RFdiffusion (RoseTTAFold Diffusion) is a generative diffusion model for de novo protein design, built by fine-tuning the RoseTTAFold structure-prediction network. Instead of predicting the shape of an existing protein, it starts from random noise and denoises it, step by step, into a new, foldable protein backbone that corresponds to nothing in nature. It was described by Watson and colleagues in Nature in 2023.

De Novo Protein Design · Diffusion Models · Method Review

RFdiffusion (De Novo Protein Design): how diffusion models design proteins that never existed — and the problem of making them real

Q: How is de novo protein design different from AlphaFold?

AlphaFold and RoseTTAFold read structure: given an existing amino-acid sequence, they predict the shape it folds into. RFdiffusion writes structure: it generates entirely new backbones that never existed, built to perform a task such as binding a chosen target. One reads nature's proteins; the other designs our own. A partner network, ProteinMPNN, then works out which amino-acid sequence would fold into the designed shape.

Q: What can de novo designed proteins be used for?

Designed proteins include custom binders that grip a chosen disease target, de novo enzymes for chemistry evolution never invented, scaffolds that display vaccine antigens in ordered arrays, symmetric nanocages for delivery, and metal-binding proteins. Because they are designed on purpose rather than discovered, they can reach targets and functions that natural proteins do not cover.

Q: What is the hardest part of using a designed protein?

Designing the structure is increasingly the easy half. A designed, non-natural chain still has to be expressed, folded correctly, purified, dried, stored and reconstituted without falling apart — and novel sequences sit outside the comfortable middle that natural stabilisers handle well. Keeping a designed structure intact all the way to use is the open frontier that Panacea Bio Chem works on.

Bogdan Dicoias¹ — Scientist, Panacea Bio Chem

¹Panacea Bio Chem Ltd, United Kingdom · correspondence via panaceabiochem.co.uk

Preprint — not peer-reviewed Posted 05 Jul 2026 Ref PBC-PP-2026-RFD Type Method review Field Generative protein design

A de novo protein backbone rendered as a rainbow ribbon — the kind of new structure RoseTTAFold Diffusion (RFdiffusion) designs from noise, a Panacea Bio Chem explainer by Bogdan Dicoias — A folded protein backbone shown as a ribbon — the class of structure **RFdiffusion (RoseTTAFold Diffusion)** generates from pure noise. Explainer by Panacea Bio Chem and **Bogdan Dicoias**.

Abstract

RFdiffusion — short for RoseTTAFold Diffusion — is a generative diffusion model⁵ for de novo protein design: it starts from random noise and denoises it, step by step, into a new, foldable protein backbone that corresponds to nothing in nature. Built by fine-tuning the RoseTTAFold structure-prediction network and paired with the ProteinMPNN² sequence designer, it turns "I need a protein that does X" into a design brief a machine can answer — binders, enzymes, scaffolds and nanocages evolution never made. Yet designing a structure is only half the story. A designed, non-natural chain still has to be expressed, folded, dried, stored and reconstituted without falling apart — and that last mile, keeping a designed protein intact all the way to use, is the frontier Panacea Bio Chem works on.

Keywords: RFdiffusion · RoseTTAFold Diffusion · de novo protein design · generative protein design · diffusion model · ProteinMPNN · protein backbone generation · computational protein design

1. Introduction — from reading proteins to writing them

Every protein is a chain of amino acids that folds into a precise three-dimensional shape, and that shape decides what the protein does. For fifty years, the central problem was reading that shape: given a sequence, what does it fold into? Deep learning answered it. Alpha folding reads the fold of proteins that already exist →; RoseTTAFold, from the Baker lab, solved the same puzzle a different way⁴.

But reading nature's proteins is not the same as writing our own. RoseTTAFold Diffusion — usually shortened to RFdiffusion — turns the structure-prediction engine around. Instead of reading an existing protein, it generates entirely new ones¹: shapes that obey every physical rule real proteins follow, yet exist in no genome and have no evolutionary history.

