Genomic selection in pig breeding: comparative analysis of machine learning algorithms.

Area of Science:

• Animal Breeding and Genetics
• Genomics
• Machine Learning Applications

Background:

• Genomic prediction (GP) is crucial for pig breeding progress, utilizing SNP markers to predict phenotypic values.
• Machine learning (ML) methods are effective for high-dimensional data in GP, but optimal methods for pigs are not well-established.
• Selecting appropriate ML methods is essential for effective pig genomic prediction.

Purpose of the Study:

• To compare the performance of popular ML methods for pig genomic prediction.
• To identify suitable ML methods for predicting various pig traits.
• To provide recommendations for ML approaches in pig breeding.

Main Methods:

• Utilized five common datasets from existing literature for method comparison.
• Evaluated popular ML algorithms including Stacking, Kernel Ridge Regression with RBF kernel (KRR-rbf), Support Vector Regression (SVR), and Genomic Best Linear Unbiased Prediction (GBLUP).
• Assessed prediction accuracy across different datasets and trait types (hidden, reproductive, growth).

Main Results:

• Stacking demonstrated superior performance on the PIC dataset with hidden trait information, closely followed by KRR-rbf.
• SVR excelled in predicting reproductive traits, with GBLUP as the second-best performer.
• GBLUP achieved the highest accuracy for growth traits, with SVR performing second best.

Conclusions:

• GBLUP, Stacking, SVR, and KRR-RBF are effective methods for genomic prediction in pigs.
• LR statistical analysis confirmed the stability of Stacking, SVR, and KRR methods.
• Recommends Stacking, SVR, and KRR as reliable ML approaches for phenotypic value prediction in pigs.