RFantibody is easy to file under “RFdiffusion, but for antibodies.” That label hides the one decision it actually settles. RFdiffusion and BindCraft invent a brand new protein backbone to bind your target. RFantibody does not invent a backbone at all. It holds a real antibody framework fixed and designs only the CDR loops. The output is a nanobody by construction, not a de novo mini-protein that happens to bind. If your deliverable has to be an antibody, that difference is the whole reason to reach for it.
This post walks through what RFantibody really does, then a decision rule for when it beats a generic binder tool.
A scaffold you keep, not a scaffold you invent
A generic de novo binder tool starts from noise and denoises a novel backbone against your target. The result is a small, stable protein optimized for one job, engaging the epitope. It is not an immunoglobulin. It cannot be reformatted to IgG, it has no germline to humanize against, and it inherits none of the manufacturing and regulatory infrastructure built around antibodies.
RFantibody, the RosettaCommons release of an antibody-finetuned RFdiffusion, starts somewhere else. You give it an antibody framework, and it treats that framework as a fixed scaffold. Only the complementarity-determining regions, the CDR loops that form the paratope, are diffused against your target. The framework you hand it stays put. So the thing that comes out the far end is a valid antibody, with designed loops sitting on a known, developable, humanizable scaffold.
On the Ranomics platform the hosted tool is specialized further: it designs single-domain antibodies (VHH, nanobodies). It ships with a nanobody framework and redesigns the three heavy-chain CDR loops (H1, H2, H3) against your epitope. That makes it the fastest route from a target structure to a yeast-display-ready nanobody, with no llama immunization and no phage panning campaign in front of it.
When the deliverable has to be an antibody
The mistake is treating RFantibody and RFdiffusion as two answers to the same question. They answer different questions. RFdiffusion answers “what is the best protein to bind this.” RFantibody answers “what is the best nanobody to bind this.” Only one of those is the question when a client needs an antibody.
Downstream compatibility is the whole point. A de novo binder is a great molecule and a dead end if the target product is a nanobody, because you cannot graft a novel fold into an immunoglobulin pipeline. RFantibody spends its diffusion budget staying inside the antibody world on purpose.
How it works on the platform
The hosted pipeline is three models in sequence, each finetuned for antibodies, with a filtering pass at the end.
- Antibody-finetuned RFdiffusion. The diffusion weights are trained on antibody-antigen complexes. The nanobody framework stays fixed while the CDR-H1, H2, and H3 backbones are diffused against your target and hotspots. CDR-H3 length is the dominant lever on diversity, so it is the parameter worth sweeping.
- ProteinMPNN sequence design. Each diffused backbone gets amino acid sequences. Framework positions are conditioned on the input scaffold, so they stay antibody-like; the CDR positions are designed freely to fit the new loop shapes.
- RoseTTAFold2-Antibody filtering. Every design is re-predicted in complex with the target by an antibody-tuned RF2. Designs that do not recover the intended binding mode are flagged before anything reaches the bench.
One accuracy note worth carrying, because it changes how you read the output: RFantibody does not score with ipTM. Its filter uses the interface PAE from RF2 (the predicted alignment error across the binder-target interface), where lower is better. A design that passes is one the model is confident about geometrically. That is a different and narrower claim than “this binds,” which is the trap in the next section.
A passing filter is not a hit
Here is the part most often missed. A design that clears the in silico filter is a hypothesis, not a hit.
RFantibody’s RF2 pass tells you the model re-predicts the same interface it designed, with low interface error. That is real signal, and it is worth filtering on. But it is a structure prediction re-agreeing with a structure prediction. It is not a measurement of affinity, expression, or specificity. Reading a clean interface PAE as evidence that the nanobody binds is the most expensive mistake in this workflow, because it sends designs to synthesis that were never going to work, and the synthesis and screening budget is where the real cost sits.
This is exactly why de novo antibody design and experimental display belong in the same loop. RFantibody gives you a ranked, structurally plausible shortlist. Yeast display measures which of those actually bind, at what apparent affinity, with the avidity and expression controls that turn a prediction into a validated hit. The design tool proposes; the assay disposes.
