How to use BoltzGen online

BoltzGen generates binder structures and sequences for protein, peptide, nanobody, Fab, redesign, and protein-small molecule tasks. Its main caveat is scale: useful binder discovery depends on sampling, filtering, structural review, and experiments, not on accepting one high-scoring generated structure.

To use BoltzGen online, open the BoltzGen webserver, choose the protocol that matches the molecule you intend to make, supply a checked target, define the site or design region when you have evidence for one, run a small trial, and review the complete ranked set before expanding the campaign.

What is BoltzGen?

BoltzGen is an open-source all-atom generative model for biomolecular binder design. It unifies structure prediction and design in one model and supports controls over binding sites, target structure, secondary structure, residue identity, and covalent bonds.^[1]

The model was introduced with experimental campaigns spanning proteins, nanobodies, linear and disulfide-bonded peptides, disordered targets, and small molecules. Those results show that the system can produce experimentally active candidates, but they do not make every generated design a binder or transfer one campaign's hit rate to another target.^[1]

BoltzGen differs from a sequence generator that predicts residues without an explicit complex. It generates atom-level binder geometry in the context of the target, assigns or redesigns sequence, predicts whether that sequence refolds into the intended complex, measures the proposed interface, and selects a final quality-diverse set.

How does BoltzGen work?

The browser workflow is short because the native pipeline performs several scientific stages after submission.

1. A design specification defines what may change

The target, designed entity, fixed structure, editable residues, binding site, and optional geometric constraints become a native design specification. This distinction matters: a protein binder generated from scratch is a different task from redesigning selected residues of an uploaded complex or replacing complementarity-determining regions (CDRs) in a fixed antibody scaffold.

BoltzGen uses 1-based canonical mmCIF identifiers for residue controls. In mmCIF files, label_asym_id and label_seq_id can differ from the author chain and residue numbers shown in a paper or PDB entry. A correct biological site expressed in the wrong numbering system becomes the wrong computational constraint.^[2]

2. Diffusion generates atom-level candidates

The generative model starts from a noisy representation and iteratively denoises it into binder atoms conditioned on the target and design specification. BoltzGen represents designed residues at all-atom resolution, so residue identity and geometry are part of the generation process rather than a backbone-only afterthought.^[1]

The Diverse and Adherence checkpoints favor different parts of the search. The default Both setting divides sampling across them. Step scale and Noise scale can also shift the balance between adherence and diversity, but leaving those fields empty preserves the source schedules and is the safest first run.^[1][2]

3. Inverse folding proposes compatible sequences

BoltzGen normally applies its BoltzIF inverse-folding model after structure generation. Inverse folding asks which amino-acid sequence is compatible with a proposed backbone. Generating several sequences per backbone expands sequence sampling, but also multiplies later folding work and cost.^[1]

Skipping inverse folding is useful for a diagnostic backbone-only run, not a general default for candidates you plan to synthesize. It removes a sequence-optimization stage that the authors use to improve agreement between the designed backbone and its refolded sequence.^[1]

4. Refolding tests structural self-consistency

BoltzGen predicts each designed sequence in complex with the target using Boltz-2 and compares the refolded complex with the generated design. For globular protein protocols, it can also refold the binder without the target. Agreement does not prove experimental folding, but disagreement is a strong reason not to prioritize a candidate.^[1]

Small-molecule binder design adds the Boltz-2 affinity stage. Its affinity values are model outputs for prioritization, not experimental dissociation constants or a substitute for biochemical measurement.^[1]

5. Analysis and quality-diversity selection produce the final set

The pipeline calculates confidence, refolding, interface, sequence, and developability-related metrics. Its native ranking emphasizes the weakest weighted metric rank, then a greedy quality-diversity step selects the final budget using structural and sequence similarity.^[1]

This is why Number of designs and Budget are different. The first controls the candidate pool. The second controls how many filtered, diversity-optimized designs remain in the final set. A larger budget cannot recover candidates that were never generated.

How to use BoltzGen online

1. Choose the molecule you will test

Select the protocol before preparing the rest of the form.

Do not choose a protocol only because its default binder length looks convenient. Protocols change the scaffold, sequence policy, refolding stages, filters, and metrics.

2. Prepare the target

For structure-target protocols, upload PDB, CIF, or mmCIF, or retrieve a structure by RCSB PDB ID. Check the biological assembly, intended chains, missing residues, alternate conformations, bound ligands, cofactors, and whether the structure represents the state you want to target.

