The effect of kernel selection on genome wide prediction of discrete traits by Support Vector Machine

Abstract

This study aimed to compare predictive performance of the linear, radial, polynomial and sigmoid kernel-based Support Vector Machine (SVM) in different scenarios of genetic architecture of a discrete trait (number of quantitative trait loci (QTL) and distribution of QTL effects) as well as heritability level. A 500 centiMorgans genome was simulated consisted of 5 chromosomes, one Morgan each, on which 10,000 bi-allelic single nucleotide polymorphisms (SNP) were distributed. To measures accuracy of prediction, Pearson's correlation between the true and predicted genomic breeding values (rp,t) was estimated. In all of the scenarios studied, increase in the heritability resulted to an increase in the accuracy of prediction. Radial and sigmoid based SVM predictors outperformed polynomial and linear kernels significantly (p < 0.05). The linear and polynomial kernel-based SVMs provided predictions with least accuracy and therefore were not recommended for genome wide prediction of discrete traits. Radial-based SVM had slightly higher predictive performance compared to sigmoid-based SVM, though the differences were non-significant. Therefore, both of them were recommended for predicting genomic breeding values. All of the studied SVM-based predictors had almost the same time and memory requirements.