De Novo Antibody Design: What RFdiffusion Demonstrated and What Still Needs Testing

De novo antibody design is the computational creation of a binding site for a chosen epitope without starting from an antibody that already binds it¹. The RFdiffusion antibody study reported experimentally confirmed VHH and scFv binders in defined settings, with low, target-dependent hit rates¹. A designed sequence therefore stays a candidate until physical testing says otherwise.

I’m Julia Pizzolato, PhD, Commercial Scientific Liaison at evitria in Zurich. My background spans protein engineering, antibody design, data science and biomanufacturing analytics. At evitria, I support customers and project teams through technical consultation, feasibility assessments and training. That gives me two ways to read a design paper: as protein design, and as a list of sequences that someone has to express and test.

In this article, I explain what de novo antibody design means, what the RFdiffusion study showed and where it stopped. I then follow a selected sequence to the decision it has to support. De novo design is one task within AI Antibody Discovery, so I leave model families, data requirements and the wider workflow aside. The evidence comes from one study, and every result holds for the targets, formats and screens it used. The planning checklist near the end is editorial practice, not a regulatory standard.

Throughout this article, “VHH” means the variable domain of a camelid heavy-chain antibody, a single-domain format. “scFv” means a single-chain variable fragment that links the heavy- and light-chain variable domains. Complementarity-determining regions (CDRs) are the loops that form the antigen-binding site: a VHH has three, an scFv six. An “epitope” is the patch on the target that the antibody binds. The “pose” is the position and orientation in which the antibody sits on it, including the conformation of each CDR loop. “Hit rate” is the share of tested designs that turned out to bind the target.

What is de novo antibody design?

De novo antibody design seeks a first binding solution for an epitope you choose. Conventional discovery gets there by immunization, library screening or isolation from patients¹. Earlier computational work mostly improved antibodies that already existed¹. A design method starts from none of them.

“From scratch” does not mean without inputs. The RFdiffusion workflow began with a target structure, epitope hotspot residues, a fixed antibody framework and filtering criteria¹. That is the operational sense of de novo used in the study and in this article, not a threshold for sequence novelty.

A blinded, prospective benchmark separates the tasks too. Affinity maturation and hit identification address existing discovery and optimization pipelines, while its third task represents limited de novo design⁶. Its successes did not transfer across tasks⁶.

What did the RFdiffusion antibody study demonstrate?

The researchers fine-tuned RFdiffusion on antibody structures. Given a target, epitope hotspot residues and a fixed framework, it designs CDR loops and the antibody’s docking mode¹. ProteinMPNN proposed the CDR sequences, and a fine-tuned RoseTTAFold2 model re-predicted each complex to filter for designs that matched their own models¹. Filtered designs went to yeast surface display, or to E. coli expression and single-concentration surface plasmon resonance (SPR). Cryo-electron microscopy (cryo-EM) then checked chosen complexes¹.

For VHHs, success rates per target ranged from 0% to 2%¹. The campaigns differed in size. Some screened 9,000 designs per target by yeast display: respiratory syncytial virus (RSV) sites I and III, the SARS-CoV-2 receptor-binding domain (RBD) and influenza H1 haemagglutinin¹. Others tested 95 designs per target by expression and SPR: Clostridium difficile toxin B (TcdB), IL-7Rα and haemagglutinin, which the authors screened both ways¹. Treat the figures as separate experiments, not as a ranking of targets.

The best parent VHHs bound with modest affinity: 78 nM for haemagglutinin, 262 nM for TcdB and 5.5 µM for the RBD¹. To raise affinity, the team put the parents through OrthoRep, a continuous hypermutation system in yeast. The matured VHHs acquired several mutations and improved by approximately two orders of magnitude¹. A matured sequence is therefore the product of the pipeline plus laboratory evolution, not of the pipeline alone.

