Abstract

Genomic and phenomic prediction (GP and PP, respectively) are innovative methods that allow plant breeders to increase the productivity of crops. Traditional methods for conducting GP and PP typically rely on linear regression models with predefined assumptions and cannot capture the complex relationships between genotypes and phenotypes. Deep learning models are focused on nonlinear algorithms that can potentially capture complex nonadditive effects, including dominance and epistasis, presenting an opportunity for improving GP and PP accuracy. In this study, the predictions of deep learning models (convolutional neural networks: CNN and multilayer perceptron: MLP) were compared those of Bayesian alphabet models (BayesA, BayesB, BayesC pi, Bayesian lasso, and Bayesian ridge regression) using both single nucleotide polymorphisms (for GP) and spectral information (for PP) datasets for eucalyptus trees adapted to arid environments. The deep learning models consistently outperformed the Bayesian models in predicting most traits, with accuracy estimates ranging from 0.13 to 0.80 for the MLP, 0.16-0.82 for the CNN, and 0.08-0.66 for the Bayesian models. Additionally, spectral information significantly enhanced the accuracy in predicting 50 % of the traits, mainly when applied to deep learning models, demonstrating the potential of high-throughput phenotyping techniques combined with deep learning models in the prediction of important agronomic traits. Thus, deep learning models and spectral data should be incorporated as a strategy for tree breeding programs.