Ranomics
Scientific research and computational biology
BoltzGenBindCraftRFdiffusionde novo designbinder designtools

BoltzGen vs BindCraft vs RFdiffusion: Which Binder Generator to Run

Three de novo binder generators are available on the Ranomics tools hub: BoltzGen, BindCraft and RFdiffusion. All three generators take a target structure, let you designate specific residues as a target epitope, and all three email candidates back when the run finishes. Only one of the three will run if you leave that field blank. The three generators run unique algorithms and can be viewed as alternative strategies towards defining a novel binder. In our experience, the three generators provide unique solutions. This is why we keep all three rather than standardizing on a single generator.

This post is about choosing. What each input run form asks of you, what each model does between the upload and the results email, and which one to put your first run behind.

Hotspot residues

BindCraft and RFdiffusion require more epitope guidance from the user. If a target residue is not available or a user does not input this information, the job never reaches a GPU. BoltzGen does not work that way. The hosted BoltzGen tool reads an empty hotspot field as no hotspot constraint and submits the job.

That is a practical fork, not trivia. Plenty of first campaigns begin with a structure or therapeutic rationale only. There is either not enough biological or structural information to support targeting any particular surface. At that stage BoltzGen is the one you can actually run, and where its designs land is a first read on which patches look tractable. Based on what BoltzGen may provide as suitable binding areas, a user can take hotspots and enlist them onto BindCraft or RFdiffusion.

If you arrive with hotspots already, from the literature, a mutagenesis scan or Epitope Scout, running all three and screening the union is the lowest risk route we know of.

How does each algorithm work?

RFdiffusion denoises random coordinates into a backbone constrained to contact your hotspots. ProteinMPNN then assigns a sequence to that backbone, then AlphaFold2 multimer scores the pair. RFdiffusion is actually a coordinated sequence of three different steps that give you a designed binder.

BindCraft works from the other end. It optimizes a sequence by backpropagating through AlphaFold2 multimer, pushing sequence logits toward a predicted complex that binds the residues you named, then resequences with ProteinMPNN and refolds to check the two still agree. The structure is AlphaFold2’s prediction of a sequence, not a backbone that was designed first.

BoltzGen generates the binder all atom against your target, reasoning over side chain atoms during design rather than the backbone alone, and reads the sequence off that design with its own inverse folding module.

The distinction worth stating precisely is narrow. RFdiffusion’s sequence is written onto a bare backbone after the fact, by a different model that had no part in choosing the geometry. BoltzGen’s is not: the atoms the sequence describes were placed while the binder was being built against the target. BindCraft avoids the same split from the other direction, since it never designs a bare backbone at all. So the “backbone now, sequence later” strategy belongs specifically to RFdiffusion, which our gradient vs diffusion post pulls on too.

Why we keep BoltzGen in the rotation

Key input requirements on design forms

Binder length window. BoltzGen accepts 10 to 200 residues but the form is pre filled at 50 and 70. RFdiffusion accepts 30 to 150 and is pre filled at 55 and 65. BindCraft accepts 50 to 150 and is pre filled at 50 and 100. Anything shorter than 30 residues, or longer than 150, only gets past the BoltzGen form.

The batch number. BindCraft takes 1 to 500, and that number is a target rather than a promise: it is how many final designs the run will accept, each one having to pass AlphaFold2 re-prediction above BindCraft’s default filters. The hosted run also carries its own search budget, which it spends whether or not designs are passing, so on a hard epitope you get back fewer than you asked for and sometimes none. RFdiffusion takes 1 to 1000 candidates and samples each one. BoltzGen takes 1 to 50. However, a BoltzGen run generates a fixed pool of 200 designs whatever number you ask for, and your number selects how many of that pool come back to you.

The target chain field. BoltzGen and RFdiffusion both take several chain ids, comma or space separated. BindCraft takes one chain and refuses a submission with multiple chains. An epitope that spans a dimer interface is therefore not a BindCraft job as the form stands, which is worth checking before you plan around one.

How to read what comes back

RFdiffusion and BindCraft both hand you AlphaFold2 confidence numbers on the binder and target together. BoltzGen’s headline pair, backbone RMSD and pLDDT, is a different kind of number. Where the run refolds a design from its sequence alone, the RMSD compares that refold against the backbone BoltzGen built for it and the pLDDT is the refold’s own confidence, so the pair is a self consistency check on the design rather than a statement about the complex. The interface number comes from the generator’s own per design metrics instead.

