Improved genomic prediction using machine learning with Variational Bayesian sparsity.

Area of Science:

• Agricultural Science
• Bioinformatics
• Computational Biology

Background:

• Genomic prediction is crucial for breeding programs, with linear models being accurate but computationally intensive for large datasets.
• Machine learning (ML) offers computational solutions but faces challenges with over-parameterization in deep learning architectures.
• Large-scale breeding programs require efficient prediction models to analyze extensive line and environmental data.

Purpose of the Study:

• To introduce and evaluate a novel ML architecture, Variational Bayesian Sparsity-ML (VBS-ML), for genomic prediction.
• To address the computational limitations and over-parameterization issues of existing genomic prediction models.
• To improve the accuracy and feasibility of genomic prediction in large-scale breeding populations.

Main Methods:

• Developed a machine learning architecture (VBS-ML) incorporating variational Bayesian sparsity in its initial layers.
• Implemented a feature selection mechanism to identify important genetic markers linked to traits.
• Applied the VBS-ML approach to four large Australian wheat breeding datasets with extensive genotypic information.

Main Results:

• The VBS-ML architecture demonstrated improved genomic prediction accuracy compared to traditional linear models across all tested datasets.
• Variational Bayesian sparsity effectively reduced network over-parameterization by selecting significant markers.
• The VBS-ML approach proved computationally feasible for large breeding populations and numerous genetic markers.

Conclusions:

• The VBS-ML architecture significantly enhances genomic prediction accuracy over legacy modeling approaches.
• VBS-ML offers a computationally efficient solution for genomic prediction in large breeding populations.
• This approach effectively reduces the parameter burden in machine learning models for breeding applications.