RFD4-Proteina

RFD4-Proteina

Designs protein sequences and all-atom structures together around a binding target, enzyme motif or ligand that you specify. It is a research model from Baker Lab and NVIDIA for molecular designers.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: RFD4-Proteina is a flow-matching model for designing biomolecules. It generates a protein's sequence and its all-atom structure at the same time. You choose what the design must keep, such as the target it binds, the motif of an enzyme or a small molecule bound to it. It comes from Baker Lab and NVIDIA GenAIR.

Who it suits: It suits computational biologists and protein engineers who design molecules around a target and can run Python on an NVIDIA GPU.

Industry | Who uses it | How they use it | Example
Biotech | Protein designers | Generate a sequence and structure around a binding target | Designing a binder for a chosen target
Enzyme engineering | Enzyme researchers | Keep an enzyme motif fixed and design the rest | Building a scaffold around a motif
Drug design | Structural biologists | Design around a ligand pocket | Generating candidates that fit a given ligand
Research | Machine learning researchers | Study joint sequence and structure generation | Comparing the two released checkpoints

People without a supported NVIDIA GPU and a Linux machine can skip running it locally.

Where to use it:

• Linux x86_64 with Python 3.12
• A supported NVIDIA GPU with a CUDA 13-compatible driver
• The RFD4 Python package
• Diffusers, through a separate Diffusers repository
• Hosted studio pages on Hugging Face, one on ZeroGPU and one on a dedicated GPU

Use case | Who | What they get back
Target-conditioned design | Protein designers | Designed structures in CIF format with generated residues and coordinates
Motif scaffolding | Enzyme researchers | A designed molecule built around the motif you supply
Ligand pocket design | Structural biologists | Designs shaped around the ligand you specify

How to set it up:

• A CUDA 13-compatible driver
• The RFD4 Python package
• About 1.74 GB of disk for the weights file, plus a config file
• Inputs as annotated CIF structures or YAML condition files

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install the RFD4 package and load the native weights.

pip install -U huggingface_hub

hf download nvidia/RFD4-Proteina

How to use it: Load the native weights with the RFD4 loader, then give the model a condition, either an annotated CIF structure or a YAML specification. The Diffusers version lives in its own repository and uses a pipeline class.

Proteina.from_pretrained("nvidia/RFD4-Proteina")

RFD4ProteinaPipeline.from_pretrained("nvidia/RFD4-Proteina-Diffusers")

Tool | What it does
Proteina.from_pretrained | Loads the native weights stored at the root of this repository
RFD4ProteinaPipeline.from_pretrained | Loads the same model through Diffusers, from its own repository

Example: You supply an annotated CIF file that marks a binding target and ask for a design. The model returns a designed structure written as a CIF file, listing the residues it generated and their coordinates. The card gives no sample command beyond the loaders.

For newcomers and experts: This is a specialist research tool, so a first step is to read the project documentation and download the files. Running it needs a suitable GPU and some structural biology background.

Each release pairs a flow model with the autoencoder that was trained alongside it. They are not interchangeable across releases, since mixing them yields nonsense without an error. The current flow model is the pretrained base after a rank-32 LoRA fine-tune was merged into it.

You give a CIF structure or YAML condition
Flow model and autoencoder work together
Sequence and structure are generated jointly
A designed structure comes out as CIF

Good to know:

• Third-party assets: the optional ones come with their own terms.
• License: BSD 3-Clause, covering the model and weights.
• Cost to run: none named. You need your own NVIDIA GPU, and the card names no paid service.
• Data it reaches: it downloads files from the Hugging Face Hub. Training data is a separate download.
• Not for: mixing checkpoints from different releases, which gives nonsense results.

Title: Residue vocabulary size by release
Current release: 147 classes
Previous release: 33 classes

On Hugging Face, 13 November 2026

• Library: rfd4-proteina
• Parameters: 436M
• Weights: about 1.7 GB in F32; about 218 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 14,235
• Likes: 45
• License: bsd-3-clause
• Languages: en
• Fine-tunes: 2
• Tags: rfd4-proteina, safetensors, rfd4_proteina, biology, protein-design, biomolecular-design, structure-generation, flow-matching, annotated-cif

Links

• Model: https://huggingface.co/nvidia/RFD4-Proteina
• Files: https://huggingface.co/nvidia/RFD4-Proteina/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/RFD4-Proteina

Then follow the usage in its model card. Nothing in this kit runs by itself.