MoEGP: an efficient crop genomic prediction approach based on the mixture of experts network

Genomic prediction plays a pivotal role in assessing crop germplasm resources, accelerating crop breeding, and enhancing genetic improvement. The mixture of experts network (MoE), has been successfully applied in the fields of natural language processing and image recognition. In this study, we proposed a novel, flexible, and robust approach for genomic prediction based on MoE, named Mixture of Experts for Genomic Prediction (MoEGP). This method incorporates a set of experts that process data features and a gating network that distributes samples and assigns weight score to each expert. MoEGP combines the predictions of the selected top-k experts, weighted by their corresponding weight score as final outputs. We evaluated the performance of MoEGP using 36 trait cases from five public datasets across three major cereal crops in comparison to five well-known genomic prediction methods. The results indicated that MoEGP exhibited superior performance in terms of both pearson correlation coefficient (PCC) and mean absolute error (MAE) in all 36 trait experiments, encompassing morphological, developmental, and yield traits. MoEGP achieved an average PCC improvement of 33.75%, and an average reduction of 19.55% in MAE compared with the other five reported methods. With its automated hyper-parameter search system, MoEGP offers flexibility in supporting traits governed by distinct quantitative trait locus (QTL) architectures, facilitating its broad application in crop genomic prediction. The corresponding code is available as open-source on GitHub (https://github.com/CNRRI-RGRT/MoEGP). We proposed a novel deep-learning network designed for genomic prediction, named MoEGP, which outperformed five previously reported methods and demonstrated strong potential for application in crop genomic selection.