Protein Structure · Meta AI (FAIR)
ESMFold
Predicts all-atom 3D protein structures directly from single amino acid sequences without alignments in seconds.
What ESMFold is
ESMFold is an end-to-end atomic-level protein structure prediction model developed by the Meta Fundamental AI Research Protein Team (FAIR). While traditional methods like AlphaFold2 rely on building multiple sequence alignments (MSAs) by searching large sequence databases, ESMFold generates 3D atomic coordinates directly from individual amino acid sequences without requiring any alignment search or external database lookups.
Bypassing sequence alignment search allows ESMFold to execute inference orders of magnitude faster than MSA-based predictors. This capability enables high-throughput structural analysis across hundreds of millions of sequences and makes it possible to fold orphan proteins that lack known homologous sequences, a feature demonstrated by the generation of over 600 million predicted structures in the ESM Metagenomic Atlas.
The model outputs atomic coordinate files alongside local and global quality assessments via predicted Local Distance Difference Test (pLDDT) scores. While its prediction accuracy trails AlphaFold2 on targets with deep sequence alignments, it closes much of the accuracy gap on targets lacking homologous alignments.
How it works
ESMFold couples the ESM-2 transformer protein language model with an end-to-end folding trunk. Instead of extracting evolutionary co-evolution patterns from an explicit multiple sequence alignment, the folding module translates structural and biochemical patterns learned directly by ESM-2 during self-supervised language model pre-training into 3D all-atom Cartesian coordinates.
The full architecture incorporates a 36-layer, 3-billion-parameter ESM-2 backbone coupled with a 48-layer, 690-million-parameter folding module. Training was performed using experimental structures from the Protein Data Bank (PDB) and sequence data from UniRef50 (UR50/D 2021_04), utilising iterative structure refinement recycles (defaulting to four recycles during inference).
Key facts
- Model parameters
- 690M folding trunk (+ 3B ESM-2 backbone)
- Architecture layers
- 48 folding layers (+ 36 ESM-2 layers)
- Training data
- Protein Data Bank (PDB) and UniRef50 (UR50/D 2021_04)
- Input modality
- Single protein sequence (amino acids)
- Output modality
- All-atom PDB structure and pLDDT confidence scores
- Default recycling iterations
- 4
- Hardware requirement
- 16 GB GPU memory
What you can use it for
- High-throughput metagenomic structure characterisation
- Rapidly fold millions of uncharacterised sequences from environmental sequencing datasets without the computational overhead of MSA generation.
- Structure prediction for orphan proteins
- Predict all-atom coordinates for isolated, non-conserved, or orphan sequences that lack known sequence homologs in evolutionary databases.
- Screening computationally designed sequences
- Evaluate de novo synthetic constructs or generative protein designs directly from sequence to assess predicted fold conformation and confidence.
- Protein complex and multimer modeling
- Predict multimeric protein interactions and complexes by supplying individual protein chains separated by a colon character.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Run this model
ESMFold
Strengths and limits
Good at
- Performs structure prediction orders of magnitude faster than MSA-dependent algorithms.
- Requires no external sequence databases or homology searching tools during inference.
- Effectively models orphan sequences that have few or no known homologous relatives.
- Provides per-residue pLDDT confidence metrics to highlight high-confidence structural regions (scores above 70).
- Supports memory optimisation options including axial attention chunking and CPU parameter offloading.
Watch out for
- Prediction accuracy trails AlphaFold2 on target proteins that have deep, diverse multiple sequence alignments.
- Inference on long sequences can cause out-of-memory errors on standard GPUs unless memory chunking or CPU offloading is enabled.
- Local environment setup for the native repository requires NVIDIA CUDA compilation tools (nvcc) and Python versions no higher than 3.9.
Inputs and outputs
You provide
Protein sequence
Single-letter amino acid code. Whitespace and FASTA headers are stripped automatically.
Max residuesoptional
The free public endpoint caps at 400 residues; with an NVIDIA key, longer chains route to the GPU-backed ESMFold NIM automatically.
You get back
All-atom coordinates in PDB format, viewable and downloadable.
pLDDT score per residue (0–100). Above 70 is generally reliable.
Global confidence in the predicted fold.
Licence
Commercial use permitted · MIT
ESMFold's weights and code are MIT-licensed. Runs here against the public ESM Atlas endpoint, whose own fair-use terms apply to the hosted service.
Questions
Can I use ESMFold commercially?
Commercial use permitted. ESMFold's weights and code are MIT-licensed. Runs here against the public ESM Atlas endpoint, whose own fair-use terms apply to the hosted service.
How do I run ESMFold without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 4.0s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
Does ESMFold require a multiple sequence alignment (MSA)?
No. ESMFold makes predictions directly from single sequences and does not perform database searches or build MSAs.
How should multimeric complexes be input into ESMFold?
Multimers can be modeled by supplying the individual sequence chains concatenated into a single input string separated by a colon delimiter.
How are prediction confidences interpreted in the output?
The model outputs per-residue pLDDT scores from 0 to 100 stored in the B-factor column of the generated PDB file, where values above 70 indicate generally reliable backbone conformations, along with a global mean pLDDT score.
How can memory usage be managed on long protein sequences?
Users can set the axial attention chunk size to values such as 128, 64, or 32 to reduce memory complexity from O(L^2) to O(L), or activate CPU offloading flags to avoid GPU memory exhaustion.
How to cite it
Zeming Lin et al., Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123-1130 (2023). DOI: 10.1126/science.ade2574
Written from ESMFold's own documentation (model card, code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.