Quick reference
| RFantibody | RFdiffusion | BindCraft | |
|---|---|---|---|
| Output | a nanobody (VHH) | a de novo mini-protein | a de novo mini-protein |
| Scaffold | fixed antibody framework, CDR loops designed | novel backbone, invented from noise | novel backbone, invented from noise |
| Reformats to antibody | yes, to VHH-Fc and bispecifics | no | no |
| Best for | when the deliverable must be an antibody | maximum scaffold diversity, any format | fast, hallucination-filtered binders |
| Watch out | interface PAE is a prediction, not a binding measurement | output is not an antibody | output is not an antibody |
Practical recommendations
A workflow that holds up across most antibody design tasks:
- Pick the tool by the deliverable, not the target. If the program ships a nanobody or an IgG, use RFantibody. If any stable binder will do, RFdiffusion or BindCraft give more scaffold diversity.
- Sweep CDR-H3 length. It is the dominant lever on paratope diversity. Give the run a range, not a single length.
- Define your hotspots first. RFantibody designs against the epitope residues you specify. If you are unsure where to aim, map the target surface before you design.
- Do not read the filter as binding. A passing interface PAE means the shape is plausible. Take the shortlist to display and then to the bench.
You can run RFantibody on tools.ranomics.com against a target structure, and if you need help choosing the epitope hotspots first, Epitope Scout maps the target surface upstream. For the wider picture on how the underlying backbone model works, the RFdiffusion designer’s guide covers the de novo case that RFantibody specializes.
Summary
RFantibody is not a general binder tool with antibody flavoring. It designs CDR loops onto a fixed nanobody framework, so the output is a VHH by construction, ready for the humanization, reformatting, and affinity-maturation workflows a de novo mini-protein can never enter. Use it when the deliverable has to be an antibody, and use RFdiffusion or BindCraft when any stable binder will do. Whichever you run, remember that the in silico filter ranks plausibility, not binding. The nanobody is a hypothesis until an assay says otherwise.
Related Ranomics services
- RFantibody technology page: how the antibody-finetuned diffusion stack runs on the Ranomics platform.
- Nanobody discovery by yeast display: the experimental side, where designed VHHs get measured, not just predicted.
- Binder Pilot: take a shortlist of designed nanobodies to a ranked, wet-lab-validated hit list on a fixed-scope campaign.
Frequently asked questions
RFantibody or RFdiffusion for my target?
Pick by the deliverable, not by the target. If the program has to ship a nanobody or an IgG, RFantibody is the answer, because it holds a real antibody framework fixed and designs only the CDR loops, so the output is a VHH by construction. If any stable binder will do, RFdiffusion or BindCraft give more scaffold diversity and are often faster to a hit.
A design cleared the interface filter. Is that a hit?
It is a hypothesis. The RoseTTAFold2 pass tells you a structure prediction re-agrees with a structure prediction at low interface error, which is real signal worth filtering on and is not a measurement of affinity, expression or specificity. A yeast display screen is what shows which of the shortlist actually bind, with the avidity and expression controls that turn a prediction into a validated hit.
Which parameter is worth sweeping?
CDR-H3 length. It is the dominant lever on paratope diversity in the diffusion step, so give the run a range rather than a single value. The framework stays fixed throughout and the H1 and H2 backbones are diffused alongside H3, but H3 is where the reach and most of the variation come from.
Why is there no ipTM in the output?
RFantibody does not score with it. The filter uses interface PAE from an antibody-tuned RoseTTAFold2 re-prediction of each design in complex with the target, where lower is better. A passing design is one the model is geometrically confident about, which is a narrower claim than ipTM is usually read to make.
Can I reformat an RFdiffusion minibinder into an antibody later?
You cannot graft a novel fold into an immunoglobulin pipeline. A de novo mini-protein has no germline to humanize against and inherits none of the manufacturing and regulatory infrastructure built around antibodies. RFantibody output reformats to VHH-Fc and into bispecifics because it was an antibody the whole time, and that is the reason to spend the diffusion budget inside the antibody world.
Do I need hotspot residues before I start?
The run designs against the epitope residues you specify, so yes. If you are not sure where to aim on the target, map the surface before you design rather than after, because the epitope choice constrains everything downstream. Epitope Scout maps the target surface upstream, and it is the tool this guide points to for choosing hotspots.