Enhancing nucleotide sequence representations in genomic analysis with contrastive optimization

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic and metagenomic sequence analysis is complex due to sequence divergence, variable codon usage, and unclear selective constraints.
• Existing methods often struggle with generalization to novel DNA sequences and taxa.

Purpose of the Study:

• To introduce Scorpio (Sequence Contrastive Optimization for Representation and Predictive Inference on DNA), a novel framework for nucleotide sequence analysis.
• To enhance representation learning for DNA sequences using contrastive learning.

Main Methods:

• Scorpio leverages pre-trained genomic language models and k-mer frequency embeddings.
• The framework employs contrastive learning to improve sequence embeddings.
• Tested on diverse datasets with varying DNA sequence lengths.

Main Results:

• Scorpio demonstrates competitive performance in taxonomic and gene classification.
• Achieved high accuracy in antimicrobial resistance (AMR) gene identification and promoter detection.
• Showcased strong generalization capabilities to novel DNA sequences and taxa, outperforming alignment-based methods.

Conclusions:

• Scorpio provides a versatile and robust framework for diverse genomic applications.
• The model's ability to generalize addresses a key limitation in current sequence analysis.
• Analysis reveals underlying biological information in Scorpio's representations, including correlations with gene expression and taxonomy.