Performance evaluation of support vector machine (SVM)-based predictors in genomic selection

The aim was to compare predictive performance of SVM-based predictors constructed using different kernel functions (radial, sigmoid, linear and polynomial) in different genetic architectures of a trait (number of QTL, distribution of QTL effects) and heritability levels. To this end, a genome comprised of five chromosomes, one Morgan each, was simulated on which 10,000 bi-allelic single nucleotide polymorphisms (SNP) were distributed. Cross validation employing a grid search was used to tune the meta-parameters of each kernel function. Pearson's correlation between the true and predicted genomic breeding values (r(p,t)) and mean squared error of predicted genomic breeding values (MSEp) were used, respectively, as measures of the predictive accuracy and the overall fit. Meta-parameter optimization had a significant effect on predictive performance of SVM-based predictors in such a way that by using improper meta-parameters, the predictive power of models decreased significantly. In all models, the accuracy of prediction increased following increase in heritability and decrease in the number of QTLs. In most of scenarios, radial-and sigmoid-based SVM predictors outperformed polynomial and linear models. The linear-and polynomial-based SVM had lower r(p,t) and higher MSEp and, therefore, were not recommended for genomic selection. The prediction accuracy of radial and sigmoid models was approximately the same in most of the studied scenarios; however, considering all pros and cons of radial and sigmoid kernels, radial kernel was recommended as the best kernel function for constructing SVM. All of studied SVM-based predictors were efficient users of time and memory.