Application of multiple shrinkage methods to genomic predictions.

Area of Science:

• Animal Genetics
• Quantitative Genetics
• Statistical Genomics

Background:

• Large marker panels in livestock breeding present challenges of model overparameterization.
• Shrinkage or regularization techniques are crucial for managing high-dimensional genomic data.
• Accurate genomic prediction is vital for improving livestock traits.

Purpose of the Study:

• To apply and compare Bayesian LASSO (B-L), Student-t, and semiparametric multiple shrinkage methods.
• To evaluate the performance of single versus multiple shrinkage frameworks.
• To assess these methods against routine genomic prediction techniques like Bayes-A and GBLUP.

Main Methods:

• Application of Bayesian LASSO, Student-t (thick-tailed), and semiparametric multiple shrinkage.
• Analysis using both simulated and real SNP genotype data from Holstein sires.
• Performance evaluation via correlation of true and predicted transmitting abilities and cross-validation.

Main Results:

• Shrinkage models consistently outperformed GBLUP with simulated data (1-8% advantage).
• Shrinkage models showed slight superiority over GBLUP for most traits in real data analysis.
• Models effectively handled traits with identified SNPs of large effect.

Conclusions:

• Shrinkage methods offer a viable alternative to current genomic prediction approaches.
• Multiple shrinkage models showed a small advantage, warranting further investigation.
• Further research is needed to explore the performance of multiple shrinkage methods across diverse scenarios.