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

Area of Science:

• Plant genetics
• Agricultural science
• Bioinformatics

Background:

• Genomic prediction models aim to improve crop breeding by predicting performance based on genetic markers.
• Integrating genotype-by-environment (GxE) interactions can enhance prediction accuracy, especially under diverse environmental conditions.
• Classical genomic best linear unbiased prediction (GBLUP) models are widely used, but machine learning (ML) models are increasingly explored for genomic prediction.

Purpose of the Study:

• To compare the performance of GBLUP and ML models for genomic prediction in soybean.
• To investigate the utility of decomposing GxE effects into main genetic and interaction components.
• To develop an integrated ML-GWAS approach for marker detection and improved genomic prediction.

Main Methods:

• Comparison of Linear Mixed Effects GBLUP, Bayesian GBLUP, Random Forest, and Extreme Gradient Boosting models.
• Phenotypic data from the EUCLEG soybean genotype set in Belgium and Serbia.
• Decomposition of environment-specific BLUPs and ML-GWAS on genetic and GxE components.

Main Results:

• Bayesian GBLUP and ML models showed similar performance to classical GBLUP.
• Decomposing GxE effects revealed increased predictive ability for the interaction component.
• ML-GWAS identified important markers for both main genetic effects and environment-specific interactions.
• A parsimonious model using 50 key markers achieved predictive ability comparable to models using all markers.

Conclusions:

• An integrated genomic prediction and ML-GWAS approach offers high predictive ability and effective marker detection in soybean.
• Decomposing GxE effects and utilizing ML-GWAS provides insights into genetic architecture across environments.
• This approach facilitates marker-assisted breeding for improved crop performance under varying environmental conditions.