Genome-wide prediction using Bayesian additive regression trees

Table 3 Mean squared prediction error (MSPE) for the LASSO, Bayesian LASSO (BLASSO), genomic BLUP (GBLUP), reproducing kernel Hilbert space (RKHS) regression, random forests (RF) and Bayesian additive regression trees (BART) methods evaluated on the pig PorcineSNP60 chip genotype data with one phenotype

Method	Mean squared prediction error (MSPE)
LASSO
minMSE	0.829
minMSE + 1SE	0.861
BLASSO	0.821
GBLUP	0.822
RKHS
\(h = 0.1\)	0.821
\(h = 0.5\)	0.819
\(h = 1\)	0.820
RF	M = 100	M = 200	M = 300	M = 400	M = 600	M = 800
	0.819	0.820	0.815	0.817	0.813	0.813
BART	M = 100	M = 200	M = 300	M = 400	M = 600	M = 800
\(q = 0.9\)
\(\kappa\) = 3	0.822	0.820	0.821	–	–	–
\(\kappa\) = 4	0.819	0.814	0.815	–	–	–
\(\kappa\) = 5	0.814	0.811	0.812	–	–	–
\(\kappa\) = 6	0.815	0.813	0.814	–	–	–
\(q = 0.95\)
\(\kappa\) = 3	0.826	0.820	0.821	–	–	–
\(\kappa\) = 4	0.823	0.814	0.814	–	–	–
\(\kappa\) = 5	0.815	0.812	0.812	–	–	–
\(\kappa\) = 6	0.814	0.814	0.814	–	–	–

The estimates are the mean over five random cross-validation-folds with 70 % training and 30 % test partitions. The lowest MSPE obtained with each method is highlighted in italics. h is the bandwidth of the radial basis function kernel. M is the number of trees for RF and BART, and \(q\) and \(\kappa\) are hyperparameters of the BART priors. The stopping criteria for the regularization coefficient λ in LASSO were obtained based on tenfold cross-validation both at minimum MSE and minimum MSE plus 1 standard error [42]

ISSN: 1297-9686