RFdiffusion
An SE(3) diffusion network generates 3D protein backbones, ProteinMPNN designs matching amino-acid sequences, and AlphaFold2 self-consistency filters select candidates for wet-lab validation.
The loop
- Proposes
3D protein backbone coordinates (residue translations and orientations) generated by an SE(3) diffusion model fine-tuned from RoseTTAFold, paired with ProteinMPNN sequences
- Judged by
AlphaFold2 / RoseTTAFold structure-prediction self-consistency (backbone RMSD and interface confidence pAE) followed by wet-lab expression, biolayer interferometry and cryo-EM
- Steers
Computational self-consistency metrics prune thousands of sampled trajectories down to top designs for physical synthesis and binding assays
Headline result
In Nature (2023), RFdiffusion designed de novo protein binders across five therapeutic and viral targets with an experimental hit rate of roughly 19% when testing 95 designs per target (two to three orders of magnitude higher than classical Rosetta binder design), and a cryo-EM structure of a designed Influenza HA binder (PDB: 8SK7) matched the design model at atomic accuracy (1.45 A backbone RMSD).
What this does not prove
Passing the AlphaFold2 computational filter does not ensure wet-lab binding (around four in five filtered binder designs still fail in vitro), and scaffolding full multi-step catalytic enzyme active sites in the lab requires additional physics and screening beyond backbone diffusion.
How it generalizes the harness
- A domain authority judges
- Search over a population
- The generator is not an LLM
- The output outlives the model
- Physical and lab grounding
The generator is an SE(3) geometric diffusion model over 3D protein coordinates rather than a language model, and structural fold prediction plus biophysical assays filter the population to produce physical protein sequences.
Classification
| What it proposes | Protein structure |
|---|---|
| Generative model | SE(3)-equivariant diffusion model fine-tuned from RoseTTAFold structure-prediction weights, coupled with ProteinMPNN graph neural network sequence design (not an LLM) |
| Algorithmic authority | AlphaFold2 structural self-consistency and interface pAE filtering, plus wet-lab biolayer interferometry, circular dichroism and cryo-EM |
| Feedback | Checks at several fidelities |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Original system |