Genomic DNA Language Models
Language models trained on DNA and RNA sequences to predict variant effects, detect pathogens and design sequences.
Tier Grades
Tier A: Frontier AI labs
feasible
Arc Institute with NVIDIA trained Evo 2 on 8.8 trillion tokens.
Tier B: Indonesian AI labs
stretch
Pretraining on 8.8 trillion tokens is out of reach, but the open 1B and 7B Evo 2 models can be run locally.
Tier C: Institution units such as ITB
feasible
Running open Evo 2 7B or METAGENE-1 7B for variant or pathogen studies fits a university genomics unit.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
Pathogen surveillance on open 7B models is a possible proof of technology with a health partner.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeEvo 2: 1B to 40B parameters with up to 1 million base pair context; METAGENE-1: 7B parameters.
- DataEvo 2: OpenGenome2, 8.8 trillion tokens from all domains of life; METAGENE-1: over 1.5 trillion base pairs from wastewater.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 7 authors on the METAGENE-1 paper (arXiv 2501.02045).
- Time horizonYears.
- Shelf-life signalsMETAGENE-1 reports state-of-the-art results on genomic benchmarks shortly after Evo 2.
- Pipeline roleTier A trains genomic models; Tier B and Tier C apply open ones to Indonesian genomes and surveillance; Tier D could support a health proof of technology.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Links
R&D TopicGenomic DNA Language Models
Models
Evo 2Arc Institute, scientific
METAGENE-1University of Southern California, scientificEvo 2 1BArc Institute, scientificEvo 2 20BArc Institute, scientific+4 more
Evo 2Arc Institute, scientificMETAGENE-1University of Southern California, scientificEvo 2 1BArc Institute, scientificEvo 2 20BArc Institute, scientific+4 moreUse CasesGenomicsHealthcare
Who Runs Itnone recorded
Models 8
Evo 2Arc Institute, NVIDIA, scientificexemplarEvo 2 40BArc Institute, NVIDIA, scientificproduced byEvo 2 7BArc Institute, NVIDIA, scientificproduced byEvo 2 1BArc Institute, NVIDIA, scientificproduced byEvo 2 20BArc Institute, NVIDIA, scientificproduced by
AlphaGenomeGoogle DeepMind, scientificproduced by
DeepVariantGoogle, scientificproduced byMETAGENE-1University of Southern California, Prime Intellect, scientificexemplarPapers
- HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide ResolutionEric Nguyen et al., NeurIPS 2023 (per arXiv comment), 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- raw.githubusercontent.com/ArcInstitute/evo2/main/README.md https://raw.githubusercontent.com/ArcInstitute/evo2/main/README.md
- arxiv.org/abs/2501.02045 https://arxiv.org/abs/2501.02045
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Genomic DNA Language Models", https://aurion.semai.id/topics/genomic-dna-language-models/, records of October 2026.