear and DL methods for GP in polyploid strawberry and blueberry.
• To investigate the impact of DL hyperparameters on model performance.

Main Methods:

• Compared linear models and DL techniques (convolutional neural networks) on strawberry and blueberry datasets.
• Analyzed 1,358 allopolyploid strawberry and 1,802 autopolyploid blueberry individuals with SNP markers.
• Optimized DL hyperparameters, focusing on convolutional filters and regularization.

Main Results:

• DL did not outperform linear models for GP, except in cases with significant epistasis.
• Linear Bayesian models were superior for additive architectures, while DL excelled under strong epistasis.
• Bayesian linear models, when accounting for non-linear effects, matched or exceeded DL predictive accuracy.

Conclusions:

• DL does not consistently offer an advantage over linear models for genomic prediction in polyploids.
• Careful hyperparameter tuning is critical for DL performance in GP.
• Flexible Bayesian linear models provide competitive or superior predictive accuracy, especially when incorporating non-linear effects.