Estimation of (co)variance components for very large datasets and complex single-step genomic models

Area of Science:

• Animal Breeding and Genetics
• Statistical Genetics
• Computational Biology

Background:

• Accurate estimation of variance components in linear mixed models is crucial for genetic analyses.
• Computational constraints often necessitate data subsetting or model simplification, potentially introducing bias.
• Existing Monte Carlo REML (MC-REML) methods lacked extensions for single-step genomic analyses.

Purpose of the Study:

• To extend Monte Carlo REML (MC-REML) to incorporate large genomic datasets within single-step genomic best linear unbiased prediction (ssGBLUP) models.
• To develop a computationally efficient method for estimating variance components in large-scale genetic evaluations.

Main Methods:

• Developed Monte Carlo single-step genomic REML (MC-ss-GREML) by simulating breeding values and solving mixed model equations.
• Utilized Expectation Maximization and Average Information for REML optimization.
• Validated the method using a three-trait beef cattle growth model and a large birth weight model.

Main Results:

• MC-ss-GREML showed no difference in variance component estimates compared to exact ss-GREML.
• The method reduced computing time by 86% and memory usage by 99% compared to exact methods.
• Demonstrated scalability with a large dataset of 7 million animals, converging in 11 rounds.

Conclusions:

• MC-ss-GREML effectively estimates variance components for large, genotyped populations.
• The method offers significant reductions in computational cost (time and memory).
• Enables accurate genetic evaluations with complex models and extensive genomic data.