Assessing the Capacity of Genomic Language Models to Generate Informative Embeddings of Viral Genome Fragments

DNABERT-S generates taxonomically informative embeddings of short viral genome fragments, outperforming a 4-mer profile on the reference dataset and ranking first on mriyaviruses.

• Why it matters: Metagenomic methods struggle with “dark matter” viruses, so better sequence representations are needed to cluster and identify taxa from short fragments.
• What they did: The study compared four gLMs (DNABERT-2, DNABERT-S, ViraLM, Evo2-7B) against 4-mer profiles and MinHash sketches using clustering and neighborhood metrics (ARI, precision@k, CROSS precision@k) on an ICTV-based balanced reference set.
• The result: Three of four gLMs beat the 4-mer profile on the reference dataset, with best results often from intermediate layers; DNABERT-S stayed strongest on egoviruses and mriyaviruses, and forward fragments matched reverse complements.

AI summary of the abstract · may contain errors · not medical advice · read the paper.