Boltz

Boltz-2

An open-weights cofolding model in the AlphaFold3 family, plus something AlphaFold3 does not have: a trained binding-affinity head. Give it a protein sequence and a ligand SMILES and it returns a complex structure, per-interface confidence, and a predicted potency — with weights licensed for commercial use.

Repo: github.com/jwohlwend/boltz Checked against: boltz 2.2.1 (PyPI), Python ≥3.10 0.8 | Confident interface; the pose is usable | | 0.6 – 0.8 | Plausible; check it against known site residues | | < 0.6 | The model does not believe its own interface — an affinity computed on it is meaningless |

collect_results.py --min-iptm 0.6 filters, and warns about what it dropped. Note that pde and ipde are in Angstrom, so for those alone lower is better.

Constrain the pocket when you know it

Without a pocket constraint, Boltz decides where the ligand goes — usually right for a well-defined site, less so for a shallow or multi-site protein.

python skills/boltz/scripts/make_boltz_yaml.py --protein-fasta target.fasta \
    --ligand-ccd SAH --pocket A:790,A:797,A:855 --pocket-distance 6 --out cofactor.yaml

--pocket-force makes it a hard constraint rather than a bias. Use it only when you are sure: forcing a wrong pocket produces a confident wrong answer, which is worse than an unconstrained one.

Screening a library

python skills/boltz/scripts/screen_library.py --protein-fasta target.fasta \
    --smiles library.smi --out-dir screen/ --affinity --msa-path target.a3m

boltz predict screen/ --out_dir screen/predictions --use_potentials --diffusion_samples 5
python skills/boltz/scripts/collect_results.py screen/predictions --min-iptm 0.6 --out hits.tsv

One YAML per ligand, plus a manifest. Precompute the MSA and pass --msa-path — every input shares the same protein, and rebuilding its MSA N times is the single largest waste in a screen. The script says so if you forget.

It also flags compounds above the affinity head's 128-atom limit (--skip-oversized drops them); past that limit Boltz returns a number that means nothing.

Scale honestly: a few minutes per ligand on a 24 GB GPU with a precomputed MSA. This is a hundreds-to-low-thousands method. Filter a large library with autodock-vina or medchem first and bring the survivors here.

What it is, and is not

The affinity head is trained on measured bioactivity, so it behaves like a very good structure-aware QSAR model, not a physics calculation. It reflects the chemistry and target classes in its training data; a novel scaffold against an under-studied target is extrapolation, and there is no thermodynamic cycle to check it against.

Treat agreement with an orthogonal method as the evidence — a docking score from autodock-vina, measured analogues from chembl, or a stability check in molecular-dynamics. Report the release, the ipTM, and the pIC50 with its ensemble spread.

Composing with the rest of the bundle

• uniprot-rcsb → here: the sequence, and a template CIF if an apo structure exists.
• binding-site-analysis → before: which site to focus on, and whether it is druggable.
• chembl → here: known actives against the target, to calibrate what the affinity head says

about chemistry you already have data for.

• medchem / rdkit → before: triage and standardise the library.
• autodock-vina → alongside: an orthogonal score on the same compounds.
• molecular-dynamics → after: does the predicted pose survive 10 ns?
• tamarind → instead: hosted Boltz when there is no local GPU.

Source & license

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

• Author: K-Dense-AI
• Source: K-Dense-AI/drug-discovery-agent-skills
• License: MIT
• Homepage: www.k-dense.ai

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