CropARNet: A deep learning framework for crop genomic prediction with attention and residual modules

Abstract

Genomic selection (GS) utilizes genome-wide markers to predict complex traits, thereby enhancing crop breeding efficiency. Recently, deep learning has emerged as a promising approach to improve prediction accuracy in GS. This study introduces CropARNet, a novel deep learning framework for GS that integrates a self-attention mechanism with a deep residual network. We systematically evaluated CropARNet's performance on 53 key agronomic traits across four major crops: rice, maize, cotton, and millet. When benchmarked against established models including GBLUP, DNNGP, XGBoost, and CropFormer, CropARNet ranked first in prediction accuracy for 29 of the 53 traits and consistently placed among the top performers for the remainder. Furthermore, CropARNet can successfully predict phenotypes using transcriptomic data. In summary, CropARNet represents a robust, accurate, and powerful tool for advancing the molecular breeding of complex traits in crops. The CropARNet software and illustrative examples are publicly available for download at: https://github.com/Zhoushuchang-lab/CropARNet.