RFD4-Proteina: Nvidia and UW release open AI model for protein design
The University of Washington's Institute for Protein Design and Nvidia have unveiled RFD4-Proteina, an open generative model for designing new proteins. The developers say it delivers five to ten times higher success rates and speed than earlier models. It's released under a BSD-3 license and available through many partners.
What this is about
The Institute for Protein Design (IPD) at the University of Washington School of Medicine has teamed up with Nvidia to introduce RFD4-Proteina, an open generative AI model for designing new proteins with desired functions. It's built on Nvidia's BioNeMo platform for chemistry and biology and, according to the developers, is available right away as an open-source resource under the BSD-3 license. Possible applications include proteins that interrupt disease processes, or proteins that detect signs of disease in diagnostic tests.
How the model works
RFD4-Proteina simultaneously refines a protein's sequence (the order of its building blocks) and structure (its three-dimensional shape), along with the structures of the molecules it interacts with. A change in one place triggers matching changes elsewhere. Scientists and AI agents can search for and refine designs based on molecular inputs, constraints and scoring goals.
The name comes from two predecessors:
- RosettaFold Diffusion (IPD): a "guided diffusion" model that builds proteins step by step out of noise, much like image generators.
- Proteina-Complexa (Nvidia): a generative model for binder design against protein and small-molecule targets. It combines inference-time optimization with a pretrained flow-based model built on Nvidia's La-Proteina for atom-level design.
Performance
The developers say the model achieved five to ten times higher design success rates and generates candidates five to ten times faster than earlier models. Nvidia states that it beat the compared baselines on computational success rate and throughput across four benchmark suites: protein binding, small-molecule binding, enzyme scaffolding (building a supporting framework around an active site) and DNA binding. These are vendor claims, and no independent verification is mentioned.
Two IPD researchers report real-world results: Meg Lunn-Halbert was able to precisely control which glycans (sugars important for immune recognition) attach to proteins, in under a week instead of months. Seth Woodbury designed a pesticide-degrading enzyme in a few days that performed many times better than his previous best, which had taken months.
Availability and who's involved
RFD4-Proteina is accessible through partners: Amazon Bio Discovery, Anthropic's Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio. That makes it easy to plug into agent, cloud and scientific workflows. The lead developers were UW doctoral students Nathaniel Corley, Maxwell Kazman and Rafi Brent, plus Kieran Didi (a visiting student from Oxford). Funding came from the Howard Hughes Medical Institute, the Gates Foundation, the Grantham Foundation, Schmidt Sciences and the OpenFold consortium, among others. The Jen-Hsun and Lori Huang Foundation provided a compute grant.
David Baker, director of the IPD and 2024 Nobel laureate in chemistry for computational protein design, said:
We are moving from an era in which we discover the proteins nature has created to one in which we can design entirely new ones.
He expects better medicines, as well as applications in polymer degradation and bioplastics.
Our take
Open protein design models lower the barrier for labs without large infrastructure of their own, and the broad partner list makes access easier. One caveat: the reported success rates are mostly computational figures and vendor claims. Whether designs hold up in the lab still has to be tested experimentally. According to the developers, the model helps narrow down and prioritize candidates but doesn't replace lab testing. The source gives no details on model size or hardware requirements.
Sources
- Google News (US) genengnews.com
- Bing News (en-US) hoodline.com
Try it yourself: RFD4-Proteina
Explained step by step, no prior knowledge needed – on your own computer or free in your browser.
What can I do with it?
RFD4-Proteina designs new proteins for specified tasks, such as binders for target proteins or small molecules, enzymes or DNA-binding proteins. You provide molecular inputs, constraints and scoring goals, and the model generates and refines candidates for lab testing.
Who is it worth it for?
Of interest to biotech researchers, bioinformaticians and labs that prepare protein design computationally. Not useful for laypeople without domain knowledge and a lab, since it isn't a chat model.
What your computer needs
The source doesn't mention any variants or sizes. The easiest way in is through partner platforms such as Hugging Face, Google Cloud, Databricks, Benchling or Tamarind Bio rather than running it locally.
- The source doesn't state any hardware requirements.
- Protein design models usually need a powerful Nvidia GPU; exact figures are unknown here.
