protein-design-skills

A collection of Claude Code skills for de novo protein design and biomolecular structure prediction. Drop this plugin into Claude Code and your agent gains expert-level operating knowledge of the major open-source tools in the field — installation, CLI flags, input schemas, output layouts, and how to chain them into full design campaigns.

Skills

Skill	What it does	Upstream
bindcraft	BindCraft — hallucination-based de novo binder design (AF2 backprop + ProteinMPNN + PyRosetta) with built-in 5-model AF2 reprediction, interface scoring, and a tunable filter pass. The hallucination camp's reference pipeline.	martinpacesa/BindCraft
bioemu	BioEmu — Microsoft Research's deep generative emulator of protein equilibrium ensembles: samples the Boltzmann distribution of backbone structures for a monomer at MD-replacing speed, with optional SMC / FKC steering, side-chain rebuild (HPacker), and OpenMM relax.	microsoft/bioemu
biotite	Biotite — fast, NumPy-backed toolkit to read / manipulate / write structures + sequences, compute RMSD / lDDT / TM-score / SASA / SSE, and fetch from RCSB / AlphaFold DB / UniProt / Entrez / PubChem. The "glue" library for preparing inputs and scoring outputs of the other tools.	biotite-dev/biotite
boltz	Boltz-1 / Boltz-2 — open foundation models for protein, complex, nucleic-acid, and protein-ligand structure + binding-affinity prediction.	jwohlwend/boltz
boltzgen	All-atom diffusion model for universal binder design (proteins, peptides, cyclic peptides, nanobodies, Fabs) against protein / small-molecule / nucleic-acid targets.	HannesStark/boltzgen
caver	CAVER 3.0/3.01 — Voronoi-based identification, geometric characterization, clustering, and visualization of transport tunnels in proteins (single PDB or MD ensembles). Packaged to run from an Apptainer/Singularity SIF.	caver.cz
chai-lab	Chai-1 — multi-modal foundation model for protein / ligand / nucleic-acid complex structure prediction, with restraints and MSAs.	chaidiscovery/chai-lab
disco	DISCO — diffusion model that co-designs protein sequence and 3D structure conditioned on small molecules, metals, DNA, or RNA.	DISCO-design/DISCO
esm-biohub	Biohub's esm repo — ESMC language model, ESMFold2 AF3-class structure prediction, ESM3 generative model, and ESMC SAEs for interpretability. Packaged to run from an Apptainer/Singularity SIF.	Biohub/esm
fair-esm	Meta FAIR fair-esm — ESM-2 / ESM-1b / ESM-1v / MSA Transformer embeddings, ESMFold structure prediction, ESM-IF1 inverse folding, zero-shot variant scoring, and the ESM Metagenomic Atlas.	facebookresearch/esm
foundry	Foundry toolkit — RFdiffusion3, RosettaFold3, and the ProteinMPNN / LigandMPNN / SolubleMPNN / EnhancedMPNN / ThermalMPNN family under a shared CLI. Includes extension paths (registering checkpoints, adding architectures).	RC Foundry
genie3	Genie 3 — fast all-atom SE(3)-equivariant diffusion model for unconditional generation, motif scaffolding, and hotspot-guided binder design.	aqlaboratory/genie3
invrotzyme	Build inverse-rotamer theozyme / active-site assemblies from Rosetta enzdes CST files (PyRosetta) — drop-in input for RFdiffusion All-Atom enzyme design.	ikalvet/invrotzyme
placer	PLACER — atom-level GNN for ligand pose refinement / docking-score, side-chain prediction, and conformational ensembles. Packaged to run from an Apptainer/Singularity SIF.	baker-laboratory/PLACER
protenix	Protenix — open-source AlphaFold 3 reproduction for protein / nucleic-acid / ligand / ion complex structure prediction, with MSAs, templates, covalent bonds, and pocket/contact constraints. Packaged to run from an Apptainer/Singularity SIF.	bytedance/Protenix
protflow	ProtFlow — compose multi-step design pipelines (RFdiffusion, MPNN family, ESMFold, AF2/AF3, Boltz, Rosetta, GROMACS, …) behind a Poses + Runner API with first-class SLURM array jobs.	mabr3112/ProtFlow

Each skill ships its SKILL.md plus curated references/ and examples/ so Claude has the source material on hand when it builds your pipeline.

Installation

As a Claude Code plugin (recommended)

/plugin marketplace add LiorZ/protein-design-skills
/plugin install protein-design-skills@protein-design-skills

Claude Code will auto-discover every skill under skills/ and trigger them by name or description match. Verify with /plugin list and /help.

Manual install (per-user)

Clone the repo and symlink the skills into your user skills directory:

git clone https://github.com/LiorZ/protein-design-skills.git
ln -s "$PWD/protein-design-skills/skills"/* ~/.claude/skills/

Manual install (per-project)

cd your-project
mkdir -p .claude/skills
ln -s /path/to/protein-design-skills/skills/* .claude/skills/

Using the skills

Skills are invoked automatically when your prompt matches their description. You can also invoke one explicitly:

/boltzgen design a 80-residue binder against PD-L1 (PDB 5O45 chain A)
/boltz  predict the complex from this YAML
/genie3 generate 100 backbones of length 120

A typical binder-design campaign chains several:

1. Generate backbones with genie3, boltzgen, bindcraft (hallucination), or foundry (RFdiffusion3). For enzymes, build a theozyme with invrotzyme first.
2. Assign sequences with foundry (ProteinMPNN / LigandMPNN / SolubleMPNN) — or let boltzgen / bindcraft / disco co-design (bindcraft runs ProteinMPNN internally). Score / rerank candidates with fair-esm.
3. Validate with boltz, chai-lab, protenix (open-source AlphaFold 3), or fair-esm's ESMFold — cross-checking two AF3-class co-folders is a strong signal. bindcraft already runs 5-model AF2 reprediction internally, but an independent AF3-class predictor is still worth running on the top picks. For enzyme / ligand pockets, refine and score the ligand pose and side chains with placer. For equilibrium / conformational validation of a designed monomer (does it actually stay folded? are there alternative basins?), sample the ensemble with bioemu.
4. Rank with ipSAE / pLDDT / iPTM filters described in each skill — and use biotite to parse the CIF/PDB outputs and compute the structural metrics (RMSD / lDDT / TM-score / SASA / clashes) you filter on, or to fetch and clean target structures feeding the design steps above.

For multi-step campaigns at cluster scale, drive the whole pipeline with protflow (SLURM array jobs, Poses DataFrame, motif tracking).

Repo layout

.
├── .claude-plugin/
│   ├── plugin.json          # plugin manifest
│   └── marketplace.json     # marketplace entry
├── skills/
│   ├── bindcraft/
│   ├── bioemu/
│   ├── biotite/
│   ├── boltz/
│   ├── boltzgen/
│   ├── caver/
│   ├── chai-lab/
│   ├── disco/
│   ├── esm-biohub/
│   ├── fair-esm/
│   ├── foundry/
│   ├── genie3/
│   ├── invrotzyme/
│   ├── placer/
│   ├── protenix/
│   └── protflow/
└── README.md

Contributing

To add a new skill: create skills/<name>/SKILL.md with YAML frontmatter (name, description) and any supporting references/ and examples/. Open a PR.

License

Each skill's content tracks the upstream tool's license (noted in its frontmatter). The skill packaging in this repo is MIT.