Why this evidence?
It is linked to a specific model version, scientific task, dataset, split, metric and source. That makes the claim inspectable rather than a detached marketing score.
ESM3: De novo protein or binder design
Reported multimodal generation includes an experimentally characterized designed protein.
Model versionESM3
TaskDe novo protein or binder design
DatasetGenerative protein evaluations
SplitGenerative protein evaluations and experimental fluorescent-protein example
MetricSequence, structure and function
Replicationexperimental-primary
Reported byModel developers
Review statuscurated
Evidence confidence · strong
Confidence is multidimensional, not a universal model score.
Evidence completenessstrongModel version, task, dataset, split, metric, source and provenance fields.
Independent validationstrongexperimental-primary
Source qualitystrongPeer-reviewed / developer-reported
Reproducibility evidencestrongReflects documented replication status, not a universal reproducibility score.
Version specificitystrongESM3
Context applicabilitystrongDepends on task, split and explicit caveats; users must still validate their own context.
Contradiction reviewclearNo direct contradiction signal is currently queued.
BioAtlas reports evidence dimensions separately so a strong source cannot hide weak applicability, incomplete replication or unresolved contradiction.
Why?
Why should this evidence influence a decision?
Why not a universal score?
Performance can change with dataset, split, preprocessing, metric and context of use. BioAtlas therefore keeps confidence dimensions separate.
What could change the conclusion?
Independent replication, a better matched prospective dataset, a version change, a contradictory result or a more relevant validation protocol can reopen this evidence record.
Evidence boundary
What this claim does not prove.
- A demonstration protein is not a universal measure of design performance.
- Hosted and open variants may differ in capability and access.
BioAtlas groups benchmark claims only when task, dataset, split, metric and protocol context align. This record is not a universal model score.
Model context
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.
Full evidence passport →Known model 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.