BoltzMaker

Two steps

Fully Automated Mode

Configure once, run one command at home, explore the results here.

  1. Prepare. Answer questions about your proteins, partners and ligands, then download a single bundle carrying the software, the pinned environment and the scripts.
  2. Analysis. Run the bundle on your own machine, upload the one results file it writes, and explore your campaign interactively.

Best if you want the whole pipeline handled for you.

Four tools

Stepwise Mode

Drive each stage yourself, one at a time.

  1. Wizard. Build a boltz_input.md by answering plain questions.
  2. Generate. Turn that spec into the per-target YAML files Boltz-2 needs.
  3. Preflight. Validate chemistry, chain ids and inputs before spending GPU time.
  4. Analyze. Upload a finished campaign folder for a summary and dashboard.

Best if you already have your own setup and want one piece of it.

What it does

One spec file describes a campaign: proteins, the chains they fold with, and the ligands to cross them against. BoltzMaker writes the inputs, checks the chemistry before any GPU time is spent, runs Boltz-2, and turns the results into a report you can read.

Latest runs

  • ABL1_KD
    2026-08-23 · 5 targets · explore
  • ABL_KINASE
    2026-08-23 · 5 targets · explore
  • GLP1R_GIPR_pocket_matrix
    2026-08-22 · 20 targets · explore
  • 5HT2A_GQ
    2026-08-13 · 2 targets · explore
  • 5ht2_gq
    2026-07-11 · 15 targets · explore

Where a protein family is known

A prediction tells you where the atoms went, not whether the receptor moved. Given an apo structure to compare against — experimental, or a ligand-free one it predicts for you — BoltzMaker measures that shift against the motifs the family is actually described by.

  • GPCRs. GPCRdb generic (Ballesteros-Weinstein) numbering, so the shift is reported per TM1–TM7 and H8 and per ICL1–3 and ECL1–3 rather than as one number for the chain. TM6, TM7 and ECL2 are marked binding-site adjacent, which is where an agonist-driven change shows up.
  • Kinases. KLIFS pocket numbering picks out the catalytic lysine, αC glutamate, gatekeeper, hinge, HRD catalytic loop and DFG motif, and reports whether DFG and αC changed between apo and holo — a coarse Cα–Cα distance proxy for detecting a shift, not a publication-grade dihedral classifier.
  • Anything else. Pfam domains from PDBe's SIFTS residue mapping, so a protein outside those two families still gets its shift per domain instead of per chain.
  • Whatever the family. Real protein–ligand interaction fingerprints from PLIP, binding affinity as pIC50 where you ask for it, and ligand chemistry checked for undefined stereocentres, ambiguous protonation and stray salts before a campaign starts rather than after.