Efficient genomic prediction based on whole-genome sequence data using split-and-merge Bayesian variable selection.

Area of Science:

• Genomics
• Animal Breeding
• Bioinformatics

Background:

• Whole-genome sequence data increases genomic prediction but demands significant computational resources.
• The split-and-merge Bayesian stochastic search variable selection (BSSVS) model was investigated to address computational challenges.
• BSSVS involves sequential analysis on data subsets and a merged dataset of selected variants.

Purpose of the Study:

• To evaluate the efficacy of the split-and-merge BSSVS model for genomic prediction using whole-genome sequence data.
• To assess if this approach can overcome increased computing time and model performance issues associated with large datasets.
• To compare prediction accuracies and biases against standard methods like the Bovine 50k-SNP chip.

Main Methods:

• Applied a split-and-merge BSSVS model to a large dataset of over 4 million variants.
• Performed initial BSSVS on 106 subsets, followed by analysis on selected variants (~472,492).
• Calculated genomic prediction accuracies for somatic cell score, protein yield, and interval first to last insemination.

Main Results:

• The split-and-merge BSSVS approach achieved accuracies comparable to or up to 1.1% higher than the Bovine 50k-SNP chip.
• Averaging predictions across subsets yielded the highest accuracies and least biased results.
• Parallelized split-and-merge analysis completed in 5 days, significantly faster than the 3+ months for standard analysis of all sequence variants.

Conclusions:

• The split-and-merge BSSVS approach enables efficient genomic prediction with whole-genome sequence data through parallelization.
• While not improving accuracy in this single-breed study, the method shows potential for multi-breed applications.
• This strategy effectively reduces computational time for large-scale genomic analyses.