🧬 ProteinClaw

Conversational protein design — from target to validated binder candidates, all through natural language.

ProteinClaw integrates a full computational protein design pipeline into the OpenClaw agent framework. Researchers can design protein binders, run structure predictions, score candidates, and assess stability — all by chatting with an AI assistant.

Built for the OpenClaw agent platform with skills inspired by BioClaw.

---

✨ What You Can Do

• Structure Retrieval — Fetch high-quality structures from PDB/UniProt with automatic quality filtering
• Binder Generation — Design peptide binders using BoltzGen diffusion model (all-atom, side-chain aware)
• Complex Scoring — Fast interface scoring with Boltz-2: interface pLDDT, buried surface area, ipTM, PAE
• Monomer Stability — Assess folding quality with Chai-1 and Boltz-2 refolds (pLDDT, disorder analysis)
• BindCraft Integration — Local GPU-accelerated binder optimization with PyRosetta scoring
• Full Campaign Management — Multi-stage design campaigns with automatic candidate tracking

---

🗂️ Repository Structure

ProteinClaw/
├── README.md
├── skills/                          # OpenClaw skill definitions
│   ├── SKILL.md                     # Hub router (entry point)
│   ├── binder-design/               # End-to-end binder design workflow
│   ├── boltzgen/                    # BoltzGen diffusion-based design
│   ├── boltz/                       # Boltz-2 structure prediction & scoring
│   ├── chai/                        # Chai-1 structure prediction
│   ├── bindcraft/                   # BindCraft local binder optimization
│   ├── proteinmpnn/                 # ProteinMPNN sequence design
│   ├── ligandmpnn/                  # LigandMPNN for small molecule binders
│   ├── solublempnn/                 # SolubleMPNN for solubility-optimized design
│   ├── esm/                         # ESM embeddings & language model tools
│   ├── foldseek/                    # Structure similarity search
│   ├── iggm/                        # IgGM antibody design
│   ├── ipsae/                       # ipSAE interface scoring
│   ├── mber/                        # MBER binding energy refinement
│   ├── pdb/                         # PDB structure queries
│   ├── uniprot/                     # UniProt sequence queries
│   ├── protenix/                    # Protenix structure prediction
│   ├── protein-design-workflow/     # Multi-stage campaign orchestration
│   ├── protein-qc/                  # Structure quality control
│   ├── campaign-manager/            # Design campaign tracking
│   ├── cell-free-expression/        # Cell-free expression assessment
│   ├── binding-characterization/    # Binding affinity characterization
│   └── setup/                       # Environment setup & Modal deployment
│
└── scripts/                         # Production pipeline scripts
    ├── rbx1_boltzgen_batch1.py      # Stage 2: BoltzGen batch generation
    ├── stage3_boltz_scoring.py      # Stage 3: Interface scoring (Boltz-2)
    ├── stage3_final.py              # Stage 3: Final filtered output
    ├── stage4_free.py               # Stage 4: Free monomer stability (Boltz-2)
    ├── stage4_chai_monomer.py       # Stage 4: Chai-1 monomer stability
    ├── stage4_plan_b.py             # Stage 4: Plan B batch Chai-1 runs
    ├── modal_chai1.py               # Modal cloud runner for Chai-1
    └── modal_pdb2png.py             # PDB structure visualization

---

🚀 Design Pipeline

ProteinClaw implements a 5-stage protein binder design pipeline:

Stage 1: Structure Acquisition
    └── Fetch target from PDB/AlphaFold DB, quality filter (resolution, coverage)

Stage 2: Binder Generation  
    └── BoltzGen diffusion → 500+ binder candidates (configurable length range)

Stage 3: Fast Scoring & Filtering
    └── Boltz-2 scoring → interface pLDDT ≥ 75, buried surface area ≥ 600 Ų
    └── Output: Top 250 candidates

Stage 4: Monomer Stability
    └── Chai-1 or Boltz-2 refold → pLDDT ≥ 75, disorder < 10%
    └── Output: Top 10–20 validated candidates

Stage 5: Experimental Prioritization
    └── MD simulations, BindCraft optimization, ranking for synthesis

Example: RBX1 Peptide Binder Design

Stage	Tool	Input	Output	Cost
1	PDB query	UniProt P62877	3DPL chain B (2.60 Å)	free
2	BoltzGen	RBX1 structure	500 binder candidates	~$2.00
3	Boltz-2 scoring	500 candidates	Top 250 scored	~$1.50
4	Chai-1	Top 8 candidates	5 validated binders	~$0.56

Top results:

• rbx1_binder_144: pLDDT=95.9, ipTM=0.743, 24 H-bonds, 102aa
• rbx1_binder_197: pLDDT=92.5, ipTM=0.720, 25 H-bonds, 106aa
• rbx1_binder_323: pLDDT=91.1, ipTM=0.719, 23 H-bonds, 135aa

---

⚙️ Setup

Prerequisites

• OpenClaw agent framework
• Modal account (for cloud GPU compute)
• Python 3.10+

Install Skills

# Clone into your OpenClaw skills directory
git clone https://github.com/junior1p/ProteinClaw.git ~/.openclaw/skills/protein-design

# Or copy individual skills
cp -r ProteinClaw/skills/* ~/.openclaw/skills/

Configure Modal

# Install Modal client
pip install modal

# Authenticate
modal setup

# Deploy endpoints (see skills/setup/SKILL.md for details)

---

🧠 Skills System

Each subdirectory in skills/ contains a SKILL.md that teaches the OpenClaw agent how to use that tool. The top-level skills/SKILL.md acts as a router, automatically selecting the right tool based on your natural language request.

Example queries:

• "Find me the best RBX1 structure for binder design" → pdb skill
• "Design 500 peptide binders for this target" → boltzgen skill
• "Score the top candidates for interface quality" → boltz skill
• "Check monomer stability of my top 8 designs" → chai skill

---

📦 Modal Endpoints

Endpoint	Model	GPU	Use Case
boltzgen-protein-design	BoltzGen	A100	Binder generation
boltz-structure-prediction	Boltz-2	A100	Structure prediction & scoring
chai1-structure-prediction	Chai-1	A100	Complex & monomer prediction
pdb2png-viewer	PyMOL	CPU	Structure visualization

---

📄 License

MIT License — see LICENSE for details.

---

🙏 Acknowledgments

• BoltzGen — all-atom protein design
• Chai-1 — molecular structure prediction
• BioClaw — inspiration for the conversational biology agent pattern
• OpenClaw — agent framework