How AI Is Designing Proteins That Don’t Exist in Nature: What’s Been Proven, and What Hasn’t

Published October 27, 2026. Facts checked against the linked papers, regulators’ and companies’ releases, and news reports on that date. This field moves quickly, so verify the latest status before relying on any number here.

The short version. Experimental structures now confirm that AI-designed proteins can match their computer models, and in some cases bind real targets with atomic-level accuracy. One computationally designed vaccine has been approved, though it was built with earlier, non-generative methods. No drug designed by today’s generative models has yet been approved. Most headline hit rates come from companies and preprints, and at least one fully de novo drug candidate has already been dropped after Phase 1.

Why this is new

Every natural protein was shaped by evolution, which keeps only what survives. De novo design skips that filter: a model proposes a protein with no natural ancestor, and a lab tests whether it works. The first de novo protein, Top7, came from David Baker’s lab in 2003 (BioRender timeline). In 2024, the Nobel Prize in Chemistry recognized Baker for computational protein design, and Demis Hassabis and John Jumper for protein structure prediction (Xataka).

Diffusion design works like an image generator. Start with noise and shape it, step by step, into a protein backbone. RFdiffusion did this by fine-tuning the RoseTTAFold structure-prediction network on denoising tasks (Watson et al., Nature 2023). A second tool then writes an amino-acid sequence that should fold into the backbone, and the lab makes the protein and finds out.

A note on terms

These words get mixed up in headlines, and the difference matters here.

• Computationally designed: any protein built with software, including older physics-based tools such as Rosetta.
• De novo: no natural protein used as the starting point.
• Generative AI design: deep-learning models such as RFdiffusion or ESM3 that generate new proteins.

The seven cases at a glance

Case 1: RFdiffusion shows the structures are real

The 2023 paper characterized hundreds of designed symmetric assemblies, metal-binding proteins and binders. The key result was a cryo-EM structure of a designed influenza haemagglutinin binder that was nearly identical to its design model (abstract). That is the standard to hold the field to: an experimental structure matching the model, not just a confident prediction.

Baker’s group has since released RFdiffusion2, tuned for enzymes, and RFdiffusion3, a broader tool released as open source in 2025 (GeekWire; GEN).

Case 2: Snake venom, with fine print

A 2025 Nature paper reported proteins designed to bind three-finger toxins. They neutralized all three toxin subfamilies tested in vitro and protected mice from a lethal neurotoxin challenge (Vázquez Torres et al.). Survival ranged from 80 to 100 percent depending on dose, toxin, and design (AlphaGalileo summary).

This is not an antivenom yet. The proteins don’t protect against whole venom, which is a mixture that differs by species. First author Susana Vázquez Torres estimates around ten designed proteins would be needed to cover a complete venom, and she describes a long road before human approval (Popular Science).

Case 3: Antibodies from scratch

Antibody binding loops are floppy and have been very hard to design. A fine-tuned RFdiffusion2 network plus yeast display screening produced VHHs, scFvs and full antibodies that bind user-specified epitopes. Cryo-EM confirmed the binding poses for VHHs against influenza haemagglutinin and C. difficile toxin B, and one scFv structure verified all six CDR loops (PubMed abstract). The deposited structures are public: 9NH7 and 9NFU. The paper appears in Nature 649:183–193.

The lead author notes that early RFdiffusion proteins were small minibinders studied in the lab, while therapeutic antibodies must work in the human body (GEN). Read this as a proof of concept, not a drug.

Case 4: ESM3 and the glowing protein

ESM3 is a language model that reasons over sequence, structure, and function together. Prompted to generate fluorescent proteins, it produced one bright protein with 58% sequence identity to known ones, which the authors estimate equals 500 million years of evolution (Science 387:850–858 record; ScienceAlert). The “500 million years” figure is the authors’ estimate, and esmGFP is one flagship result, so it shows what is possible, not how often.

Case 5: SKYCovione, an approved design with a caveat

This COVID-19 vaccine carries 60 copies of the spike receptor-binding domain on a designed nanoparticle. The University of Washington calls it the first computationally designed protein medicine, and it was approved in South Korea and later in the U.K. (IPD). The Phase 3 trial enrolled 4,037 adults (GeekWire).

