Inverse folding when the design surface carries a ligand, a nucleic acid or a metal. LigandMPNN puts those atoms into the graph, so the network sees what vanilla ProteinMPNN ignores. Its run.py also threads the designed sequence back onto the input coordinates and writes the PDBs next to the FASTA, which is why it is the more convenient runner even for protein-only jobs — it accepts the ProteinMPNN and soluble weights too. The recipe points at the external LigandMPNN repository and its checkpoints; no executable and no weights are bundled here.
LigandMPNN designs around the ligand coordinates it was handed, and it treats that pose as fact. Hand it a docked or modelled placement that is off and you get a pocket shaped to a ligand position that may not be the real one — and the run looks exactly as healthy as a correct one. The ligand_confidence in each header is the model scoring the sequence it just wrote, not a measurement of binding. Chain, fixed-residue, context-atom and model-type choices all have to be checked against the actual input structure, and the sequences and threaded structures that come back are design candidates until the bench says otherwise.
A Skill is a directory of files, so installing one is copying that directory
somewhere an agent looks. With Node 18+ and git on PATH (npm resolves a
github: spec through it), and nothing cloned:
npx github:PKU-YuanGroup/OpenAI4S install ligandmpnn --target claude
--target claude writes to ~/.claude/skills, claude-project to
./.claude/skills, openai4s to <data_dir>/user-skills, and --dir <path>
to anywhere you name. The resolved absolute path is printed before anything is
written there, and --dry-run stops at that plan. A reinstall refuses to
overwrite a copy you have edited or one it did not install, and uninstall
removes only the files it wrote. For a recipe included in the npm release,
npx @pku-yuangroup/openai4s-skills install ligandmpnn --target claude installs
the published copy. The npm catalog can differ from this repository.
Without Node, take the directory itself — and turn it into a .zip if an upload
field wants one. The tarball is the whole repository (over 100 MB), and the pipe
is a POSIX shell recipe (macOS, Linux, WSL) that extracts only this directory:
curl -L https://codeload.github.com/PKU-YuanGroup/OpenAI4S/tar.gz/refs/heads/main \
| tar -xz --strip-components=2 OpenAI4S-main/skills/ligandmpnn
python3 -m zipfile -c ligandmpnn.zip ligandmpnn
The click-through form of that same download is the
repository zip,
which is also the route to take on Windows: extract it and copy
skills/ligandmpnn/ out (unzip main.zip 'OpenAI4S-main/skills/ligandmpnn/*'
unpacks only that directory). If you already run OpenAI4S there is nothing to
install — the wheel ships every bundled Skill, and a bundled Skill takes
precedence over a same-named copy in <data_dir>/user-skills. Targets,
provenance, and what the installer refuses to do:
tools/skills-installer/.
| File | Responsibility |
|---|---|
SKILL.md |
Covers repository setup, the model types and their checkpoints, how the PDB and its ligand context are parsed, which chains and residues get designed or held fixed, sampling, the batch outputs, and the threaded PDBs. It also collects the ligand-aware traps, including the one where a stripped HETATM record leaves the model pocket-blind without complaining, and the validation that comes after. |