HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Virtual Cell ChallengeArc Institute—

Evo genomic foundation models

by Arc InstituteUS

PilotStage 3 of 5

Evo 2 released openly in Feb 2025; used to predict mutation effects and design DNA.

Updated Feb 2025Checked 25 Sep0 updates this week

Milestones

Next · Extend design to complex (eukaryotic) genomes
  1. Evo 1 preprintFeb 2024Complete.
  2. Evo published in ScienceNov 2024Complete.
  3. Evo 2 released with NVIDIAFeb 2025Complete.
  4. AI-designed virus genomes reportedSep 2025Complete.
  5. Extend design to complex (eukaryotic) genomesNowCurrent milestone.

Most important updates

  • Feb 2025
  • Jun 2024

Current obstacles

  • BiosecurityAI that writes genomes could be misused; Arc left viruses that infect people, animals and plants out of Evo 2's training data.

Physics limits

  • Genomes are long; context is limitedEvo 2 reads up to about a million DNA letters at once, but the human genome has 3.1 billion, and switches can act from a million letters away. Whole-genome reasoning remains out of reach.
  • DNA says what, rarely how wellMost sequence data has no label for what it does or how strongly. Models learn what looks natural, but predicting precise function needs lab measurements that are scarce.
  • Designs must still be built and testedA generated genome is only a hypothesis. It must be synthesised and grown to see if it works, and most of biology's complexity shows only in living cells.

How it works

3 parts
NVIDIA H100 GPU boards inside a supercomputer: the same kind of chip the Evo 2 DNA models were trained on
NVIDIA H100 GPU boards inside a supercomputer: the same kind of chip the Evo 2 DNA models were trained onPhoto: Fukumoto · CC BY-SA 4.0 (opens commons.wikimedia.org)
Data

DNA from all of life

Trained on trillions of DNA letters from bacteria, plants, animals and viruses.

Model

Reads very long stretches

It can take in about a million DNA letters at once, seeing how genes fit into whole genomes.

Use

Score and generate

It flags harmful mutations and writes new sequences, which are then made and tested in the lab.

Papers & demos

3 items
  1. Feb 2025paper
    Genome modeling and design across all domains of life with Evo 2 (preprint) (opens doi.org)Scaled the model up and released it openly with NVIDIA.
  2. Nov 2024paper
    Sequence modeling and design from molecular to genome scale with Evo (Science) (opens doi.org)First Evo model: showed a DNA AI can design working CRISPR systems.
  3. Jun 2024paper
    Bridge RNAs direct programmable recombination of target and donor DNA (Nature) (opens doi.org)Introduced a new way to program DNA rearrangements.

Update log

2 updates

Feb 2025

  • Major: BlogResearch

Jun 2024

  • Minor: PaperResearch

About Arc Institute

The team behind Evo genomic foundation models

Arc Institute

Evo DNA AI, virtual cell, bridge editing

Nonprofit research institute in Palo Alto founded in 2021 by Patrick Hsu, Silvana Konermann and Patrick Collison, with Stanford, Berkeley and UCSF. It funds scientists without grant applications and builds shared tools like the Evo DNA models.

  • Evo 2 (Feb 2025, with NVIDIA) learned from DNA of all known life and can predict effects of mutations and write new sequences.
  • In 2024 Arc researchers showed 'bridge RNAs' that guide enzymes to insert, delete or flip DNA.
  • The second Virtual Cell Challenge launched on 20 Aug 2026.

arcinstitute.org (opens arcinstitute.org)Company profile →

Founded
20215 yrs
Headquarters
United States
Status
Private
Valuation
private
Works in
BiotechSynthetic biologyAI
Coverage
2 programs · 4 updateslatest 20 Aug 2026checked 25 Sep
Partners
NVIDIA (opens nvidia.com)Stanford University (opens stanford.edu)
People
Patrick Hsu (opens arcinstitute.org)Co-founder and core investigatorSilvana Konermann (opens arcinstitute.org)Co-founder and executive directorPatrick Collison (opens en.wikipedia.org)Co-founder
Primer