Variable selection models for genomic selection using whole-genome sequence data and singular value decomposition.

Area of Science:

• Quantitative Genetics
• Genomic Prediction
• Bioinformatics

Background:

• Non-linear Bayesian genomic prediction models, like BayesC, often rely on computationally intensive Markov chain Monte Carlo (MCMC) algorithms.
• Whole-genome sequence (WGS) data exacerbate computational costs in genomic prediction.
• Singular value decomposition (SVD) offers a computationally efficient alternative for analyzing large genomic datasets.

Purpose of the Study:

• To develop and evaluate a direct, non-iterative method for estimating marker effects within the BayesC genomic prediction model.
• To leverage SVD for efficient computation of marker effects and prediction error variances (PEV).
• To compare the predictive accuracy and computational efficiency of the SVD-based BayesC method against traditional MCMC-based approaches and SNP-BLUP.

Main Methods:

• Developed a direct, non-iterative method for BayesC marker effect estimation using SVD of the genotype matrix.
• Calculated marker effects and their PEV via SVD, using posterior probabilities to approximate BayesC estimates.
• Conducted simulation studies to compare prediction accuracies over 10 generations of forward prediction against MCMC-based BayesC and SNP-BLUP.

Main Results:

• SVD-based posterior probabilities for marker effects showed comparable, though sometimes lower, values than MCMC-based probabilities, effectively identifying QTL-rich regions.
• Prediction accuracies for breeding values using SVD-based and MCMC-based BayesC were similar across 10 generations.
• The SVD-based BayesC model showed slightly higher accuracies than SNP-BLUP for intermediate prediction generations (2-5), while SNP-BLUP excelled with reduced marker density in the short term.

Conclusions:

• A direct SVD-based method for BayesC marker effect estimation is feasible and achieves similar prediction accuracies to MCMC methods.
• The SVD-based BayesC method offers comparable computation times to SNP-BLUP, particularly if SVD is pre-computed.
• This non-iterative approach provides a computationally efficient alternative for genomic prediction using large WGS datasets.