Embed-Search-Align: DNA sequence alignment using Transformer models.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• DNA sequence alignment is crucial for genomics, traditionally involving genome indexing and read searching.
• Large Language Models (LLMs) show promise in encoding DNA sequences, but performance in classification doesn't guarantee genome-wide alignment.
• Existing methods face challenges in efficiently searching extensive reference genomes for short DNA reads.

Purpose of the Study:

• To bridge the gap between DNA sequence classification and genome-wide alignment using LLMs.
• To develop a novel framework for accurate and efficient DNA sequence alignment.
• To introduce a reference-free DNA embedding model capable of genome-scale search.

Main Methods:

• Developed the "Embed-Search-Align" (ESA) framework incorporating a Reference-Free DNA Embedding (RDE) Transformer model.
• Utilized contrastive loss for self-supervised training to generate rich, reference-free DNA sequence embeddings.
• Implemented a DNA vector store for efficient, global-scale search across reference genome fragments.

Main Results:

• The RDE model achieved 99% accuracy in aligning 250-length reads to a 3-gigabase human reference genome.
• The ESA framework's performance rivals established alignment tools like Bowtie and BWA-Mem.
• RDE significantly outperformed recent DNA-Transformer baselines (e.g., Nucleotide Transformer, Hyena-DNA) and demonstrated cross-species and cross-chromosome transferability.

Conclusions:

• The ESA framework and RDE model offer a powerful new approach to DNA sequence alignment.
• This method provides a viable, high-accuracy alternative to conventional algorithmic alignment techniques.
• The reference-free embedding strategy shows potential for advancing genomic analysis and large-scale sequence comparison.