[R] Genomic Large Language Models
A genomic LLM trained on 9.3 trillion nucleotides spotted a functional link between two genes with zero sequence similarity.
Arc Institute's genomic large language model, Evo2, is demonstrating an ability to see biological connections that traditional sequence analysis tools cannot. Trained on a massive dataset of 9.3 trillion nucleotides, the model was tested by extracting embeddings from its intermediate layers for 512-base-pair windows across 25 human genes. Researchers then compared Evo2's similarity judgments against those from BLAST, the standard sequence alignment tool. While most matches were driven by common repeat elements, a significant discovery emerged after strict filtering: the model found a very high similarity (cosine = 0.948) between sections of the VIM (vimentin) and DES (desmin) genes, despite there being zero detectable sequence match between them.
This finding is biologically meaningful. Both identified regions are active promoters in muscle and connective tissue cells and are known to share key regulatory proteins. VIM and DES are related genes often expressed together, suggesting Evo2 has learned to recognize patterns of gene regulation and functional context, not just the raw 'letters' of the DNA sequence. This represents a potential leap beyond current bioinformatics methods. However, the research notes that this kind of clear signal is still rare and difficult to isolate from noisy data, indicating that turning this capability into a practical, reliable tool for genomic discovery will require significant further development.
- Evo2, trained by Arc Institute on 9.3 trillion nucleotides, identified a functional link (cosine=0.948) between the VIM and DES gene promoters.
- The linked regions share biological function in muscle cells but have zero detectable sequence similarity, a connection BLAST could not find.
- The model appears to learn patterns of gene regulation, not just sequence, though extracting clear signals requires heavy filtering and more work.
Why It Matters
This could unlock new discoveries in genomics by revealing functional gene relationships invisible to current sequence-alignment tools, accelerating biomedical research.