PCAGP, a parallel convolutional attention network-based method for crop genomic prediction

Genomic selection (GS) has emerged as a transformative breeding paradigm, leveraging genome-wide marker data to predict phenotypic outcomes and accelerate crop improvement cycles. Traditional statistical models face fundamental limitations in modeling the higher-order, non-linear genotype-phenotype relationships inherent in complex agronomic traits. Deep learning architectures have recently shown remarkable predictive performance in GS applications. Current deep learning approaches in GS predominantly employ a serial structure convolution kernel, this may lead to gradient vanishing and computational efficiency challenges. In this study, we propose a method based on a parallel convolutional attention network for crop genomic prediction (PCAGP), a novel deep learning framework for crop genomic prediction. The architecture uses a convolution kernel to transmit information through a parallel structure, combined with a coordinate attention mechanism that effectively integrates inter-channel relationships and spatial genomic information. This approach significantly improves the accuracy of phenotypic prediction. We compared PCAGP with six widely used GS methods on four benchmark datasets, including statistical method (GBLUP), machine learning approaches (SVR, RF), and deep learning models (DeepGS, DNNGP, and SoyDNGP). The experimental results demonstrate that PCAGP outperforms all baseline methods, and the average prediction accuracy is improved (0.4%-53.1%).