Documentation
How the studio is put together — what each mode does, which tool to reach for, and what the numbers in your results table mean.
Experiment modes
Every run starts as an experiment in one of three modes. The mode determines which tools are offered and what inputs are required.
Bind
Design a de novo binder against a target structure. You supply the target and mark the residues you want the binder to engage; the tool generates and scores candidate binders.
Fold
Predict the structure of sequences you already have — a monomer, or a complex of several chains. No binder/target distinction.
Edit
Start from an existing binder and refine it — affinity maturation, CDR redesign, or reseeding a design from another tool's output.
Design tools
Binder design
- BindCraft — hallucination-based binder design with an AlphaFold2 backbone and ProteinMPNN sequence design. The default starting point for most targets. Up to 500 total residues.
- ProteinHunter — iterative search that runs in cycles, reporting per-cycle metrics. Strong on hard targets where single-shot methods stall. Up to 550 residues.
- Proteina-Complexa — complex-aware generative design.
Structure prediction
- Boltz-2 — co-folding of protein complexes with confidence metrics. Uses a cached target MSA. Up to 1000 residues.
- ESMFold2 — fast single-sequence folding, no MSA required. Best when you want a structure in minutes rather than a maximally accurate one.
- Protenix — an AlphaFold3-family model, for complexes involving ligands, nucleic acids or modified residues.
A note on MSAs
Where a tool supports it, an MSA is used for the target and never for the designed binder — de novo binders have no meaningful homologs, so they always fold single-sequence. Target MSAs are computed once and cached, so a large campaign against one target doesn't recompute the alignment for every design.
Preparing a target
- Search — type a PDB ID or a protein name into the search bar above the viewer. The biological assembly is fetched from the RCSB.
- Upload — drag a
.pdb,.cif,.mmcifor.fastafile anywhere onto the page. - Paste — paste a raw sequence to fold it first, then design against the result.
Trimming
Cost and runtime scale super-linearly with residue count, so trimming a target down to the domain you actually care about is the single highest-leverage thing you can do. Select the chains and residue ranges to keep in the viewer; the structure is renumbered and your hotspots are carried across automatically.
Hotspots
Hotspots tell the design tool where on the target to bind. Click residues in the 3D view or the sequence track to toggle them. The interface-detection helper can propose a set from an existing complex if your target is already bound to something.
Reading your results
All binder tools share the same six leading columns so runs can be compared across tools, with tool-specific metrics to the right.
| Metric | What it measures | Rule of thumb |
|---|---|---|
| pLDDT | Per-residue confidence in the predicted structure (0–100) | Higher is better; >80 is a confident fold |
| ipTM | Predicted accuracy of the interface between chains (0–1) | The primary binder screen; >0.8 is strong |
| ipSAE | Interface score corrected for chain size and PAE asymmetry | More discriminating than ipTM on small binders |
| Hotspot % | Fraction of the hotspots you marked that the binder actually contacts | Low values mean it bound somewhere else |
| pTM / PAE | Global fold confidence and predicted positional error | Higher pTM, lower PAE |
| RMSD | Deviation from the reference or, for self-consistency, from the redesign | Lower is better; under 2 Å is tight |
| ΔG / SASA | Estimated binding energy and buried surface area | More negative ΔG, larger buried SASA |
| Rg / asphericity | Binder compactness and shape | Asphericity above ~0.45 flags a rod-like single helix |
Results can be filtered and sorted on any column, and exported as CSV with the structures attached.
Public API
Everything the studio does is available over HTTP at public.seq.design/v1 — submit runs, poll them, pull the results table, and mint viewer links. The full reference is a live, try-it-in-the-browser page:
API reference (OpenAPI)GraphQL explorerllms.txt (for agents)
Get a key
Create one on the API keys page. Keys carry scopes (jobs:submit, jobs:read, tools:read, structures:write) and optional per-run and daily spend caps, so an automated pipeline can't run away with your balance. The secret is shown once at creation.
What a run costs, before you run it
Pricing is readable without a key — you shouldn't have to sign up to find out what something costs. Same numbers the estimator on the pricing page uses, and the same function that bills a real submission:
curl https://public.seq.design/v1/pricing curl "https://public.seq.design/v1/pricing/estimate\ ?tool=bindcraft&total_residues=260&num_designs=50"
Submit a run
curl -X POST https://public.seq.design/v1/jobs \
-H "Authorization: Bearer $SEQDESIGN_API_KEY" \
-H 'Content-Type: application/json' \
-d '{"tool":"bindcraft","target":{"pdb_id":"5C3T"},
"hotspots":"A56,A58,A115","num_designs":20}'
# then poll, and pull every QC column as CSV
curl -H "Authorization: Bearer $SEQDESIGN_API_KEY" \
https://public.seq.design/v1/jobs/<id>
curl -H "Authorization: Bearer $SEQDESIGN_API_KEY" \
"https://public.seq.design/v1/jobs/<id>/dataframe?format=csv"Share a structure without anyone signing in
POST /v1/structures is the one endpoint that needs no key at all. It turns a .pdb or .cif into a link that opens in the 3D viewer, hotspots highlighted, on any device. POST /v1/collections does the same for a whole run and gives you one sortable table — also downloadable as CSV or FASTA. Send it coordinates (100 per call), or the codes of links you already minted (500 per call) to re-cut them into a table without uploading anything again.
Agents and MCP
There is an MCP server on npm: SEQDESIGN_API_KEY=… npx @seqdesign/mcp. For any other agent, llms.txt is the whole API in one plain-text fetch, with worked examples and the traps worth knowing.