What researchers are trying
ESM3 tokenizes all three modalities in a shared masked-generative transformer and supports conditional generation from partial sequence, structural motifs or functional descriptions.
Multimodal protein programming
Can one generative model reason jointly over protein sequence, structure and function and create functional proteins from mixed prompts?
Why it matters
A unified model could let scientists specify what a protein should do and constrain how it should look without stitching together separate tools.
Evidence boundary
One striking protein demonstration does not establish general success across enzymes, therapeutics or complex multi-objective design tasks.
Organizations represented
EvolutionaryScale
Demonstrated evidence
What has actually been shown.
- Laboratory synthesis and fluorescence testing of esmGFP after two design rounds.
Unresolved questions
- How frequently do generated functions survive experimental testing?
- Can the model optimize potency, stability and safety together?
- How should synthetic training labels affect confidence?
Signals to watch next
- Prospective functional panels
- Specialized drug-design variants
- Open-model capability gaps
- Safety evaluations
Connected evidence graph
Related BioAtlas model passports.
ESM3
A generative model that reasons over sequence, structure & function at once.
4/7 evidence fields documentedChroma
Programmable protein generation with a diffusion 'grammar'.
3/7 evidence fields documentedProGen / ProGen2 / ProGen3
GPT-style language models that write functional proteins.
3/7 evidence fields documentedAlphaProteo
High-affinity binder generation from DeepMind.
2/7 evidence fields documentedPrimary and evaluation sources