ESMFold

ESMFold online server overview

ESMFold predicts a protein's 3D structure from its amino acid sequence alone, with no multiple sequence alignment to build first. ProteinIQ runs ESMFold online on GPUs with Meta's public esmfold_v1 model from fair-esm 2.0.0. Paste one sequence or up to 10 chains, and the server returns a PDB model with per-residue pLDDT, pTM and an optional full PAE matrix.

The model pairs the 3-billion-parameter ESM-2 protein language model with a folding trunk and structure module. Skipping the alignment search is what makes it fast, and also what limits it: on hard targets with few known relatives, MSA-based AlphaFold2 is still more accurate.

ESMFold server and the ESM Metagenomic Atlas

Meta's ESM Metagenomic Atlas hosts a public fold page and an API that run the same esm.pretrained.esmfold_v1 model. The fold page takes a single sequence of up to 400 residues, and the Atlas notes that the server limits sequence length and the number of requests per user. The foldSequence/v1/pdb/ endpoint takes a sequence by POST and returns a PDB file.

Meta describes those APIs as rate limited and not appropriate for folding a large set of proteins; for that it recommends downloading the model from GitHub. The Atlas confidence endpoint, which returns PAE, pTM and pLDDT, serves the precomputed Atlas entries by MGnify ID rather than new sequences.

ProteinIQ is not affiliated with Meta, and runs are paid in credits. In return it folds up to 2,000 residues per job, accepts multimers of up to 10 chains, lets you set the number of recycles, and returns pTM and an optional PAE matrix for your own sequence.

Jobs are saved in your workspace history and open in a 3D viewer. A predicted model can go straight to MolProbity, fpocket or a Ramachandran plot, and on paid plans the API quickstart submits an ESMFold job from a short Python script.

Pricing

ESMFold jobs start at 50 credits, and any input up to 800 residues costs 50 credits at the default 4 recycles. Above that, the quote grows with the total residue count, which also selects a larger GPU, and changes slightly with the number of recycles. The exact quote is calculated before submission.

Total residues	0 recycles	4 recycles (default)	8 recycles
600	50	50	50
800	50	50	56
900	50	54	65
1,000	73	86	103
1,250	96	113	136
1,500	174	205	246
2,000	249	293	352

For a multimer, the residues of all chains are added together, so two 300-residue chains cost the same as one 600-residue chain. Save PAE matrix does not change the price.

Inputs

Input	Accepted forms	Limits
Protein	Pasted sequence or FASTA, .fasta, .fa or .txt file, or a sequence fetched by RCSB PDB ID	Up to 10 chains and 2,000 residues in total; files up to 10 MB

Each chain is a separate entry. ESMFold joins them internally with its standard poly-glycine linker, leaves the linker out of the PDB, and writes each chain under its own chain ID.

Letters outside ESMFold's amino-acid alphabet are converted to X rather than deleted, so residue numbering stays aligned with the input. Gaps (- and .), spaces and other non-letter characters are removed. Inputs longer than 2,000 residues are rejected before submission; AlphaFold2 handles larger proteins.

Settings

Parameter	Type	Default	Description
Number of recycles	integer	4	Refinement passes, 0 to 8. 0 runs a single forward pass; 4 is the model's training value. Each step away from 4 changes the quote by 5 percent, within 0.85 to 1.2 times.
Save PAE matrix	boolean	false	Writes the full predicted aligned error matrix as a JSON file next to the structure.

Recycling refines the model ESMFold already has; it does not add the evolutionary information an alignment would. A protein that stays low-confidence at 4 recycles is a candidate for an MSA-based model rather than for more recycles.

Outputs

Results open in three tabs. Viewer shows the model in 3D, Data lists the prediction with its scores, and Files holds the downloads. Each job returns one model.

Output	Format	Contents
prediction_rank01.pdb	PDB	Predicted structure with per-residue pLDDT, on a 0 to 100 scale, in the B-factor column.
prediction_rank01_pae.json	JSON	Optional predicted_aligned_error matrix in ångströms, written when Save PAE matrix is on.
Prediction scores	Table	Mean plddt, ptm and max_predicted_aligned_error.

