Genomic evaluation of support vector machine and common genomic prediction methods in different prevalence of threshold phenotype- A simulation study

Background and objectives: Many prominent traits in livestock including disease resistance and dystocia, present a classification distribution of phenotypes. These traits are important in animal breeding due to importance of animal welfare and human tendency for healthy and high quality products. Therefore, identifying and characterizing the genetic variants that impact threshold traits, ranging from disease susceptibility, is one of the central objectives of animal genetics. In this regard, genomic selection can have an important role in increasing the genetic progress of the threshold traits. The objective of current study was genomic evaluation of area under receiver operating characteristic curve (AUROC) of support vector machine (SVM), GBLUP and Bayes LASSO methods for different rates of binary phenotype distribution in training set.
Materials and methods: A population of 1000 animals genotyped for 10,000 markers was simulated using QMSim software. Genomic population were simulated to reflect variations in heritability (0.05 and 0.2), number of QTL (100 and 1000) and linkage disequilibrium (low and high) for 29 chromosomes. In order to create different rates of discrete phenotype, the animal’s phenotype of training set was coded as 1 (inappropriate phenotype) depending on whether their phenotype residuals was less than the average of residuals (e ̅), e ̅- 1〖SD〗_eor e ̅+ 1〖SD〗_efor the first, second and third approaches, respectively, and other individuals was defined as code 0 (appropriate phenotype). Three statistical models were implemented to analyze the simulated data including SVM, GBLUP and Bayes LASSO methods.
Results: Optimal training sets were characterized by inappropriate phenotype rate that were similar to the population real, leading to the highest AUROC in SVM, GBLUP and Bayes LASSO methods, in which concluded for e ̅- 1〖SD〗_e threshold point to the training set. The highest (0.813)and lowest(0.521) AUROC were observed for SVM method.Generally, heritability of trait was a factor affecting on genomic AUROC of SVM, GBLUP and Bayes LASSO methods; so that we recognized an increase in genomic AUROC with increase in heritability in all three statistical methods. Average r2 in the low and high LD scenarios was 0.221 and 0.435 at distances of 0.05 cM and the results showed an increase in genomic AUROC using GBLUP, Bayes LASOO and SVM methods with increasing in linkage disequilibrium. The result of current study showed that high level of LD between SNP and QTLs increased the probability of adjacent markers sampling for re-sampling methods. Therefore, this resulted in a positive performance of SVM. Despite of the higher AUROC of GBLUP and Bayes LASSO methods at different scenarios, SVM method showed a better performance when discrete traits were controlled by a large number of QTLs.
Conclusions: Despite the important role of different rates of binary phenotype distribution in training set, SVM method to predict genomic AUROC of discrete traits depends on genetic basis of the population analyzed and cost parameter.