Boltzgen

BoltzGen — Universal All-Atom Binder Design

What this is

BoltzGen is an MIT-licensed (code + weights) all-atom diffusion model from the Barzilay / Jaakkola groups at MIT for universal binder design. A single YAML specification produces a ranked, filtered set of binders covering an unusually broad design space:

You design …	Against …	Protocol
Mini-proteins (~30 – 200 aa)	Protein, peptide, IDR, or disordered chain	protein-anything
Linear / cyclic / disulfide / stapled peptides	Anything	peptide-anything
Mini-proteins	Small molecule (SMILES or CCD code)	protein-small_molecule
Nanobody / VHH CDR loops	Protein target	nanobody-anything
Antibody Fab CDR loops (heavy + light)	Protein target	antibody-anything
Sequence redesign of an existing backbone	Itself (symmetric, monomer, complex)	protein-redesign

Other things BoltzGen can do that most "binder" tools cannot:

• Bind DNA / RNA (e.g., de novo zinc fingers against B-DNA).
• Condition on specific binding residues and anti-binding residues.
• Condition on secondary structure of designed regions (H/E/L).
• Insert variable-length designed loops into existing backbones (design_insertions), used heavily for antibody / nanobody scaffolding.
• Apply per-position residue whitelists / blacklists (residue_constraints).
• Sample variable-length designs in a single run (sequence: 80..140 randomly samples the length).
• Honor arbitrary covalent bonds for staples, disulfides, head-to-tail cyclization, and covalent small-molecule chemistry (SG–CK, SG–CH, OG–C1, …).
• Use antibody / nanobody scaffold libraries as drop-in files.
• Optionally fold the binder without the target as an extra confidence check (design_folding, on by default for protein targets).
• Predict small-molecule binding affinity with Boltz-2 (affinity step, on by default for protein-small_molecule).
• Filter the result by biophysical liabilities (cysteines, ALA / GLY / GLU / LEU / VAL composition outliers, hydrophobic patches).
• Diversity-optimize the final set (--alpha between 0=quality only and 1=diversity only).
• Rerun filtering without redoing GPU work (--steps filtering, ~15 sec) or with the bundled filter.ipynb.

Underneath, BoltzGen ships five checkpoints that are downloaded on first use (~6 GB total, into ~/.cache or $HF_HOME):

Checkpoint	Step	Notes
boltzgen1_diverse.ckpt	design	Diffusion backbone — diversity-tuned
boltzgen1_adherence.ckpt	design	Diffusion backbone — site-adherence-tuned
boltzgen1_ifold.ckpt	inverse_folding	Inverse-folding model trained jointly with diffusion
boltz2_conf_final.ckpt	folding (refold)	Boltz-2 structure (binder + target re-prediction)
boltz2_aff.ckpt	affinity	Boltz-2 affinity head (log10 IC50 + binder probability)

By default the design step alternates between the two diffusion checkpoints (half each) — pass --design_checkpoints <one.ckpt> to use only one, or list more.

Prerequisites

Requirement	Minimum	Notes
OS	Linux	Docker image is nvidia/cuda:12.2.2-cudnn8-devel-ubuntu22.04
Python	≥ 3.11 (requires-python = ">=3.11")	Project pins numpy==2.0.2, numba==0.61.0 — use a fresh env
GPU	NVIDIA, CUDA 12.x, bfloat16	A100 / H100 / L40S recommended; CUDA-capability ≥ 8 auto-enables cuequivariance kernels — older GPUs: --use_kernels false
VRAM	≥ 24 GB practical	OOMs go up roughly with target size + binder length + diffusion batch
Disk	~6 GB cache + design output	Weights into ~/.cache (override with --cache or $HF_HOME)
Network	First weight download + first moldir pull	Then can run offline

CPU-only is technically possible but pointless for production runs — treat that as unit-test mode only.

