boltz

name	boltz
description	Cofold protein-ligand, protein-protein, and nucleic-acid complexes with Boltz-2, and predict binding affinity with its trained affinity head. Use this skill to build Boltz input YAML, run structure prediction with MSAs, pocket constraints, templates, and modified residues, screen compound libraries by cofolding, and interpret confidence scores (pLDDT, pTM, ipTM, PDE) and affinity output (binder probability and log10 IC50). Also trigger on Boltz, Boltz-1, Boltz-2, cofolding, boltz predict, affinity_pred_value, affinity_probability_binary, ipTM, or open-weights AlphaFold3 alternatives.
license	MIT
allowed-tools	Read Write Edit Bash
compatibility	The bundled scripts need only Python 3.10+ and the standard library. Running a prediction needs boltz 2.2+ (pip install boltz, Python >=3.10 and <3.13) and an NVIDIA GPU; 24 GB is a practical minimum for a typical protein-ligand complex, CPU works but is 50-100x slower. Weights download to ~/.boltz on first run. --use_msa_server sends sequences to the public ColabFold server.
metadata	{"version":"1.0","skill-author":"K-Dense Inc.","openclaw":{"emoji":"⚡","homepage":"https://github.com/jwohlwend/boltz"},"hermes":{"category":"research"}}

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 <3.13.

Read references/yaml-schema.md before writing an input, references/confidence-and-affinity.md before believing an output, and references/running.md for install, GPU sizing, flags, and errors.

When to reach for this

Situation	Use
No experimental structure, and you need a ligand in the pocket	Boltz — folding and docking in one step
Experimental structure exists, defined site, need to rank thousands	autodock-vina
Experimental structure exists, need a pose without a box	diffdock
Need a defensible free energy for a congeneric series	Alchemical FEP, not this
No GPU	tamarind runs Boltz as a service

The loop

# 1. write the input
python skills/boltz/scripts/make_boltz_yaml.py \
    --protein-fasta target.fasta \
    --ligand-smiles "Cc1ccc(cc1Nc1nccc(n1)c1cccnc1)NC(=O)c1ccc(CN2CCN(C)CC2)cc1" \
    --affinity --out complex.yaml

# 2. predict (external, needs a GPU)
boltz predict complex.yaml --out_dir predictions/ --use_msa_server \
    --use_potentials --diffusion_samples 5

# 3. read the results with the units converted
python skills/boltz/scripts/collect_results.py predictions/

name  sample  confidence_score  iptm  ligand_iptm  binder_probability  pIC50  IC50_uM  dG_kcal_mol
lig1  0       0.84              0.82  0.79         0.93                8.10   0.0079   -11.05
lig2  0       0.51              0.42  0.31         0.21                4.60   25.1189  -6.27

make_boltz_yaml.py catches at write time the four things that otherwise fail after the GPU has already spun up: non-amino-acid characters in a sequence, an affinity binder that is not a ligand chain, with no ligand, and an out-of-range . It also quotes SMILES — an unquoted (an alkyne) starts a YAML comment and silently truncates the molecule.