Can I run BoltzGen online without setting up the model locally?

Yes. Ranomics Tools runs BoltzGen on a dedicated GPU through your browser. Upload a target PDB, pick a scaffold class (mini-protein, nanobody, antibody or peptide), and candidates come back with a structure and the generator's interface score per hit — plus pLDDT and a refolding RMSD on the mini-protein, nanobody and antibody protocols.

Which modalities does BoltzGen design against the same target?

BoltzGen can design mini-proteins, nanobodies, antibodies (scFv-class), or peptides against the same target with glycan and PTM support. The form switches scaffold class per run, so you can A/B different modalities cheaply before committing to wet-lab.

How long does a BoltzGen trial run take?

Runtime depends on target size and the binder format you picked. You pay for the GPU time your run uses.

BoltzGen

BoltzGen is a no-install online multi-format binder design tool you can run through tools.ranomics.com on a dedicated GPU. One model covers four binder formats against the same site, so you can weigh a mini-protein against a nanobody, an antibody or a peptide without changing tools — and it reads the sugars and modified residues on your target instead of ignoring them.

Example run — Human PD-L1, the IgV domain, taken as chain A of PDB 4ZQK — the solved PD-1/PD-L1 complex. See what came back →

Design inputs

Design

Binder length min

Binder length max

How long the new binder should be, in residues. Every design draws its own length from this window, so a wide window buys variety and a narrow one buys consistency. What counts as sensible depends on the format you pick below — the Protocol help lists a typical window for each. Within a format, aim short for a small flat patch and longer for a broad face or one sitting in a groove a short binder cannot reach across.

Protocol

What shape of binder to design. Mini-protein is the general-purpose choice and the one to start with — a small de novo protein with no scaffold to inherit. Nanobody and antibody build on real antibody frameworks, so pick one of those when your downstream work is display, hybridoma or an IgG format. Peptide is for short linear or cyclic binders. The choice sets what a sensible binder length is above: mini-protein 50 to 100, nanobody 110 to 130, antibody 110 to 200, peptide 10 to 30 residues (10 is the shortest this tool accepts).

Runs against your own target. Upload PDB, optionally pick hotspot residues — the numbered residues on the patch of your target you want the binder to sit on — and set binder length. Results emailed when the run completes; A100-40GB.

Target

Target PDB

Upload the target structure as .pdb, .cif, or .mmcif.

Target chain

One chain (A), or several for an oligomeric target (A,B). Every named chain reaches the model; designing against one protomer of a dimer aims at half the epitope.

Hotspot residues

Optional. Leave blank for unconstrained. Residues the binder should contact, in original PDB numbering. A plain number is read as your first target chain; prefix the chain to name another (A296,B264). Click residues in the 3D viewer below to toggle them in and out of this list.

Not sure which residues? Score your target's surface with Epitope Scout first.

3D residue picker

Surface

Upload a target PDB above to enable click-to-pick. Typed entry in the Hotspot residues field above also works.

Budget (final candidates)

Top-N final candidates returned (1 to 50). Each refolded to check the fold holds — mini-protein, nanobody and antibody — and ranked on the generator's interface score. Need more than one job's worth? Run it as a full-size run: wallet-bounded, at any scale, split into pieces that run on our GPUs and bill as they finish.

First time on this target? 4 is enough to read through while you confirm your binder length and protocol, and raising it on a later run returns more candidates out of the same search. You pay for the GPU time your run uses; cancellation does not refund time already used.

Group

Group related runs on your jobs page. Leave blank to file under Uncategorized.

You pay for the GPU time your run uses. Cancelling does not refund GPU time already used.

About BoltzGen

What it is

Designs a binder against your target, then refolds each candidate from its own sequence so you can see whether it holds the shape it was designed as. The interface score is the generator's own read, not a second opinion, so rank on it and re-fold a shortlist before you trust it. Four formats share one target — a small de novo protein, a nanobody, an antibody, or a short peptide — so you can compare formats on the same epitope rather than guessing which to commit to. It is the only design tool here that handles sugars, post-translational modifications and non-standard residues on the target natively. BoltzGen, Stark et al., bioRxiv 2025.

When to use

You have not settled on the binder format and want to aim mini-proteins, nanobodies, antibodies and peptides at the same site.
Your target carries sugars, modified residues or other chemistry a protein-only model would silently drop.
You want each design refolded on its own, so you can see whether it folds back to the shape it was designed as.
You can leave the run to finish rather than needing an answer immediately.

Prerequisites

Target structure (.pdb / .cif).
Chain ID of the target.
Optionally, hotspot residues — target-chain residues the binder should contact. Leave the field empty and the run starts with no hotspot constraint.

