ESM3
Multimodal masked generative model over protein sequence, 3D structure, and function, aligned with geometric and wet-lab fluorescence feedback.
The loop
- Proposes
A 98B multimodal protein model iteratively unmasks sequence, 3D backbone coordinates, and catalytic/functional tokens.
- Judged by
All-atom geometric consistency metrics (TM-score, active-site RMSD), preference alignment rewards, and wet-lab chromophore fluorescence assays judge candidates.
- Steers
Chain-of-thought structural decoding and direct preference alignment on verified structures steer generation toward folded functional proteins.
Headline result
Generated esmGFP, a glowing fluorescent protein sharing only 58% sequence identity with the closest known fluorescent protein (~500 million years of evolutionary divergence).
What this does not prove
Wet-lab expression success rates vary strongly by target protein fold and cellular host environment.
How it generalizes the harness
- A domain authority judges
- The output outlives the model
- Feedback retrains the generator
- The generator is not an LLM
3D geometric consistency verifiers and wet-lab functional assays filter and align a multimodal protein generator.
Classification
| What it proposes | Protein structure |
|---|---|
| Generative model | ESM3 98B multimodal masked protein model |
| Algorithmic authority | 3D structural consistency verifier and wet-lab fluorescence assay |
| Feedback | Reward signal |
| What gets updated | Model weights and population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |