책임운영기관 농촌진흥청 국립축산과학원 (로고)

This study was conducted to test genomic prediction using machine learning and to compare predictions with those of existing techniques. In this study, DL, which is a type of machine learning, and GBLUP and Ensemble, which are integrated techniques, were used. Data were predicted and analyzed using 7,324 genotype data (37,712 SNPs) of Korean cattle and data of four carcass traits. Accuracies were predicted using 5-fold cross-validation and correlations were calculated using test data. Heritability estimated using REML was highest at 0.44±0.02 for MS. Regarding calculated correlations, the strongest relationship was observed between CWT and EMA, which was 0.79±0.01 for genetic and 0.52±0.02 for phenotypic correlations. Single analysis showed prediction accuracies were 0.34±0.01 for BFT, 0.41±0.01 for CWT, 0.37±0.01 for EMA, and 0.41±0.01 for MS, but Ensemble produced significantly better coefficients, that is, 0.35±0.01 for BFT, 0.42±0.01 for CWT, and 0.38±0.01 for EMA. We conclude that the accuracy of genomic prediction can be improved by using various techniques, and consider that the findings of this study may be of considerable value to the breeding industry, as they demonstrate the feasibility of developing highly accurate prediction models.