Input glossary

Protocol. BoltzGen design protocol. protein-anything for general mini-protein binders, nanobody-anything for VHH scaffolds, antibody-anything for antibody scaffolds, peptide-anything for short cyclic or linear peptides.
Hotspot residues. Comma-separated target-chain residue indices the binder should contact. Click residues in the 3D viewer to toggle.
Binder length (min/max). Residue-count window for the generated binder. Typical starting ranges: mini-protein 50 to 100, nanobody 110 to 130, antibody 110 to 200, peptide 10 to 30. 10 is the shortest binder this tool accepts.
Budget (designs). How many of the ranked candidates come back to you (1 to 50). This only chooses how many you receive: BoltzGen is asked for the same 200 candidates at every setting, so the budget does not change the size of the search.

Ranked candidate binders with ipTM, pLDDT, refolding RMSD, and structures downloadable from a run of your own — mmCIF or PDB, whichever the container wrote. Refolding RMSD at or under 1.5 Å clears the RMSD leg of the pass bar, which also needs pLDDT at or above 80; the number is the design measured against its own refold. That says the binder folds as designed, not that it binds — re-fold a shortlist against your target to check that.

References

A guided first run

Check that your target and the face you picked produce designs at all, on the one model that also handles glycans, modified residues and non-canonical chemistry. This is a guided first run with the settings filled in — the same run the form produces on its defaults, not a reduced-scope trial.

What you need

A structure file for your target (.pdb or .cif), the chain ID, and at least one residue on the face you want bound.

Hotspot residues are the numbered residues on the patch of your target you want the binder to sit on, written in your PDB’s own numbering — this tool does not require them, and aims the design at the residues you give it. If you do not know yours yet, score your target’s surface with Epitope Scout first — it is free, and its results hand the target and the residues back into this form.

What it costs

You pay for the GPU time your run uses.

Then what

Read the interface scores on what comes back. budget only chooses how many of the candidates are returned to you — the run is asked for the same 200 either way — so raising it on a later run buys no bigger search.

Load these settings →

A past run we actually did

The target

Human PD-L1, the IgV domain, taken as chain A of PDB 4ZQK — the solved PD-1/PD-L1 complex. 115 modelled residues, crystal numbering 18-132. Hotspots Ile54, Tyr56 and Met115.

It is a target you can mark your own homework on. Chain B of the same file is PD-1, the natural partner, so the three hotspots are not a guess — they are residues a real binding protein is known to cover. That makes one question answerable here that is usually not: never mind whether these designs are good, did the model aim where it was pointed?

What we put in the form

Target PDB — the 4ZQK file, uploaded whole: Both chains, exactly as it downloads from the PDB. No trimming and no renumbering, so chain A keeps its crystal numbering 18-132 — which is the numbering the hotspot field expects.
Target chain — A: This is what restricts the design to PD-L1. Chain B in the same file is PD-1, and the run ignores it — otherwise the model would be designing into an occupied site.
Hotspot residues — 54, 56, 115: Ile54, Tyr56 and Met115 in the file's own numbering. Three is a normal number: enough to aim the binder at one patch, few enough that you have not drawn the interface yourself.
Binder length min — 50: A window rather than a fixed number, so the generator chooses within it.
Binder length max — 70: The top design came back at 57 residues, in the middle of the window rather than pinned to either end.
Budget (final candidates) — 5: How many designs to keep. BoltzGen generated and scored 200 to return these 5, ranked by ipTM — you are seeing the top of a much larger pile. The guided trial run on this page uses 4.

What came back

Read the banner above the table first: every one of these five falls short of the bar, and the banner is right. The pass bar wants pLDDT at or above 80 and none of these reaches it — they run 69.1 to 78.2. Nothing here is a design you would order. What the run did do is aim. The delivered complex puts the binder on Ile54, Tyr56 and Met115 — all three residues we asked for, at 3.8, 2.3 and 3.3 Å — along with Glu58, Arg113 and Tyr123, part of the same face PD-1 itself covers. On a target whose natural partner sits in the next chain of the same file, that is checkable rather than asserted, and it is what a trial run is for: the epitope is reachable, so a bigger run is worth paying for. Four of the five also refold to under 1.5 Å of the pose they were designed in, the tightest at 0.51 Å. 200 designs against a 115-residue target is the heavy end of a trial run. You pay for the GPU time your run uses.

Candidates (5)

Every design here falls short on at least one of pLDDT 80 and Refolding RMSD 1.5 Å. The pipeline ran cleanly. All 5 designs are shown so you can see the score distribution; each row names the measurement that fell short. The fastest next step is a second-opinion fold: it re-scores a shortlist with an independent predictor. Boltz-2 additionally re-folds against your target, which is a different measurement rather than more of the same one. Re-running with a bigger budget or a wider length range only helps if the bar is one these designs could move.

