Predictive Ability of Statistical Genomic Prediction Methods When Underlying Genetic Architecture of Trait Is Purely Additive

A simulation study was conducted to address the issue of how purely additive (simple) genetic architecture might impact on the efficacy of parametric and non-parametric genomic prediction methods. For this purpose, we simulated a trait with narrow sense heritability h2= 0.3, with only additive genetic effects for 300 loci in order to compare the predictive ability of 14 more practically used genomic prediction models based on four criteria (mean squared error (MSE), Bias, γy,GEBV and γGEBV,TBV). Results suggested that parametric genomic prediction models have greater superiority over non parametric genomic models under a simple purely additive genetic architecture. Our result also showed that, all parametric methods, other than ridge-regression BLUP (RR-BLUP), could explain most of phenotypic variation because they showed lower MSE, higher predictive correlation (γy,GEBV), the least amount of bias (by,GEBV) and the higher correlations between true breeding values and the estimated genomic breeding values (γTBV,GEBV). Random forest regression had the worst performance among non parametric methods. The simulation results suggested that there is a large difference between performances of non parametric methods in comparison with parametric methods when underlying architecture is purely additive. But this may not happen when dominance and epistatic genetic effects contributing to both additive and non-additive genetic variances.