"Genome modelling and design across all domains of life with Evo 2" is the technical paper describing Evo 2, published open access in Nature 652(8112):1349–1361 on March 4, 2026 (DOI 10.1038/s41586-026-10176-5). It carries 62 authors, with Garyk Brixi first and Brian L. Hie last; the other senior authors listed are Dave P. Burke, Hani Goodarzi and Patrick D. Hsu. Author affiliations span the Arc Institute and the "Evo 2 Core Team" in Palo Alto, Stanford University, the University of California, Berkeley, NVIDIA, Goodfire, Liquid AI, OpenAI, Columbia University, the University of Washington, the University of California, San Francisco, and Johns Hopkins University. Greg Brockman and Patrick Collison are among the co-authors.
The paper introduces Evo 2 as "a biological foundation model trained on 9 trillion DNA base pairs from a highly curated genomic atlas spanning all domains of life to have a 1 million token context window with single-nucleotide resolution," and states that the authors "have made Evo 2 fully open, including model parameters, training code, inference code and the OpenGenome2 dataset."
Architecture, training and data
Two model sizes were trained: Evo 2 7B, with 7 billion parameters on 2.4 trillion tokens, and Evo 2 40B, with 40 billion parameters on 9.3 trillion tokens. The training corpus, OpenGenome2, was compiled from curated, non-redundant nucleotide sequence data totalling more than 8.8 trillion nucleotides drawn from bacteria, archaea, eukarya and bacteriophage.
Training proceeded in two phases. Pretraining used a context length of 8,192 tokens with data weighting focused on genic windows; a multi-stage midtraining phase then extended the context to 1 million tokens. The paper frames this as matching practice in natural-language modelling, where shorter-context pretraining improves efficiency and model quality. On a modified needle-in-a-haystack evaluation, the model recalls a specific 100-base-pair sequence hidden within 1 million base pairs of random DNA.
The architecture is StripedHyena 2, described as a convolutional multi-hybrid combining three variants of input-dependent convolution operators with attention. At 40 billion parameters it is reported to deliver up to a 3× speedup at 1-million context length over optimized Transformer baselines, and to improve loss scaling on DNA against both Transformers and StripedHyena 1.
Prediction results
The paper reports zero-shot prediction — no task-specific fine-tuning — across several benchmark families.
| Task | Reported result |
|---|---|
| ClinVar coding SNVs | 7B and 40B competitive, ahead of zero-shot methods including ESM-2, behind ESM-1b, GPN-MSA and some PhyloP variants |
| ClinVar non-SNV coding variants (insertions, deletions) | Both Evo 2 models outperformed all other methods tested; the paper notes AlphaMissense and GPN-MSA cannot score these variants at all |
| ClinVar noncoding SNVs | Evo 2 40B first among unsupervised models, trailing supervised models |
| ClinVar noncoding non-SNVs | Evo 2 40B outperformed all models tested |
| SpliceVarDB (exonic and intronic) | 40B and 7B first among unsupervised models; on intronic variants competitive with supervised models, trailing SpliceAI and CADD but ahead of Pangolin |
| BRCA1 saturation mutagenesis | Strong on coding SNVs; outperformed all other models on BRCA1 noncoding SNVs |
| BRCA1, supervised ridge regression on Evo 2 embeddings | AUROC 0.95, AUPRC 0.88 on the test set, exceeding zero-shot Evo 2 40B |
| Human gene essentiality (zero-shot) | Evo 2 40B AUROC 0.66, AUPRC 0.15, against other genomic language models at AUROC 0.50–0.59 and AUPRC 0.09–0.12, and within the range of four PhyloP conservation scores (AUROC 0.65–0.71) |
| Exon classification, eight held-out species | AUROC 0.91–0.99, outperforming Nucleotide Transformer, Evo 1, conservation baselines and ab initio AUGUSTUS |
| DART-eval regulatory tasks | caQTL AUROC 0.58 and dsQTL AUROC 0.66, against Nucleotide Transformer at 0.52 and 0.61, but trailing ChromBPNet at 0.77 and 0.89 |
The paper is explicit about limits as well as results: Evo 2 "underperforms state-of-the-art models on protein DMS," its fitness-prediction performance "begins to saturate and can decrease at the largest model scales," its human gene-essentiality performance "remains modest," and it "falls behind supervised models for distal regulatory variants." The authors note that Evo 2 is not trained on any human genetic variation or functional genomics data.
