Deep learning versus parametric and ensemble methods for genomic prediction of complex phenotypes.

Area of Science:

• Animal breeding and genetics
• Genomic prediction
• Machine learning in agriculture

Background:

• Predicting complex traits from genomic data is crucial for animal and plant breeding.
• Machine learning, including deep learning (MLP, CNN), is increasingly explored for genomic prediction.
• Comparing various ML and parametric methods is essential for optimizing prediction accuracy.

Purpose of the Study:

• To compare the predictive performance of deep learning (MLP, CNN), ensemble learning (RF, GB), and parametric (GBLUP, Bayes B) methods.
• To evaluate these methods using real genomic data and simulated datasets with varying genetic architectures.
• To determine the conditions under which deep learning excels in genomic prediction.

Main Methods:

• Utilized a real dataset of 11,790 Holstein bulls with 58k SNPs and sire conception rate (SCR) records.
• Conducted simulations with heritability (0.30), additive/non-additive gene effects, and varying numbers of quantitative trait nucleotides (100, 1000).
• Assessed predictive performance using correlation and mean squared error of prediction.

Main Results:

• Gradient boosting (GB) achieved the highest predictive correlation (0.36) in the real dataset.
• Parametric methods outperformed others under purely additive gene action.
• Gradient boosting was superior for non-additive gene action; deep learning's advantage depended on sample size and genetic architecture.

Conclusions:

• Gradient boosting is a robust method for predicting traits with non-additive gene action.
• Deep learning methods are not inherently superior for genomic prediction unless substantial non-additive genetic variance is present.