Cryo-EM supplied the structural evidence. A designed VHH bound to influenza haemagglutinin matched its design model closely in a 3.0 Å structure, and its CDR3 loop deviated by 0.8 Å from the design¹. It was the only one of four tested designs that bound the glycosylated trimer effectively¹. For TcdB, maps at modest resolution (4.6 and 5.7 Å) placed the parent and the matured VHH on the intended epitope¹. The matured VHH kept its original pose¹. In the scFv work, structure-aware combinatorial libraries yielded epitope-specific binders where fixed-pair libraries did not¹. For the best scFv (72 nM), cryo-EM at 3.6 Å matched the design for all six CDR loops, and as full-length IgG1 it kept a comparable affinity of 68 nM¹. The authors chose these complexes for cryo-EM, and the results do not establish atomic accuracy for every generated design.

The SARS-CoV-2 case shows why a binder, binding to the intended epitope and reproducing the designed pose are three separate claims. Competition experiments indicated that the parent VHH bound the intended epitope. Cryo-EM of an affinity-matured descendant then confirmed epitope targeting. It showed a different, predominantly framework-mediated pose, and the authors classified the design as a failure¹. Binding the intended epitope is a claim about location: the footprint overlaps a chosen patch of surface. Adopting the designed pose also fixes the rigid-body placement, meaning approach angle and rotation, which parts of the antibody make contact and the conformation of each CDR loop. Antibodies can cover the same patch from many directions, so the pose is the more specific claim.

The authors locate the key advance in targeting a specified epitope, since binders against a target are often achievable by purely experimental methods¹.

Why does designed not mean validated?

A generated or filtered sequence is a hypothesis until physical evidence answers the next decision. The authors attribute the low success rates and the inaccurate SARS-CoV-2 design probably to the limited filtering power of their RoseTTAFold2 filter¹. Models also aim beyond affinity, at specificity, stability, viscosity and manufacturability¹², and each of those needs measurement on real protein. That is the case for the physical validation of AI-designed antibodies.

Between a selected sequence and a decision, I see five stages. The exact tests depend on the program.

1. Define the claim and design constraints. Name the target, epitope, intended format and the evidence that counts as success.
2. Select a physical test panel. Down-select exact sequences and controls, and keep model scores separate from experimental evidence.
3. Express and purify. Produce each selected construct in the agreed format and record failures as results.
4. Run analytical QC. Confirm sample identity, condition and comparability with the agreed package.
5. Test and decide. The team or its assay partners assess binding, specificity or function, then advance, redesign or stop each candidate.

Structural confirmation belongs to stage 5 when the designed pose is itself the claim, as in the SARS-CoV-2 case. It is not a universal requirement.

How selected designs move into expression and QC

The handoff starts after computational down-selection, with a precise record of what you will make. For each design, fix the exact sequence and chain pairing, the format, the construct boundaries, a sample identifier and the controls. One line per design that states what you produce, in which format and who tests it is enough, and it lets every later result map back to its design.

Next, set the panel size by the decision the panel has to support. Agree the expression and purification scope, the sample amount, the buffer and the aliquot plan, then the analytical QC outputs the comparison needs. Concentration, purity and monomeric state are typical examples, and the final package depends on the project. Binding, specificity, functional and structural work stays with the team or its assay partners.

The AIntibody benchmark compared methods under the same conditions, because independent laboratories tested designs under uniform protocols⁶. A production panel works the same way. Expressibility, aggregation and chain-pairing behaviour only read as signal when process, vector system and analytics stay constant across variants. A transient titer ranking is an early filter, not a prediction of stable-line manufacturability.

I recommend spending the first production wave on comparability, not coverage. Run a small, deliberately diverse set of designs, including a known difficult parent and a well-behaved benchmark, through the vector system, process and analytics planned for the full campaign. That shows where the scaffold works, what material per variant is realistic and whether the analytics can separate good from bad at that scale. Changing format or analytics midway splits the dataset into two halves that you cannot compare.

A VHH or scFv selected by yeast display or E. coli expression changes context in CHO expression or as full-length IgG. Agree construct feasibility and the test format first. QC describes the material, while affinity, specificity and pose come from the downstream tests.

