ProteinMPNN

ProteinMPNN is the default inverse-folding step in the binder pipeline: a message-passing network that sees backbone geometry only, so it is the right choice when the design surface is protein–protein and the wrong one as soon as a ligand, nucleic acid, or metal is part of the interface — ligandmpnn adds those atoms to the graph with a near-identical CLI, and solublempnn swaps in weights trained on soluble structures for an expression-biased prior. Code and weights are MIT (github.com/dauparas/ProteinMPNN). The model is small enough to run on CPU — for a handful of sequences on one backbone that is seconds and usually faster than dispatching a remote job; a GPU helps for batched campaigns (hundreds of backbones or large --numseqper_target). Either way the repo is cloned in-job — there is no PyPI dist and the checkpoints are bundled in the repo.

Running it

pip install torch numpy   # if not already present
git clone --depth 1 https://github.com/dauparas/ProteinMPNN.git proteinmpnn
cd proteinmpnn
python protein_mpnn_run.py \
  --pdb_path backbone.pdb --pdb_path_chains "A" \
  --out_folder out --num_seq_per_target 16 --sampling_temp "0.1"

Two flags trip almost everyone the first time. --samplingtemp is parsed as a space-separated string so one run can sweep several temperatures; a single value needs no quoting, but a multi-value sweep must be quoted ("0.1 0.2 0.3"), and commas never split — "0.1,0.2" fails the float cast. --pdbpath_chains is also space-separated inside one quoted argument ("A B"); a comma is kept as part of the chain ID.

Designs land in out/seqs/<pdbstem>.fa. The first record is the input sequence; each design header carries score= (mean negative log-likelihood — lower is more confident), globalscore=, and seqrecovery=. ProteinMPNN writes sequences only — it does not thread them back onto the backbone; if you need designed-sequence PDBs, the ligandmpnn runner writes them to backbones/ automatically and accepts --modeltype protein_mpnn for the same weights.

A flat chain map in --fixedpositionsjsonl silently redesigns every residue

--fixedpositionsjsonl expects one JSON object per line keyed by the PDB stem first, then chain, then a list of 1-indexed residue numbers: {"backbone": {"A": [10, 11, 12], "B": []}}. Passing the inner {"A": [...]} directly — the obvious guess — is silently treated as "no PDB matched," and every position is redesigned. The bundled helperscripts/makefixedpositionsdict.py writes the correct shape from a chain and range string and is worth the extra call; the same outer-stem rule applies to --chainidjsonl and --tiedpositionsjsonl.

Checkpoints — which one to pick

Errors worth recognizing

Next: fold the designs in complex with the target via boltz, chai1, or esmfold2 and filter on ipTM.