“First” here is the developers’ own claim, and the scaffold came from earlier physics-based methods, years before diffusion models. It shows designed proteins can clear regulators, not that generative AI has.

Case 6: NL-201, the sobering one

Neoleukin believed NL-201 was the first fully de novo protein evaluated in clinical trials. Phase 1 data showed target engagement and no significant immunogenicity after multiple cycles, yet the company discontinued it in November 2022 (company release). Its CEO said the data wouldn’t be dramatically differentiated from other IL-2 molecules (BioSpace), and the company later cut most of its remaining staff (Fierce Biotech). Working and safe wasn’t enough in a crowded field.

Case 7: The hit-rate race

Chai Discovery reports that Chai-2 reached a 16% binding rate across 52 targets with no known antibody binders in the database it used, with at least one binder for 50% of targets and a 68% success rate on miniproteins (Chai Discovery). Its technical report says earlier de novo antibody success rates were typically below 0.1% (technical report). Latent Labs reports low immunogenicity for de novo VHHs in a ten-donor human panel, and says animal studies and trials remain ahead (Latent-X2 preprint).

Reported lab success rates (percent of tested designs)

Earlier de novo antibody methods, typical best: below 0.1%

Chai-2 antibodies: 16% (≤20 designs per target, 52 targets)

AlphaDesign retroviral inhibitors: 17 of 88 selected designs (19%)

Chai-2 miniproteins: 68%

Sources: Chai Discovery; DenovAI/Molecular Systems Biology report. Each study used different targets, formats, and definitions of a “hit,” so read the bars as scale, not as a ranking. Chai-2’s figures are company-reported and, per this review, had not cleared peer review at last check.

Where this still breaks

• Confidence scores can mislead. In one 2025 preprint, a Gremlin-1 design campaign produced a picomolar binder, but several other designs with similarly high AlphaFold-derived confidence showed no detectable binding (bioRxiv).
• Enzymes are harder. A sequence-design tool reports raising in silico success on a hard enzyme benchmark from 6.56% to 17.57% (listing). That is a big relative gain from a low base.
• A binder isn’t a therapy. A 2026 study of de novo binders as CAR-T targeting domains found three problems: tonic signalling, occluded epitopes, and off-target activity (PMC).
• Biosecurity needs attention. A 2025 Science study found that protein-design tools could rewrite the DNA of known toxins so that the screening software used by DNA synthesis companies missed them. This was a test of specific tools and specific screening systems, and a patch was developed before publication (NPR). It does not show that every design model is equally risky.

What I’d conclude

• The core science is real: experimental structures match designs, which is the strongest evidence in the field.
• Some binders are becoming tractable to design. Dependable, broadly applicable design is still unproven, and hit rates vary widely by target.
• No generative-AI-designed medicine has been approved. The one approved vaccine was designed using earlier methods.
• The bottleneck is shifting from finding binders to making them developable, safe, and better than what already exists.
• Many of the biggest numbers are company-reported or preprints, so independent replication is the thing to watch.

Practical tips

If you’re reading a headline

• Check the evidence level: peer-reviewed paper, preprint, or press release.
• Look for an experimental structure (cryo-EM or X-ray).
• Ask how many designs were tested for each success.
• Treat “in mice” and “in vitro” as early-stage.

If you build or fund this technology

• Budget for wet-lab validation, because confidence scores alone mislead.
• Plan for stability, immunogenicity, and manufacturing from the start.
• Benchmark against the incumbent drug, as NL-201’s story suggests.
• Build biosecurity review into your workflow.

Sources

1. Watson et al., RFdiffusion, Nature 2023: paper PDF
2. Vázquez Torres et al., snake venom toxins, Nature 2025: nature.com
3. Bennett et al., de novo antibodies, Nature 649:183–193: PubMed; structures 9NH7, 9NFU
4. Hayes et al., ESM3, Science 387:850–858, 2025: record
5. SKYCovione: University of Washington IPD; GeekWire
6. Neoleukin NL-201, November 14, 2022: company release
7. Chai-2: announcement; technical report (preprint)
8. Latent-X2 (preprint): arXiv 2512.20263
9. Biosecurity screening study, Science, October 2025: NPR report