Fold & Studio
Two workspaces for protein structures, next to the Workbench. Fold predicts a structure from a sequence and lets you question it. Studio keeps a structure as a project, with labels, figures and a stack of edits you can switch on and off.
01Fold
Paste an amino-acid sequence, between 8 and 400 residues. Fold predicts its structure with ESMFold and shows it in 3D, coloured by per-residue confidence (pLDDT). A nucleotide sequence is recognised and refused with an explanation.
Before it explains the structure, Fold tries to identify the protein by exact sequence match against UniProt. When the sequence is a known protein, the explanation leads with its real name. Without an exact match, it explains the structure from the sequence alone.
After a fold the page splits in two: the structure on the left, a chat on the right that explains what you are looking at.
- Click a residue to label it and attach it to your next question, so the answer is about that exact site.
- Double-click anywhere on the model to pin your own label.
- Ask the chat to label something — "label the disulfide bonds", "label residues 40–52" — and the labels appear on the matching residues.
- Export the view as a PNG. Labels keep their place in 3D and appear in the image.
Fold does not save anything. Labels last for the session. To keep an annotated structure, use Studio.
02Studio
Studio is a dashboard of protein projects. Each project holds one structure, the labels pinned to it, a conversation about it, and the figures you save from it. Each project has its own link, and projects can be cloned and deleted.
Starting a project
| Start from | What happens |
|---|---|
| PDB entry | An experimental structure fetched from the RCSB PDB by its four-character ID, with its method and resolution. |
| UniProt accession | The AlphaFold DB model for that accession, coloured by per-residue confidence. |
| Amino-acid sequence | A sequence or FASTA record, folded with ESMFold (up to 400 residues). It is identified first where possible. |
| Upload a file | A .pdb, .cif, .ent or .fasta file. Coordinates are parsed without sending the file to an outside service. |
| Describe what it does | A plain description, such as "an enzyme that cuts DNA". Studio finds the closest real protein and loads its structure. If nothing matches exactly, you get the closest match and it says so. |
| Start empty | A blank viewer and chat. Ask the assistant for a protein. |
Labels from measured features
When a structure loads, Studio measures it: secondary structure, buried and exposed residues, cavities, disulfides, metal sites, salt bridges, catalytic triads and common sequence motifs. Where a ligand is bound, its contact shell is the binding site.
Requests like "label the active site", "label the cysteines" or "label helix 3" are resolved against those measured features first, with no model involved. A model is consulted only when that finds nothing, and even then it can only choose from features that exist in the structure. Every label is pinned to a real residue.
Tweaks: editing a protein
Tweaks are edits stacked on top of the original structure. The original is never changed, so every edit can be switched off, reordered or removed, and you always have a wild-type to compare against.
| Edit | How the structure is updated |
|---|---|
| Substitution | Side chains rebuilt in place on the existing backbone. |
| Deletion, truncation | Atoms removed from the existing coordinates. |
| Disulfide staple | Exactly two residues made cysteines to form a new disulfide, on the existing backbone. |
| Insertion, tag | These change the backbone, so the construct is refolded from its sequence and renumbered. |
Tags include His6, His8, FLAG, Myc, Strep-II, HA, a TEV site, a sortase motif, and flexible and rigid linkers. A project can hold up to 60 tweaks.
Common requests such as "mutate Ser195 to alanine" are parsed by rule and cannot target a residue that does not exist. Open-ended requests such as "make it more soluble" go to a model, which is given the measured features and whose suggestions are checked against the real residues. Suggestions arrive as cards to accept or discard; nothing is applied until you accept it.
Labels are anchored to residues, not positions. When an edit renumbers the chain, labels follow their residues. A label on a residue that an edit removed says so instead of moving to a neighbour.
What a tweak is likely to do
Each tweak gets three readings, built from features measured on the structure:
| Panel | What it is |
|---|---|
| Stability | A banded estimate, from stabilising to likely disruptive, with every contributing term listed. Never a kcal/mol figure. |
| Functional-site impact | Whether the edit touches a measured site. A lookup, not a prediction. |
| Developability | Liabilities found by exact sequence and structure rules, such as a new deamidation motif. |
Exporting
From a project you can download the view as a PNG with its labels, save it as a figure in the project, or export the current construct as FASTA, as PDB, or as a plain-text report of the tweak stack.
03Structures in the Workbench
The Workbench agent builds interactive 3D models too, from a PDB id, a UniProt accession, a sequence or inline coordinates, and it can find binding sites and dock molecules. Those models appear in the chat and in Artifacts. See Figures, models and artifacts and Engines & discovery.