Three-step quick start

1) Install

# Recommended — fresh conda env with Python 3.12
conda create -n bg python=3.12 -y
conda activate bg
pip install boltzgen

# Or editable install from a clone
git clone https://github.com/HannesStark/boltzgen
cd boltzgen && pip install -e .

# Or Docker
docker build -t boltzgen .
docker run --rm --gpus all \
  -v "$(realpath workdir)":/workdir \
  -v "$(realpath cache)":/cache \
  -v "$(realpath example)":/example \
  boltzgen boltzgen run /example/vanilla_protein/1g13prot.yaml \
    --output /workdir/test --protocol protein-anything --num_designs 2

A boltzgen console script is installed with five subcommands: run, configure, execute, download, check, merge.

2) Write a .yaml spec and verify it

# binder.yaml — design a ~100-aa mini-protein against chain A of 1g13.cif
entities:
  - protein:
      id: B
      sequence: 80..140         # random length in [80, 140]
  - file:
      path: 1g13.cif            # relative to *this* yaml's directory
      include:
        - chain:
            id: A

# Validate the spec and visualize the constraints
boltzgen check binder.yaml
# (open the resulting *.cif in https://molstar.org/viewer/ —
#  the binding-site / design regions are colored differently)

3) Run the full pipeline

boltzgen run binder.yaml \
  --output workbench/test_run \
  --protocol protein-anything \
  --num_designs 50      # use 10,000 – 60,000 for a real campaign
# Output → workbench/test_run/final_ranked_designs/{intermediate_ranked_N_designs, final_<budget>_designs, results_overview.pdf}

If the run is interrupted, restart with --reuse — no GPU work is lost.

If you only want to retune filters after generation: boltzgen run binder.yaml --steps filtering --output workbench/test_run ... (~15 sec).

⚠️ The single rule that catches almost everyone

All residue indices are 1-based and use the canonical mmCIF label_asym_id, NOT the author chain ID / residue number. Verify in Molstar: hover over a residue and read the index on the bottom right. The author IDs will mislead you if you trust PyMOL, ChimeraX, or your downloaded PDB blindly.

And the one that gets everyone next:

File references inside a YAML are resolved relative to the YAML file's directory, not your CWD.

CLI at a glance

Subcommand	Purpose
boltzgen run SPEC …	Run the full design → IF → folding → analysis → filtering pipeline. The main entry point.
boltzgen check SPEC …	Validate one or more .yaml specs and write a colored mmCIF you can open in Molstar to sanity-check binding sites.
boltzgen configure SPEC …	Generate resolved Hydra step configs in --output without executing them (so you can hand-edit and then execute).
boltzgen execute DIR	Run a pipeline from a pre-configured directory produced by configure.
boltzgen download …	Pre-fetch checkpoints / moldir into --cache (not normally needed — run lazily downloads).
boltzgen merge DIRS… --output …	Combine outputs from multiple parallel runs (e.g., SLURM array tasks) before running --steps filtering on the merged set.

Most-used run flags:

Flag	Effect
--protocol P	One of protein-anything (default), peptide-anything, protein-small_molecule, antibody-anything, nanobody-anything, protein-redesign.
--num_designs N	Total intermediate designs to generate (defaults 10000; aim for 10k–60k for real campaigns).
--budget B	Size of the diversity-optimized final set produced by filtering (default 30).
--alpha A	Quality-vs-diversity knob for filtering (0=quality, 1=diversity; default 0.001, or 0.01 for peptide).
--steps STEPS…	Run only specific steps from {design, inverse_folding, design_folding, folding, affinity, analysis, filtering}.
--reuse	Re-attach to an existing --output and only generate the missing designs.
--devices N	GPUs to use; multi-GPU within a step works via Lightning DDP.
--cache PATH	Where to download / cache weights (defaults to ~/.cache; same as $HF_HOME).
--diffusion_batch_size N	Per-trunk diffusion batch (auto-set to 1 if num_designs<100, else 10; large batches sample fewer lengths).
--design_checkpoints A.ckpt B.ckpt …	Use a custom or single checkpoint (defaults to both diverse and adherence).
--inverse_fold_num_sequences K	Sequences sampled per generated backbone (default 1; raise to e.g. 5 if --num_designs is small).
--skip_inverse_folding	Use the diffusion sequence directly without IF.
--only_inverse_fold	Skip diffusion — IF an existing structure end-to-end (replaces ProteinMPNN).
--inverse_fold_avoid LETTERS	Disallowed residues, e.g. KEC; default is none for protein, C for peptide / antibody / nanobody.
--step_scale X / --noise_scale X	Replace the diffusion sampling schedule with a constant (1.8 / 0.98 are typical hand-set values).
--filter_biased true/false	Default ALA/GLY/GLU/LEU/VAL composition caps; turn off if you genuinely want biased designs.
--additional_filters 'X>0.3' 'Y<2.5'	Hard filters on any analyzed metric; single-quote because of </>.
--metrics_override k=w …	Re-weight ranking — larger weight = less important; k=none removes a metric entirely.
--size_buckets 10-20:5 20-30:10	Cap the number of designs per length bucket in the final set.
--refolding_rmsd_threshold X	RMSD threshold for "design folds as predicted" hard filter (lower is better).
--use_kernels {auto,true,false}	cuEquivariance kernels. auto enables on capability ≥ 8. Use false on old GPUs.
--no_subprocess	Run steps in-process — required for some debugging, but breaks devices>1.

