GS-Impute: A neural network framework for accurate imputation of low-density markers in across-population genomic selection.

Genomic selection (GS) holds great promise for accelerating breeding progress in plants, and the advancement of across-population GS is essential to realize its full potential. However, conventional across-population GS heavily relies on precisely aligned markers across dense genotypes, whereas the feasibility of using flexible low-density markers remains underexplored. This study developed GS-Impute, a residual convolutional denoising autoencoder-based neural network framework that enables accurate genotype imputation for low-density across-population GS. A key breakthrough of GS-Impute is an automatic matching algorithm that resolves the persistent challenge of targeted training in the presence of both sporadic and systematic missing data. Additionally, GS-Impute incorporates a data augmentation strategy and several advanced techniques to enhance imputation accuracy, including residual blocks, dynamic learning-rate optimization, and layer normalization. Comprehensive evaluations across rice and maize breeding populations demonstrated that GS-Impute outperforms the latest versions of established benchmark tools, including Beagle5.4, Minimac4, and STICI. Importantly, the results indicate that GS-Impute makes across-population GS feasible with low-density markers, establishing a resource-efficient strategy with the potential to transform genomic breeding programs.