The two sides are therefore not directly comparable, and a BoltzGen design that is internally consistent is still a guess about binding, exactly as a high ipTM is. None of the three measure binding. All three return a model’s opinion, which is why the next step is a screen rather than a bigger sweep.

A worked example

Say the target is a cytokine with a published structure, two copies in the file, and a receptor blocking surface named in the literature.

  1. Confirmation batch on BindCraft. Four designs, hotspots set to the blocking residues, the pre-set length window, and one chain named as the target. Look at whether the returned binders touch the residues you named.
  2. Breadth on RFdiffusion. Same hotspots, both chains named this time, a larger design count, for backbone diversity narrow sampling will not give you.
  3. A different proposal set from BoltzGen. Same hotspots, a wider length window than the default.
  4. Screen the top sets across all three generators, rather than ranking the three against each other on paper.

Reverse the order when you do not know the surface. BoltzGen first with the hotspot field empty, read where the designs land, then carry those residues into the other two.

If you can only run one

Pick on what you are short of. Short of a hotspot list, run BoltzGen. Short of confidence that your inputs are right, run a small BindCraft batch, because its filtering during generation means a weak batch is informative rather than merely large. Short of candidate diversity, run RFdiffusion at volume. The detailed head to head on those last two is already written up in BindCraft vs RFdiffusion, and where PXDesign fits covers how many generators to keep in the kit at all.

All three run in the browser on the Ranomics tools hub, with no local install and no GPU of your own, and results arrive by email when the run completes. The hub is funded from a prepaid wallet and billed by the second of cloud GPU time, so on BindCraft and RFdiffusion a small batch costs a fraction of a sweep. Not on BoltzGen, for the reason above: the pool is the same size whatever number you ask for.

When the deliverable has to be a validated hit rather than a scored model, the generator choice stops being the interesting question. That is what a binder discovery campaign is for: the sweep run across generators, the survivors screened on display, a ranked hit list at the end.

Summary

In our runs the three hosted generators do not return interchangeable candidate sets. RFdiffusion diffuses a bare backbone and has ProteinMPNN write the sequence afterwards. BindCraft optimizes a sequence through AlphaFold2 until the predicted complex grips your hotspots. BoltzGen builds the binder all atom and reads the sequence off its own design. The hotspot field is the fork that decides most first runs, because only BoltzGen will run without one. The length windows and batch numbers do not transfer between the forms. And none of the numbers you get back is a measurement of binding.

  • BoltzGen technology page: how the all atom generator runs on the Ranomics platform, and what each design arrives with.
  • BindCraft vs RFdiffusion: the detailed head to head on the two most used generators, and how we chain them.
  • Binder Pilot: a single round campaign when you want the sweep run and the survivors validated on yeast display rather than scored on paper.

Frequently asked questions

Which binder generator should I run first on my target?

If you already know which residues the binder has to target, start with BindCraft for a narrow high confidence batch, or RFdiffusion when you want breadth. If you do not know the patch yet, start with BoltzGen: it is the only one of the three whose hosted form will run without a hotspot list, so it can show you where designs land before you commit to a surface.

Do I need hotspot residues to run BoltzGen?

No. The hosted BoltzGen tool treats an empty hotspot field as no hotspot constraint and submits the job anyway. BindCraft and RFdiffusion both refuse a submission with no hotspot residues, before anything reaches a GPU. Naming hotspots on BoltzGen still focuses the run, it is simply not required.

What is the difference between BoltzGen and RFdiffusion?

Where the sequence comes from. RFdiffusion denoises a bare backbone against your target and hotspots, and ProteinMPNN then writes a sequence onto that backbone afterwards. BoltzGen generates the binder with its side chain atoms in place and reads the sequence off that same design with its own inverse folding module, so the sequence belongs to the structure instead of being fitted to it later.

What binder lengths can each tool design?

The hosted forms accept different windows. BoltzGen takes 10 to 200 residues and arrives pre filled at 50 and 70. RFdiffusion takes 30 to 150 and arrives at 55 and 65. BindCraft takes 50 to 150 and arrives at 50 and 100. Anything shorter than 30 residues, or longer than 150, is only accepted by the BoltzGen form.

Share

Ready to design your binder?

Single target on a fixed scope, start a Binder Pilot. Multi-target program or pharma-grade timeline, discuss a Custom Campaign.

Not sure which fits? Answer five quick questions for a recommendation.