version: "1.0.1" name: boltz description: > Structure prediction using Boltz-1/Boltz-2, an open biomolecular structure predictor. Use this skill when: (1) Predicting protein complex structures, (2) Validating designed binders, (3) Need open-source alternative to AF2, (4) Predicting protein-ligand complexes, (5) Using local GPU resources.

For QC thresholds, use protein-qc. For AlphaFold2 prediction, use alphafold. For Chai prediction, use chai. license: MIT category: design-tools tags: [structure-prediction, validation, open-source] biomodals_script: modal_boltz.py

Boltz Structure Prediction

Prerequisites

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

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

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

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

Boltz-1 vs Boltz-2

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.