ESMFold predicts structure without alignments and folds 617 million proteins

Lin and colleagues at Meta published ESMFold in Science on 16 March 2023. It predicts a protein's structure from its sequence alone, with no search for related sequences, because the evolutionary information such a search supplies is already inside a language model of 15 billion parameters. They used that speed to fold over 617 million metagenomic proteins.

Background

Predicting a protein’s shape had depended on finding its relatives. The method gathers sequences of similar proteins from other organisms, lines them up, and reads the evolutionary constraints out of which positions vary together. That alignment is where the information comes from.

It has a cost and a limit. The cost is a database search for every protein. The limit is that a protein with no known relatives has no alignment, and returns nothing.

Metagenomic proteins are exactly that case. They come from organisms sampled out of soil or seawater and never grown in a laboratory, so the proteins we understand least were the ones the method could not reach.

What happened

Meta published ESMFold in Science on 16 March 2023, having released the preprint the previous year.

They scaled a language model trained on protein sequences to 15 billion parameters, and found that at that size a picture of atomic structure emerged in what the model had learned. The evolutionary information an alignment would have supplied was already inside the weights, paid for once during training.

So the alignment could be dropped. A structure comes from a single forward pass over one sequence, about an order of magnitude faster than searching a database first.

They spent that speed on scale. The ESM Metagenomic Atlas, released in November 2022, holds predicted structures for over 617 million metagenomic proteins, of which more than 225 million are predicted with high confidence.

What followed

The trade is accuracy for reach. ESMFold is generally not as accurate as AlphaFold 2, and the entry does not claim otherwise. What it can do is answer at all for sequences that have no relatives to align against, and answer fast enough to run over hundreds of millions of them.

The code was released under a permissive licence and the structures under Creative Commons, and the Atlas grew by another 150 million structures in an update with the European Bioinformatics Institute.

Why it mattered

Existing methods needed an alignment of related sequences, which meant a database search for every protein and no answer at all where relatives were unknown. That is precisely the case for organisms sampled from soil and seawater, so the proteins least understood were the ones the method could not reach.

Sources

• Evolutionary-scale prediction of atomic-level protein structure with a language model. science.org. Primary source
• Evolutionary-scale prediction of atomic level protein structure with a language model. biorxiv.org. Primary source
• Evolutionary Scale Modeling. github.com. Official

Cite this page

AI Achievements. (2023). ESMFold predicts structure without alignments and folds 617 million proteins. Retrieved 2026-08-29, from https://achievements.ai/milestone/esmfold-metagenomic-atlas

@misc{achievements_esmfold_metagenomic_atlas,
  title  = {ESMFold predicts structure without alignments and folds 617 million proteins},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/esmfold-metagenomic-atlas}
}