Dimension reduction and variable selection for genomic selection: application to predicting milk yield in Holsteins.

Area of Science:

• Animal Genetics
• Genomic Selection
• Statistical Genomics

Background:

• Accurate prediction of genetic merit is crucial for livestock breeding.
• Genome-assisted prediction requires models handling massive marker data and fewer observations.
• Dimension reduction and marker selection are key strategies for genomic selection.

Purpose of the Study:

• To evaluate supervised Principal Component Regression (PCR) and sparse Partial Least Squares (PLS) for predicting dairy bull genomic breeding values (BV) for milk yield.
• To assess the performance of these methods in combining dimension reduction and variable selection using single-nucleotide polymorphisms (SNPs).

Main Methods:

• Applied supervised PCR (Methods I and II) and sparse PLS to predict genomic BV for milk yield in dairy bulls.
• Supervised PCR II utilized multiple-SNP analyses, outperforming single-SNP analysis (Method I).
• Evaluated predictive ability using varying SNP subset sizes and compared results to a full SNP set PCR model.

Main Results:

• Supervised PCR II demonstrated superior predictive ability compared to supervised PCR I.
• Sparse PLS showed intermediate performance between the two supervised PCR methods.
• Supervised PCR II with 300-500 SNPs achieved 80-87% of the predictive correlation obtained with all 32,518 SNPs.
• Predictive correlation plateaued around 0.68 with 3500 SNPs using supervised PCR II.

Conclusions:

• Combining dimension reduction and variable selection offers a powerful approach for accurate and cost-effective genomic BV prediction.
• Supervised PCR II and sparse PLS are promising methods for genomic selection in dairy cattle.
• These methods can reduce the number of SNPs required for accurate prediction, lowering costs.