What should teams decide before producing a panel?

Questions 2 and 4 matter most for production, because they define what you express and what comes back. Questions 1, 3 and 5 give the project its context. A common assumption treats expression as a neutral, interchangeable step. Sequence files go in, protein comes out at a fixed rate, and yield is the only output that matters. In practice, expression is part of the experiment, and comparable process and analytics matter more than a yield number.

1. What target, epitope and downstream decision define the panel?
2. Which exact sequences, chain pairings, formats and controls does the panel include?
3. How many designs will the team compare, and why did those designs make the cut?
4. What output do you need per variant, and why: material amount, buffer, aliquot plan and analytical QC outputs?
5. Who owns downstream testing, and how will each result map back to its design?

What evitria does with selected antibody designs

At evitria, we pick up where the computational work ends, and we like that moment: a list of sequences becomes protein your team can measure. You send us the antibody sequences your team has selected, and we express them by CHO-based transient expression in our recombinant antibody expression service. We purify each construct and run analytical quality control. Standard analytics cover purity, quantification, concentration and endotoxin level. Depending on your project, we add titer measurement, HPLC-SEC for purity and aggregation, and CE-SDS for structural integrity.

When designs run as a panel, every construct goes through the same standardized workflow, so differences between designs trace back to the sequence and not to the process. That is what makes a first production wave comparable. Your team or your assay partners run the binding, specificity, functional and structural tests that follow, and we confirm construct and format feasibility with you during scoping. Our services cover research and preclinical work, and our products are not for human use.

Key takeaways

• De novo antibody design proposes a first binding solution for a chosen epitope, and the RFdiffusion study still began from a target structure, hotspots and a framework.
• The study reported experimentally confirmed VHH and scFv binders in defined settings. VHH success rates ranged from 0% to 2% across targets, from screens of 9,000 or 95 designs per target.
• A binder, the intended epitope and the designed pose are three claims. The SARS-CoV-2 descendant bound the intended epitope but not in the designed pose.
• Matured VHHs are the product of the pipeline plus laboratory evolution.
• A designed sequence is a test candidate until it exists as purified, characterized material that downstream tests can compare.
• evitria expresses and purifies client-supplied sequences by transient CHO expression and delivers them with analytical QC for research and preclinical work.

FAQs about de novo antibody design

References

1. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
2. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
3. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
4. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
5. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
6. Erasmus MF, Bedinger D, Hopkins E, et al. A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability. Nat Biotechnol (2026). https://doi.org/10.1038/s41587-026-03238-6 ↩︎
7. Erasmus MF, Bedinger D, Hopkins E, et al. A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability. Nat Biotechnol (2026). https://doi.org/10.1038/s41587-026-03238-6 ↩︎
8. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
9. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
10. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
11. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
12. Matsunaga R, Tsumoto K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. J Biomed Sci 32, (2025) 46. https://doi.org/10.1186/s12929-025-01141-x ↩︎
13. Matsunaga R, Tsumoto K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. J Biomed Sci 32, (2025) 46. https://doi.org/10.1186/s12929-025-01141-x ↩︎
14. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
15. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
16. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
17. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
18. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
19. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
20. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
21. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
22. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
23. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
24. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
25. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
26. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
27. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
28. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
29. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
30. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
31. Matsunaga R, Tsumoto K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. J Biomed Sci 32, (2025) 46. https://doi.org/10.1186/s12929-025-01141-x ↩︎
32. Erasmus MF, Bedinger D, Hopkins E, et al. A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability. Nat Biotechnol (2026). https://doi.org/10.1038/s41587-026-03238-6 ↩︎
33. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
34. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
35. Matsunaga R, Tsumoto K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. J Biomed Sci 32, (2025) 46. https://doi.org/10.1186/s12929-025-01141-x ↩︎
36. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
37. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎
38. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, (2026) 183-193. https://doi.org/10.1038/s41586-025-09721-5 ↩︎