Structure: Boltz-2 Affinity
Sequence-only protein–ligand structure and binding-affinity prediction using Boltz-2 v2.0.2. Provide a protein amino-acid sequence and a ligand SMILES, get back a predicted complex plus binding probability and pIC50.
What it simulates
- Joint structure and affinity prediction from sequence-only inputs.
- Pinned
boltz[cuda]==2.0.2driven by the upstream Boltz CLI on a GPU runner. - Diffusion + recycling pipeline with optional MSA via the configured server.
- Bundled example protein and tyrosine-derivative ligand so a fresh run produces a renderable complex without setup.
Run it on the Hub
- Open the Boltz: Boltz2AffinityPredictor Lab on the public Hub.
- Click Run. The bundled defaults predict the example complex without any parameter editing.
Inputs you can tune
| Input | Meaning |
|---|---|
protein_sequence | Amino-acid sequence string. |
ligand_smiles | SMILES string for the ligand. |
msa_path | Optional path to a pre-computed MSA (.a3m). When unset and use_msa_server: true, Boltz queries the configured MSA server. |
run_options.recycling_steps | Recycling iteration count. |
run_options.sampling_steps | Diffusion sampling steps. |
run_options.diffusion_samples | Number of diffusion samples. |
run_options.accelerator | Compute accelerator (gpu recommended; CPU is plumbing-only). |
What results to expect
- 3D structure view: predicted protein–ligand complex assembled from the top-ranked Boltz output. Sanity-check that the ligand lands in a plausible pocket on the predicted fold.
- Affinity summary: binding probability (0–1) and predicted affinity in pIC50. Useful for ranking related candidates against the same target; less useful as absolute numbers.
- Confidence summary: pTM and ipTM (global / interface fold confidence, 0–1) and pLDDT (per-residue confidence, 0–100). Low ipTM with reasonable pTM usually means the fold is fine but ligand placement is uncertain.
- Run metadata: Boltz version, output directory, truncated stdout/stderr, and
status: ok/status: error.
Reference
Boltz-2 v2.0.2.
Source on GitHub: models-boltz. Boltz-2 is GPU-bound: remote runs use the GPU-enabled runtime image; local runs need a CUDA device.