Genomic prediction of feed efficiency in boars by deep learning

Pork is the most widely consumed meat globally, and the industry has achieved substantial genetic advancements for several traits using genomic selection. However, traditional linear genomic prediction models may be inadequate for predicting complex traits, such as feed efficiency, as they primarily capture additive genetic effects and overlook nonadditive effects, including dominance and epistasis. Deep learning (DL) has the potential to address this limitation due to its ability to model nonlinear patterns inherent in genomic data. The objectives of this study were to compare the predictive ability of DL models to the linear models for predicting feed efficiency (FE) trait in 2 boar populations, estimate the nonadditive genetic variance captured by DL, and assess its effect on predictive ability. Our results showed that the DL models using the averaged-prediction method had the highest predictive ability in the sire line test population (0.381 for multilayer perceptron [MLP] and 0.377 for convolutional neural network [CNN]), compared to 0.366 for linear models. DL models also showed higher abilities in the dam line test population, with MLP achieving a predictive ability of 0.364. Additionally, we showed that DL models captured nonadditive variance; however, this did not significantly improve predictive ability. In conclusion, DL models, particularly MLP, demonstrated the highest predictive ability for FE, improving performance by approximately 4.1% for the sire line and 2.8% for the dam line compared to the traditional linear models. Therefore, DL models are recommended for predicting phenotypes and for estimating total genetic effects, including nonadditive components. However, this comes at a significant increase of computational cost.