VIGoR: Software for Variational Inference for Genome-Wide Regression

Bayesian genome-wide regression is a reliable method for genome-wide prediction and association studies. Because current software for the Bayesian regression usually adopts Markov chain Monte Calro (MCMC), it is computationally expensive for daily use in plant and animal breeding. We developed novel software VIGoR based on variational Bayesian inference, which is much faster than MCMC but makes fully Bayesian inference possible. VIGoR is a user-friendly and cross-platform application. The methods available are Bayesian Lasso (Park and Casella 2008), extended Bayesian Lasso (Mutshinda and Sillanpaa 2010), weighted Bayesian Shrinkage regression (wBSR, Hayashi and Iwata 2010), random regression BLUP, and stochastic search variable selection (George and McCulloch 1997). wBSR can be seen as an analogous of BayesB or BayesA (Meuwissen et al. 2001). Three procedures are possible: 1) model fitting: VIGoR returns the estimates of marker effects and breeding values, and provides Manhattan plots. 2) Tuning of hyperparameters: VIGoR performs cross-validation (CV) to select the best setting of the hyperparameters, and fits the model under the setting. With 1) and 2), users also can predict the breeding values of new genotypes. 3) Evaluation of prediction accuracy: VIGoR evaluates the prediction accuracy via CV. When a user specifies multiple hyperparameter settings, it automatically performs the nested CV and predicts with the best setting. For any procedure, VIGoR can analyze multiple traits simultaneously via parallel computing. Fitting of Bayesian Lasso to QTLMAS 2010 data consisting of 10,031 markers and 2,326 observations takes less than 1 min. with an ordinary laptop computer.