ligandmpnn

LigandMPNN

LigandMPNN extends the ProteinMPNN graph with non-protein atoms — small molecules, nucleic acids, and metals are visible to the network — so it is the right inverse-folding tool whenever the design surface includes a bound ligand or cofactor that vanilla proteinmpnn would ignore. The same run.py is also the most convenient runner for the other MPNN families because, unlike the original ProteinMPNN script, it threads designs back onto the input structure and writes PDBs alongside the FASTA. Code and weights are MIT (github.com/dauparas/LigandMPNN). The model is small enough to run on CPU — for a handful of designs on one structure that is seconds and usually faster than dispatching, so the normal path is local with pip install torch numpy biopython ProDy ml_collections dm-tree; a GPU helps for batched campaigns.

Running it

pip install torch numpy biopython ProDy ml_collections dm-tree
git clone --depth 1 https://github.com/dauparas/LigandMPNN.git ligandmpnn
cd ligandmpnn
sed -i 's/np\.int\b/np.int64/g' openfold/np/residue_constants.py   # repo pins numpy 1.23; alias removed in >=1.24
bash get_model_params.sh ./model_params
python run.py \
  --model_type ligand_mpnn \
  --checkpoint_ligand_mpnn ./model_params/ligandmpnn_v_32_010_25.pt \
  --pdb_path complex.pdb \
  --out_folder out \
  --batch_size 8 --number_of_batches 4 \
  --temperature 0.1 \
  --fixed_residues "A45 A46 A47 A48"

Residue selections are space-separated {chain}{resnum} tokens inside one quoted string ("A45 A46 B10"; insertion codes append directly, "B82A"). That is the format for --fixed_residues and --redesigned_residues; --bias_AA_per_residue and --omit_AA_per_residue instead take a path to a JSON file whose keys use the same {chain}{resnum} form, and --chains_to_design is comma-separated ("A,B"). If you want to redesign only the pocket, naming the pocket residues in --redesigned_residues is usually shorter than fixing everything else.

Under --out_folder you get seqs/<stem>.fa (headers carry overall_confidence and ligand_confidence), backbones/<stem>_{1..N}.pdb with the designed sequence threaded onto the input coordinates, and — with --pack_side_chains 1 — full-atom packed models in packed/. The threaded PDBs are the reason to prefer this runner even for protein-only jobs.

Model types — which one to pick

Each model type has its own --checkpoint_<type> flag; the wrong pairing is caught at load time, but the default checkpoint path is relative to the repo, so run from inside the clone or pass the absolute path.

ProDy compiles from source on py3.11 — pip install fails without a C compiler

run.py imports ProDy unconditionally for ligand atom parsing. On py3.11 the prebuilt wheel is missing on PyPI, so pip install ProDy compiles from source and needs a working C/C++ compiler. On an unprivileged SSH context, prefer an existing compiler module or conda-provided toolchain and export CC=gcc CXX=g++; do not assume system package installation is allowed. On most CPU-local Python distributions the sdist builds in ~10 s if no wheel matches.

Turning ligand context off changes the answer, not the model

--ligand_mpnn_use_atom_context 0 keeps the ligand-aware weights but masks the ligand atoms at inference. That is useful for an ablation — the difference between context-on and context-off tells you how much the ligand is shaping the design — but it is not equivalent to running protein_mpnn, which uses a different checkpoint trained without those features. For a fair protein-only baseline, switch --model_type.

Stripped HETATM or a chain filter silently drops the ligand — the design comes back pocket-blind

LigandMPNN does not warn when no ligand atoms are found; it just runs as if --model_type protein_mpnn had been picked. The two common ways this happens are an input PDB whose HETATM records were stripped by an upstream clean-up step, and --parse_these_chains_only naming the protein chains but not the ligand's. If ligand_confidence in the FASTA header is missing or zero across every design, the model never saw the ligand — fix the input, do not trust the sequences.

Wisp execution

Use python only for bounded interactive checks. For a long or GPU-backed workload, require a selected and probed ssh:<alias> context and load remote-compute-ssh. Put the documented invocation in a self-contained project script, activate the remote environment explicitly, stage only small files with input_paths, and make the command write to a known absolute remote result path. Submit it with run_in_context and register that exact ssh:// path in output_specs. Call monitor_run once when waiting is needed, get_run once for a snapshot, or cancel_run to stop. Do not send a scheduler submission through the SSH-direct runner.

Errors worth recognizing

Next: fold the designs in complex with the ligand via boltz or chai1 (both accept SMILES/CCD) and filter on ipTM and ligand placement.