What it is
ESM3 is a frontier multimodal generative protein model trained on 2.78B proteins that can be prompted across sequence, structure and function simultaneously — 'programming' biology like code. To showcase it, the team generated esmGFP, a fluorescent protein ~58% distant from any natural one, equivalent to ~500M years of evolution. Founded by ex-Meta FAIR researchers; acquired by Chan Zuckerberg Biohub in 2025.
Evidence trail
BioAtlas keeps the path from source to decision visible. A connection records provenance; it does not imply that evidence is sufficient for every context.
Model passport
How ESM3 represents biology
Category is navigation. These fields describe the model-specific computational transformation and deliberately override broad category defaults.
Biological scale
Modalities & tasks
Registry, claims and frontier intelligence
ESM3
1 version record · latest curated year 2024. Model-family identity remains separate from capability and access changes.
Explore version lineage →1 normalized claim
De novo protein or binder design · Generative protein evaluations
Open claim intelligence →2 connected frontiers
Protein programming · Peer-reviewed capability
Inspect research horizon →Connected research frontiers
These records describe active research directions, not guaranteed capabilities of this model. Evidence stages and unresolved questions are preserved separately.
Multimodal protein programming
EvolutionaryScale · 2025-01-16Can one generative model reason jointly over protein sequence, structure and function and create functional proteins from mixed prompts?
Evidence boundary and unresolved questions
One striking protein demonstration does not establish general success across enzymes, therapeutics or complex multi-objective design tasks.
- How frequently do generated functions survive experimental testing?
- Can the model optimize potency, stability and safety together?
- How should synthetic training labels affect confidence?
multimodal · protein language model · function generation · synthetic biologyOpen frontier record →Prospective de novo binder generation
Google DeepMind · 2024-09-05Can AI generate high-affinity protein binders for diverse targets with fewer rounds of experimental optimization?
Evidence boundary and unresolved questions
Performance varies by target; the system did not succeed on every attempted target and is not publicly released for unrestricted reproduction.
- Which target properties predict designability?
- How transferable are success rates to membrane and flexible targets?
- Can developability, immunogenicity and function be optimized jointly?
protein binders · prospective validation · generative design · wet labOpen frontier record →Inputs and outputs
Inputs
Protein sequence, structure or functional annotationsOutputs
Protein sequenceStructureFunction-conditioned completionsScientific and technical profile
Scientific principles
Technology
Scientific lineage
These are transparent concept matches—not claims that one scientist alone caused this model. Each connection is based on the model’s recorded domain, scientific principles, technical terms or an explicit lineage link.
Anfinsen’s dogma—the thermodynamic hypothesis
Christian B. AnfinsenProtein structure prediction, inverse folding and generative protein design all assume that sequence strongly constrains structure and function.
Transformer self-attention
Ashish Vaswani and colleaguesProtein, genome, molecule and single-cell foundation models use attention to learn dependencies across biological sequences and multimodal inputs.
Information, entropy and communication
Claude E. ShannonSequence modelling, cross-entropy training, language models, mutual information and representation learning all use Shannon’s framework.
Protein sequence databases, evolutionary substitution matrices and computational comparison
Margaret Oakley DayhoffProtein language models, homology inference, multiple-sequence alignments and evolutionary priors inherit her conversion of sequence biology into computable data.
Atomic structures of biologically important molecules by X-ray crystallography
Dorothy Crowfoot HodgkinStructure-based drug design depends on the experimental structural tradition she helped establish.
Reading the sequences of proteins and DNA
Frederick SangerBiological foundation models exist because proteins and genomes became readable, comparable and computable at scale.
Evaluation evidence
Task-specific evidence only; not comparable as a universal leaderboard score.
Generative protein evaluations
ESM3 · Generative protein evaluations and experimental fluorescent-protein exampleReported multimodal generation includes an experimentally characterized designed protein.
Claim caveats
- A demonstration protein is not a universal measure of design performance.
- Hosted and open variants may differ in capability and access.
Known limitations
- Performance depends on the evaluation dataset and operating conditions.
- Task-specific benchmark results should not be compared across unlike domains.
- Outputs require task-specific scientific and experimental validation.
Milestones
Generated the novel esmGFP fluorescent protein.
$142M seed led by Nat Friedman, Daniel Gross, Lux Capital.