Megavariate methods capture complex genotype-by-environment interactions.

Area of Science:

• Quantitative genetics
• Computational biology
• Genomic prediction

Background:

• Genomic prediction models are crucial for predicting crop performance across diverse environments.
• Capturing genotype-by-environment (GxE) interactions enhances prediction accuracy but poses computational challenges.

Purpose of the Study:

• To introduce and evaluate scalable algorithms for genomic prediction models incorporating GxE interactions.
• To benchmark the accuracy and runtime of novel computational approaches against existing methods.

Main Methods:

• Developed two latent GxE models: MegaLMM and MegaSEM.
• Implemented an efficient multivariate mixed-model solver, Pseudo-expectation Gauss-Seidel (PEGS).
• Fitted models with unstructured, extended factor analytic (XFA), and Heteroskedastic compound symmetry (HCS) covariance structures.

Main Results:

• MegaLMM and PEGS-based XFA/HCS models achieved highest accuracy in sparse testing (100 environments).
• PEGS unstructured models were significantly faster than REML-based GBLUP with comparable accuracy.
• MegaSEM demonstrated exceptional speed, fitting large-scale models rapidly.

Conclusions:

• Scalable genomic prediction models, particularly MegaLMM and MegaSEM with PEGS, offer efficient and accurate solutions for GxE interaction analysis.
• The choice of covariance structure (XFA, HCS) impacts prediction accuracy.
• Environment-specific genomic estimated breeding values (GEBVs) derived from averaged models provide robust predictions.