De novo AI-designed binders
Peptides and small proteins invented from nothing by diffusion models rather than found in nature, now routinely producing picomolar-to-nanomolar binders against targets that had no drug-like chemistry at all.
Also known as RFdiffusion binders, AlphaProteo peptides, computationally designed minibinders, de novo protein binders
In vitro only — Cell or tissue studies. A mechanism, not yet an effect in a living body.
Genuinely transformative in-vitro and structural results with crystallographic confirmation that designed binders adopt their intended fold, plus a growing set of mouse efficacy studies. Zero human clinical data. Treat published affinity numbers as real and published therapeutic claims as aspiration.
How it works
RFdiffusion applies a denoising diffusion model to protein backbone geometry: starting from noise, it generates a backbone shaped to a specified target epitope, and ProteinMPNN then designs a sequence predicted to fold into that backbone, with AlphaFold2 or RoseTTAFold used as an in-silico filter before anything is expressed. DeepMind's AlphaProteo reported 3- to 300-fold better binding affinities and substantially higher experimental hit rates than prior methods across seven targets. Newer atom-level models such as Latent-X have pushed hit rates and affinities further still. The practical consequence is that the rate-limiting step in biologics discovery has shifted from finding a binder to validating it: designed minibinders have been produced against SARS-CoV-2 spike, PD-L1, IL-7 receptor alpha, HER2, VEGF-A, TrkA and snake-venom toxins, several with in-vivo efficacy in mice. Compared with antibodies, minibinders are small, hyperstable, cheap to make, and can be produced synthetically or in bacteria — but they are also immunogenic in unpredictable ways precisely because they have no evolutionary relationship to anything in the human proteome.
Targets: Arbitrary - the method chooses the target epitope, Demonstrated examples include PD-L1, HER2, VEGF-A, IL-7Ralpha and SARS-CoV-2 spike
Dosing
| Protocol | Dose | Frequency | Route |
|---|---|---|---|
| No clinical protocol existsNot applicable. | — | not established | subcutaneous |
- · As of mid-2026 no de novo designed binder has a published phase 2 readout, and this is a method rather than a compound. There is no dose because there is no drug — only a growing catalogue of validated preclinical leads.
Cycling
Not applicable.
Pharmacology
- Half-life
- Very short unless engineered — a 60-residue mini-protein is below the renal filtration cutoff and clears in minutes to hours without Fc fusion, albumin binding or PEGylation.
- Onset
- Not defined as a class.
- Routes
- subcutaneous, intravenous, inhaled, topical
- Molecule
- Computationally designed peptides and mini-proteins, typically 12-80 residues, with no natural sequence ancestor
Handling
- Diluent
- Not applicable
- Lyophilised
- Freezer for research material; designed mini-proteins are often unusually thermostable by construction.
- Reconstituted
- Refrigerated.
Mixing
Produced by bacterial expression or long-chain synthesis in research settings. Nothing in this category is available to consumers, and any vendor claiming to sell 'AI-designed peptides' is selling a marketing story.
Side effects
- commonImmunogenicity and anti-drug antibody formation— The central unresolved risk: sequences with no human homologue are novel to the immune system by definition.
- commonUnknown human profile— No published human exposure data for any designed binder.
Do not use if
- Any non-research human use — there is no characterised molecule here to have a contraindication about.
What to monitor
- · Anti-drug antibody titres would be the first-order safety readout in any future trial.
Legal status
Research materials. No approved product; no consumer availability.
References
- Watson et al. 2023, Nature — de novo design of protein structure and function with RFdiffusion (preclinical)
- Zambaldi et al. 2024 — de novo design of high-affinity protein binders with AlphaProteo (preclinical)
- Dauparas et al. 2022, Science — robust deep-learning-based protein sequence design with ProteinMPNN (preclinical)
Mechanism in depth
Not a compound but a design method: diffusion models and inverse-folding networks generate backbone geometries and sequences for peptides or mini-proteins that bind a chosen target surface, without starting from a natural binder. Success rates have improved by orders of magnitude in a few years, and designed binders now reach picomolar affinity against selected targets.
What usually goes wrong
Binding is not efficacy. A designed binder can have exceptional affinity and no useful biological effect, because occupying a surface is not the same as changing what a protein does. Immunogenicity of non-natural sequences is also a real and under-discussed risk, and computational success rates are quoted from binding assays rather than from cells or animals.
Receptor targets
- Whatever the design specified — Varies from micromolar to picomolar
Binding is designed; function is not automatic
Genuinely uncertain
- Very few designed binders have entered clinical trials.
- Immunogenicity of de novo sequences in humans is largely uncharacterised.
- Reported success rates come from in vitro binding, not from therapeutic effect.