Reducing dimensionality for prediction of genome-wide breeding values

Area of Science:

• Quantitative genetics
• Genomic selection
• Statistical modeling

Background:

• Genome-wide breeding values (EBV) are crucial for animal and plant breeding.
• Traditional methods like Partial Least Square Regression (PLSR) and Principal Component Regression (PCR) are used when predictors exceed records.
• Comparing these methods with advanced Bayesian approaches is essential for optimizing genomic selection.

Purpose of the Study:

• To compare the accuracy of EBV estimated using PLSR and PCR against a Bayesian method (BayesB).
• To evaluate the impact of varying marker densities on the accuracy of these methods.
• To assess the computational efficiency and bias of each method.

Main Methods:

• Utilized PLSR and PCR for EBV estimation.
• Employed the Bayesian method BayesB for comparison.
• Simulated data with marker densities of 1, 2, 4, and 8 Ne markers/Morgan at an effective population size of 100.

Main Results:

• PLSR and PCR showed lower accuracy (correlation 0.611-0.681 and 0.604-0.658) compared to BayesB (0.690-0.860).
• BayesB demonstrated greater responsiveness to increased marker density, with its advantage growing by 17%.
• PLSR and PCR exhibited higher prediction bias than BayesB across all marker densities.

Conclusions:

• BayesB significantly outperforms PLSR and PCR in accuracy and responsiveness to marker density for genomic prediction.
• Despite computational advantages, the reduced accuracy of PLSR and PCR makes them less desirable.
• Incorporating prior information on gene effect distributions is beneficial for improving genomic prediction accuracy.