Application of deep learning in crop research: From genomics to phenomics.

Abstract

Deep learning, as a pivotal branch of machine learning, has demonstrated remarkable potential in advancing crop science by effectively integrating genomics and phenomics. This review systematically outlines the application of diverse deep learning architectures-such as convolutional neural networks, recurrent neural networks, and transformers-across key crop genomic tasks, including gene expression prediction, alternative splicing analysis, cis-regulatory element identification, epigenomic profiling, and genome-based trait prediction. In phenomics, these models facilitate high-throughput extraction of crop phenotypic traits from multispectral, unmanned aerial vehicle, and ground-based imagery, supporting yield forecasting, disease diagnosis, and stress response monitoring. We critically evaluate the performance and limitations of each model type across tasks, considering trade-offs between complexity, accuracy, and interpretability, to offer practical guidance for crop researchers. Additionally, the review addresses major challenges in deploying deep learning-such as data scarcity, model transparency, and computational demands-and proposes future pathways to enhance model generalizability, multimodal data integration, and applications in intelligent breeding and sustainable agriculture.