BoltzGen

BoltzGen designs proteins, peptides, nanobodies, and other molecules that bind to target structures or small molecules. Achieves 60-70% success rates for nanobody and protein binders against novel targets with only 15 designs tested.

​

Experimental Validation

Validated across 8 independent wetlab campaigns spanning diverse targets and modalities.

​

Novel Targets (No Similar Bound Structures in PDB)

Nanobody & Protein Design: 66% success rate (6/9 targets with nM binders) testing ≤15 designs per target

​

Benchmark Targets

80% success rate for both nanobodies and proteins against 5 clinical targets (IL-7Rα, InsulinR, PDGFR, PD-L1, TNFα)

​

Other Modalities

Bioactive Peptides: nM-μM binders for melittin, indolicidin, protegrin (6 designs per target) RagC Linear Peptides: 7/29 designs bound (best: 3.5 μM) RagA:RagC Cyclic Peptides: 14/24 designs bound (best: 80 μM) NPM1 Disordered Region: 1/5 designs showed nucleolar localization in live cells Small Molecules: Weak binding (30-250 μM) for rucaparib and rhodamine derivative GyrA Antimicrobials: 19.5% of 1,808 designs inhibited E. coli growth >4×

​

Design Tasks

​

Nanobody Design

Design single-domain antibodies against protein targets. Required:

• Target PDB file
• Target chains (optional, defaults to all)

Optional:

• Binding site residues
• Scaffold type: Default (de novo) or Custom (optimize existing framework)

Custom Framework Options:

• Framework input type: Structure or Sequence
• For structure input: Framework PDB file, framework chain, CDR regions to design
• For sequence input: Framework sequence with CDR placeholders using number ranges (e.g., EVQLVESGGGLVQPGGSLRLSCAASG5..10WVRQAPGKGLEWVS8..12RFTISRDNSKNTLYLQMNSLRAEDTAVYYC10..20WGQGTLVTVSS)
• CDR exclude counts: Residues to remove from start of each CDR (e.g., “3,3,7”)
• CDR insertion lengths: Ranges for new residues per CDR (e.g., “1-5,1-5,1-14”)

​

Antibody Design

Design antibodies with heavy and light chains against protein targets. Required:

• Target PDB file
• Target chains (optional, defaults to all)

Optional:

• Binding site residues
• Scaffold type: Default (de novo) or Custom (optimize existing framework)

Custom Framework Options:

• Framework input type: Structure or Sequence
• For structure input: Framework PDB file, heavy chain ID, light chain ID, heavy/light CDR regions
• For sequence input: Heavy chain framework sequence and light chain framework sequence with CDR placeholders
• Heavy/Light CDR exclude counts and insertion length ranges

​

Protein Binder Design

Design general protein binders. Required:

• Target PDB file
• Length range (e.g., 100-150)

Optional: Target chains, binding site, number of designs, batches, budget

​

Peptide Design

Design linear or cyclic peptides. Required:

• Target PDB file
• Length range (e.g., 10-20)

Optional: Target chains, binding site, cyclic flag, number of designs, batches, budget

​

Cyclotide Design

Design cyclic peptides with disulfide bonds. Required:

• Target PDB file
• Cyclotide sequence specification with cysteines for disulfide bonds (e.g., “3C8C6C5C3C1C2” means 3 residues, Cys, 8 residues, Cys, etc.)
• Disulfide bond pairs (positions of cysteine pairs to form bonds)

Optional: Target chains, binding site, number of designs, batches, budget

​

Small Molecule Binder Design

Design proteins that bind small molecules. Required:

• Target SMILES string
• Length range (e.g., 100-150)

Optional: Number of designs, batches, budget Note: Currently achieves weak (30-250 μM) binding; consider 10,000+ designs for better results

​

Custom Design

Advanced mode for complex designs with multiple entities and constraints. Required:

• Entities: List of proteins and ligands with:
  • Type: “protein” or “ligand”
  • Chain ID
  • Sequence (use numbers for designed regions: “15..20”, letters for fixed: “AAVTT15”)
  • CCD Code or SMILES (for ligands)

