Predicting GxE through phenomic prediction: A benchmark study across three species

Abstract Phenomic prediction (PP) is a genetic value prediction method inspired by genomic prediction (GP). One key feature of PP compared to GP is its potential ability to predict interactions between genotypes and environments (GxE). Using kernel-based models, many relationship matrix calculation methods for the G and GxE terms have been proposed but never fairly compared. In this study, we exhaustively tested different combinations of published methods to compute these matrices, and implemented a new method supposed to better split the genetic (G) and GxE information in their respective matrices. Using three datasets on maize, rice, and triticale, results show that the G effect matrix is the main driver of predictive abilities (PA). Computing the G effect matrix using BLUPs of wavelengths across environments was better when predicting new genotypes in known environments. To predict known genotypes in new environments or in a sparse testing scenario, it was better to compute one relationship matrix per environment and average them into a single G effect matrix. We found little differences between models including a GxE term and models without it. Nevertheless, dense GxE matrices including relationships within and across environments were generally the best choice. Finally, combining genomic information for the G effect, and phenomic information for the GxE effect gave the best results. Clear guidance of the best ways to take advantage of phenomic prediction to predict GxE will likely require a simulation approach to better control the GxE information in datasets.