ProteinMPNN
Message-passing graph neural network for inverse protein folding whose designed sequences are ranked by the model log probability, with AlphaFold2 structure prediction and wet-lab crystallization used to measure whether they fold.
The loop
- Proposes
An SE(3)-invariant message-passing neural network decodes amino-acid sequences conditioned on a target 3D protein backbone.
- Judged by
The model log probability ranks designed sequences; AlphaFold2 / ESMFold single-sequence structure prediction (pLDDT and backbone RMSD) plus wet-lab SEC, cryo-EM, and X-ray crystallography measure whether they refold.
- Steers
Tied-position symmetry constraints shape decoding, and designed sequences are ranked by the model log probability.
Headline result
Achieved 52.4% native sequence recovery (vs. 32.9% for Rosetta) and rescued previously failed de novo protein cyclic oligomers, nanoparticles, and binders in wet-lab crystallography.
What this does not prove
Assumes a designable target backbone; cannot fix a backbone with severe steric clashes on its own.
How it generalizes the harness
- A domain authority judges
- The output outlives the model
- Search over a population
- The generator is not an LLM
A message-passing 3D graph network proposes sequences that are ranked by the model log probability; AlphaFold refolding RMSD and X-ray crystallography measure whether they fold.
Classification
| What it proposes | Protein structure |
|---|---|
| Generative model | ProteinMPNN 3D message-passing neural network |
| Algorithmic authority | The model log probability for ranking; AlphaFold2 refolding and X-ray crystallography / cryo-EM as measures of whether designs fold |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |