BGLR

BGLR stands for Bayesian Generalized Linear Regression, a software package designed for Bayesian inference in high-dimensional linear and generalized linear models, with a strong emphasis on genomic selection and genomic prediction. It provides a flexible framework for estimating effects of many markers or features under priors that encourage shrinkage and variable selection, enabling robust prediction in settings with more predictors than observations.

Implemented in R, BGLR centers around the BGLR function, which takes a response vector, a list of design matrices (ETA) describing fixed and random effects, and MCMC settings such as the number of iterations, burn-in, and thinning. It supports several prior formulations for marker effects, including Bayesian shrinkage priors appropriate for genomic data, and can model Gaussian, binomial, and Poisson responses. The package uses Gibbs sampling and, when necessary, Metropolis steps to obtain posterior samples, and can incorporate non-additive models via multiple random effect terms and correlated error structures.

Beyond single-trait Gaussian models, BGLR supports multi-trait and multi-environment modeling, as well as kernel-based approaches that capture nonlinear relationships. It provides facilities for cross-validation, prediction on new data, and estimation of breeding values or genomic estimated breeding values (GEBVs).

Applications span plant and animal breeding, genetics research, and other fields requiring high-dimensional predictive modeling from marker data. The package is open-source and available on CRAN and GitHub, with documentation and vignettes to guide users through model specification, data preparation, and interpretation of results.

Limitations include computational demands for very large datasets and the need for careful convergence assessment and prior choice. Users should ensure adequate MCMC convergence and assess model fit and predictive performance.