Boltz

Boltz-2: Biomolecular Structure and Binding Affinity Prediction

Overview

Boltz-2 is a biomolecular foundation model that jointly predicts complex 3D structure and binding affinity in a single inference pass. It is the first fully open-source model to approach AlphaFold 3-level structural accuracy while also providing quantitative affinity predictions 1000× faster than physics-based FEP methods.

| Capability | Boltz-2 | AlphaFold 3 | |---|---|---| | Proteins, RNA, DNA, ligands | ✓ | ✓ | | Binding affinity prediction | ✓ | ✗ | | Pocket/contact constraints | ✓ | ✗ | | Cyclic peptides | ✓ | ✗ | | License | MIT (commercial OK) | CC-BY-NC-SA 4.0 | | Model weights | Freely available | Must apply to Google | | Input format | YAML | JSON | | Output format | mmCIF or PDB | mmCIF only |

Installation

pip install boltz[cuda] -U

For CPU-only (slow, for testing):

pip install boltz -U

Requirements: Python ≥ 3.10, / ├── model0.cif ← predicted structure (best sample) ├── confidence__model0.json ← confidence metrics ├── affinity.json ← binding affinity (if requested) ├── pae__model0.npz ← PAE matrix ├── pde_model0.npz ← predicted distance error └── plddt_model_0.npz ← per-residue pLDDT

With `--diffusion_samples 5`: output files for `_model_0` through `_model_4`.

### Confidence metrics

From `confidence__model_0.json`:

| Metric | Range | Interpretation |
|---|---|---|
| `confidence_score` | 0–1 | Primary: `0.8 × complex_plddt/100 + 0.2 × iptm` |
| `iptm` | 0–1 | Interface confidence (>0.6 = acceptable, >0.8 = high) |
| `ptm` | 0–1 | Global fold confidence |
| `complex_plddt` | 0–100 | Mean per-residue local confidence |
| `complex_iplddt` | 0–100 | pLDDT at protein-ligand interface |
| `complex_pde` | Å | Predicted Distance Error (lower = better) |

### Affinity output

From `affinity_.json`:

```json
{
  "affinity_probability_binary": 0.85,
  "affinity_pred_value": -2.1
}

| Field | Use for | Interpretation | |---|---|---| | affinity_probability_binary | Hit screening (binder vs. non-binder) | 0–1; >0.5 = predicted binder | | affinity_pred_value | Lead optimization (ranking actives) | log₁₀(IC₅₀ in μM); lower = stronger |

affinity_pred_value scale:

• −3 → IC₅₀ = 1 nM (very potent)
• 0 → IC₅₀ = 1 μM
• 3 → IC₅₀ = 1 mM (very weak)

Convert to kcal/mol: ΔG ≈ (6 − y) × 1.364

Important: Use affinity_probability_binary for screening (distinguish binders from non-binders). Use affinity_pred_value only when comparing active compounds — do not use it to compare actives against inactives.

When to Use Boltz-2 vs AlphaFold 3

Use Boltz-2 when:

• You need binding affinity estimates alongside structure
• Commercial use is required (MIT license vs AF3's non-commercial restriction)
• You want pocket constraints to guide ligand placement
• You are working with cyclic peptides
• You need fast iterative screening

Use AlphaFold 3 when:

• You have access to model weights and purely academic use is fine
• You need post-translational modifications (Boltz-2 supports them via CCD but more limited)
• Reproducing AF3 benchmark results is required

Scripts

# Build a YAML input from sequences and ligands
python scripts/build_input.py --name my_job \
  --protein MKTAYIAKQRQISFVK \
  --ligand-smiles "CC(=O)Nc1ccc(O)cc1" \
  --affinity B

# Build from a FASTA file
python scripts/build_input.py --name complex \
  --from-fasta sequences.fasta \
  --ligand-ccd ATP

# Parse and summarize an output directory
python scripts/build_input.py --report predictions/my_job/

Resources

• GitHub: https://github.com/jwohlwend/boltz
• Boltz-2 paper: Passaro et al., bioRxiv 2025 — https://doi.org/10.1101/2025.06.14.659707
• Boltz-1 paper: Wohlwend et al., bioRxiv 2024 — https://doi.org/10.1101/2024.11.19.624167

References

• references/input-format.md — full YAML schema: all entity types, constraints (pocket, bond, contact), affinity properties, modifications, templates, MSA options
• references/outputs.md — parsing mmCIF, confidence JSON, affinity JSON, PAE/pLDDT arrays; batch ranking

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: naity
• Source: naity/FM4Life
• License: MIT

Install and usage instructions live in the source repository linked above.