EBMGP: a deep learning model for genomic prediction based on Elastic Net feature selection and bidirectional encoder...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic estimated breeding values accelerate breeding programs.
• Deep learning (DL) approaches are increasingly used for genomic prediction (GP).

Purpose of the Study:

• Introduce a novel DL framework, EBMGP, for enhanced genomic prediction.
• Improve predictive accuracy and reduce computational burden in GP.

Main Methods:

• EBMGP uses Elastic Net for feature selection.
• Employs bidirectional encoder representations from transformers embeddings and multi-head attention pooling.
• Models SNPs as 'words' and adjacent SNP groups as 'sentences' to capture genetic interactions.

Main Results:

• EBMGP demonstrated superior predictive performance across diverse plant and animal datasets.
• Outperformed competing models in 13 out of 16 tasks, with accuracy gains of 0.74-9.55% over the second-best model.
• Achieved marked improvement compared to the one-hot representation.

Conclusions:

• EBMGP offers a robust and effective approach for genomic prediction.
• Highlights the potential of DL, particularly transformer-based methods, in life sciences applications.
• Accelerates genetic gains through enhanced early selection.