BoltzGen conditions on the supplied structure. It does not establish that an apo conformation, predicted model, or crystallographic state is the biologically relevant binding state. If a structure needs routine repair, use PDBFixer before design, but preserve cofactors or partners that define the intended site.

For Protein-small molecule, submit a chemically correct SMILES string or CCD code. Confirm stereochemistry, protonation, charge, and tautomer state before treating the ligand identity as settled.

3. Decide whether to constrain the target site

Leave Binding site empty when the purpose is exploratory surface sampling. Specify residues when structural biology, mutagenesis, competition data, or a known ligand defines the desired epitope.

A site constraint narrows the question; it does not guarantee a productive interface. Include enough residues to describe the surface rather than one isolated atom-level contact, and use Target chains to remove irrelevant chains only when you are sure they should not participate.

Before submitting mmCIF residue constraints, verify label_asym_id and label_seq_id in a structure viewer. Author numbering is a frequent source of valid-looking but incorrect design tasks.^[2]

4. Set binder size and scaffold controls

For direct peptide, protein, and protein-small molecule design, use one exact binder length for the first trial. A range increases structural diversity and the runtime estimate, but makes it harder to distinguish a numbering or target problem from a length effect.

For nanobodies, the bundled library is the simplest start. A custom VHH framework requires the framework chain and explicit CDR regions. ProteinIQ does not infer antibody numbering or CDR boundaries. Preserve CDR lengths initially; enable variable segments only when you have a defensible deletion and insertion plan.

For Fab design, All scaffolds samples the bundled scaffold collection. Choosing one scaffold makes the experimental format more controlled. The returned native chain identifiers should not be interpreted as verified heavy and light chain labels unless the scaffold metadata proves those roles.

5. Choose campaign size

Use a small trial to validate the task, then scale only after inspecting the output. The source documentation recommends tens of thousands of generated designs for serious campaigns, while ProteinIQ supports 1 to 100 candidates per hosted job.^[2][3]

The source chart reports core native stage time per design, not total ProteinIQ job duration. Hosted quotes also include setup, post-processing, calibration, and a safety margin, so use the quote shown before submission rather than deriving credits from this chart.

That hosted limit changes how you should use the webserver:

• Use 1 to 10 designs to confirm parsing, residue numbering, protocol behavior, and approximate runtime.
• Use a larger job only after the trial produces sensible structures and metrics.
• For broader exploration, run independent jobs and compare the complete candidate pools rather than assuming one batch of 100 exhausts the design space.
• Keep Budget at or below Number of designs; a smaller budget asks the native selector for a tighter quality-diverse shortlist.

6. Keep advanced controls purposeful

The native defaults are a good baseline for Quality vs diversity, Refolding RMSD threshold, Step scale, Noise scale, filtering, and checkpoint selection. Change one scientific choice at a time and record why.

Useful advanced controls include:

• Sequences per backbone when sequence sampling is the intended experiment.
• Avoid amino acids for a documented synthesis or chemistry constraint.
• Secondary structure, disulfide, staple, or fixed-sequence constraints when the molecular format requires them.
• Custom filters and Metrics weights after the native metric table shows which failure mode needs attention.
• Design insertions only for protein redesign.

Avoid adding many constraints because they are available. An overconstrained task can suppress useful diversity or encode mutually incompatible requirements.

7. Submit and inspect every retained candidate

The result viewer links each ranked complex to its sequence and native metrics. Compare the top design with lower-ranked but structurally distinct candidates. Inspect the interface, target conformation, binder fold, clashes, exposed hydrophobics, unstructured termini, and whether the candidate actually occupies the requested site.

Download the complex CIFs, binder-only refolds when present, per-design FASTA files, native CSV tables, PDF overview, and YAML provenance. The in-tool BoltzGen reference lists every returned artifact and setting.

A reproducible first exercise with KRAS G12D

ProteinIQ includes a form preset for a 12-residue peptide binder against RCSB PDB 8JJS, using 20 generated designs and a final budget of 20. The PDB entry is a 1.53 Å X-ray structure of human KRAS G12D bound to GDP and the cyclic peptide inhibitor AP10343.^[4]

Use this preset as an interface and task-definition exercise:

1. Load the Peptide binder (12 residues) example.
2. Inspect chain A and identify whether the existing peptide and heterogens belong in your intended target model.
3. Run a small trial before using all 20 candidates.
4. Compare automatic site selection with a second run constrained to a justified KRAS surface.
5. Check whether the returned candidates occupy the intended region and whether rankings are supported by several metric families.

