A Feature Engineering Method for Whole-Genome DNA Sequence with Nucleotide Resolution.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Feature engineering for whole-genome DNA is crucial for plant trait prediction.
• Current SNP-based methods have limitations in genetic information capture and locus coverage.
• These limitations hinder accurate prediction of phenotypic traits.

Purpose of the Study:

• To develop a whole-genome DNA sequence feature engineering method (FE-WDNA) for improved plant trait prediction.
• To leverage large language models (LLMs) for nucleotide-level genome-wide feature representation.
• To overcome the shortcomings of SNP-based approaches by considering nucleotide relationships.

Main Methods:

• Proposed FE-WDNA, a whole-genome DNA sequence feature engineering method.
• Utilized HyenaDNA to fine-tune the model on whole-genome data from 1000 soybean samples.
• Derived genome-wide feature vectors by analyzing contextual and long-range dependencies among nucleotide sites.

Main Results:

• Achieved significant improvements in agronomic trait prediction compared to SNP-based methods.
• Evaluated the impact of context window length, feature vector dimensions, and prediction methods.
• Demonstrated FE-WDNA's capability for high-quality DNA sequence feature engineering at nucleotide resolution.

Conclusions:

• FE-WDNA offers a superior approach to DNA sequence feature engineering for plant trait prediction.
• The method provides comprehensive genome-wide feature vectors by capturing nucleotide-level dependencies.
• FE-WDNA is adaptable to other plant species and applicable to various computational breeding tasks.