bindcraft-reference

BindCraft Reference Skill

Full documentation: docs/plans/2026-01-15-bindcraft-comprehensive-skill.md

Quick Reference

Pipeline Overview

Target PDB → AF2 Hallucination → MPNN Sequences → AF2 Validation → PyRosetta Relax → Filtering

Key Difference from RFdiffusion

• BindCraft: AF2 backpropagation (sequence+structure simultaneously)
• RFdiffusion: Diffusion model (backbone first, then MPNN)

Design Algorithms

Target Settings (JSON)

{
    "design_path": "/path/to/output/",
    "binder_name": "MyBinder",
    "starting_pdb": "/path/to/target.pdb",
    "chains": "A",
    "target_hotspot_residues": "56,58,62",
    "lengths": [65, 150],
    "number_of_final_designs": 100
}

Key Loss Weights

Key Filters

PyRosetta Interface Metrics

• dG: Binding free energy
• dSASA: Buried surface area
• ShapeComplementarity: Geometric fit
• PackStat: Packing quality
• InterfaceHbonds: H-bond count
• InterfaceUnsatHbonds: Buried polar without H-bonds

Presets

Concepts for Banta Lab

1. Multi-stage optimization: Soft → temp → hard → greedy
2. Combined loss functions: Multiple objectives
3. Integrated AF2 validation: After MPNN design
4. PyRosetta interface scoring: Add to validation pipeline
5. Filter-based selection: Multi-metric thresholds
6. Acceptance rate monitoring: Auto-terminate poor campaigns

When to Use BindCraft vs RFdiffusion