BoltzGen

BoltzGen webserver overview

BoltzGen v0.3.2 designs peptide, protein, nanobody, and Fab binders against protein structures, redesigns selected regions of protein complexes, and designs proteins around small-molecule targets. ProteinIQ runs the complete native pipeline and returns ranked complexes, native metrics, binder-only refolds when available, sequence files, reports, and run provenance.

A run turns the target, the designed entity, and any constraints into a native design specification, generates binder atoms by diffusion, assigns sequences with the BoltzIF inverse-folding model, refolds each designed sequence in complex with the target using Boltz-2, then measures, filters, and selects a quality-diverse final set. Small-molecule design adds the Boltz-2 affinity stage, whose values are model predictions used for ordering candidates.

Because generation and selection are separate stages, Number of designs sets the candidate pool and Budget sets how many candidates survive selection: a large budget cannot recover diversity that was never generated.

Pricing

BoltzGen is billed for the compute it uses, at 45 credits per minute of metered runtime. A run's length depends on the target, the protocol, and how many candidates pass the native filters, so the final cost cannot be known before the job runs. The quote shown at submission is an estimate of the final cost, and real runs can finish above or below it.

To start the job, ProteinIQ holds credits from your balance, with a minimum of 500. While the job runs, the hold is extended in small steps so that it always covers the compute used so far. When the job finishes, you pay only for the metered runtime and the unused part of the hold is returned. For example, a run of 1,417 seconds is charged 1,063 credits, even if the quote was lower or higher. Metered runtime is capped at 12 hours, equivalent to 32,400 credits at the current rate. To cap the cost of a run, turn on Limit spend for this job in the submit bar. The job then stops at that limit and is charged for the compute used, and results may be incomplete.

The table shows example quotes for one 200-residue protein target, a uniform binder length, one sequence per backbone, inverse folding enabled, and native sampling and filtering defaults. These are estimates, not prices or maximums.

Protocol	Binder assumption	1 design	10 designs	30 designs	100 designs
Peptide binder	12 residues	500	611	1,320	3,804
Protein binder	100 residues	500	736	1,811	5,572
Nanobody (~130 residues)	Bundled scaffold	500	549	1,191	3,437
Antibody / Fab CDR (~450 residues)	All bundled scaffolds	500	678	1,746	5,481

The quote can increase with target size, binder length, candidate count, Sequences per backbone, variable binder length, and protocol-specific folding or affinity stages. Budget changes how many candidates are retained but does not currently increase the runtime estimate independently of Number of designs. Binding-site and structure constraints, checkpoint choice, filters, metric weights, and size buckets do not currently change the quote. Skipping inverse folding lowers the estimate.

Inputs

Input	Accepted value	Limit	Used by
Target protein	.pdb, .ent, .cif, or .mmcif upload, or RCSB PDB ID	One structure, 50 MB, protein atoms required	Peptide, protein, redesign, nanobody, and Fab protocols
Nanobody framework	.pdb, .ent, .cif, or .mmcif upload	One structure, 10 MB, protein atoms required	Custom nanobody framework only
Target ligand	SMILES, CCD code, or PubChem result	One ligand	Protein-small molecule protocol

For mmCIF inputs, chain and residue controls use label_asym_id and 1-based label_seq_id, not author numbering.

Choosing a protocol

Protocol decides what BoltzGen designs and which scientific stages run, so pick it from the molecule that will be synthesized or expressed.

Protocol	What it designs	Pick it when
Peptide binder	A linear or backbone-cyclic peptide	The product is a short peptide against a protein structure
Protein binder	A new folded protein	The product is a de novo miniprotein or larger binder
Protein redesign / optimization	Selected residues or inserted regions of an uploaded structure	A complex or scaffold already exists and needs optimizing
Nanobody (~130 residues)	CDRs in a bundled or uploaded VHH framework	The experimental format is a single-domain antibody
Antibody / Fab CDR (~450 residues)	CDRs in a bundled Fab scaffold	The experimental format is a paired Fab
Protein-small molecule	A protein around a ligand	The target is supplied as SMILES or a CCD code rather than a PDB

Each protocol brings its own sequence policy, refolding stages, filters, and metrics, so a convenient default binder length is not a reason to choose one. For antibody formats, the bundled libraries are the simplest start: a custom VHH framework additionally needs Framework chain and explicit CDR design regions, and ProteinIQ infers neither antibody numbering nor CDR boundaries.

Preparing the target

BoltzGen conditions on the structure as supplied. Before submitting, check the biological assembly, the intended chains, missing residues, alternate conformations, and the bound ligands or cofactors that define the site, because an apo form, a predicted model, or one crystallographic state is not automatically the state worth targeting. Routine repair belongs to PDBFixer beforehand, with cofactors and partners preserved.

Author numbering is the most common cause of a valid-looking but wrong design task. Confirm every chain and residue identifier in a structure viewer before entering Binding site (optional), Target chains (optional), Redesign regions, or custom CDR ranges. Leave Binding site (optional) empty for exploratory surface sampling, and when an epitope is known from structural, mutagenesis, or competition data, describe it with enough residues to cover the surface rather than one contact.