Understanding results

pLDDT

ESMFold writes pLDDT into the B-factor column, so most viewers can color by it directly. The plddt score in the Data tab is the mean over the model.

pLDDT	Confidence	Interpretation
Above 90	Very high	Backbone and side chains accurate
70 to 90	Good	Reliable backbone prediction
50 to 70	Low	Treat with caution
Below 50	Very low	Likely disordered or unstructured

For ESMFold, pLDDT is a good guide to whether to trust the model at all. On the CAMEO benchmark, ESMFold predictions with pLDDT above 70 reached an LDDT of 0.83, against 0.85 for AlphaFold2. Very confident predictions, above 90, had a median backbone RMSD95 of 0.94 Å.

Low pLDDT has two causes here. A region may be intrinsically disordered and only fold on binding a partner, which shows up as extended ribbons in an otherwise confident model. Or the whole protein may score low because the language model has little to go on.

Lin and colleagues found that accuracy tracks the language model's perplexity on the sequence, with a correlation of -0.55 against TM-score on CAMEO. A natural protein that scores low throughout is the case where an MSA-based model is worth running.

pTM and PAE

pTM estimates the TM-score of the whole model. Above 0.5 suggests the overall fold is right. For a multimer, pTM covers all chains together; ESMFold reports no interface score such as ipTM.

max_predicted_aligned_error is the top of ESMFold's PAE scale, 31.75 Å, rather than a property of your model. It is the ceiling that every value in the PAE matrix sits under.

To judge domain packing or a multimer interface, turn on Save PAE matrix and read the off-diagonal blocks: low error between two chains supports their arrangement, high error says it is a guess. ESMFold2 adds a pair-chain ipTM, and the ipTM vs pTM explainer covers how to read it.

Accuracy relative to AlphaFold2

ESMFold trades some accuracy for speed. In the ESMFold paper, it reached an average TM-score of 0.83 on CAMEO and 0.68 on CASP14, where AlphaFold2 reached 0.88 and 0.85. The gap is concentrated in the harder CAMEO targets, which have shallower alignments and are better predicted by AlphaFold2.

The comparison changes when AlphaFold2 has no alignment to use. With MSAs removed, other models' performance fell well below ESMFold's, so for a single sequence with nothing to align, such as a designed protein, ESMFold is a fair first choice. For a natural protein whose family is well represented in sequence databases, AlphaFold2 with an MSA usually gives the better model.

How ESMFold works

ESM-2 is trained by masked language modeling: random amino acids are hidden and the model learns to predict them from the rest of the sequence. Across millions of sequences, its internal representations come to encode evolutionary constraints, and its attention patterns correspond to residue-residue contacts in 3D structures.

ESMFold feeds those representations into a folding trunk and a structure module similar to AlphaFold2's, which produce atomic coordinates. A recycling phase passes the output back through the network to refine it.

Skipping the alignment search is where the speed comes from. A 384-residue protein took 14.2 seconds in the paper's benchmark, six times faster than a single AlphaFold2 model, and the authors report a speedup of up to 60 times on the inference forward pass alone.

When to use ESMFold vs other tools

Tool	Best for	Key distinction
ESMFold	Fast single-sequence protein folding	No MSA, one model per job, pLDDT, pTM and optional PAE
ESMFold2	Protein chains and complexes	Newer ESMFold-family model with optional PAE and pair-chain ipTM
MiniFold	Faster monomer prediction	Smaller single-sequence model
AlphaFold2	Natural proteins with many homologs	MSA-assisted monomer and multimer prediction with ranked models
Boltz-2	Protein-ligand complexes and affinity	Handles ligands, DNA and RNA; returns affinity estimates
Chai-1	Guided multi-component complexes	Handles ligands, DNA and RNA with contact and pocket inputs
Protenix	All-atom complexes	Open AlphaFold 3 reproduction for proteins, nucleic acids, ligands and ions

Use ESMFold when you need a structure quickly, are working with designed proteins that lack natural homologs, or want to screen many sequences before running more expensive predictions. Running it side by side with ESMFold2 and MiniFold is the approach described under single-sequence protein structure prediction. Use Boltz-2 or Chai-1 when ligands, DNA or RNA are part of the question.

Limitations

ESMFold predicts proteins only. It cannot place ligands, nucleic acids or ions, and it returns one model per job, so there is no ensemble of samples to compare. Its accuracy depends on how well the language model knows the sequence family, which is why low-confidence natural proteins are better served by an MSA-based method. Multimer predictions come without an interface confidence score.

---

References

• Lin, Z., et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. https://doi.org/10.1126/science.ade2574
• Lin, Z., et al. (2022). Evolutionary-scale prediction of atomic level protein structure with a language model. bioRxiv preprint. https://doi.org/10.1101/2022.07.20.500902