Genomic Prediction with Genotype-by-Environment Interaction Using a Residual Network and Transformer Framework

Genotype, environment, and their interactions collectively determine crop phenotypes. While integrating genotypic and environmental data has the potential to improve genomic prediction performance, most existing methods do not adequately incorporate environmental factors or the actual growth conditions of crops, and their prediction accuracy remains suboptimal. In this study, we propose Resformer, a genotype-by-environment interaction genomic prediction model that combines residual networks with Transformer architecture. The model employs gated convolution and Transformer to capture genetic features, residual network and fully connected layers to extract environmental features, and a cross-attention mechanism to model genotype-by-environment interactions. We evaluated Resformer on the G2F dataset against three traditional machine learning methods and two state-of-the-art genotype-by-environment interaction approaches. Experimental results demonstrate that Resformer consistently achieves better predictive performance, confirming its effectiveness and generalizability for genomic selection in crop breeding.