Resumen
Genomic Prediction (GP) is a powerful approach for inferring complex phenotypes from genetic markers. GP is critical for improving grain yield, particularly for staple crops such as wheat and rice, which are crucial to feeding the world. While machine learning (ML) models have recently started to be applied in GP, it is often unclear what are the best algorithms and how their results are affected by the feature selection (FS) methods. Here, we compared ML and deep learning (DL) algorithms with classical Bayesian approaches, across a range of different FS methods, for their performance in predicting wheat grain yield (in three datasets). Model performance was generally more affected by the prediction algorithm than the FS method. Among all models, the best performance was obtained for tree-based ML methods (random forests and gradient boosting) and for classical Bayesian methods. However, the latter was prone to fitting problems. This issue was also observed for models developed with features selected by BayesA, the only Bayesian FS method used here. Nonetheless, the three other FS methods led to models with no fitting problem but similar performance. Thus, our results indicate that the choice of prediction algorithm is more important than the choice of FS method for developing highly predictive models. Moreover, we concluded that random forests and gradient boosting algorithms generate highly predictive and robust wheat grain yield GP models.