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

LOCALE achieves 4.5 times higher AUPRC than MetaGraph on petabase-scale DNA search with 10% genomic variation.

• Why it matters: Efficiently searching massive sequencing repositories is crucial for biological discovery, but existing methods struggle with scalability and genomic variation.
• What they did: Researchers trained LOCALE, a dense retrieval encoder using an InfoNCE objective on biologically informed augmentations, across four encoders and three tokenization schemes, improving recall by 7–39 points at 10% variation.
• The result: LOCALE maintains 63.2% recall at 10% variation on a 50-accession SRA benchmark, surpasses existing methods, and generalizes to real biological variance, reaching 0.813 recall in a cross-genotype Hepatitis B task.