Genomic evaluation of threshold traits in different scenarios of threshold number using parametric and non-parametric statistical methods

Abstract

AbstractThe aim was to study the effect of the threshold number on the accuracy of genomic evaluation of the threshold traits using support vector machine (SVM), genomic best linear unbiased prediction (GBLUP) and Bayesian method B (BayesB). For this purpose, a genome consisting of three chromosomes was simulated for 1000 individuals on which 3000 bi-allelic single nucleotide polymorphism markers were evenly distributed. Genomic breeding values were predicted in different scenarios of threshold number (1–6 thresholds), QTL number (30 and 300 QTLs) and heritability level (0.1, 0.3 and 0.5). By increasing the number of thresholds from 1 to 6 thresholds, especially at higher levels of heritability, the accuracy of genomic evaluation increased; however, the increase in accuracy was not linear so that it was much more noticeable when the number of thresholds increased from 1 to 2 thresholds. In the most studied scenarios, SVM showed a very poor performance compared to other methods. BayesB ranked first regarding prediction accuracy, though in some cases the observed differences with GBLUP was not significant. While increase in heritability increased the accuracy of genomic evaluation, change in the QTL number had a slight effect on the prediction accuracy. According to the results, the SVM is not recommended for genomic evaluation of threshold traits, especially those which have only one threshold and instead, use of GBLUP and BayesB is recommended. For traits with more than one threshold, fortunately we can achieve accuracy similar to continuous traits by applying traditional genomic evaluation methods.