GENOME-ENABLED PREDICTION OF GENETIC VALUES FOR USING RADIAL BASIS FUNCTION NEURAL NETWORKS

Abstract

The objective of this work is to evaluate the efficiency of genomic selection (GS) of genome-enabled prediction by Radial Basis Function Neural Networks (RBFNN) in the prediction of genetic values considering dominance effects and diferent degrees of heritability. In addition, the results were compared with those obtained by G-BLUP. An F1
population with 500 individuals genotyped with 1000 SNP-type markers was simulated. The phenotypic traits were determined by adopting two different gene action models: additive and dominance admitting heritability levels (h2) 30 and 60%, each is controlled by 50 loci, considering two alleles per loco. The accuracy and the mean squared error root (MSER) were estimated using a five-fold cross-validation scheme. For the low heritability scenario, h2 = 0.3 in the additive scenario, the accuracy of validation was 31% for RBFNN, 58% for RR-BLUP, and in the complete dominance scenario the values were 28% e 25%, respectively. Additionally, when analyzing the MSER the difference in performance of the techniques is even greater. For additive scenario, the estimates were 97.33 RR-BLUP and 5.80 for RBFNN, in the most critical scenario, 91.31 GBLUP and 14.55 for RBFNN. Overall, a RBFNN shows accuracy lower than those obtained through G-BLUP. On the other hand, the RBFNN has low prediction bias. Finally, the adjustment of the GBLUP to dominance models despite increasing the complexity of the model also increased the predictive accuracy compared to the model without considering the dominance effect.