Full flag list per subcommand is in references/cli.md.

Choosing the right reference doc

You want to…	Read
Install (pip / conda / Docker / dev), configure cache, fetch weights manually	references/installation.md
See every CLI flag for run, configure, execute, download, check, merge	references/cli.md
Author the YAML (entities, sequence ranges, binding, structure groups, design / not_design, insertions, residue constraints, scaffolds, file include / exclude / proximity / assembly / fuse / reset_res_index)	references/yaml-spec.md
Pick a protocol and understand its per-step config differences	references/protocols.md
Understand each pipeline step (what runs, what it writes, when to skip it)	references/pipeline.md
Tune filtering / ranking, rerun in the Jupyter notebook, use --metrics_override, --additional_filters, --alpha, --size_buckets	references/filtering.md
Read the output directory (CIF, NPZ, refold_cif, refold_design_cif, metrics CSVs, results_overview.pdf)	references/outputs.md
Run on SLURM as a job array + boltzgen merge + refilter	references/slurm.md
Train BoltzGen / inverse-fold from scratch on your own data	references/training.md
Diagnose OOMs, weight-download / HF auth issues, kernel errors, MSA fallbacks, wrong residue indices, etc.	references/troubleshooting.md

Example library

Each example in examples/ is a self-contained, fully commented YAML you can drop into a boltzgen run invocation. Coverage:

File	Use case
vanilla_protein_binder.yaml	Minimal protein binder against a target chain, no binding site specified
protein_with_binding_site.yaml	Same, but conditioning on specific binding-site residues
peptide_with_binding_site.yaml	Linear peptide binder with a binding-site spec
cyclic_peptide.yaml	Head-to-tail cyclic peptide
disulfide_peptide.yaml	Multi-disulfide-stapled peptide + secondary-structure conditioning
helicon_with_staple.yaml	Helical peptide stapled with the small-molecule WHL "staple" via two covalent bonds
cyclotide.yaml	Cyclic peptide with three disulfide bridges (e.g., 3IVQ)
nanobody.yaml	Nanobody CDR design using a library of scaffolds
nanobody_scaffold.yaml	A single nanobody scaffold YAML (referenced from nanobody.yaml)
antibody_fab.yaml	Fab CDR design against a target with a scaffold library
antibody_fab_scaffold.yaml	A single Fab scaffold YAML
protein_against_small_molecule.yaml	Protein binder against a SMILES ligand (de novo small-molecule binder)
small_molecule_via_ccd.yaml	Protein binder against a ligand specified by CCD code (e.g., TSA chorismite)
covalent_small_molecule.yaml	Covalent small-molecule chemistry — bonds from both CCD and SMILES atoms
zinc_finger_against_dna.yaml	De novo zinc finger redesigned against DNA (uses design, not_design, design_insertions, reset_res_index)
disordered_target.yaml	Bind a disordered region (target shown as fixed sequence, no structure)
disordered_peptide_as_string.yaml	Bind a fixed peptide given as a sequence (no CIF needed)
flexible_target_assembly.yaml	Use use_assembly: true and structure_groups to hide flexible loops of the target
redesign_symmetric_dimer.yaml	protein-redesign protocol with symmetric_group to tie chains during inverse folding
inverse_folding_only.yaml	--only_inverse_fold — IF an existing complex from a CIF
residue_constraints.yaml	Per-position whitelist / blacklist of amino acids
design_spec_kitchen_sink.yaml	One file demonstrating every feature listed in YAML reference

Quick decision tree

• I have a target protein and want a binder. Nothing fancy. → protein-anything. Spec: one designed protein with a length range, one file entity for the target. See examples/vanilla_protein_binder.yaml.