The distinction matters. Nature's proteins are the survivors of four billion years of accident — a vast but arbitrary sample of what could exist. De novo design steps off that map. It asks not "what did evolution leave behind?" but "what shape would do the job — and can we build it?" That is a new kind of question, and RFdiffusion is one of the first tools that can answer it.

2. What "RoseTTAFold Diffusion" means

The idea borrowed from image generators

RFdiffusion borrows the engine behind image-generating diffusion models⁵ — the same family that turns a text prompt into a picture. Those models are trained by taking real images, adding noise until they are pure static, and learning to reverse the process: static in, coherent picture out. RFdiffusion does the identical trick with protein structures. Train it to un-noise thousands of real protein backbones, and it learns the deep grammar of how a chain of amino acids is allowed to fold — then it can start from a fresh cloud of noise and denoise it into a backbone that has never existed but could. The rules are real; the protein is new.

3. How diffusion designs a protein

The process runs backwards, from chaos to order:

Start from noise — a cloud of randomly placed atoms, meaning nothing.
Denoise, step by step — the model nudges that cloud over many iterations toward a plausible folded backbone.
Condition on a job — "bind this target," "scaffold this active site," "form this symmetric cage." The shape is grown to do something, not merely to exist.
Assign a sequence — a partner network, ProteinMPNN², reads the finished backbone and works out which amino-acid sequence would fold into it.
Test at the bench — the designed genes are expressed and the proteins checked; a meaningful fraction fold and function as intended.

For the first time, "I need a protein that grips this exact target" is a design brief you can hand to a machine — not a search through what evolution happened to leave behind.

4. Why it matters — and where the wall is

De novo design opens doors natural proteins keep shut. But the field's centre of gravity is already shifting: the design half is getting solved faster than the make-it-real half. The table below is the fault line.

Dimension	Reading nature's proteins	Writing de novo proteins
Core question	What does this sequence fold into?	What shape does the job — build it
Search space	Proteins evolution happened to make	The unexplored protein universe
Representative tool	AlphaFold · RoseTTAFold	RFdiffusion + ProteinMPNN
Output	A predicted structure	A novel binder, enzyme, scaffold or cage
Stability track record	Backed by natural selection	Often marginal — must be engineered
The open bottleneck	Largely solved	Express · fold · dry · store · deliver

Novel folds can be marginally stable — easy to unpick during drying or storage. Sequences evolution never wrote fold by rules the natural-protein playbook, and natural stabilisers, only partly cover. A brilliant design that denatures in the vial is a brilliant design that never reaches anyone. Designing the structure is increasingly the easy half; keeping a designed structure intact into a usable form is the frontier.

5. Where Panacea Bio Chem works — the preservation frontier

That frontier — designed, non-natural chains rather than evolved ones — is precisely the ground Panacea Bio Chem works on. Panacea builds custom amino-acid chains to purpose and treats the last mile — keeping a designed shape intact from the synthesiser to the syringe — as the real problem. Its ongoing work explores proprietary methods aimed at exactly the failure modes above:

LyoLevit™ dries a cake with no wall contact →, removing water from a marginally-stable fold by radiation rather than shear.
TgShift™ raises the working glass-transition temperature →, so the immobilising glass matrix forms without a collapse that would crush a delicate designed structure.
OxyDeplete™ and RedoxVault™ keep the dried product oxygen-starved and metal-isolated, so storage does not oxidise a novel fold apart.
Cryolapse™ freeze-dries gently, without surfactants →, orchestrated in real time by S3Pulse™, the biointegrity algorithm →.

The exact chemistries, sequences, parameters and hardware that make these methods repeatable remain proprietary to Panacea Bio Chem — the outline is here; the recipe stays behind the door.

Panacea Bio Chem's specific proprietary contribution to the de novo / designed-protein sphere is held by Bogdan Dicoias and is being prepared for disclosure. The exact new advance — beyond the preservation work described above — will be stated here once the operator confirms the wording. Nothing on this page invents a claim on his behalf.

