Abstract
To enhance the implementation of genomic selection (GS) in plant breeding, we conducted a comprehensive comparative analysis of deep learning (DL) models and genomic best linear unbiased predictor (GBLUP) methods across 14 real-world datasets derived from diverse plant breeding programs. We evaluated model performance by meticulously tuning hyperparameters specific to each dataset, aiming to maximize predictive accuracy and reliability. Our results demonstrated that DL models effectively captured complex, non-linear genetic patterns, frequently providing superior predictive performance compared to GBLUP, especially in smaller datasets. However, neither method consistently outperformed the other across all evaluated traits and scenarios. The analysis revealed that the success of DL models significantly depended on careful parameter optimization, reinforcing the importance of rigorous model tuning procedures. In the discussion, we emphasize the complementary nature of DL and GBLUP methods, highlighting that the choice between these models should be driven by the specific characteristics of the traits under study and the evaluation metrics prioritized in breeding programs. These insights contribute practical guidelines for selecting and optimizing genomic prediction models to achieve robust outcomes in plant breeding contexts.