For Protein-small molecule, the ligand identity is settled only once stereochemistry, protonation, charge, and tautomer state have been checked in the SMILES string or CCD entry.

Campaign size

The BoltzGen documentation suggests a first run of about 50 designs, and 10,000 to 60,000 designs for a real campaign. ProteinIQ accepts 1 to 100 candidates per job, so a hosted job is a trial or a focused search rather than a discovery-scale pool. Start with 1 to 10 designs to confirm that parsing, numbering, constraints, and runtime behave as expected, then raise the count, and run independent jobs when the question needs broader sampling.

Keep the native defaults for Quality vs diversity (alpha), Refolding RMSD threshold (Å), Step scale, Noise scale, Model checkpoint, filters, and metric weights on a first run, and change one at a time afterwards with a reason recorded. Skip inverse folding is a diagnostic for backbone-only output, not a saving to apply to candidates that will be synthesized, since it removes the stage that improves agreement between a designed backbone and its refolded sequence.

The bundled Peptide binder (12 residues) example loads RCSB entry 8JJS with 20 designs and a budget of 20. It is an input exercise: inspect which chains, heterogens, and the existing cyclic peptide belong in the intended target model, run a smaller trial first, and compare automatic site selection against a run constrained to a justified surface. No published result job accompanies it, so the output has to be judged from the actual run.

Settings

Protocol and campaign

Parameter	Type	Default	Description
Protocol	enum	Peptide binder	Selects peptide, protein, protein redesign, nanobody, Fab, or protein-small molecule design.
Antibody scaffold	enum	All scaffolds (diverse sampling)	Selects all bundled Fab scaffolds or one named therapeutic scaffold; Fab protocol only.
Nanobody framework	enum	BoltzGen scaffold library	Uses bundled scaffolds or an uploaded VHH framework; nanobody protocol only.
Framework chain	string	conditional	Required for a custom nanobody framework; accepts an alphanumeric chain ID.
CDR design regions	string	conditional	Required for a custom framework; comma-separated 1-based canonical residues or closed ranges such as 26..34,52..59,98..118.
Sample CDR lengths	boolean	false	Allows selected custom CDR segments to be replaced by sampled-length insertions.
Variable CDR segments	string	conditional	Required when CDR length sampling is enabled; uses removed_range:replacement_range, such as 26..28:1..5.
Number of designs	integer	30	Generates 1 to 100 candidates before filtering.
Budget	integer	30	Retains 1 to 100 candidates after quality and diversity selection; a budget above the generated count keeps every design.

Binder geometry and target selection

Parameter	Type	Default	Description
Uniform binder size	boolean	true	Uses one exact binder length for direct peptide, protein, and protein-small molecule design.
Binder length	integer	12	Exact direct-binder length from 5 to 500 residues.
Minimum length	integer	10	Minimum sampled direct-binder length from 8 to 150 when uniform size is off.
Maximum length	integer	14	Maximum sampled direct-binder length from 8 to 150 and must not be below the minimum.
Binding site (optional)	string	optional	Target residues in chain:residue_spec form, such as A:12,14,61; empty allows automatic site selection.
Cyclic peptide	boolean	false	Enables backbone cyclization for the peptide protocol.
Target chains (optional)	string	optional	Comma-separated chains such as A,B; empty includes all target chains.
Redesign regions	string	conditional	Required for protein redesign; one chain per line using A:10, A:10..20, A:..30, or A:55...

Sequence generation and final selection

Parameter	Type	Default	Description
Skip inverse folding	boolean	false	Skips sequence redesign and its runtime stage.
Sequences per backbone	integer	1	Generates 1 to 10 inverse-folded sequence variants for each backbone.
Avoid amino acids	string	optional	Excludes supplied one-letter amino-acid codes; empty preserves the protocol policy.
Quality vs diversity (alpha)	number	native	Optional value from 0 to 1; empty preserves 0.01 for peptide design and 0.001 for other protocols.
Filter biased compositions	boolean	true	Applies the native amino-acid composition filter.
Refolding RMSD threshold (Å)	number	native	Optional value from 0.5 to 5.0 Å; empty preserves 2.0 Å for peptides and 2.5 Å for other protocols.
Custom filters	string	optional	One strict metric<value or metric>value expression per line.
Metrics weights	string	optional	One metric=value override per line; metric=none removes a metric from ranking.
Size buckets	string	optional	One min-max:count cap per line for the final selected set.

