How to design a protein with AI

To design a protein with AI, you run RFdiffusion, an open-weight model from the Baker Lab that generates new protein shapes from scratch. You tell it a target or a motif, it draws a 3D backbone, a second model (ProteinMPNN) writes the amino acid sequence, and AlphaFold2 checks whether that sequence would fold back into the shape. You can do all of it in a free Google Colab notebook before touching a lab.

What is RFdiffusion?

RFdiffusion is a diffusion model for proteins, built by fine-tuning the RoseTTAFold structure-prediction network. It works the way image models like DALL-E do. It starts from random noise and denoises it step by step into a real protein backbone.

The Baker Lab at the University of Washington built it with collaborators at Columbia and MIT. It was posted as a preprint in December 2022, made open source in March 2023, and published in Nature in July 2023. It can design single proteins, binders that stick to a target, symmetric assemblies, metal-binding proteins, and scaffolds around an enzyme active site.

Sources: Institute for Protein Design | RFdiffusion on GitHub

Can I really run it for free?

Yes. The RFdiffusion repository links a Google Colab notebook maintained by Sergey Ovchinnikov as part of ColabDesign. Colab gives you a free GPU in the browser, so you do not need to install anything or own a graphics card.

The free tier has limits on session time and GPU access, but it is enough to design and screen your first batch of proteins. If you want it local instead, the repo ships a Docker image and a conda environment, and you download the model weights with a set of wget commands from the lab's file server.

The license matters here too. RFdiffusion is released under a BSD license, so it is free for both non-profit and for-profit use. That means you can build a commercial product on it without paying a license fee.

Sources: RFdiffusion on GitHub

What are the steps to design a binder?

A binder is a small protein that grabs onto a target, like a receptor on a cell. Here is the loop the Baker Lab pipeline uses. I call it the draw-write-check loop, because each model does one job and hands off to the next.

1. Pick a target. Get the structure of the protein you want to bind, usually a PDB file. Mark the patch on its surface you want your binder to sit on (the hotspot residues).
2. Draw the backbone. RFdiffusion generates many candidate backbones that fold against that hotspot. This is the shape only, with no sequence yet.
3. Write the sequence. ProteinMPNN takes each backbone and writes an amino acid sequence likely to fold into it.
4. Check the fold. AlphaFold2 predicts whether that sequence actually folds back into the intended shape and sits on the target. You keep only the designs where the prediction is confident and matches.
5. Rank and order. Sort the survivors by confidence scores, pick the top few dozen, and only then have genes synthesized and tested in a lab.

The whole first pass is compute, not chemistry. That is the point. You throw away thousands of bad designs on a server before you spend money on the good ones.

Sources: Institute for Protein Design

How well does it actually work?

In the Nature paper, the RFdiffusion pipeline hit an experimental success rate of about 19% for binder design. That means roughly one in five designs that passed the computer screen actually bound the target in the lab.

The team reported this was around two orders of magnitude higher than older Rosetta-based methods on the same targets. They counted a design as a success only if it bound at or above 50% of the response from a known positive control, so the real number may be a bit higher.

That is strong, but it is not magic. An independent 2025 preprint reported a low success rate when they used RFdiffusion to design binders for a specific biochemical detection task. The lesson is that success depends heavily on the target and what you are asking for. Treat the model as a way to narrow thousands of options down to dozens, not as a guarantee.

Sources: RFdiffusion paper (Nature 2023) | 2025 biorxiv preprint

Can it design enzymes, not just binders?

Yes, with a caveat. RFdiffusion ships a special "active site" model for enzyme work. That job is called motif scaffolding.

You give it the exact geometry of a catalytic site, the handful of amino acids that do the chemistry, and it builds a new protein scaffold that holds those residues in place.

But designing a working enzyme is harder than designing a binder, because catalysis needs the geometry to be nearly perfect. Expect more design-build-test rounds than you would for a simple binder.

Sources: RFdiffusion on GitHub

What does this cost, and where does the money go?

The design step is close to free. The expensive part is the wet lab: gene synthesis, protein expression, and the assays that tell you if a design binds. So the smart order is compute first, chemistry second.

Run hundreds of designs on a free or cheap GPU, keep the top few dozen by AlphaFold2 confidence, and only order those. This is the same pattern showing up across science. AI is getting cheap and fast at proposing candidates, while the physical test is still the bottleneck.

It is the same story behind Google DeepMind's GNoME materials database, where a model proposed millions of new crystals and the hard part was making and measuring them.

One more thing to plan for: these models run on GPUs, and GPUs run in data centers that draw real power. If you scale from a few designs to a screening campaign, your compute bill and its energy footprint both grow.

The short version

• RFdiffusion draws protein backbones. ProteinMPNN writes the sequence. AlphaFold2 checks the fold.
• You can run the whole thing free in a Google Colab notebook, no GPU of your own.
• It is BSD-licensed, so commercial use is allowed.
• Reported binder success was about 19% in the Nature paper, roughly 100x better than the older Rosetta method, but results vary a lot by target.
• Design on the server first. Spend the wet-lab budget only on the top designs.

If you want help wiring a design-then-screen pipeline into your own stack, that is the kind of thing I build at nomadtechnologist.com.

---

Not legal, financial, or tax advice.