Binder Design

Binder Design Tool Selection

Decision framework for choosing between BoltzGen, RFdiffusion, BindCraft, and other tools for protein binder design campaigns.

Decision Tree

What do you need?
│
├── All-atom output with side-chain awareness?
│   └── YES → BoltzGen (recommended default)
│
├── Backbone-only + sequence design separately?
│   └── YES → RFdiffusion → ProteinMPNN/LigandMPNN
│
├── End-to-end with built-in AF2 validation per design?
│   └── YES → BindCraft (slower but higher hit rate)
│
├── Ligand/small molecule in binding site?
│   └── YES → BoltzGen or RFdiffusion + LigandMPNN
│
├── Peptide binder (< 30 residues)?
│   └── YES → BoltzGen (peptide-anything protocol)
│
├── Symmetric oligomer scaffold?
│   └── YES → RFdiffusion (symmetry support)
│
└── Very large target (> 500 residues)?
    └── YES → RFdiffusion (better memory efficiency)

Tool Comparison

Campaign Scale Guide

Pipeline Templates

Standard Binder Campaign

# 1. Generate backbones
python3 RFdiffusion/scripts/run_inference.py \
    inference.output_prefix=designs/binder \
    inference.input_pdb=target.pdb \
    'ppi.hotspot_res=[A45,A67,A102]' \
    'contigmap.contigs=[A1-200/0 60-100]' \
    inference.num_designs=200

# 2. Design sequences
python3 LigandMPNN/run.py \
    --model_type proteinmpnn \
    --pdb_path designs/ \
    --out_folder seqs/ \
    --number_of_batches 8

# 3. Predict structures (ColabFold batch)
colabfold_batch seqs/all_sequences.fasta predictions/ \
    --model-type alphafold2_multimer_v3 \
    --num-recycles 3 \
    --num-models 1

# 4. Rank by ipSAE
ipsae rank predictions/ --chain-lengths 200 80 --output rankings.csv

BoltzGen All-in-One

# binder.yaml
version: 1
sequences:
  - protein:
      id: A
      sequence: MTEYKLVVVGAGGVGKS...  # target
  - protein:
      id: B
      length: 80
      design: true
constraints:
  hotspots:
    - chain: A
      residue: [45, 67, 102]

boltzgen design binder.yaml --out_dir designs/ --num_designs 100

BindCraft High-Quality

{
  "target_pdb": "target.pdb",
  "hotspot_residues": "A45,A67,A102",
  "binder_length": [60, 100],
  "num_designs": 100,
  "protocol": "default",
  "af2_validation": true,
  "output_dir": "bindcraft_out/"
}

python3 bindcraft.py --config binder_config.json

Hotspot Identification

from Bio.PDB import PDBParser
import numpy as np

def identify_hotspots(pdb_path: str, chain: str = "A",
                      binding_site_center: tuple = None,
                      radius: float = 10.0) -> list:
    """Find surface residues near a binding site for hotspot specification."""
    parser = PDBParser(QUIET=True)
    structure = parser.get_structure("target", pdb_path)
    model = structure[0]

    hotspots = []
    for residue in model[chain]:
        if residue.id[0] != " ":  # Skip HET
            continue
        # Use CA atom as representative
        if "CA" not in residue:
            continue
        ca = residue["CA"].get_vector()

        if binding_site_center:
            center = np.array(binding_site_center)
            dist = np.linalg.norm(ca.get_array() - center)
            if dist < radius:
                hotspots.append(f"{chain}{residue.id[1]}")

    return hotspots

Avoiding Common Failures