Generative Protein Design: What RFdiffusion Has Actually Built

Generative protein design uses diffusion models to draw entirely new protein shapes, then software to find amino acid sequences that should fold into them. Since 2023 it has produced real, lab-tested proteins: binders for viral and toxin targets, metal-binding proteins, self-assembling cages, and in late 2025 the first antibody-like molecules aimed at chosen sites on a target. The catch is the hit rate. Many designs fail, and the best early binders are often good rather than drug-grade.

This piece walks through the main published milestones in order, then sets out what is still hard.

Timeline: three papers that define the field so far

All three come from the group of David Baker, who shared the 2024 Nobel Prize in Chemistry for computational protein design. Other labs and companies work on the same problem with their own tools, but these papers are the best documented and peer reviewed, which is why they anchor this article.

How the method works

Think of it as two steps. First, a model proposes a backbone, the three-dimensional scaffold of a protein. RFdiffusion does this by taking the RoseTTAFold structure-prediction network and retraining it as a diffusion model, so that each design begins as random noise and is refined step by step into a plausible structure. According to the 2023 Nature paper, the process can be steered by a target to bind, by a symmetry, or by a small functional motif that must be preserved.

Second, a separate program called ProteinMPNN works out an amino acid sequence expected to fold into that backbone. The designs are then ranked, usually by checking whether a structure predictor agrees that the sequence folds as intended, and the best few are produced in the lab and tested.

That last sentence hides the key economic fact. Software can produce thousands of candidates cheaply. Testing them is slow and costly, so the success rate decides how practical the whole approach is.

Why designing a protein is harder than it sounds

A protein is a chain of amino acids that folds into a shape, and the shape determines what it can do. The number of possible chains is astronomically large, and only a sliver of them fold into anything stable. Evolution has sampled a tiny part of that space. Design is an attempt to walk into the unexplored remainder on purpose.

For decades the main route was physics-based: build a scaffold, search for sequences that suit it, score the result with an energy function. It worked, but it was labor-intensive and hit rates were low. Deep learning changed the arithmetic in two ways. Structure predictors gave designers a fast, reasonably accurate filter for discarding sequences that would not fold. Diffusion models then gave them a way to generate the scaffolds themselves, rather than starting from shapes borrowed from nature.

That is the sense in which the technique is generative. It does not look up a protein that exists. It samples one that never has, conditioned on what you ask for.

How to read a protein-design claim

1. Find the number of designs tested, not just the number that worked. A single success out of thousands is a different story from 19 percent.
2. Look for an experimental structure. A crystal or cryo-EM map that matches the design is the strongest evidence that the computer got the geometry right.
3. Check the affinity units. Micromolar binding is weak, nanomolar is useful, and sub-nanomolar is strong.
4. Ask what was tested: purified protein, cells, or animals. Each step up is harder to pass.
5. Separate peer-reviewed papers from preprints and company announcements, and treat the latter as provisional.

What the 2023 paper showed

The RFdiffusion paper tested several design types, with numbers that are useful for calibration:

• Symmetric assemblies. Of 608 selected designs, at least 87 had oligomerization behavior consistent with the models in size-exclusion chromatography. Electron microscopy confirmed the shapes of several, including an icosahedral assembly.
• Metal-binding proteins. Of 36 designs tested for nickel binding, 18 bound it. Where the designed histidine was mutated, binding vanished or dropped sharply in all 17 designs that expressed successfully.
• Binders. Across five targets the overall experimental success rate was 19 percent. A cryo-EM structure of a binder to influenza hemagglutinin closely matched its computed model, and designed binders to a segment of p53 reached sub-nanomolar affinity for MDM2.

A 19 percent hit rate would have been startling a decade ago, when computational binder design usually meant screening large libraries and hoping. It is still far from guaranteed, and it varies by target. The enzyme work in that paper, as far as I could see in the text I accessed, was computational rather than tested for activity.

Case study: antivenom proteins

The January 2025 paper is a good example of design aimed at a concrete medical problem. The group designed small proteins of about 100 amino acids to bind three-finger toxins from elapid snake venoms, hoping for something cheaper and easier to manufacture than antivenoms derived from animals. The Nature paper reports:

• Binding affinities of 0.9 nM for a short-chain neurotoxin binder and 1.9 nM for a long-chain one, and crystal structures that closely matched the models.
• Full neutralization of the toxins in the lab at a 1:1 molar ratio.
• In mice given three times a lethal dose, 100 percent protection when binder and toxin were premixed at a 1:10 ratio and when the binder was given 15 minutes later. At 30 minutes, the short-chain binder still protected every mouse, and the long-chain binder protected 60 percent.

It was not a clean sweep. The binder for the tissue-damaging cytotoxins protected human cells in culture but did not significantly reduce tissue lesions in a preliminary mouse model, so its affinity likely needs work. This is a mouse and cell-culture study. Nobody has tested these proteins as a treatment in people.

Case study: designing antibodies from scratch

Antibodies matter because they are the largest class of protein drugs, and they are normally found by immunizing animals or screening huge libraries. Designing one on a computer, aimed at a specific spot on a target, is a long-standing goal. The paper “Atomically accurate de novo design of antibodies with RFdiffusion”, published online November 5, 2025, retrained the model on antibody structures. It generated single-domain antibodies (VHHs), single-chain fragments (scFvs) and full antibodies directed at chosen epitopes, then screened designs by yeast display.

The structural proof is strong: cryo-EM confirmed the predicted binding poses for VHHs against influenza hemagglutinin and the Clostridioides difficile toxin TcdB, and for scFvs against TcdB. The functional numbers are more modest.

Nanomolar binders straight from a computer are real progress. Therapeutic antibodies, though, typically need to bind far more tightly, resist aggregation, avoid triggering immune responses and be manufacturable at scale. None of those were the point of this paper.

What is still hard

Four limits recur across the literature:

1. Hit rates. Per-target success can range from respectable to near zero, and it is hard to predict in advance which target you have.
2. Affinity. First-round designs usually need experimental maturation to reach drug-level strength.
3. Function beyond binding. Designing a molecule that sticks is easier than designing one that does something, such as catalyzing a reaction efficiently.
4. Developability. Binding is the first filter. Stability, safety and manufacturing come later and are not solved by the generative step.

None of this makes the technology less interesting. It does mean that a headline claiming a model “designed a drug” usually describes a binder in a dish or a mouse, a long way from a medicine. When you see such claims, look for the success rate, the affinity and what organism or cell type the test used.

What to watch next

The next evidence will come from whether groups other than the original authors can reproduce these hit rates on new targets, and whether designed proteins reach human trials. For antibodies in particular, the interesting question is whether computational design can become a first step in a pipeline, narrowing the search before experiments, rather than a replacement for them. The papers so far support the first reading better than the second.

For readers outside the field, the takeaway is calibration. The technology is real, peer reviewed and improving quickly, and it has moved protein design from a specialist craft toward something closer to an engineering workflow. But a workflow with a 0 to 2 percent hit rate on some targets is still a research tool. Expect progress to show up first as faster discovery of starting points, and only later, if at all, as medicines.

Sources

• Nature, De novo design of protein structure and function with RFdiffusion (2023)
• Nature, De novo designed proteins neutralize lethal snake venom toxins (2025)
• Nature, Atomically accurate de novo design of antibodies with RFdiffusion (2025)
• NobelPrize.org, The Nobel Prize in Chemistry 2024