Impact of scale parameter for marker variance prior in some Bayesian whole-genome regression methods.

Abstract

In this study, traditional single-trait Bayesian whole-genome regression methods (BayesA, BayesB, and BayesCπ), which treat the scale parameter of the marker variance prior as fixed, were compared to alternative versions (BayesSA, BayesSB, and BayesSCπ) that estimate this parameter from the data using a Gamma prior. Analyses were conducted using a publicly available dataset of 3,534 animals from a commercial PIC pig population with genotypic and phenotypic records. Model performance for each trait was evaluated through six-fold cross-validation, using the correlation between genomic estimated breeding values (GEBV) and pseudo estimated breeding values (pEBV) as the measure of predictive accuracy. The alternative BayesSA model achieved higher accuracy than BayesA, whereas BayesSB showed no improvement, likely because its estimated scale parameters were similar to the fixed values in BayesB. For BayesCπ, results confirmed that the limited influence of the prior on marker variance led to similar performance between the classical and alternative versions. Heritability was estimated for five traits (T1–T5), and model specification affected these estimates slightly. The alternative models, which estimate prior variance parameters from the data rather than assuming them as fixed, generally produced lower heritability estimates than their classical counterparts. Considering that the reference heritability values from the original PIC study were based on extensive pedigree and high-density genotype data, the alternative models’ estimates can be regarded as reasonably consistent. These findings indicate that hyperparameter estimation in Bayesian models can improve prediction accuracy and the reliability of heritability estimates, providing a more suitable framework for traits with complex genetic architectures.