• I want the binder to hit a specific pocket. → Same protocol, add binding_types: - chain: id: A: binding: 5..7,13 to the target entity. Run boltzgen check first and visualize.

• I want a peptide (not a mini-protein). → peptide-anything. By default no cysteines are inverse-folded (--inverse_fold_avoid C) and design_folding is skipped. Set cyclic: true on the protein entity for head-to-tail cyclization. Add explicit Cys positions in the sequence string and bond: constraints for disulfides or staples.

• I want a stapled peptide using WHL (or any chemistry). → Declare the staple as a ligand: ccd: WHL (or SMILES), add two bond: constraints between designed Cys SG atoms and the staple atom names. See examples/helicon_with_staple.yaml.

• I want a nanobody / antibody. → nanobody-anything / antibody-anything. Use the scaffold libraries in example/nanobody_scaffolds/ and example/fab_scaffolds/ from the upstream repo as inspiration. The pattern is: include + design (the CDRs)

  • structure_groups visibility 0 over the CDRs + exclude + design_insertions
  • reset_res_index. CDRs get random insertion lengths.

• Small-molecule binder. → protein-small_molecule. Spec: designed protein + ligand (ccd or smiles). Affinity prediction is automatically on. For covalent inhibitors add bond: constraints to the ligand.

• Bind DNA or RNA. → Treat the nucleic acid as a file entity (CIF or PDB). For new zinc fingers see examples/zinc_finger_against_dna.yaml.

• Bind a disordered peptide / region. → Two options: (a) provide a CIF for the target and hide structure with structure_groups: visibility: 0, OR (b) declare the target as a protein entity with a fixed sequence string (no structure at all). See examples/disordered_*.yaml.

• Redesign sequences on an existing backbone (no diffusion). → --only_inverse_fold (one diffusion-free call to the IF model). For symmetric multimers use the protein-redesign protocol and tag chains with symmetric_group: 1.

• Resume an interrupted run. → Same command + --reuse. Already-finished designs are kept.

• I generated designs and want to tune filters / ranking only. → boltzgen run SPEC --steps filtering --output OLD_OUT … (~15 s) or open the bundled filter.ipynb.

• I want to run a 60,000-design campaign across many GPUs. → SLURM job array of boltzgen run … --num_designs 1000 tasks + boltzgen merge task-* + boltzgen run … --steps filtering. See references/slurm.md.

• I want to validate my picks further. → Re-fold the final binders with Boltz-2 directly (boltz skill, --use_msa_server), or with Chai-1 / AF-Multimer; then re-rank with ipsae for a much stronger binder-quality score than ipTM alone.

If BoltzGen is not the right tool:

• RFdiffusion (rfdiffusion skill) — fast backbone-only diffusion; pair with ProteinMPNN. Less side-chain aware but cheaper.
• BindCraft (bindcraft skill) — AF2-hallucination with built-in validation; high empirical success on protein targets, less flexible on chemistry.
• DiscoDiff / Disco (disco skill), Genie3 (genie3 skill), Chroma, Proteus — other diffusion backbone generators.

Hard rules / gotchas

1. 1-indexed, label-asym residues everywhere. sequence, res_index, bond.atom1, bond.atom2, binding, not_binding, design, not_design, design_insertions.res_index, etc. Check in Molstar.

2. Relative paths. Any path: file.cif inside a YAML is resolved relative to the YAML's directory, not your CWD. path: can also be a list of YAML files (for scaffold libraries — each scaffold is itself a file-shaped YAML).

3. Bond atom names are CCD-standard, case-sensitive. SG, CK, CH, OG, C1, N, … For a SMILES ligand D, atom names are element + index, e.g. C6 is the 6th carbon as written in the SMILES.

