2026-03-18

Evo 2: the AI that reads, understands and writes DNA

Blue-lit DNA double helix merging with a digital neural network, symbolising AI applied to genomics

In March 2026, the Arc Institute, NVIDIA, and Stanford published in Nature the results of a model that could redefine genomics research: Evo 2.

This is not just another tool in the bioinformatics ecosystem. Evo 2 is a biological foundation model, trained on 9 trillion DNA base pairs spanning all of life: bacteria, archaea, and eukaryotes.

To put it simply: it is a language model comparable to ChatGPT, but its language is neither English nor French. It is DNA.

And like large language models, it does not merely "read". It understands the grammar of life. And it can write.

A model of unprecedented scale

Evo 2 comes in two versions: 7 billion and 40 billion parameters. Both have a context window of one million tokens with single-nucleotide resolution, meaning each A, T, C, G base is an individual token.

The training dataset, called OpenGenome2, compiles over 8.8 trillion nucleotides from bacterial, archaeal, eukaryotic, and bacteriophage genomes, carefully curated to avoid redundancy and bias.

The underlying architecture, StripedHyena 2, is not a standard Transformer. It combines three types of convolution operators with attention, delivering up to three times the throughput of optimised Transformers on long sequences — a critical advantage when working at genome scale.

Result #1: predicting mutation impact without specific training

One of Evo 2's most remarkable capabilities is zero-shot prediction of mutation effects. In practice: you present it with a DNA sequence, introduce a mutation, and the model estimates whether that mutation is deleterious or benign — without any prior fine-tuning.

The model was evaluated on human clinical variants from the ClinVar database:

On non-SNV variants (insertions, deletions, duplications) in coding and non-coding regions, Evo 2 40B outperforms all tested models, including supervised models like AlphaMissense and CADD.
On non-coding variants, it ranks first among unsupervised models.
On BRCA1 variants (the most studied gene in oncogenetics), the model achieves state-of-the-art performance, particularly on non-coding variants where it surpasses even specialised supervised models.

For rare diseases, where annotated data is by definition limited, this zero-shot capability represents a considerable lever. A model able to score mutation impact without requiring a specific training dataset is a game-changer for genetic diagnosis and drug repurposing.

Result #2: the model learned biology on its own

By applying mechanistic interpretability techniques (Sparse Autoencoders), researchers decomposed Evo 2's internal representations into interpretable "features". The finding is striking: without ever being given biological annotations, the model spontaneously learned to recognise:

Exon/intron boundaries in eukaryotic genomes
Transcription factor binding sites in humans (70% of known motifs recovered)
Protein secondary structures (alpha helices, beta sheets)
Mobile genetic elements such as prophages and CRISPR spacers
Open reading frames (ORFs), intergenic regions, tRNAs, and rRNAs

Even more remarkably, this knowledge transfers across species. Features identified on the human genome also work on a woolly mammoth genome fragment 52,000 years old.

This result illustrates a phenomenon already observed in text-based LLMs: large models develop internal representations that correspond to real semantic concepts, without explicit supervision.

Result #3: generating experimentally validated functional DNA

Evo 2 is not just a predictive model. It is also a generative model. It can complete genes from genomic context and, most importantly, generate entire DNA sequences at genome scale.

The researchers demonstrated the generation of:

Complete mitochondrial genomes (~16 kb) with the correct number of coding genes, tRNAs, and rRNAs
Prokaryotic genomes (~580 kb, M. genitalium-like) where 70% of annotated genes have known functional homologs
Eukaryotic sequences (~330 kb, yeast chromosome-like) with genes, introns, promoters, and tRNAs

But the most spectacular result concerns chromatin accessibility design. By coupling Evo 2 with chromatin prediction models (Enformer and Borzoi) via inference-time beam search, the authors generated multi-kilobase DNA sequences with controlled chromatin accessibility profiles.

To prove the concept in a memorable way, they literally wrote messages in Morse code in the epigenome of mouse embryonic stem cells: "LO", "ARC", and "EVO2".

The synthetic sequences were manufactured, inserted into mouse cell genomes, and the experimentally measured chromatin profiles matched predictions with 92 to 95% accuracy (AUROC).

We are no longer talking about prediction. We are talking about programmable biological design.

What this concretely changes for biotech and pharma

The implications for industry are direct:

Accelerated pathogenic variant screening. Evo 2's zero-shot capability allows scoring mutation impact on understudied genes without requiring specific training datasets. This is a major asset for rare diseases and variants of uncertain significance (VUS).

Genomic annotation without relying on existing databases. Evo 2 embeddings, combined with lightweight classifiers, outperform classic tools like AUGUSTUS for exon classification, including on non-model organisms.

Custom regulatory sequence design. The generative model + scoring model coupling paves the way for rational design of enhancers, promoters, and synthetic regulatory elements for gene therapy.

A generalisable framework. Evo 2's key paradigm (generative model + scoring function + inference-time search) is applicable to any phenotype for which a predictive model exists. This is a biological design architecture, not a one-off tool.

Biosafety: a responsibly designed model

An important point: genomes of eukaryote-infecting viruses were deliberately excluded from the training data. Evaluations show that the model has high perplexity on these sequences and essentially random performance for human viral protein generation.

This is one of the most comprehensive risk mitigation efforts for an open-source biological foundation model.

Fully open source

The model is entirely open source:

7B and 40B model weights on Hugging Face
Training and inference code on GitHub
Complete OpenGenome2 dataset
Evo Designer web interface for generation and scoring
SAE feature exploration tool: Evo Mech Interp

Conclusion

Evo 2 marks a turning point. For the first time, a language model trained solely on raw DNA achieves state-of-the-art performance in human variant prediction, spontaneously learns complex biological concepts, and generates experimentally validated functional DNA sequences.

We are entering an era where AI no longer processes just human language, but the very language of life.

Reference: Brixi, G. et al. Genome modelling and design across all domains of life with Evo 2. Nature (2026). DOI: 10.1038/s41586-026-10176-5

Working in biotech or pharma and want to understand how AI can accelerate your R&D processes? Contact SG AI Solutions for an AI audit tailored to your challenges.