Comparison of Genomic Best Linear Unbiased Prediction and Bayesian Regularization Neural Networks for Genomic Selection

Abstract

This study assessed the predictive ability of genomic best linear unbiased prediction (GBLUP) and Bayesian regularization for feed-forward neural networks (BRNN-s1-s3-neuron) with one to three neurons using genomic relationship based on single nucleotide polymorphisms markers. Simulated and actual datasets were used to compare models and choose the better fit model. A five-generation simulated dataset consisted of 3,226 individuals with 10,031 single nucleotide polymorphism (SNP) were taken from the 14th QTL-MAS workshop. Actual mice dataset included body weights (BW) at the age of 6 weeks (g) obtained from 1904 animals genotyped at 10345 SNP loci (aa, Aa, and AA, genotypes were coded as 0, 1 and 2, respectively) and variables of gender of animal, month of birth, year of birth, coat color, cage density, litter. Predictive performance of GBLUP and BRNN-s1-s3-neuron models was investigated by examining the correlations from the cross-validation datasets. In the 14th QTL-MAS validation dataset, the correlations between the simulated true genetic and predicted phenotypic values were 0.607 for GBLUP model and 0.559, 0.353, and 0.288 for BRNN-s1-s3-neuron models. In the 10-fold cross-validation mice datasets, the overall predictive ability of models was low and average of the correlations were 0.419 for GBLUP, 0.336 for BRNN-s1, 0.256 for BRNN-s2, and 0.250 for BRNN-s3-neuron models. In this study, correlation results from the BRNN-s2 and BRNN-s3-neuron models indicated overfitting problem in training datasets as the number of neurons and parameters rises and this led to worse predictions in the validation datasets. The correlations from the GBLUP and BRNN-s1-s3-neuron models for the simulated and actual mice datasets indicated that there was no superiority of the BRNN-s1-s3-neuron models over the GBLUP model for predictive performance. The BRNN model with one neuron had less parameters and resulted in predictive performance similar with those from the GBLUP model.