Fold: predict protein structure from sequence

Protein sequences to complexes, visualized

  • Copies
Model
Sequence Length0 / 2048 max

Fold is an online protein structure prediction tool powered by the ESMFold2 model that predicts high-resolution, all-atom 3D structures of proteins and biomolecular complexes directly from sequence. Paste an amino acid, DNA, or RNA sequence, optionally with small-molecule ligands or modified residues, and the Fold server returns a full 3D structure with per-residue confidence scores. You can zoom in or rotate the resulting interactive model to view specific residues, features, molecular interactions or confidence scores, and you can download your result to export for downstream analysis.

How to predict a protein structure

  1. Paste a protein, DNA, or RNA sequence (up to 768 residues).
  2. Choose a model and adjust sampling settings if needed.
  3. Select "Run Fold" to generate the 3D structure.
  4. Inspect the structure and confidence metrics, or download the coordinates.

FAQ

What types of molecular sequences does Fold accept?

Fold accepts protein (amino acid), DNA, and RNA sequences. To fold ligands or more complex molecules that require specifying covalent bonds and other modifications, use ESMFold2 over the API.

What is the maximum sequence length that can be inputted into Fold?

Fold currently supports sequences up to 768 residues per prediction. To fold sequences beyond the app’s length limit, use ESMFold2 through Biohub’s hosted API, subject to its limits, or run the model on your own or a cloud GPU.

How can I run multiple Fold jobs?

To run multiple jobs, use the Batch API to run ESMFold2.

What does Fold return?

Fold outputs all-atom 3D coordinates (backbone and side chains), sparse autoencoder (SAE) features, confidence metrics including pLDDT, pAE, and pTM.

Which model powers Fold?

Fold runs on ESMFold2, Biohub’s state-of-the-art structure prediction model. ESMFold2 takes biomolecular sequences including protein, DNA, and RNA, processing them using deep learning models trained on structural data from the Protein Data Bank (PDB) and the AlphaFold DB (AFDB). Instead of querying databases of known structures, it generates 3D molecular predictions directly from sequence data. See the model page for architecture, benchmarks, and the research paper: Language Modeling Materializes a World Model of Protein Biology.

Do I need to install or download anything to use Fold?

No, Fold is an online protein folding tool that runs in your browser so there is no need to download software. For running longer sequences, multiple jobs, or incorporating into a workflow, the same model is available via the Biohub API and the esm Python package.

Is Fold free to use?

The Fold server is free to use. For additional details, see the Biohub.ai Terms of Use.

How do I choose what loops and sampling steps to run with my sequence?

Loops are iterative refinement passes through the folding trunk of the ESMFold2 model that Fold uses to improve the predicted structure of the sequence. Three loops is a good default value for monomers, while 10 is a good default value for multimers. For best results, use 20 loops. The sampling steps refer to the diffusion ordinary differential equation (ODE) solver steps of the ESMFold2 model. The default and recommended setting is 100 steps. Lower step settings can be chosen for lower quality folds that are generated faster.

I want to run multiple jobs or incorporate folding into my workflow. How can I access the ESMFold2 model variants and weights?

The ESMFold2 models can be accessed through the Biohub Platform API or through Hugging Face Transformers. The Hugging Face implementation is recommended for experimentation and rapid prototyping, while we recommend the esm package for production deployments.

Can I use MSA with the Fold app?

The Fold app does not support MSA (multiple sequence alignment), but users can use ESMFold2 over the API or through local inference with the option to input MSA. Optional MSA input can improve model performance on challenging sequences.

What is MSA?

MSA, or multiple sequence alignment, provides evolutionary data for the sequence you want to fold by giving the model more context for how similar proteins evolved over time. Adding MSA information to your sequence enhances folding predictions for more complex sequences, which can be done using the ESMFold2 model only (not ESMFold2-Fast).

How do I interpret the confidence metrics in the resulting prediction?

The pTM (predicted TM-score) indicates the accuracy of the overall fold (0–1, where higher values are better). The mean pLDDT score indicates the per-residue confidence values (0-1, where higher values are better). Higher scores means the model highly trusts its local atom prediction.

How many jobs can I run daily using the Fold app?

You can run up to 100 Fold jobs per day.

Can I use Fold if my sequence has unknown residues?

ESMFold2 supports protein sequences with the 20 canonical residues plus X (unknown) and 4 non-canonical residues: B (Aspartic acid or Asparagine), Z (Glutamic acid or Glutamine), U (Selenocysteine) and O (Pyrrolysine). DNA/RNA sequences support N for unknown residues.