LOCALE: Local-Alignment Embeddings for Noise-Robust DNA Search at SRA Scale.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Searching large-scale raw sequencing data repositories like the NIH Sequence Read Archive (SRA) is crucial for biological discovery.
• Current search methods struggle with scalability and are sensitive to sequencing errors and biological variations due to reliance on exact k-mer matching.

Purpose of the Study:

• To develop a scalable and robust sequence search method for petabase-scale repositories.
• To improve the accuracy of sequence retrieval in the presence of sequencing errors and biological divergence.

Main Methods:

• Recasting sequence search as a dense retrieval problem using vector embeddings.
• Training a DNABERT-2 encoder with an InfoNCE objective on biologically informed data augmentations (corrupted sequence crops).
• Evaluating the LOCALE method on SRA benchmarks with varying dataset sizes and mutation rates.

Main Results:

• LOCALE achieved 62.4% average Recall@Rq at a 10% mutation rate on a 50-accession SRA benchmark, outperforming baselines in noisy-query settings.
• On a larger 500-accession, 15-Gbp benchmark, LOCALE achieved an AUPRC of 0.508 at 10% mutation, significantly higher than MetaGraph's 0.129.
• The method demonstrates effective retrieval by ranking locally aligned sequences higher than unaligned ones.

Conclusions:

• LOCALE offers a scalable and accurate solution for searching large sequencing data repositories.
• The dense retrieval approach effectively handles sequencing errors and biological divergence, outperforming traditional methods.
• This method has the potential to transform biological discovery by enabling efficient exploration of vast genomic datasets.