bindcraft

name	bindcraft
description	End-to-end binder design using BindCraft hallucination. Use this skill when: (1) Designing protein binders with built-in AF2 validation, (2) Running production-quality binder campaigns, (3) Using different design protocols (fast, default, slow), (4) Need joint backbone and sequence optimization, (5) Want high experimental success rate. For backbone-only generation, use rfdiffusion. For QC thresholds, use protein-qc. For tool selection guidance, use binder-design.
license	MIT
category	design-tools
tags	["structure-design","sequence-design","binder","pipeline"]
proteinbase_slug	bindcraft
proteinbase_url	https://proteinbase.com/design-methods/bindcraft
biomodals_script	modal_bindcraft.py

BindCraft Binder Design

Prerequisites

Requirement	Minimum	Recommended
Python	3.9+	3.10
CUDA	11.7+	12.0+
GPU VRAM	32GB	48GB (L40S)
RAM	32GB	64GB

How to run

First time? See Getting started to set up Modal and biomodals.

Option 1: Modal (recommended)

cd biomodals
modal run modal_bindcraft.py \
  --input-pdb target.pdb \
  --target-chains A \
  --target-hotspot-residues "45,67,89" \
  --lengths "70,100" \
  --number-of-final-designs 50

GPU: L40S (48GB) | Timeout: 300 min default

Option 2: Local installation

git clone https://github.com/martinpacesa/BindCraft.git
cd BindCraft

# BindCraft is configured with JSON files, not flags
python -u ./bindcraft.py \
  --settings ./settings_target/mytarget.json \
  --filters ./settings_filters/default_filters.json \
  --advanced ./settings_advanced/default_4stage_multimer.json

The target PDB, chains, hotspots, and binder length range are set inside the --settings JSON. See the BindCraft repo for the settings schema.

Key parameters (Modal wrapper)

Parameter	Default	Description
--input-pdb	required	Target structure
--target-chains	A	Target chain(s)
--target-hotspot-residues	""	Target hotspots (e.g. "45,67,89")
--lengths	50,130	Binder length range
--number-of-final-designs	1	Passing designs to return
--max-trajectories	none	Cap on trajectories

Output format

output/
├── design_0/
│   ├── binder.pdb         # Final design
│   ├── complex.pdb        # Binder + target
│   ├── metrics.json       # QC scores
│   └── trajectory/        # Optimization trajectory
├── design_1/
│   └── ...
└── summary.csv            # All metrics

Metrics Output

{
  "plddt": 0.89,
  "ptm": 0.78,
  "iptm": 0.62,
  "pae": 8.5,
  "rmsd": 1.2,
  "sequence": "MKTAYIAK..."
}

Sample output

Successful run

$ modal run modal_bindcraft.py --input-pdb target.pdb --target-chains A --target-hotspot-residues "45,67,89" --number-of-final-designs 50
[INFO] Loading BindCraft model...
[INFO] Target: target.pdb (chain A)
[INFO] Hotspots: 45, 67, 89
[INFO] Generating designs...

Design 1/50:
  Length: 78 AA
  pLDDT: 0.89, ipTM: 0.62
  Saved: output/design_0/

Design 50/50:
  Length: 85 AA
  pLDDT: 0.86, ipTM: 0.58
  Saved: output/design_49/

[INFO] Campaign complete. Summary: output/summary.csv
Pass rate: 32/50 (64%) with ipTM > 0.5

What good output looks like:

• pLDDT: > 0.85 for most designs
• ipTM: > 0.5 for passing designs
• Pass rate: 30-70% depending on target
• Diverse sequences across designs

Decision tree

Should I use BindCraft?
│
├─ What type of design?
│  ├─ Production-quality binders → BindCraft ✓
│  ├─ High diversity exploration → RFdiffusion
│  └─ All-atom precision → BoltzGen
│
├─ What matters most?
│  ├─ Experimental success rate → BindCraft ✓
│  ├─ Speed / diversity → RFdiffusion + ProteinMPNN
│  ├─ AF2 gradient optimization → ColabDesign
│  └─ All-atom control → BoltzGen
│
└─ Compute resources?
   ├─ Have L40S/A100 → BindCraft ✓
   └─ Only A10G → RFdiffusion + ProteinMPNN

Typical performance

Campaign Size	Time (L40S)	Cost (Modal)	Notes
50 designs	2-4h	~$15	Quick campaign
100 designs	4-8h	~$30	Standard
200 designs	8-16h	~$60	Large campaign

Adaptyv's own tests of these models showed BindCraft costing about $2.90 per accepted design, averaged across 7 targets.

Experimental success rate (BindCraft paper): 10 to 100%, averaging 46.3% across 12 targets; strongly target-dependent.

---

Verify

find output -name "binder.pdb" | wc -l  # Should match num_designs

---

Troubleshooting

Low ipTM scores: Check hotspot selection, increase designs Slow convergence: Use fast protocol for screening OOM errors: Reduce num_models, use L40S GPU Poor diversity: Lower sampling_temp, run multiple seeds

Error interpretation

Error	Cause	Fix
RuntimeError: CUDA out of memory	Large target or long binder	Use L40S/A100, reduce binder length
ValueError: no hotspots	Hotspots not found	Check residue numbering
TimeoutError	Design taking too long	Use fast protocol

---

Next: Rank by ipsae → experimental validation.