4. Sequence length ranges sample once per diffusion batch. sequence: 80..140 with --diffusion_batch_size 10 and only 10 intermediate designs gives you 10 designs all of the same length. For balanced sampling across lengths, keep batch size small or num_designs large.

5. --num_designs is intermediate designs, before filtering. The final set has --budget designs. Typical ratio is 10000 : 30.

6. Inverse folding avoids Cys by default for peptide / antibody / nanobody protocols. Override with --inverse_fold_avoid "" if you want them back (e.g., for new disulfide chemistry).

7. Visibility hides structure, not residues. structure_groups: visibility: 0 for a residue range tells the model the structure of those residues is unknown (they still exist as sequence). Useful for flexible loops, disordered tails, and "use this peptide but don't anchor its position".

8. include_proximity crops the target to residues within radius Å of a reference selection. Indispensable for large targets to keep GPU memory / accuracy reasonable.

9. use_assembly: true instantiates biological assemblies in a CIF (chains get suffix numbers like A1, B1, A2, …). Use it when the functional unit is a multimer and the asymmetric unit isn't.

10. reset_res_index renumbers chain residues consecutively from 1 after include / exclude operations. Critical for antibody / nanobody scaffolds where you cut out the CDRs (exclude) and then insert design loops back in (design_insertions).

11. design marks residues for redesign on a file-loaded chain; not_design removes residues from design. Useful when an entire chain should be redesigned except for catalytic / structural ones.

12. fuse: A on a file entity grafts the included segment onto a sibling protein entity with the same chain id — handy for splicing fixed N-terminal anchors onto a designed body.

13. Filtering is fast and cheap. Always plan to run it 2–5 times with different --metrics_override, --additional_filters, --alpha, --refolding_rmsd_threshold settings (or just use filter.ipynb).

14. The output directory is intentionally reusable. Pass --reuse to top up an existing run; pass --steps filtering to refilter only. To re-generate designs from scratch, point --output somewhere new or delete the existing intermediate_designs/.

15. --devices > 1 requires --no_subprocess to be OFF (the default). --no_subprocess keeps everything in the main process — only useful for single-GPU debugging.

16. Large --diffusion_batch_size × small --num_designs trades coverage of the length range for speed. Default heuristic (1 if num_designs<100, else 10) is reasonable; raise to 16–32 on H100s if VRAM permits and you have many designs to generate.

17. For protein targets with no MSA available, BoltzGen still runs (the design step does not call ColabFold) — the folding step uses Boltz-2 with target features built from the input CIF directly. There is no --use_msa_server flag (unlike Boltz standalone).

Installing this skill

# Symlink (recommended — picks up edits live)
mkdir -p ~/.claude/skills
ln -s "$(pwd)" ~/.claude/skills/boltzgen

# Or copy
cp -R . ~/.claude/skills/boltzgen

After that, an agent can invoke it via Skill(skill="boltzgen").

Citation

@article{stark2025boltzgen,
  author  = {Stark, Hannes and Faltings, Felix and Choi, MinGyu and Xie, Yuxin and Hur, Eunsu and O'Donnell, Timothy John and Bushuiev, Anton and Uçar, Talip and Passaro, Saro and Mao, Weian and Reveiz, Mateo and Bushuiev, Roman and Pluskal, Tomáš and Sivic, Josef and Kreis, Karsten and Vahdat, Arash and Ray, Shamayeeta and Goldstein, Jonathan T. and Savinov, Andrew and Hambalek, Jacob A. and Gupta, Anshika and Taquiri-Diaz, Diego A. and Zhang, Yaotian and Hatstat, A. Katherine and Arada, Angelika and Kim, Nam Hyeong and Tackie-Yarboi, Ethel and Boselli, Dylan and Schnaider, Lee and Liu, Chang C. and Li, Gene-Wei and Hnisz, Denes and Sabatini, David M. and DeGrado, William F. and Wohlwend, Jeremy and Corso, Gabriele and Barzilay, Regina and Jaakkola, Tommi},
  title   = {BoltzGen: Toward Universal Binder Design},
  year    = {2025},
  doi     = {10.1101/2025.11.20.689494},
  journal = {bioRxiv}
}