Obscured-ensemble models for genomic prediction.

Area of Science:

• Agricultural Science
• Genetics
• Machine Learning

Background:

• Genomic Prediction (GP) utilizes whole-genome markers for predicting crop traits.
• Deep learning (DL) models achieve state-of-the-art results in GP but are sensitive to data and architecture.
• Investigating shortcut learning in GP is crucial for model reliability.

Purpose of the Study:

• To evaluate an obscured model for Genomic Prediction (GP) without genomic content.
• To assess the potential for shortcut learning in GP.
• To develop and evaluate a deep learning ensemble method for GP using an obscured model.

Main Methods:

• Utilized an obscured model for GP, withholding genomic information.
• Tested GP performance with reduced marker sets (20% of obscured markers).
• Developed and applied a deep learning ensemble method based on the obscured model.

Main Results:

• GP was successful with the obscured model using only 20% of markers, enhancing efficiency without compromising accuracy.
• Eliminating markers confirmed the model did not rely on linkage for shortcut learning.
• The obscured ensemble model demonstrated success in GP, even with limited training genotypes and random subset selection.

Conclusions:

• Obscured models can perform effective GP, reducing reliance on extensive genomic data.
• The developed deep learning ensemble method offers a robust approach to GP.
• Selective feature usage and ensemble methods enhance GP efficiency and reliability.