RosettaCommons/RFD4-Proteina

What it solves

Research teams need to design new proteins and biomolecular complexes (like protein binders, antibodies, enzymes, or small-molecule pockets) with specific functional roles. Traditional methods are slow and require many rounds of trial and error. This project provides a single foundation model that can generate novel biomolecules from high-level design conditions, cutting down the guesswork and enabling tasks like motif scaffolding, partial refinement, and reward-guided optimization.

How it works

It is a flow-matching foundation model that performs sequence–structure co-design: it generates both the amino-acid sequence and the 3D structure together. You describe your design intent either through an annotated CIF file or a text-based YAML specification, where you can set conditions such as fixed motifs, hotspots, burial requirements, hydrogen bonds, or entire chain constraints. The model also supports partial diffusion—re-noising and refining an existing structure—and reward-guided search, where candidate designs are branched and pruned based on scores from structure-prediction oracles or custom metrics. Inference is fast by default (using torch.compile) and can optionally run on a BioIR backend. You interact with the model via command-line tools (rfd4-generate, rfd4-evaluate), a Python API, or a Hugging Face Diffusers pipeline.

Who it’s for

This is aimed at protein designers, structural biologists, and computational researchers who want to generate novel biomolecules (monomers, binders, antibodies, nanobodies, enzymes, ligand pockets) with controlled properties. It suits teams comfortable with Linux, Python 3.12, and NVIDIA GPUs, and who need a flexible, scriptable design workflow. The model is also packaged for easy use on Hugging Face, including browser-based studios for no-install experimentation.

Highlights

• One model handles many design tasks: monomers, protein and DNA binders, antibodies and nanobodies, enzymes, motif scaffolding, and small-molecule pockets.
• Rich conditioning via annotated CIF inputs or simple text-based YAML specs (fixed motifs, hotspots, burial, hydrogen bonds, and more).
• Reward-guided search over generation paths, scored by structure-prediction oracles and custom metrics.
• Partial diffusion lets you refine or diversify existing structures.
• Fast inference with torch.compile by default, and an optional BioIR backend.
• Available as a Hugging Face model, a Python API, command-line tools, and a Diffusers pipeline.