Spaces:
Running on Zero
Running on Zero
| title: RFD4-Proteina Diffusers | |
| emoji: 🧬 | |
| colorFrom: blue | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 6.31.0 | |
| app_file: app.py | |
| python_version: "3.12" | |
| startup_duration_timeout: 1h | |
| short_description: Joint sequence-structure protein design from annotated CIF | |
| # RFD4-Proteína · Diffusers | |
| Interactive demo of [`nvidia/RFD4-Proteina-Diffusers`](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers) — the | |
| [Diffusers](https://huggingface.co/docs/diffusers/api/pipelines/rfd4_proteina) release of the RFD4-Proteína | |
| flow-matching model (Baker Lab × NVIDIA GenAIR), which designs protein **sequences and all-atom structures together**, | |
| conditioned on the molecular context you want to preserve: a binding target, an enzyme motif, a ligand, or selected | |
| structural features. | |
| Upload an **annotated CIF** describing what to hold fixed and what to design (or pick one of the author examples below), | |
| choose a design mode, and press **Design**. The output is a designed structure rendered in 3D plus a downloadable | |
| annotated CIF written back with the generated sequence and coordinates. | |
| ## Design modes | |
| - **Full design** — the tokens marked as designable are generated from pure noise: sequence and structure jointly. | |
| - **Inverse folding** — the input backbone is kept and only the sequence is redesigned. | |
| - **Partial diffusion** — the input is re-noised to the point where a chosen fraction of the schedule remains, then all | |
| 60 steps run over that range (the released example uses 0.4). Useful for relaxed re-designs of an existing structure. | |
| ## Inputs and outputs | |
| The input CIF uses [AtomWorks Studio](https://github.com/atlas-labinit/atomworks) annotations to say which residues, | |
| atoms, ligands or motifs are context (`is_in_context`) and which are to be designed. The demo prepares it with the | |
| authors' [`rfd4-proteina-prep`](https://huggingface.co/nvidia/RFD4-Proteina-prep) adapter (vendored verbatim into this | |
| Space as `rfd4_proteina_prep/`, because the Hub's git server does not serve partial clones to pip) and samples with | |
| `RFD4ProteinaPipeline` (the released `thinking_medium` recipe: 60 steps, 180 trunk passes, BF16). | |
| ## Examples | |
| The bundled example inputs are the authors' own, taken from the | |
| [`examples/` folder](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers/tree/main/examples) of the model repository | |
| (protein binder · PD-L1, enzyme catalytic motif, motif scaffolding · 1QJG, small-molecule binder · SAM, unconditional, | |
| inverse folding · 2KL8). They demonstrate the conditioning workflow and are not experimental binding or | |
| catalytic-activity results. | |
| ## Biosafety note | |
| The official RFD4-Proteína deployment applies [SecureDNA](https://secure-dna.org) screening to every design before it is | |
| returned. **This demo does not screen designs.** Use it for computational biomolecular design research only, and follow | |
| designs with appropriate structural and experimental evaluation. | |
| ## Acknowledgements | |
| Model, weights and example inputs: RFD4-Proteína, developed by the Baker Lab (Institute for Protein Design, | |
| University of Washington) and NVIDIA GenAIR — [BSD-3-Clause](https://huggingface.co/nvidia/RFD4-Proteina-Diffusers/blob/main/LICENSE). | |
| When reporting results, cite the RFD4-Proteína project, the exact model and code revisions, and any source datasets or | |
| evaluation tools used. | |