7 BoltzGen Project Setup

7.1 Overview#

BoltzGen is an all‑atom generative model for universal binder design. On Ariax, you can launch the following BoltzGen project types:

• VHH (nanobody)
• Miniprotein
• Peptide
• Cyclic‑Peptide
• Helicon
• Miniprotein + Small Molecule (design against a small‑molecule ligand via CCD or SMILES)

7.2 Inputs#

• Protein targets: Upload a .pdb or .cif, or search RCSB by PDB ID/keyword directly from the setup page.
• Small‑molecule targeting (miniprotein + small molecule):
  • Provide exactly one: CCD (3‑letter) or SMILES.
  • No target structure file is required in this mode.

Canonical indexing and residue masks#

For mmCIF targets, choose protein label_asym_id chains and use their positive label_seq_id positions in residue masks. Solvent chains are excluded, even when their label IDs match a protein's author ID. The viewer and logs display shift messages when author and label indices differ.

You can define binding rules per chain:

• “Bind anywhere” (equivalent to all)
• binding and/or not_binding masks using:
  • comma lists: 5,13,21
  • ranges: 50..60
  • the keyword all
    Validation will flag invalid tokens, out‑of‑range residues, and overlaps between binding and not_binding.

7.3 Project Type–specific fields#

• VHH / Helicon: Length fields are hidden (not required).
• Miniprotein / Peptide / Cyclic‑Peptide: Configure Min Length and Max Length.
  • Miniprotein: min ≥ 31 (typical defaults 65–150)
  • Peptide/Cyclic‑Peptide: 8–30

7.4 Design Settings#

• # Generated Designs (num_designs): Requests the generation volume (e.g., 10,000+ recommended by authors for robust down‑selection), not the number of final designs.
• Budget (budget): Must be ≤ num_designs. Requests the final diversity selection size; selection does not guarantee that every selected design passes pipeline filters.

7.5 GPU Selection & Turbo Mode#

• Mode: Performance (fastest) or Cost (cheapest).
• Primary GPUs: H100, A100 80GB, L40/L40S, and RTX 6000 Pro. Select at least one for availability.
• Additional GPUs: B200, H200, and RTX 6000 Ada are optional selections.
• Turbo Mode (2×/4×/8×): Runs in parallel across multiple GPUs to accelerate campaigns. Compute charges depend on the allocated GPUs and duration; see current allocation pricing. Turbo increases concurrent GPU use.

7.6 Validate & Start#

Click Validate Settings to check structure parsing, chain selection, mask syntax, length rules, GPU availability, and Turbo options. Once validated, click Start BoltzGen. Your job appears on the dashboard immediately, streaming logs and results as they become available.

7.7 Outputs & Files#

Results stream into organized folders:

• Final Ranked Designs: A timestamped directory containing the final ranked CIFs; a subfolder final_<N>_designs is selected based on your budget.
• Intermediate Designs and Inverse Folded Designs: Optional folders for additional inspection (when produced by the pipeline).
• ZIP archives: Per‑tab .zip bundles are created for quick download; large files include CLI download suggestions (cURL/wget/aria2c).

Per‑step progress bars indicate pipeline stages and percent complete. For viewing structures, CIF files may be converted to PDB for a consistent viewer experience.

7.8 Helpful Links#

• BoltzGen Getting Started Tutorial: /resources/boltzgen-getting-started — complete workflow from target prep to campaign results
• Announcement: /resources/boltzgen-arrives-at-ariax-bio
• Boltz website: https://boltz.bio/
• BoltzGen announcement: https://boltz.bio/boltzgen
• GitHub: https://github.com/HannesStark/boltzgen
• Manuscript (PDF): https://hannes-stark.com/assets/boltzgen.pdf