The computer-generated sequence for esmGFP is only 58% similar to its closest known fluorescent counterpart — a human-modified protein derived from bubble-tip sea anemones (Entacmaea quadricolor). The remaining 42% of the sequence, created by the AI model, would have required 96 genetic mutations and 500 million years to evolve naturally.
ESM3: The AI Behind esmGFP
Researchers at EvolutionaryScale unveiled esmGFP using an AI model called ESM3, introduced in a preprint study in 2024.
Now peer-reviewed and published in Science, ESM3 operates differently from natural evolutionary processes. Instead of following evolution’s constraints, ESM3 solves protein-design problems by filling in gaps in incomplete protein sequences.
“We’ve found that ESM3 learns fundamental biology, and can generate functional proteins outside the space explored by evolution,” said Alex Rives, co-founder and chief scientist of EvolutionaryScale, in an email to Live Science.
