ResDeepGS: A deep learning-based method for crop phenotype prediction

Abstract

Genomic selection (GS) is a breeding technique that utilizes genomic markers to predict the genetic potential of crops and animals. This approach holds significant promise for accelerating the improvement of agronomic traits and addressing food security challenges. While traditional breeding methods based on statistical or machine learning techniques have been useful in predicting traits for some crops, they often fail to capture the complex interactions between genotypes and phenotypes. Additionally, these methods struggle to handle large-scale data, limiting their predictive performance. Recent advancements in deep learning offer a promising solution by better capturing nonlinear relationships and gene interactions.

In this study, we propose a novel crop phenotype prediction method, ResDeepGS, which leverages deep learning techniques. The model consists of two main components: the feature selection module and the phenotype prediction module. The feature selection module employs an incremental recursive feature elimination method, combining the strengths of recursive feature elimination and incremental learning to improve both the efficiency and reliability of feature selection. The phenotype prediction module integrates an enhanced multi-layer convolutional neural network with residual structures and dropout strategies to better capture complex relationships in gene data, accelerate convergence, and reduce overfitting. Through extensive experimentation, we demonstrate that ResDeepGS outperforms current state-of-the-art methods on three datasets: wheat, maize, and soybean. Notably, on the wheat dataset, ResDeepGS improved prediction accuracy by 5% to 9%, highlighting its superior performance in genomic selection tasks. These results underscore the robustness and adaptability of ResDeepGS, offering a promising solution for enhancing crop breeding efficiency and addressing future food security challenges.