Structure prediction · Meta AI · 2023

ESMFold

Predicts a structure from a single sequence with no alignment search at query time. A frozen ESM-2 supplies sequence representations to a folding trunk and a structure module derived from AlphaFold2.

In plain language

It folds a protein from the sequence alone. No search for relatives, no waiting on a database. The language model already read enough proteins that its internal state stands in for the alignment.

One way to picture it

A translator who has read the whole library does not stop to look words up. Faster, and slightly worse on the rare sentence where looking it up would have helped.

Commonly misread as

Faster does not mean better. When a deep alignment exists, an alignment-based predictor usually wins, and giving up that evidence is a trade you are choosing to make.

How it is put together

Select a component to read it

Sequence and pair features

Refined representations

Try it

Predict a structure and write a PDB file

import torch
from transformers import AutoTokenizer, EsmForProteinFolding

tok = AutoTokenizer.from_pretrained("facebook/esmfold_v1")
model = EsmForProteinFolding.from_pretrained("facebook/esmfold_v1").eval()

seq = "MALWMRLLPLLALLALWGPDPAAAFVNQHLCGSHLVEALYLVCGERGFFYTPKT"
inputs = tok([seq], return_tensors="pt", add_special_tokens=False)
with torch.no_grad():
    output = model(**inputs)

pdb = model.output_to_pdb(output)[0]
open("prediction.pdb", "w").write(pdb)

These snippets have not been executed here. Versions move; check the model card before trusting a line of it.

01 / Why it is here

Standing

Made the language-model route to structure practical at the scale of metagenomic surveys. It is a structure predictor built on ESM-2, not another name for the language model.

02 / What sets it apart

Distinctions

The alignment search disappears, which is where most of the speed comes from.
The language model is frozen. Only the folding trunk is trained.
Accuracy trails alignment-based prediction on targets with deep, informative alignments.

03 / Where it stops

Limits

When a rich alignment exists, the evidence it carries is real and a model that ignores it gives that up. Single-sequence folding is a trade, not a free improvement.

Weights and code

Checked against the registry, not from memory

Repository	Size	Licence	Note
facebook/esmfold_v1	ESM-2 3B plus folding trunk	MIT	—

Sources

Each number above comes from one of these

Lin et al. 2023, Science. Evolutionary-scale prediction of atomic-level protein structure with a language modelPeer reviewed ESMFold weightsModel card

Same task, other answers

Structure prediction

Structure prediction2021

AlphaFold 2

Evolution in, coordinates out.

Predicts the structure of a protein from its sequence, an alignment of its homologues and optional templates.

Explore the architecture Structure prediction2024

AlphaFold 3

One model for the whole complex.

Predicts the joint structure of complexes containing proteins, nucleic acids, small molecules, ions and modified residues.

Explore the architecture Structure prediction2026

ESMFold2

A language model encoder with an all-atom diffusion decoder.

Predicts all-atom structures of proteins and their complexes from ESM C representations, with an optional alignment for difficult targets.

Read the entry

Back to the catalogueCompare it with its neighbours