Learning Life's Language
at Scale.
Welcome to the interactive deep dive into Evo 2, the world's first generalist genomic foundation model. Trained on 9.3 trillion letters of raw DNA across all domains of life, it learns biology without human labels.
The Scale of the Breakthrough
Evo 2 reads, understands, and writes DNA across bacteria, archaea, eukaryotes, and phage. It treats DNA like a language, and it just became fluent.
How to Use This Report
Use the navigation to explore the 6 critical benchmarks. Interact with the charts and visualizers to understand how Evo 2 turns theoretical biology into programmable code.
Primary Source — Peer-Reviewed Research
Sequence modeling and design from molecular to genome scale with Evo 2
Arc Institute & Stanford University · Nature, 2026 · DOI: 10.1038/s41586-026-10176-5
This interactive report is a technical analysis and commentary derived directly from the above peer-reviewed publication. All primary findings, benchmark results, and experimental data belong to the original authors.
01. The 1-Million Letter Memory
The Analogy: Imagine reading the entire Lord of the Rings trilogy in one sitting, and instantly recalling a specific three-word phrase buried on page 412. That is what a 1-million token context window means for AI.
What this section shows: This section demonstrates the "Needle-in-a-Haystack" benchmark. Previous DNA models "forgot" context after about 8,000 to 128,000 bases. Real biology operates at massive scales (like distant enhancers regulating a gene 500,000 bases away). Evo 2 can hold an entire small bacterial genome in its memory at once, perfectly recalling a 100-base-pair "needle" sequence hidden within 1,000,000 random bases.
Needle Recall Accuracy by Sequence Length
Interactive ChartHover over data points to see the dramatic drop-off of legacy models versus Evo 2's sustained memory.
02. Polyglot: Genetic Dialects
The Analogy: It's like someone learning multiple regional dialects of a language just by listening to crowds, without ever being handed a dictionary. They instantly know a word that means "Stop" in London means "Keep going" in Sydney.
What this section shows: While most life uses the same genetic code, some organisms use slight variations. For example, the codon TGA usually tells the cellular machinery to STOP making a protein. But in Ciliates, TGA codes for an amino acid (Cysteine). Without human labeling, Evo 2 learned these dialects purely from raw sequence statistics. Interact below to see how Evo 2 shifts its interpretation of the exact same DNA based on the organism's context.
Test Sequence: ATG GGC TGA CCA
Evo 2 Translation Prediction:
In standard genomes, TGA is a premature stop codon. Evo 2 correctly identifies this as highly detrimental to the protein structure.
03. The Zero-Shot Clinician
The Analogy: Imagine a master editor who has read millions of books. You hand them a single sentence with a subtle mutation — a wrong word, a misplaced clause — and they catch it instantly, no reference book needed. Now imagine that "grammatical error" isn't a misplaced comma. It's cancer.
What this section shows: This section visualizes Evo 2's ability to evaluate human genetic variants (like mutations in the BRCA1 cancer gene). "Zero-shot" means it scores these mutations purely by predicting how "surprising" or "unlikely" the mutated sequence is, based on its training on healthy genomes. It matches or beats models that were explicitly trained on clinical hospital data.
BRCA1 Variant Pathogenicity Scoring (AUROC)
Performance ComparisonHigher scores approach 1.0 (perfect prediction). Notice Evo 2 (Unsupervised) competing closely with models that required expensive clinical labeling.
04. The Internal Textbook (SAE)
The Analogy: We didn't give the AI a biology textbook; we gave it an ocean of letters. By analyzing its own internal brain waves (using Sparse Autoencoders), we found it wrote its own textbook. It independently invented concepts like "paragraphs" (exons) and "punctuation" (promoters).
What this section shows: Sparse Autoencoders (SAEs) allow researchers to look inside the model's neural network to see what features it is paying attention to. Below is a simulated sequence. Click the "biology concepts" to see what parts of the raw sequence Evo 2 automatically identifies as having functional meaning, all without human annotation.
Probe the AI's Brain:
GATCTCGTAGCTAGCCTAGTACGATCATGGCCGATTACTATAGCGCGTACCTAGGCATTTTAAGCATGC
CTAGCAATCGTAGCGCTAGCTAGCTAGCTAGCTAGGCTAGCTAGCTAGCGGGCGGCTAGCTAGCTAG
05. Genome Architect
The Analogy: Older DNA AI could write a sensible five-word sentence. Evo 2 can generate a fully formatted, 300-page functional technical manual from scratch — complete with working schematics. Working schematics, for life.
What this section shows: This visualizes the sheer scale of de novo (from scratch) generation capabilities. Evo 2 isn't just predicting mutations; it can generate entire functional sequences that pass in-silico structural tests (AlphaFold verifies the generated proteins actually fold properly). It scales from organelles up to complete eukaryotic chromosomes.
Human Mitochondria
~16 kbGenerates correct number of tRNA/rRNA genes and complex synteny.
M. genitalium (Bacteria)
~580 kb70% of predicted genes have valid Pfam protein hits.
S. cerevisiae (Yeast) Chr III
~330 kbFirst generation of complex eukaryotic chromosomes with introns and promoters.
06. Epigenetic Morse Code
The Analogy: Until now, we could observe the orchestra — cataloguing every instrument, every note they've ever played. Evo 2 hands us the conductor's baton. We compose the score. We hand it to a living cell. It plays exactly what we wrote — on demand. The team proved it by spelling "EVO2" in chemical signals across living mammalian cells. The cell didn't approximate it. It executed it.
What this section shows: This is the section where biology becomes programmable. Epigenetics govern whether genes are switched on or off — not by changing the DNA sequence itself, but by controlling how tightly it is wound. For decades this layer felt almost unreachable: we could read it, but not reliably write to it. Evo 2 changes that. Researchers designed synthetic DNA sequences that forced live human and mouse cells to open their chromatin in precise, pre-specified patterns — verified by ATAC-seq in real lab conditions. This is not stochastic. Not approximate. The model produced sequences that drove the cell to a chosen state. We are no longer waiting to see what biology randomly does. We are telling it what to do.
Designed ATAC-seq Peaks: "EVO2" Pattern
Live Cell ValidationEach peak is a designed sequence forcing chromatin to open at a specific location. This is not a natural signal — it is a human-specified command, executed by a living cell.
☶ Appendix
Glossary of Terms
Foundation Model
A giant neural net trained on massive data to predict the next "letter". Becomes a Swiss-army knife for many tasks without needing to be retrained from scratch.
Zero-Shot
When the AI answers questions it was never explicitly taught (e.g., "is this mutation bad?") simply by relying on the deep patterns it learned during general training.
Sparse Autoencoder (SAE)
A mathematical tool used by researchers to compress the model's complex internal "thoughts" into human-readable concepts, revealing how the AI understands biology.
ATAC-seq
A real-world lab experiment that measures which parts of DNA are "open" and readable inside a living cell. Used here to prove the AI's designs actually work physically.
StripedHyena 2
The specific, highly efficient neural network architecture that allows Evo 2 to process sequences 1 million letters long without crashing standard computing clusters. It mixes convolutions with attention mechanisms.
Sources & Citations
This report is a single-source analytical commentary on the Evo 2 paper (Nature, 04 Mar 2026).
Report structure informed by OSINT Research Protocol v1.0
— March 2026