Optional: Binding types, secondary structure, cyclic flag, constraints (disulfide bonds), number of designs, batches, budget Use Cases: Disulfide-bonded cyclic peptides, stapled peptides, multi-chain complexes

​

YAML Configuration

Provide a custom YAML configuration file with associated structure files for advanced control. Required:

• YAML configuration file
• Structure file(s) (PDB or CIF format)
• Protocol selection: Nanobody, Protein, Small Molecule Binder, or Peptide

​

Common Parameters

Number of Designs:

• Default: 10
• Range: 1-100,000
• For large runs, design batching automatically splits into multiple jobs

Design Batching:

• Auto-enabled when number of designs > 10
• Configurable batch size (default: 10 designs per job)

Budget:

• Final number of designs optimized for diversity and quality
• Default: 2

Omit Amino Acids:

• Amino acids to exclude from the design
• Default: C (cysteine) for peptide and nanobody design, none for others
• Use “empty” to include all amino acids

​

Best Practices

Starting Out:

• Test run: 10-50 designs
• Production: 10,000-60,000 designs for challenging targets
• Budget: 2-100 final diverse designs

Target Selection:

• Binding sites should have ≥3 hydrophobic residues
• Avoid heavily glycosylated regions
• Specify binding site when possible

Expected Success Rates:

• Nanobodies/Proteins: 60-70% (nM affinity) for novel targets
• Peptides: Lower affinity (μM range) but good hit rates
• Small molecules: Challenging (μM affinity)

Troubleshooting:

• No successes: Try different binding sites or longer runs
• Low expression: Check for hydrophobic patches
• Avoid lengths 73-76: Known memorization issue (generates ubiquitin)

​

Pipeline Output

When the pipeline completes, your output directory will contain:

​

Configuration Files

• config/ — Configuration files
• steps.yaml — Pipeline steps configuration

​

Initial Designs

intermediate_designs/ — Output of design step

• *.cif — CIF structure files for designed proteins and targets before inverse folding
• *.npz — Metadata files for designs

​

Processed Designs

intermediate_designs_inverse_folded/ — Output of inverse folding, folding, and analysis steps

• *.cif — CIF files after inverse folding (designed residues have backbone atoms only; sidechain coordinates are 0,0,0)
• *.npz — Metadata files
• refold_cif/ — Refolded complex structures (target + binder). Primary input for analysis and filtering
• refold_design_cif/ — Refolded binder structures without target
• aggregate_metrics_analyze.csv — Aggregated metrics across all designs
• per_target_metrics_analyze.csv — Metrics per target

​

Final Results

final_ranked_designs/ — Output of filtering step

• intermediate_ranked_[N]_designs/ — Top-N quality designs (CIFs copied from refold_cif/)
• final_[budget]_designs/ — Final quality + diversity set (CIFs copied from refold_cif/)
• all_designs_metrics.csv — Metrics for all designs considered by filtering
• final_designs_metrics_[budget].csv — Metrics for selected final set
• results_overview.pdf — Visualization plots

​

Key Metrics in CSV Files

Quality Metrics:

• design_ptm — Predicted TM-score for designed structure (higher = better, >0.75 recommended)
• design_iptm — Predicted TM-score for design-target interactions (higher = better)
• filter_rmsd — RMSD to refolded structure (lower = better, <2.5Å recommended)

Interface Metrics:

• plip_hbonds_refolded — Number of hydrogen bonds between design and target
• plip_saltbridge_refolded — Number of salt bridge interactions
• delta_sasa_refolded — Change in solvent accessible surface area (higher = better burial)

Developability Metrics:

• design_hydrophobicity — Hydrophobicity score of designed residues
• design_largest_hydrophobic_patch_refolded — Area of largest hydrophobic patch (lower = better)
• liability_score — Overall developability score (lower = better)

Ranking:

• final_rank — Final ranking position (1 = best)
• pass_filters — Binary flag indicating whether design passed all filters

​

Runtime

Approximate time per design for ~200 residues:

• ~60 sec (generation)
• ~5 sec (inverse folding)
• ~60 sec (structure prediction)
• ~20 sec total (filtering all designs)

Try BoltzGen