Genomic prediction of feed efficiency in boars by deep learning

Area of Science:

• Animal Genetics
• Genomic Selection
• Machine Learning in Animal Breeding

Background:

• Genomic selection has advanced pork production, but linear models struggle with complex traits like feed efficiency due to overlooking nonadditive genetic effects.
• Deep learning (DL) offers a potential solution by modeling nonlinear patterns in genomic data.

Purpose of the Study:

• Compare the predictive ability of DL models against linear models for feed efficiency (FE) in two boar populations.
• Estimate nonadditive genetic variance captured by DL models.
• Assess the impact of nonadditive variance on predictive ability.

Main Methods:

• Utilized deep learning models, including multilayer perceptron (MLP) and convolutional neural network (CNN), for genomic prediction.
• Employed an averaged-prediction method within DL models.
• Compared predictive abilities with traditional linear genomic prediction models.

Main Results:

• DL models, specifically MLP (0.381) and CNN (0.377), demonstrated higher predictive ability than linear models (0.366) in the sire line population.
• MLP achieved a predictive ability of 0.364 in the dam line population.
• DL models captured nonadditive variance, but it did not significantly enhance predictive performance.

Conclusions:

• Deep learning models, particularly MLP, offer superior predictive ability for feed efficiency in pigs, outperforming linear models.
• DL models are recommended for predicting phenotypes and estimating total genetic effects, including nonadditive components.
• The enhanced predictive power of DL comes with a substantial increase in computational cost.