#	ipTM ?	pLDDT ?	Refolding RMSD (Å) ?	vs. quality bar ?	3D	Structure
1 Top	0.653	78.2	0.51	pLDDT 78.2, below 80		.pdb

2	0.610	69.1	1.77	pLDDT 69.1, below 80; Refolding RMSD 1.77 Å, above 1.5 Å	—	—

3	0.597	73.1	0.58	pLDDT 73.1, below 80	—	—

4	0.586	71.6	1.12	pLDDT 71.6, below 80	—	—

5	0.538	73.6	0.53	pLDDT 73.6, below 80	—	—

Validating in the lab is optional Designs from your own run are yours to download and take anywhere. If you do want them made and measured, Ranomics runs yeast display, mammalian display, BLI and DMS on the same data loop.

Test binders in the lab AI Binder Sprint

How to read that

Refolding RMSD is the column with a real bar on it. BoltzGen re-folds each design on its own, with no target present, and measures the drift from the pose it was generated in. Under 1.5 Å the table calls it self-consistent: the design wants to be that shape whether or not the target is there. Four of these five clear it; rank 2, at 1.77 Å, does not. pLDDT is the one that fails here. Above 80 is confidently folded and 0 of 5 make it. A design can sit on the right epitope and still not be a protein worth synthesising, and that is exactly this run. Do not judge the ipTM against 0.7. That bar comes from cofolding, where a model is handed both partners and asked whether they dock. BoltzGen's is its generator's own score, not on that scale, which is why this site shows it with no pass bar. The 0.538 to 0.653 here is not “nearly good” — it is a reading you can rank designs by and cannot compare against a cofold number, including the 0.98 quoted further down this page from a different measurement.

What happened next

Treated all five as failed, because they are, and kept the epitope. What 200 designs bought is the knowledge that the site is reachable — the generator put binders on the residues we named — and that a 50-70 residue budget produced nothing foldable enough to order. The next run is a larger budget on the same hotspots, then an independent cofold of whatever clears pLDDT 80, and only then SPR or BLI. Nothing on this page is evidence of binding.

Recorded from that one past run. That is what it did, not a prediction for yours.

What good looks like

Each tool reports a subset of these depending on whether it does design, sequence recovery, or structure prediction.

ipTM: Predicted confidence in the contact between two chains, on a 0 to 1 scale. Higher is better, but BoltzGen's number is not on the shared scale the other tools are read against, so there is no cross-tool band to compare it to — rank on it, then re-fold a shortlist to confirm. On a multi-chain target the number may cover the target's own chain–chain interface as well as the binder's.
pLDDT: Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.
i_pAE and pAE: Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.
ProteinMPNN recovery: Fraction of native residues recovered when ProteinMPNN redesigns a known sequence on its native backbone — the folded shape on its own, before a sequence is chosen for it. Higher is better: > 0.4 usable; > 0.6 excellent.

Related tools on Ranomics

If you are picking between BoltzGen and a sibling algorithm, these run on the same hub against the same target.

Run RFantibody online: You have a target structure and want nanobodies — single-domain antibodies you can carry straight into yeast display, mammalian display or a hybridoma workflow. Each candidate is refolded and scored against the target before you see it. For non-antibody mini-proteins use RFdiffusion or BindCraft; for sugar-coated targets use BoltzGen.
Run IgGM online: You have an antibody or nanobody and an antigen, and you want to redesign its binding loops, humanise its framework, raise its affinity, or just see how the two dock — one model does all of it, aimed at the epitope you name. For a nanobody from scratch use RFantibody; for a paired heavy and light antibody fragment use ESMFold2 design.
Run RFdiffusion online: You have a target structure and a patch of its surface you want gripped, and you want brand-new binders of whatever shape works — the general-purpose starting point for de novo design. Every candidate comes back with a real AlphaFold2 confidence score against your own target. For antibody or nanobody formats use RFantibody or IgGM instead.

Questions people ask about BoltzGen

Can I run BoltzGen online without setting up the model locally?: Yes. Ranomics Tools runs BoltzGen on a dedicated GPU through your browser. Upload a target PDB, pick a scaffold class (mini-protein, nanobody, antibody or peptide), and candidates come back with a structure and the generator's interface score per hit — plus pLDDT and a refolding RMSD on the mini-protein, nanobody and antibody protocols.
Which modalities does BoltzGen design against the same target?: BoltzGen can design mini-proteins, nanobodies, antibodies (scFv-class), or peptides against the same target with glycan and PTM support. The form switches scaffold class per run, so you can A/B different modalities cheaply before committing to wet-lab.
How long does a BoltzGen trial run take?: Runtime depends on target size and the binder format you picked. You pay for the GPU time your run uses.