ProteinMPNN designs sequence from backbone geometry and nothing else. Every ligand, nucleic acid and metal in the input PDB is invisible to it, which is why ligandmpnn exists; solublempnn keeps the same architecture but swaps in weights trained on soluble structures. So this is the right inverse-folding step for a protein-protein design surface, and the wrong one the moment a cofactor joins the interface. What is here is a runbook for the external repository: how to clone and drive it, which flags actually matter, and how to read the FASTA that comes back. The repo is cloned in-job — there is no PyPI dist, and the checkpoints ride inside the clone.
What comes out is a sequence and a likelihood. The score= in each FASTA header is a mean negative log-likelihood — how ProteinMPNN-like the sequence looks given that backbone — and seq_recovery= is agreement with the sequence the input already had. Neither says the design folds, expresses, or binds; that needs a fold model first and a bench after. Small runs are fine on CPU; real runtime, and whether a GPU buys you anything, depend on the size of the campaign and the environment it runs in.
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 proteinmpnn --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 proteinmpnn --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/proteinmpnn
python3 -m zipfile -c proteinmpnn.zip proteinmpnn
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/proteinmpnn/ out (unzip main.zip 'OpenAI4S-main/skills/proteinmpnn/*'
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 |
The runbook, built around the two things that trip people first. --sampling_temp and --pdb_path_chains are space-separated strings inside one quoted argument, and a comma does not split them. And a --fixed_positions_jsonl file missing its outer PDB-stem key is read as "no PDB matched": every position gets redesigned, with no warning — the bundled helper script writes the right shape. Around those: the noise-level checkpoints (v_48_002 for close-to-native redesign through v_48_030 for rough backbones), when CPU is enough, and why the output is FASTA only — ProteinMPNN never threads the designed sequence back onto the backbone, which is one reason to run even protein-only jobs through the ligandmpnn runner. |