Model Book/RFdiffusion

RFdiffusion

A-

UW Baker Lab

Protein DesignDe Novo GenerationBinder Designopen
Updated 1 month ago
NextIn take

RFdiffusion changed what is possible in protein design — designing binders that actually work in the lab at double-digit success rates. Pair it with ProteinMPNN for the full design pipeline.

Specifications

ArchitectureDenoising diffusion on RoseTTAFold frames for protein backbone generation
Parameters~300M
Training DataPDB structures — trained to reverse the noising process on known backbones
LicenseBSD license. Fully open weights and code.
Hardware1x A100 40GB, CPU-feasible for small designs
Inference CostSelf-hosted — ~$0.01/design on cloud GPU
API AvailableNo
Weights AvailableYes

Benchmark Performance

BenchmarkScore
ProteinGymN/A (generative)

When to Use This

  • +De novo protein binder design
  • +Scaffold generation for therapeutic protein engineering

When NOT to Use This

  • Structure prediction of natural proteins — use Boltz or AF3
  • Sequence design — pair with ProteinMPNN downstream

Production Readiness

pilot

Known Users

  • Baker Lab and collaborators
  • Multiple biotech companies including Generate Biomedicines

Grade Rationale

A-

The most impactful protein design tool released in the last five years. Experimentally validated de novo binders with success rates that were unthinkable before diffusion models. A- because generative design is inherently harder to benchmark than prediction.

Sources

Update History

2026-02-20Initial entry