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 --num_seq_per_target). Either way
the repo is cloned in-job — there is no PyPI dist and the checkpoints are
bundled in the repo.
Running it
bashpip 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. --sampling_temp 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. --pdb_path_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/<pdb_stem>.fa. The first record is the input
sequence; each design header carries score= (mean negative log-likelihood —
lower is more confident), global_score=, and seq_recovery=. 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 --model_type protein_mpnn for the
same weights.
A flat chain map in --fixed_positions_jsonl silently redesigns every residue
--fixed_positions_jsonl 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
helper_scripts/make_fixed_positions_dict.py writes the correct shape from a
chain and range string and is worth the extra call; the same outer-stem rule
applies to --chain_id_jsonl and --tied_positions_jsonl.
Checkpoints — which one to pick
--model_name | training noise | use |
|---|---|---|
v_48_002 | 0.02 Å | highest recovery; close-to-native redesigns |
v_48_020 (default) | 0.20 Å | de novo backbones — tolerates RFdiffusion imperfection |
v_48_030 | 0.30 Å | very rough backbones; lowest recovery |
--use_soluble_model | — | swaps to the soluble-trained set; see solublempnn |
Errors worth recognizing
| You see | It means / do this |
|---|---|
KeyError: 'A' | Chain letter not in the PDB — grep '^ATOM' file.pdb | cut -c22 | sort -u to see what is. |
JSONDecodeError on a *_jsonl flag | The flag wants a file path, not inline JSON; write the file first. |
All positions redesigned despite --fixed_positions_jsonl | Outer PDB-stem key missing — see the gotcha above. |
ModuleNotFoundError for relative imports | Script run from the wrong cwd — cd into the cloned repo first; the imports are repo-relative. |
Next: fold the designs in complex with the target via boltz, chai1, or
esmfold2 and filter on ipTM.