- Cloud access through partners such as Google Cloud, Nebius or Tamarind Bio is an alternative to local hardware.
How do I find out what my computer has?
- Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
- Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
- Linux: Type
nvidia-smiin the terminal (shows your graphics card and its memory) andfree -hfor system memory.
Guide
This model runs with its developers' own program (RosettaCommons/RFD4-Proteina on GitHub). The commands below come from the Model Card and the developers' instructions.
In the browser (no installation)
The developers have an online demo on Hugging Face. You can try the model right away there without installing anything. Sometimes you have to wait a moment for the demo to wake up.
Windows 10 / 11
Windows 10 / 11: not supported. According to the README, Windows and macOS are not supported. Alternatives: use the Hugging Face Studio Spaces in your browser (ZeroGPU or dedicated GPU, Hugging Face login required), or WSL2 on Windows (Linux inside Windows), though that is untested.
macOS
macOS: not supported. According to the README, macOS and Windows are not supported; you need an NVIDIA graphics card with CUDA. Alternative: Studio in your browser on Hugging Face (Spaces "rfd4-proteina-zerogpu" or "rfd4-proteina-space", login required).
Linux
Linux: supported. Officially supported is Linux x86_64 with Python 3.12 and an NVIDIA graphics card with a CUDA 13-compatible driver. The "supported hardware" table in the documentation lists which cards have been tested. Without a suitable card, use the browser-based Studio Space on Hugging Face.
- Check prerequisites. You need Linux (64-bit), Python 3.12, Git, make and an NVIDIA graphics card with a current driver (CUDA 13 compatible). Run nvidia-smi to see whether your card and driver are detected. Plan for about 7 GB for the environment plus 1.7 GB for the model.
nvidia-smi python3.12 --version git --version - Download the code. This clones (copies) the code from GitHub. --depth 1 makes it faster by skipping the old history.
git clone --depth 1 https://github.com/RosettaCommons/RFD4-Proteina.git cd RFD4-Proteina - Set up. make setup creates the environment and downloads the model (about 1.7 GB) into the cache. This takes a while. If it asks whether you want AlphaFold environments, you don't need them for a first test – you can use bash setup.sh --skip-search-envs instead.
make setup - Activate the environment. This makes your terminal use the newly created Python environment. You need to repeat this in every new terminal.
source .venv/bin/activate - Save your design request as a file. Save this text as binder.yaml in the RFD4-Proteina folder. It asks for a binder (a protein that docks onto a target) of 50–120 residues for the protein PD-L1. The input file is already in the repository.
cat > binder.yaml <<'EOF' binder: input: docs/examples/inputs/5o45_cropped.cif.zst contig: "50-120, /0, A17-131" conditions: {C_CRDGRP: {1: [{select: "sel('A/*/*/[N,CA,C,O]')"}]}, C_SEQ: {true: [{select: "sel('A')"}]}} EOF - Generate your first design. This command produces two designs. '~metrics' turns off scoring, and compile_lengths=null saves warm-up time. The quotes around '~metrics' must stay.
rfd4-generate inputs=binder.yaml num_replicates=2 out_dir=./runs/first '~metrics' compile_lengths=null - View the results. The finished structures are in the output folder as CIF files (a format for 3D molecular structures). You can open them with a molecule viewer such as PyMOL or ChimeraX.
ls runs/first/inference_samples/ - Result. The folder runs/first/inference_samples/ contains two generated protein structures as CIF files.
Try it
Design a protein binder for a given target protein with high binding affinity.Generate an enzyme scaffold around a given active site to break down a pesticide.Design a protein that binds a specific DNA sequence.If something goes wrong
This is not a text model; inputs are molecular structures and target specifications, not free-form chat prompts.
The exact Hugging Face repository and installation details aren't in the source; check the official documentation.
Computational success rates don't guarantee function in the lab; candidates must be tested experimentally.
“out of memory,” a crash, or your computer freezes
The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.
It runs, but extremely slowly
Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).
“ollama” or “python” is not found
Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.
Windows: "Running scripts is disabled on this system"
Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.
Error message when loading the model (version, "trust_remote_code", unknown architecture)
Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.
Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: bsd-3-clause.
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.