A new approach fits multivariate genomic prediction models efficiently.

Area of Science:

• Animal Breeding and Genetics
• Quantitative Genetics
• Bioinformatics

Background:

• Accurate genomic estimated breeding values (GEBV) are crucial for efficient animal breeding decisions.
• Multivariate genomic prediction enhances GEBV accuracy by leveraging genetic correlations between traits and environments.
• Current methods like REML are computationally intensive for estimating genetic parameters in complex models.

Purpose of the Study:

• To develop and evaluate fast, memory-efficient multivariate genomic prediction algorithms.
• To assess the performance of new Gauss-Seidel based methods (THGS, PEGS) against REML.
• To compare GEBV accuracy and genetic parameter estimation for balanced and unbalanced designs.

Main Methods:

• Proposed a multivariate randomized Gauss-Seidel algorithm for simultaneous estimation of model effects and genetic parameters.
• Combined two existing methods with a Gauss-Seidel solver, creating Tilde-Hat-GS (THGS) and Pseudo-Expectation-GS (PEGS).
• Simulated balanced and unbalanced experimental designs to compare runtime, GEBV accuracy, and parameter estimation biases and standard errors.

Main Results:

• PEGS and THGS demonstrated significantly faster runtimes compared to REML.
• GEBV accuracies from THGS and PEGS were slightly lower than REML but superior to univariate methods.
• Estimates of genetic parameters showed small biases, though standard errors for genetic correlations were higher than REML, decreasing with increased sample size.

Conclusions:

• THGS and PEGS are scalable, fast, and memory-efficient for multivariate genomic prediction in various experimental designs.
• The proposed methods achieve GEBV accuracy comparable to REML.
• While genetic parameter estimates are largely unbiased, further research is needed for datasets with selection.