Disentangling soybean GxE effects in an integrated genomic prediction and machine learning-GWAS workflow

Integrating genotype-by-Environment (GxE) interactions into genomic prediction models has been demonstrated to enhance the accuracy of predictions for crops exposed to unfavourable environmental conditions. However, despite the increasing complexity of machine learning models in genomic prediction, no model or approach has been found to be overall superior in comparison to a classical genomic best linear unbiased prediction (GBLUP) model. In this paper, we compared two GBLUP models (Linear Mixed Effects model and Bayesian GBLUP) with two machine learning models (Random Forest and Extreme Gradient Boosting) on the EUCLEG soybean genotype set phenotyped in Belgium and Serbia. We found similar performance for the Bayesian GBLUP and the two machine learning methods. However, using a workflow that decomposed the environment-specific BLUPs into a main genetic and an interaction GxE effect, we found increased predictive ability for the interaction component compared to a single-component approach. Furthermore, conducting a machine learning-genome wide association study (ML-GWAS) on both components allowed us to identify important markers for the main genetic effect, as well as environment-specific markers. These could then be associated with correlated markers in other environments. By constructing a small random forest model using only 50 uncorrelated, important markers we constructed a genomic prediction model with similar predictive ability over all scenarios when compared to the large models including all markers. The results demonstrate a new, integrated genomic prediction and machine learning-genome-wide association study (ML-GWAS) approach, aimed at high predictive ability and coupled marker detection in the soybean genome for traits phenotyped in different environments.