Ensemble supervised learning for genomic selection

Abstract

To meet the world's growing food and nutrition demands, agricultural breeders need to grow crops with improved phenotypes and create varieties that allow increased production. Genomic selection enables the breeders to select individuals with improved phenotypes even before growing them. Existing genomic selection is made mostly through statistical methods that do not accurately predict complex non-linear traits. Deep learning and other machine learning methods have been applied, but most of the deep learning methods are not specifically designed for genomic selection. Also, there has been relatively little comparison between different machine learning methods. We propose three ensemble learning methods: i) ensemble support vector regression; ii) ensemble deep convolutional neural networks and iii) random forests; to predict phenotype with high accuracy. We also propose a feature selection strategy that identifies important markers and contributes to improved phenotype prediction. The proposed marker selection strategy is independent of machine learning methods; thus, the markers that are selected remain the same when the machine learning model is changed. We employed our methods to Iranian wheat landraces. The result shows that ensemble learning methods are better than the single machine learning methods with the lowest PCC 0.339 for plant height and the highest PCC 0.747 for grain length. Our models are also robust as they rank both top twenty and bottom twenty individuals well with nDCG@20 ranges from 0.188 to 0.712.