Evaluation of neural network for computing accuracy of genomic breeding value

Today, parametric and nonparametric statistical methods are used for genomic selection oftraits with additive and epistatic genetic architectures. This study was conducted to investigate the effectof non-parametric method of neural network compared with of genomic best linear unbiased prediction(GBLUP) for calculating genomic breeding value using simulated data. Therefore, a genome containing۴ chromosomes, each ۱۰۰ CM long, was created. Then, for supplying variation, genomic data weresimulated with ۱۲۰۰ evenly distributed single nucleotide markers (SNP) and ۱۲۰ randomly distributedquantitative trait loci (QTL) on each chromosome. In addition, additive allelic effects of QTL weredetermined with gamma distribution. Finally, shape, scale, and heritability level were considered ۰.۳,۱.۷۲ and ۰.۲۵ respectively. Results showed that by considering only additive allelic effects, usingnonparametric neural network approach had lower accuracy of genomic breeding value compared withgenomic best linear unbiased prediction (۰.۷۲ ± ۰.۰۶ for neural network versus ۰.۷۶ ± ۰.۰۴ for GBLUP).While, including other effects more than additive effects resulted in higher accuracies of genomicbreeding value in neural network method rather than genomic best linear unbiased prediction (۰.۷۱ ±۰.۰۳ for neural network versus ۰.۶۹ ± ۰.۰۵ for GBLUP). Although in both approaches, predictionaccuracy decreased significantly in next generations after reference population, but this decrease washigher for neural network. In conclusion, non-parametric neural network method can be used as well asgenomic best linear unbiased prediction in conditions such as considering more effects than onlyadditive allelic effects in genomic selection.

کلیدواژه ها:

accuracy ، breeding value ، genomic selection ، neural network ، simulation