Boltz

Boltz Structure Prediction

Prerequisites

| Requirement | Minimum | Recommended |

|-------------|---------|-------------|

| Python | 3.10+ | 3.11 |

| CUDA | 12.0+ | 12.1+ |

| GPU VRAM | 24GB | 48GB (L40S) |

| RAM | 32GB | 64GB |

How to run

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

Option 1: Modal

cd biomodals
modal run modal_boltz.py \
  --input-faa complex.fasta \
  --out-dir predictions/

GPU: L40S (48GB) | Timeout: 1800s default

Option 2: Local installation

pip install boltz

boltz predict \
  --fasta complex.fasta \
  --output predictions/

Key parameters

| Parameter | Default | Range | Description |

|-----------|---------|-------|-------------|

| --recycling_steps | 3 | 1-10 | Recycling iterations |

| --sampling_steps | 200 | 50-500 | Diffusion steps |

| --use_msa_server | true | bool | Use MSA server |

FASTA Format

>protein_A
MKTAYIAKQRQISFVK...
>protein_B
MVLSPADKTNVKAAWG...

Output format

predictions/
├── model_0.cif       # Best model (CIF format)
├── confidence.json   # pLDDT, pTM, ipTM
└── pae.npy          # PAE matrix

Note: Boltz outputs CIF format. Convert to PDB if needed:

from Bio.PDB import MMCIFParser, PDBIO
parser = MMCIFParser()
structure = parser.get_structure("model", "model_0.cif")
io = PDBIO()
io.set_structure(structure)
io.save("model_0.pdb")

Comparison

| Feature | Boltz-1 | Boltz-2 | AF2-Multimer |

|---------|---------|---------|--------------|

| MSA-free mode | Yes | Yes | No |

| Diffusion | Yes | Yes | No |

| Speed | Fast | Faster | Slower |

| Open source | Yes | Yes | Yes |

Sample output

Successful run

$ boltz predict --fasta complex.fasta --output predictions/
[INFO] Loading Boltz-1 weights...
[INFO] Predicting structure...
[INFO] Saved model to predictions/model_0.cif

predictions/confidence.json:
{
  "ptm": 0.78,
  "iptm": 0.65,
  "plddt": 0.81
}

What good output looks like:

• pTM: > 0.7 (confident global structure)
• ipTM: > 0.5 (confident interface)
• pLDDT: > 0.7 (confident per-residue)
• CIF file: ~100-500 KB for typical complex

Decision tree

Should I use Boltz?
│
├─ What are you predicting?
│  ├─ Protein-protein complex → Boltz ✓ or Chai or ColabFold
│  ├─ Protein + ligand → Boltz ✓ or Chai
│  └─ Single protein → Use ESMFold (faster)
│
├─ Need MSA?
│  ├─ No / want speed → Boltz ✓
│  └─ Yes / maximum accuracy → ColabFold
│
└─ Why Boltz over Chai?
   ├─ Open weights preference → Boltz ✓
   ├─ Boltz-2 speed → Boltz ✓
   └─ DNA/RNA support → Consider Chai

Typical performance

| Campaign Size | Time (L40S) | Cost (Modal) | Notes |

|---------------|-------------|--------------|-------|

| 100 complexes | 30-45 min | ~$8 | Standard validation |

| 500 complexes | 2-3h | ~$35 | Large campaign |

| 1000 complexes | 4-6h | ~$70 | Comprehensive |

Per-complex: ~15-30s for typical binder-target complex.

Verify

find predictions -name "*.cif" | wc -l  # Should match input count

Troubleshooting

Low confidence: Increase recycling_steps

OOM errors: Use MSA-free mode or A100-80GB

Slow prediction: Reduce sampling_steps

Error interpretation

| Error | Cause | Fix |

|-------|-------|-----|

| RuntimeError: CUDA out of memory | Complex too large | Use --use_msa_server false or larger GPU |

| KeyError: 'iptm' | Single chain only | Ensure FASTA has 2+ chains |

| FileNotFoundError: weights | Missing model | Run boltz download first |

| ValueError: invalid residue | Non-standard AA | Check for modified residues in sequence |

Boltz-1 vs Boltz-2

| Aspect | Boltz-1 | Boltz-2 |

|--------|---------|---------|

| Speed | Fast | Faster |

| Accuracy | Good | Improved, notably antibody-antigen |

| Ligands | Basic | Better support |

| Affinity prediction | No | Yes (small-molecule binding) |

| Release | 2024 | 2025 |

Boltz-2 is the current default. Boltz-1 is still used where a design pipeline

inverts the v1 model.

Affinity prediction (Boltz-2)

Boltz-2 adds an affinity-prediction module that approaches free-energy-perturbation

accuracy at a fraction of the cost. It is trained on small-molecule binding data, so

use it for protein-ligand and small-molecule work. It does not predict

protein-protein binding affinity; for protein binders, rely on interface confidence

(ipTM, ipSAE) instead.

Next: protein-qc for filtering and ranking.