Crop phenotype prediction using SNP context and whole-genome feature embedding based on DNABERT-2.

Area of Science:

• Genomics
• Plant Breeding
• Machine Learning

Background:

• Traditional genomic prediction models overlook the influence of SNP sequence context on crop phenotypes.
• Elite crop breeding requires precise genomic prediction for accelerated genetic gain.

Purpose of the Study:

• To develop novel feature embedding methods for genomic prediction that incorporate SNP sequence context.
• To evaluate the performance of these methods compared to traditional approaches.

Main Methods:

• Utilized DNABERT-2, a cross-species genomic foundation model, for SNP-context and whole-genome feature embedding.
• Applied dimensionality reduction techniques (PCA, PLS) to high-dimensional embeddings.
• Tested methods on rice and maize datasets, analyzing prediction accuracy across varying feature dimensions and context lengths.

Main Results:

• Machine learning models with SNP-context embeddings showed higher accuracy and lower MAEs than traditional SNP features.
• Optimal context lengths (1000-3000 bp) improved predictions for low-to-moderate heritability traits.
• Whole-genome embeddings enhanced prediction accuracy for highly heritable traits, outperforming existing deep learning models.

Conclusions:

• DNABERT-2 based feature embeddings effectively capture SNP context, overcoming limitations of traditional models.
• SNP-context mode is optimal for low-to-moderate heritability traits; whole-genome mode is superior for highly heritable traits.
• Provides a flexible framework for plant breeders to select prediction methods based on trait complexity, accelerating crop improvement.