An Upper Bound for Accuracy of Prediction Using GBLUP

Area of Science:

• Quantitative Genetics
• Genomics
• Statistical Genetics

Background:

• Genomic prediction R2 (R-squared) is crucial for estimating breeding values.
• Understanding the asymptotic behavior of prediction accuracy with increasing reference population size is vital for optimizing genomic selection strategies.
• Simulations are essential for exploring scenarios not feasible in real-world experiments.

Purpose of the Study:

• To characterize the asymptotic behavior of genomic prediction R2 as reference population size increases.
• To compare the performance of different genomic prediction methods (GBLUP, BayesB, BayesC) under varying conditions.
• To provide an upper bound for prediction reliability and R2.

Main Methods:

• Simulated whole-genome haplotypes from 85 individuals (1000 Genomes Project) to create LD structure.
• Simulated random mating in a population of 10,000 individuals for 100+ generations.
• Focused simulations on 0.5M genome length (5 chromosomes) with 4,200 markers and 70 QTL for h2=0.8, mimicking human height.

Main Results:

• A 0.5M genome simulation requires a reference population 60-fold larger than a 30M genome for equivalent prediction R2.
• Variable selection methods (BayesB, BayesC) showed no advantage over GBLUP with small reference populations (<6,000 individuals) for a 30M genome.
• BayesB and BayesC outperformed GBLUP as reference population size increased, with all methods becoming asymptotically equivalent around 480,000 individuals.

Conclusions:

• Genomic prediction accuracy is highly dependent on reference population size.
• The choice of genomic prediction method's advantage is contingent on population size and genome length.
• Prediction R2 approaches genomic heritability with sufficiently large reference populations, highlighting the importance of data scale.