BindCraft is available as a browser tool at tools.ranomics.com/tools/bindcraft. Upload a target structure, name the chain, list your hotspot residues, and filtered de novo binders come back by email. No GPU of your own, no conda environment, no PyRosetta license.
That last point is why many people arrive here searching for FreeBindCraft rather than BindCraft. Ranomics runs the MIT-licensed FreeBindCraft build, which reproduces the BindCraft design loop without the PyRosetta dependency, with interface scoring handled by FreeSASA and OpenMM.
This post covers the one mistake that makes first-time users think the tool is broken, what you actually control, and how to run it properly.
The mistake: treating one run as a verdict
The most common way a BindCraft run goes wrong has nothing to do with BindCraft. It is that people run it once, at the default settings, look at what comes back, and decide the tool does not work on their target.
BindCraft is not a function that returns an answer. It is a sampler over a design space, and you define that space. One run at defaults samples a single point in it. When that point is weak, almost nothing has been learned about the target, and quite a lot has been concluded anyway.
This is the same error as judging a co-folding prediction from a single sample. The fix is the same. Sample the space, then decide.
What you actually control
The hosted form exposes five things, and only five. Knowing the full list is what makes a sweep possible.
Target structure. A .pdb, .cif or .mmcif file.
Target chain. A single chain ID, which defaults to A. Everything else in the file is discarded before the run, along with waters and other non-protein records. If the surface you care about is formed by two chains, the hosted pilot is not the right shape for that job.
Hotspot residues. A comma separated list of residue numbers on the target chain, with no default. This is the only input the tool cannot guess for you, and it is the one that most changes the outcome.
Binder length window. A minimum and a maximum, anywhere from 50 to 150 residues, defaulting to 50 and 100.
Trajectory count. How many design trajectories to run, from 1 to 500, defaulting to 4.
There is no temperature, no seed, and no sampler setting. The design space is the hotspot set and the length window. That is the whole board.
Hotspot numbering is author numbering
BindCraft references your hotspot residues by the numbering in the file you uploaded, exactly as deposited. The hosted pipeline preserves that numbering on purpose, precisely so your hotspot list keeps meaning what you intended all the way through the design loop.
The practical consequence: read your residue numbers off the structure you are about to upload, not off a UniProt entry, a renumbered copy, or a figure from a paper. A wrong number does not raise an error. It designs a perfectly good binder against a patch you did not ask for, and the run looks like a failure when it was a typo.
Two related traps in the same family. Your hotspots must sit on the chain you named, because every other chain is stripped before the run. And if your structure uses modified residues, they are converted to their standard parents on the way in, so check that the positions you listed survive that conversion.
If you are not confident which patch to target in the first place, score the surface before you spend any GPU time. Epitope Scout ranks candidate epitope patches and exports a residue list you can paste straight into the hotspot field.
How to actually run it
First, a confirmation batch. Run about four trajectories. This takes roughly thirty to forty five minutes and it is not really a design run, it is a check that your chain ID, your hotspot numbering and your length window all mean what you think they mean. Look at whether the returned binders contact the residues you named. If they do not, stop and fix the inputs rather than scaling up.
Then vary the specification, not just the volume. This is the step almost everyone skips. Once the confirmation batch looks sane, do not simply rerun the same settings with a bigger trajectory count. Run several jobs that differ in their specification:
- The same hotspots at a short length window and again at a long one. A 50 to 70 residue binder and a 110 to 150 residue binder are solving different geometric problems against the same patch.
- Different hotspot subsets drawn from the same epitope. Three residues at the centre of a patch and five residues spread across it pull the design loop toward different binding modes.
- Adjacent patches, if the surface offers more than one plausible site.
Then scale the winner. When one specification is clearly producing better interfaces than the others, put the trajectory count behind that one. Scaling a specification you have not tested against alternatives is how a campaign spends its budget confirming an arbitrary first guess.
What comes back, and when
Runs are long by design and the session allows up to four hours, so the tool does not ask you to sit and watch it. Submit the job, close the tab, and the results arrive by email when the run completes.
Output is a set of filtered candidate binders with AlphaFold2 ipTM and pLDDT, shape complementarity, and downloadable structures. BindCraft filters during generation rather than handing you everything it tried, which is the main practical difference from a broad diffusion sweep. You get fewer designs with stronger per candidate interface scores.
What those scores are not is evidence of binding. They are the model’s confidence in its own prediction. Before you order genes, put the survivors through a developability pass, because the cheapest liability to remove is the one that never gets synthesised.
Choosing BindCraft in the first place
BindCraft is the better default when your target structure is clean and you have literature or experimental hotspots to point at. It refines narrower, higher confidence designs. RFdiffusion is the better default when the target is unfamiliar or the surface is flat with several plausible binding modes, because it samples backbone geometry far more broadly.
For a full treatment of that choice, see BindCraft vs RFdiffusion. For production campaigns the honest answer is usually to run both on the same target and compare interfaces before committing anything to the bench.
What it costs
The tools hub is funded from a USD wallet and billed by the second of GPU time, so a short confirmation batch costs a fraction of a large sweep. New accounts start with five dollars in the wallet and no credit card. Top ups start at twenty dollars, and the wallet is prepaid, so you can only spend what you have funded. Current rates are on tools.ranomics.com.
Run BindCraft now: tools.ranomics.com/tools/bindcraft.
Related Ranomics services
- BindCraft tool: The FreeBindCraft build, hosted self serve. Funded from your wallet, five dollars to start on signup.
- Epitope Scout: Free epitope patch scoring. Decide which residues belong in the hotspot field before you spend GPU time.
- Binder Pilot: A full campaign when you want the sweep run and the hits validated on yeast display rather than scored on paper.
Frequently asked questions
Can I run BindCraft online without installing it?
Yes. Ranomics hosts BindCraft as a browser tool at tools.ranomics.com/tools/bindcraft. You upload a target structure, name the target chain, list hotspot residues, set a binder length window and a trajectory count, and results are emailed when the run finishes. There is no conda environment, no GPU of your own, and no command line.
Do I need a PyRosetta license to run BindCraft?
Not on the hosted tool. Ranomics runs the MIT-licensed FreeBindCraft build, which reproduces BindCraft without the PyRosetta dependency. Interface scoring is handled by FreeSASA and OpenMM instead. If you install upstream BindCraft yourself, the PyRosetta licensing question comes back.
Why did my first BindCraft run return disappointing designs?
Almost always because it was one run at default settings. BindCraft gives you a hotspot set, a binder length window and a trajectory count, and a single run samples one point in that space. Vary the hotspot set and the length window across several runs before concluding anything about your target.
How many BindCraft trajectories should I run?
Start with a small confirmation batch of about four trajectories to check that your chain ID and hotspot numbering are right, which takes roughly thirty to forty five minutes. Once that batch looks sane, scale up. The hosted tool accepts up to 500 trajectories in a single run.