RFdiffusion / ProteinMPNN
Generative models for designing novel protein and peptide structures from scratch — complementing AlphaFold's prediction role with de novo design capability.
What it does
RFdiffusion and ProteinMPNN are generative design models that work together: where AlphaFold predicts the structure of an existing protein sequence, these models design new protein/peptide structures from scratch to meet a target function. RFdiffusion generates 3D backbone structures; ProteinMPNN then designs amino acid sequences that fold into those backbones. The resulting designed proteins can be validated with AlphaFold and eventually tested experimentally.
Both are open source and come from the Baker Lab at UW Medicine — one of the leading protein design groups globally.
Best for
Synthetic biology and peptide-drug researchers doing de novo design, not just structure lookup. If you need a protein that binds a specific target, inhibits a specific interaction, or performs a new catalytic function, these are the primary generative tools for doing that computationally. Pairs naturally with AlphaFold: design with RFdiffusion/ProteinMPNN, then validate predicted structure with AlphaFold.
Pricing
Open source / research access. Both tools are freely available via GitHub. Running them at scale requires GPU infrastructure (local or cloud).
Strengths
- State-of-the-art de novo protein design — experimental success rates for designed proteins have improved substantially using these tools
- Open source with active development and community
- Natural complement to AlphaFold: a complete pipeline from design to structure validation exists using open-source tools
- ProteinMPNN is particularly effective at generating multiple diverse sequence solutions for a given backbone
Limitations
- Design tools produce candidates, not guarantees — designed proteins must be synthesized and tested experimentally; failure rates in wet-lab testing are still significant even for well-designed candidates
- Requires computational infrastructure and ML expertise to run effectively — not a web-based point-and-click tool
- Results need AlphaFold or experimental validation before claims about structure can be made
- Rapidly evolving field — newer models (e.g., RFdiffusion2) may have superseded parts of this pipeline by the time you read this; check the Baker Lab GitHub for the latest releases
How it compares
| vs. | Key difference |
|---|---|
| AlphaFold | AlphaFold predicts structure of existing proteins; RFdiffusion/ProteinMPNN generate new sequences and structures |
| ESM (Meta AI) | ESM is a protein language model used for representation and some design tasks; RFdiffusion uses diffusion-based generative modeling — different architectures for overlapping problems |
| Rosetta | Rosetta is the long-standing protein design platform; RFdiffusion/ProteinMPNN have largely superseded it for de novo backbone design in recent benchmarks |
Related content
- Field Guide: AI Tools for Structural Biology
- Related tool: AlphaFold