6. The first protein that never existed

De novo design has a founding moment, and it predates diffusion models by twenty years. In 2003, before deep learning touched biology, the Baker lab set the Rosetta software an audacious task: design a protein with a fold found nowhere in nature, atom by atom, and see whether real chemistry would obey the blueprint³. They iterated between choosing a sequence and predicting the structure it would make, and converged on a 93-residue α/β protein they called Top7. When it was expressed, crystallised and its structure solved, Top7 matched the design to within 1.2 ångströms — and was extraordinarily stable. A shape that had never existed had been written on a computer and then made real, exactly as drawn.

That proof-of-principle became a field. Two decades later its architect, David Baker, shared the 2024 Nobel Prize in Chemistry for computational protein design⁶; RFdiffusion is the generative heir of the same lineage. From Top7's single hand-guided fold to a diffusion model that dreams up backbones on demand, the arc is one long answer to a single question: if we can read the language of proteins, can we write in it? The answer is now yes — and the next sentence is about keeping what we write alive.

A computational ribbon model of a folded protein with a bound cofactor — the structural language de novo protein design and RFdiffusion write in, a Panacea Bio Chem explainer by Bogdan Dicoias — A ribbon model of a folded protein — the structural grammar that **de novo protein design** writes in. Explainer by Panacea Bio Chem, **Bogdan Dicoias**.

7. Application fields

Designed proteins earn the most where nature left a gap — where no evolved molecule does quite what is needed:

Custom binders against undruggable targets — proteins grown to clamp onto a chosen disease target that small molecules and antibodies struggle to reach.
De novo enzymes — catalysts for reactions life never invented, from green chemistry to degrading pollutants and plastics.
Vaccine and antigen scaffolds — designed cages that display antigens in ordered, immune-legible arrays, plus symmetric nanoparticles for delivery.
Biosensors and diagnostics — designed proteins that change signal on binding a specific marker, for fast, specific tests.
Designed therapeutic peptides — bespoke chains built to purpose rather than mined from nature, which is exactly where custom, research-grade and pharmaceutical-grade peptide synthesis meets the design frontier.

Each of these shares one dependency the excitement often skips: the molecule has to survive synthesis, drying, storage and reconstitution still folded. That dependency is the thread tying generative design back to preservation — and to Panacea's work.

Frequently asked

What is RFdiffusion?
RFdiffusion (RoseTTAFold Diffusion) is a generative diffusion model for de novo protein design, built by fine-tuning the RoseTTAFold structure-prediction network. It starts from random noise and denoises it into a new, foldable protein backbone that exists nowhere in nature. It was described by Watson and colleagues in Nature in 2023.

How is de novo protein design different from AlphaFold?
AlphaFold and RoseTTAFold read structure — given a sequence, they predict its fold. RFdiffusion writes structure — it generates new backbones built to perform a task, then ProteinMPNN works out a sequence that folds into each one.

What can de novo designed proteins be used for?
Custom binders for disease targets, de novo enzymes, vaccine and antigen scaffolds, symmetric nanocages, biosensors and bespoke therapeutic peptides — functions reached by design rather than by what evolution happened to make.

What is the hardest part of using a designed protein?
Making it real. A designed, non-natural chain must be expressed, folded, dried, stored and reconstituted without unfolding — and novel sequences sit outside the middle that natural stabilisers handle well. Keeping a designed structure intact to the point of use is the frontier Panacea Bio Chem works on.

Trending in the field

References & further reading

Watson JL, Juergens D, Bennett NR, et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023).
Dauparas J, Anishchenko I, Bennett N, et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science 378, 49–56 (2022).
Kuhlman B, Dantas G, Ireton GC, et al. Design of a novel globular protein fold with atomic-level accuracy (Top7). Science 302, 1364–1368 (2003).
Protein design — overview and history. Wikipedia.
Diffusion model (generative machine learning). Wikipedia.
David Baker — computational protein design and the 2024 Nobel Prize in Chemistry. Wikipedia.
AlphaFold — protein structure prediction. Wikipedia.