Bagging Strong Bayesian Learners for Genomic Prediction

Whole-genome regression models in Bayesian framework represent the main family of prediction methods, where there is much space for improvement. Over the past decade, researches have tried to overcome the tradeoff of genome-wide prediction modeling associated increase in computational burden and gain in prediction accuracy.

Here, we propose the use of machine learning ensemble algorithm with potential to reduce the computational burden while increasing the model predictive ability. The strategy consists in weakening strong learners from the Bayesian alphabet using bootstrap aggregating (aka. bagging) in each round of Monte Carlo Markov Chain (MCMC), working as a cross-validation as the model fits to generate a more stable distribution of parameters a posteriori.

To evaluate the efficacy of bagging strong learners, we evaluated the prediction mean square error (PMSE) and correlation between prediction and observation in a five-fold cross validation scheme, increasing the proportion of resampling data using to fit the model. Bagging was performed without replacement. The Gibbs sampling algorithm was implemented in C++ to run in R. The dataset in study is a wheat panel from CIMMYT, with 599 lines and 1279 markers available in R package BGLR.

By ensemble bagging and Gibbs sampler, we observed enhanced predictive ability with reduced the computational burden associated to the model fitting. Our results suggest that data-resampling techniques can and shall be incorporated in the Gibbs sampling.