ProteinMPNN
Upload a backbone — a structure with no sequence decided yet — and get back candidate sequences that should fold into it, each with a score and a native-recovery figure. About 30 s per run.
What it is for
You have a backbone — a 3D shape with no sequence decided yet — and need amino-acid sequences that will fold into it. Ranked candidates come back in about 30 seconds. To generate the backbone in the first place, run a binder design tool and feed its PDB in here.
You give it a backbone — a 3D protein shape with no sequence decided yet — and it proposes an amino acid for every position, chosen so the sequence should fold back into that exact shape. It reads only the backbone atoms, so side chains in your file are ignored. Each candidate comes back with a score and, on a natural backbone, the fraction of the real sequence it recovered. ProteinMPNN, Dauparas et al., Science 2022.
When it fits:
- You have a backbone and need sequences for it.
- You want to re-sequence a binder another tool here designed, before folding or ordering it.
- You want several alternative sequences threaded through a curated structure so you can choose between them.
Inputs
You will need:
- Backbone PDB or mmCIF (only Cα and backbone atoms are used).
- Chain ID(s) of the region(s) to redesign. Other chains stay fixed as context.
Each run uses a preset that sets the scale and scope:
- Standalone with your backbone
- Upload a backbone PDB, pick chain(s) to redesign, get up to 1000 candidate sequences. ~30 to 60 s on A10G-24GB.
Parameters you set on the form:
- Chains to design
- Which chains in the PDB MPNN should redesign (e.g.
A,A B,H L). Other chains are held fixed as context. - Number of sequences
- How many independent samples to draw (1 to 1000). Each sample is independent; rank by score and ProteinMPNN recovery rate.
- Sampling temperature
- Lower means more conservative (closer to argmax); higher means more diverse. Defaults to 0.1 per the upstream README.
- Fixed positions (optional)
- Positions to hold fixed inside a designed chain; everything else in that chain is redesigned. Written as
CHAIN:listgroups with single positions or ranges —A:1-44,46-66 B:5,7. Positions are 1-indexed within their chain, not author residue numbers, so the chain must be numbered from 1 with no gaps or insertion codes. Use this to redesign a liability patch while leaving a binding interface untouched: list the complement of the patch. Leave blank to redesign the whole chain.
Typical runtime:
- standalone
- ~1 min
How to read the results
Ranked candidate sequences with per-position score and overall ProteinMPNN recovery, downloadable as FASTA. Pair downstream with AlphaFold2 or ColabFold to confirm the predicted fold.
Where a tool reports them, the scores mean:
- ipTM
- Predicted confidence in the contact between two chains, on a 0 to 1 scale. Higher is better: > 0.75 strong; > 0.65 acceptable. Individual tools set their own pass bar a little either side of that — this guide's own results summary above states this tool's. On a multi-chain target the number may cover the target's own chain–chain interface as well as the binder's, so read the per-tool note on the results table before comparing designs on it.
- pLDDT
- Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.
- i_pAE and pAE
- Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.
References
Dauparas et al., Science 2022