ResDeepGS: A Deep Learning-Based Method for Crop Phenotype Prediction

Genomic selection (GS) is a breeding technique that uses genomic markers to predict the genetic value of crops and animals. It is an effective method to accelerate the improvement of plant agronomic traits and alleviate food security issues. Although some traditional breeding methods based on statistics or machine learning can obtain meaningful predictions for some crops, the complex relationship between genotype and phenotype cannot be captured accurately and impacts the prediction performance. In recent years, deep neural networks have shown significant advantages in capturing nonlinear relationships. This study proposes a deep learning-based plant breeding method called ResDeepGS for predicting crop phenotypes. First, an improved multilayer convolutional neural network is employed to learn the features. Next, the features were processed through normalization and flattening layers. Finally, a dense layer with residual structures is integrated for phenotype prediction. Residual structures and dropout strategies were introduced to avoid overfitting and improve the convergence speed. Extensive experiments show that the proposed ResDeepGS has superior performance compared to other state-of-the-art methods such as DeepGS, DNNGP, RF, GBLUP, Bayes lasso, and BayesB. In particular, for the wheat dataset, the prediction accuracy of ResDeepGS improved by 5%–9%.