Comparison of regression tree-based methods in genomic selection

Abstract

The aim of this study was to compare the predictive performance of tree-based methods including regression tree (RT), random forest (RF) and Boosting (BT) in genomic selection. To do this, a genome comprised of five chromosomes was simulated for 1000 individuals on which 5000 single-nucleotide polymorphisms were evenly distributed. Comparison of methods was made in different scenarios of genetic architecture (number of QTL and distribution of QTL effects) and heritability level (0.1, 0.3 and 0.5). Computing time and memory requirement of the studied methods were also measured. In all the scenarios studied, the RT had the lowest accuracy, one-half to one-third of that was observed for RF and Boosting. Therefore, while RT was most efficient user of time and memory, because of its low accuracy, it was not recommended for genomic selection. Comparing RF and Boosting, at low levels of heritability (0.1 and 0.3), the prediction accuracy of RF was significantly higher than Boosting, but at heritability of 0.5, their accuracy was almost equal. In addition, RF was significantly superior to Boosting regarding computing time and memory requirement. While, heritability had a significant impact on the accuracy of prediction, the effect of number of QTL and distribution of QTL effects were not very dramatic. According to the overall performance of the studied methods, RF is recommended for genomic selection.