Using Bayesian regularized neural networks (BRNN) for predicting DRP of Holstein sires by including different SNP marker effects

Abstract

Artificial neural networks (ANNs) are computational algorithm in the field of machine learning (ML) prediction methods. The aim of this research, was using two-layers Bayesian Regularized Neural Network (BRNN) by including both additive and dominance effects of SNP markers for prediction of reproductive traits’ de-regressed proofs (DRP) of Holstein sires. The studied reproductive traits were the age at first calving (AFC), calving interval (CI), days open (DO) and daughter pregnancy rate (DPR). The genotypic information related to 2419 sires with 41,099 SNPs. DRP of these sires calculated using an animal model and the Garrick method (DRP_G). The predictive accuracy for the AFC_DRP trait ranged from 0.85 with additive effects to 0.89 when both additive and dominance components were considered. Similarly, the predictive accuracy for CI_DRP trait varied between 0.54 for additive effects and 0.55 when both dominance and additive effects were included. For DO_DRP trait the predictive accuracy ranged from 0.53 with additive effects to 0.55 with the inclusion of both dominance and additive effects. The accuracy for DPR_DRP trait ranged from 0.51 with additive to 0.53 both additive and dominance effect. For all studied traits accuracy of ANN were increased by adding dominance effect of SNP markers.