Diffusion and structural constraints

Parameter	Type	Default	Description
Step scale	number	native	Optional value from 0 to 3; empty preserves the native per-stage schedule.
Noise scale	number	native	Optional value from 0 to 1; empty preserves the native per-stage schedule.
Model checkpoint	enum	Both (diverse + adherence)	Uses both native checkpoints or only the diverse or adherence checkpoint.
Secondary structure	string	optional	One chain:start-end:HELIX, SHEET, or LOOP constraint per line for direct binders.
Disulfide bonds	string	optional	One chain:residue,chain:residue pair per line for direct binders.
Staple bonds	string	optional	One chain:residue:atom,chain:residue:atom pair per line for direct binders.
Fixed sequence regions	string	optional	One chain:start-end:sequence region per line; requires a uniform direct-binder length and non-overlapping ranges.
Binding residues	string	optional	Binder positions that must contact the target in chain:residue1,residue2 form.
Non-binding residues	string	optional	Binder positions that must not contact the target in chain:residue1,residue2 form.
Design insertions	string	optional	Protein-redesign insertions in chain:position:min..max[:secondary_structure] form.
Residue constraints	string	optional	Inverse-folding rules in chain:position:allowed|disallowed:amino_acids form.

Direct binder constraints use chain B for peptide and protein targets and chain A for small-molecule targets. Malformed or protocol-inapplicable constraints are rejected before compute starts.

Outputs

Output	Format	Description
Ranked binder-target complex	.cif	Native final refold for each retained design in final-rank order.
Binder-only refold	.cif	Native isolated-binder refold when the selected protocol produces it.
Designed chain sequence	.fasta	One additive per-design FASTA projected from native full-chain sequence columns. Fab runs use source-proven heavy_chain and light_chain roles; other modes retain neutral native asym labels.
Final and all-design metrics	.csv	Native final-set and all-design tables without renamed columns.
Aggregate and per-target analysis	.csv	Native analysis tables when produced.
Results overview	.pdf	Native BoltzGen plots when produced.
Run provenance	.yaml	Native steps.yaml and generated configuration files.
Run log	.log	Curated protocol, campaign, scientific phase, resolved Fab scaffold, result metric, warning, and failure summary.
Reference inputs	Original format	Uploaded target and custom framework structures, or ligand reference.

Completed runs omit high-volume intermediate CIF and NPZ families. If a campaign reaches its reserved runtime limit, completed designs and native artifacts produced before execution stops are returned when available.

Understanding results

Field	Meaning
Binder pTM	Native design_ptm, a predicted quality score for the designed component; higher is better.
Binder-target iPTM	Native design_to_target_iptm, confidence for interactions between designed residues and the target; higher is better.
Minimum binder-target PAE (Å)	Native min_design_to_target_pae, the most confident predicted design-target contact; lower is better.
Delta SASA (Ų)	Native delta_sasa_refolded, the change in solvent-accessible surface area after complex formation; larger values indicate more buried surface.
Normalized rank score	Native quality_score used by final ranking; compare it only within the same run and settings.
Sequence	Native designed residue sequence for the retained candidate.

No single confidence or interface score proves binding, affinity, specificity, expression, solubility, or biological activity. Inspect the complete native metrics, structures, sequence diversity, and independent experimental evidence before selecting candidates.

Confidence fields

Minimum binder-target PAE (Å) is the field most often overread. It names the single most confident predicted contact, not the uncertainty of the interface as a whole, so read it beside Binder-target iPTM and the full native table rather than on its own. The PAE and ipTM and pTM guides cover what those measures do and do not say; neither has a threshold that transfers between targets.

Refolding agreement

The native tables carry filter_rmsd, which compares the refolded complex with the generated design, and designfolding_filter_rmsd, which does the same for the binder folded without its target where the protocol produces it. Lower is better inside one protocol and atom selection. A candidate with an attractive generated interface and poor refolding agreement is internally inconsistent, and one that holds its fold only in the complex is a weaker choice when the molecule has to express and stay folded on its own.

Interface and rank

Buried surface and contact counts describe an interface without judging it. Clashes, exposed hydrophobic patches, unsatisfied polar groups, strained geometry, or simply the wrong epitope outweigh a large Delta SASA (Ų), and MolProbity catches the geometric problems that a confidence score does not. Normalized rank score aggregates weighted metric ranks by emphasizing the weakest component and then applies the quality-diversity step, which makes it meaningful inside one run and its settings and meaningless across campaigns.

Sequence diversity

Compare sequence identity and structural similarity across the shortlist before ordering anything, using the RMSD calculator for pairwise structural comparison. A budget of 20 is not 20 options when most of the set collapses onto one near-duplicate family. For multi-chain designs outside the Fab protocol, the asym identifiers in the FASTA files are neutral labels, so chain order carries no information about which chain is the target or which is a variable domain.

Sources

• Stark H, Faltings F, Choi M, et al. "BoltzGen: toward universal binder design." bioRxiv (2025). doi:10.1101/2025.11.20.689494
• BoltzGen source code and documentation. HannesStark, GitHub. github.com/HannesStark/boltzgen
• 8JJS: human K-Ras G12D (GDP-bound) in complex with cyclic peptide inhibitor AP10343. RCSB Protein Data Bank (2023). rcsb.org/structure/8JJS