Interpretability
Sparse autoencoders were trained on Evo 2 representations without prior biological annotations. The paper reports latent dimensions corresponding to open reading frames, intergenic regions, transfer RNAs and ribosomal RNAs in the E. coli genome, and, in the abstract, to exon–intron boundaries, transcription factor binding sites, protein structural elements and prophage genomic regions. Transcription-factor motifs recovered this way were matched against the HOCOMOCO v.12 CORE database using the TOMTOM comparison tool. The authors present this against "the common critique of large language models as black box systems," and conclude that the model "contains distinct internal representations of noncoding regulatory elements."
Generation and design
The abstract states that Evo 2's generative abilities "produce mitochondrial, prokaryotic and eukaryotic sequences at genome scale with greater naturalness and coherence than previous methods." The paper qualifies this directly: "these evaluation metrics do not guarantee functional or replication-competent genomes, and our genome-scale generations lack important elements, such as some essential genes."
The design result the paper validates in the laboratory is epigenomic rather than genomic. Using predictive models and an inference-time beam search — sampling 30 or more 128-base-pair chunks and keeping the top two at each step — the authors designed sequences with specified chromatin-accessibility profiles, including patterns spelling the Morse code messages "LO", "ARC" and "EVO2". The designed DNA was synthesized, integrated site-specifically into the genomes of mouse embryonic stem cells, and measured by ATAC-seq. The three Morse code patterns achieved experimental AUROCs of 0.92–0.95, and simpler peak patterns exceeded 0.89. Where the design varied the level of accessibility within a sequence, 33 of 36 designs (92%) achieved AUROC above 0.8. On the harder task of designing differential accessibility between two cell types, 4 of 24 designs (17%) achieved greater than twofold differential accessibility and 1 of 24 (4%) greater than threefold.
Open release and biosafety
The paper's safety discussion is short and is the passage most often cited in policy contexts. The authors state that they aligned with the Responsible AI × Biodesign commitments (responsiblebiodesign.ai) and "preemptively assessed and mitigated potential concerns prior to open source publication," acknowledging that fully open-source models "may also be used in unanticipated ways that could lead to accident or misuse risks." They report collaborating with multidisciplinary experts on data exclusion measures, safety and security evaluations, and population bias evaluations.
The operative control is a training-data exclusion carried over from Evo 1: "By excluding genomic sequences of viruses that infect eukaryotes from our training data, we aimed to ensure our openly shared model did not disseminate the capability to manipulate and design pathogenic human viruses." The authors report two checks that the exclusion took effect — high perplexity on eukaryotic viral genomic sequences, and no correlation between Evo 2 likelihood and viral protein fitness — and state that red teaming "meant to directly elicit pathogenic human viral proteins showed generations were effectively random in this domain."
The paper also states the exclusion's known limit in its own words: "Task-specific post-training may circumvent this risk mitigation measure and should be approached with caution." No system card or responsible-scaling-style framework accompanies the model. For governance context the paper cites Bloomfield et al., "AI and biosecurity: The need for governance," Science 385:831–833 (2024).
Lineage and reception
Evo 2 follows Evo 1, described in Nguyen et al., "Sequence modeling and design from molecular to genome scale with Evo," Science 386 (2024), and a companion line of gene-design work published as Merchant, King, Nguyen and Hie, "Semantic design of functional de novo genes from a genomic language model," Nature 649:749–758 (2026).
Five months after publication, the same group used Evo 1 and Evo 2 to generate complete bacteriophage genomes, reported in Generative design of bacteriophages with genome language models. That paper drew an accompanying Science Perspective from the Johns Hopkins Center for Health Security arguing that the governance for this capability does not yet exist (AI-designed viral genomes).
Provenance
The body was retrieved from the Nature open-access HTML. The initial retrieval covered the paper through the opening of the interpretability section; the later sections, including the full safety and open-release discussion quoted above, were retrieved in a second targeted pass as verbatim quotations. Methods, Supplementary Information, figures and the reference list were not retrieved, so claims resting on Methods detail — including the specific composition of the safety and security evaluations and the population bias evaluations — are not made here. Publisher metadata confirms the DOI, volume 652, issue 8112, pages 1349–1361, the online date of March 4, 2026, open-access status, and the 62-author list in order.
Relationships
- depends-on: Foundation Models, Scaling Laws
- supports: Evo 2 (Arc Institute / NVIDIA), AI for Science
- instance-of: Open-Weight Frontier Models
- related: AI Biosecurity, Dual-Use Frontier AI, Mechanistic Interpretability, CBRN Uplift
- contradicts: AI-designed viral genomes — on whether open release with a training-data exclusion is a sufficient governance posture