Build a pipeline from your request
Describe the target and design constraints. Review the proposed settings, then use Start to run the pipeline.
User settings
Pipeline settings to apply
1 Target & binding site
Whatever you type, this step ends with one thing: a 3-D structure of the target, which it hands to Step 2. Three paths get there — name → id: a protein name is resolved against UniProt (hit the search button and pick a match); id → structure: a UniProt accession or PDB ID is looked up and its structure downloaded (best experimental PDB, else the AlphaFold DB model); sequence → structure: a bare sequence has nothing to look up, so it is folded here into a structure.
Shown only when UniProt says the target is a membrane protein. A peptide in the bloodstream can only reach the ectodomain — the cytoplasmic tail and the transmembrane helix are behind the membrane, so designing a binder against them produces something that can never engage the target in a cell. The dropdown groups domains by side and offers the whole extracellular region as one choice; tick this to hide the unreachable ones entirely. Sides come from UniProt's own topology annotation and are never guessed — a multi-pass protein with no topology annotation shows no sides rather than a coin-flip.
By default hotspots are searched over the whole chain, which lets a campaign commit to a site you may not care about. Pick a domain here and discovery is confined to it — with a binding partner, only the contacts on that domain are kept; without one, the surface search picks the best patch inside the domain instead of the biggest patch anywhere, and the receptor is trimmed to it (a smaller receptor is also faster per BindCraft iteration). Linear sequence is always excluded: residues outside every annotated domain, and residues the structure model itself is unsure of (pLDDT < 70), are dropped — flexible linkers and disordered tails have no persistent surface for a binder to grip. The list is InterPro annotation, so it appears only for annotated targets; leave it on Whole sequence to keep the previous behaviour.
2 Binding partner (optional)
Skip this step and the binder is aimed at the target's own surface — the domain and hotspots chosen in Step 1. Give a partner and the site is instead read off the interface of the target+partner complex: the experimental co-complex when the PDB has one, otherwise a predicted one. PyMOL analyses that interface — contacts within 4 Å, angle-filtered H-bonds, bond-type classification. This step mirrors Step 1 (input, domain, hotspots) because it is the same question asked about the other side of the interface.
Which part of the partner to fold into the predicted complex. A large multi-domain partner makes the prediction both slower and vaguer: which of its domains ends up facing the target becomes the predictor's guess, and the hotspots read off that interface are only as good as the guess. Naming the domain states the answer instead. Appears only when the partner resolves to a UniProt entry with annotated domains — an uploaded partner PDB keeps its own residue numbering, so a UniProt range would cut it in the wrong place and is deliberately ignored.
If the target has no experimental co-complex in the PDB, give its binding partner — a protein name or gene name, a UniProt accession, an amino-acid sequence, or an uploaded PDB / sequence file (the same input kinds as the target in Step 1). A name/sequence is turned into a structure with the selected Structure predictor (Boltz-2 / AF2); PepDesigner then predicts the target+partner complex and reads the hotspots off its interface with PyMOL. Needed whenever there is no experimental complex — otherwise there is no interface to find hotspots on. (Adds a GPU prediction run; de novo / auto mode only.)
3 Peptide design (BindCraft)
No. designs — how many de novo peptide candidates BindCraft should accept (passing AF2/Rosetta filters). More designs explore more of the space but take longer on the GPU.
Peptide length — binder length range (residues) BindCraft hallucinates. Longer binders fold more stably (more designs pass the Binder_RMSD ≤ 2.5 check); very short ones often fail to fold consistently, very long ones cost more GPU. 65–150 (a small protein binder, not a short peptide) is the default and matches BindCraft's own original design regime (see settings_target/PDL1.json) — if a run finds no accepted design, retries automatically try a bigger binder rather than a shorter one.
Max GPUs — upper bound on GPUs to shard across (leave blank to auto-use all eligible, one shard per design). BindCraft prefers idle (empty) GPUs — low utilization plus enough free VRAM — and only shares a busy GPU as a fallback when not enough free ones are available. The running step shows how many GPUs (and which) are in use, live.
Per-epitope design — run a separate BindCraft design per spatial epitope (from the PyMOL interface annotation) and merge the results, instead of one design against all hotspots. Targets each binding patch independently; GPU cost scales with the number of epitopes. Falls back to a single run when fewer than 2 epitopes are found.
Model that validates each designed binder. AlphaFold2 is the default and unchanged. Boltz-2 validates with an independent structure predictor (its MSA for the target is fetched once and cached, then reused) — it validates on one GPU per attempt, so racing several attempts still uses one GPU each, and it runs a higher trajectory budget with a looser early-stop to allow for its lower per-trajectory pass rate.
4 MD & binding free energy
Number of hits — how many of the top-ranked candidates to carry forward into docking / MD and into the final report.
5 Closed-loop redesign
Loop iters — 1 = single pass. Higher repeats redesign → refold → MD until no penalty residues remain or the cap is reached. Each pass costs GPU hours.
⚠ De novo design occupies a GPU for hours and runs MD per accepted design. Start it only when a GPU is free.
Pipeline progress
Start an analysis to see step-by-step progress here.
Results
Ranked candidates appear here when the run completes.