There is no maintained public ProteinIQ result job for this preset, so this guide does not provide invented sequences, scores, screenshots, or expected winners. The exercise is reproducible at the input level; its output must be evaluated from the actual run.

How to interpret BoltzGen results

Read confidence as a panel

Binder pTM reports predicted quality for the designed component. Binder-target iPTM reports predicted interface confidence between designed residues and the target. Minimum binder-target PAE is the lowest predicted aligned error for any design-target residue pair. Higher pTM and iPTM and lower PAE are favorable directions, but none has a universal experimental threshold.^[1]

Minimum PAE is especially easy to overread. It identifies the most confident predicted contact, not the uncertainty of the entire interface. Inspect the full native table and complex, not only the best pairwise value.

Check whether the sequence refolds to the design

filter_rmsd compares the refolded complex with the generated structure. For protocols with isolated binder folding, designfolding_filter_rmsd checks the binder without its target. Lower RMSD is better within the same protocol and atom selection.^[1]

A candidate with a plausible generated interface but poor refolding agreement is internally inconsistent. A candidate that refolds only in the complex may also be less attractive when the intended product must express and remain folded on its own.

Inspect the physical interface

Hydrogen bonds, salt bridges, and delta_sasa_refolded describe aspects of the predicted interface. Larger buried surface or more contacts are not automatically better: clashes, exposed hydrophobic patches, unsatisfied polar groups, strained geometry, and the wrong epitope can outweigh a large contact area.^[1]

Preserve the native rank

Normalized rank score is the native quality_score. BoltzGen aggregates weighted metric ranks by emphasizing the weakest component, then applies quality-diversity selection. The final rank is meaningful within that candidate pool and configuration, not as an absolute score to compare across unrelated targets or campaigns.^[1]

Treat sequences as design candidates

The FASTA files are lossless projections of BoltzGen's native full-chain sequence columns. For multi-chain designs, internal asym IDs are neutral identifiers. Do not infer VH, VL, target, or other biological roles from their numeric order.

Before ordering candidates, check sequence identity and structural similarity across the shortlist. A nominal budget of 20 is not useful if most candidates collapse onto one near-duplicate family.

How to validate a BoltzGen shortlist

Validation has two separate meanings: checking the computational campaign and testing the molecular claim.

For computational review:

• Confirm the exact target state, chains, residue numbering, and constraints.
• Inspect generated and refolded structures together.
• Require support from confidence, refolding, interface, and sequence metrics rather than one score.
• Cluster candidates by sequence and structure and preserve diversity in the experimental set.
• Compare independent runs to detect unstable rankings or repeated motifs.
• Screen obvious developability risks with tools appropriate to the molecule class.

For experimental validation, express or synthesize a diverse panel and measure binding with a suitable assay such as surface plasmon resonance or biolayer interferometry. Confirm clean concentration-dependent behavior, repeat promising measurements, include negative controls and off-targets, and test the intended biological function separately from binding. The BoltzGen paper distinguishes initial screening signals from confirmed binders for exactly this reason.^[1]

Binding is only one gate. Expression, folding, aggregation, solubility, specificity, stability, manufacturability, and biological activity require their own evidence.

Common BoltzGen mistakes

Using author residue numbers for mmCIF constraints

BoltzGen expects canonical mmCIF numbering. Verify the exact chain and residue identifiers before entering a binding site, redesign mask, or custom CDR range.^[2]

Treating a small trial as a discovery-scale search

A trial answers whether the task is configured correctly. It does not sample enough candidates to estimate hit rate or exhaust the design space. The source workflow was built around generating large pools and filtering them down.^[1][2]

Ranking by one confidence number

High interface confidence can coexist with poor isolated folding, biased sequence composition, an implausible epitope, or weak experimental behavior. Use the native panel and inspect structures.

Confusing the final budget with compute

Budget selects from generated candidates. Setting a budget of 30 does not create diversity that a Number of designs value of 30 never sampled.

Overriding native schedules on the first run

An explicit Step scale, Noise scale, alpha, RMSD threshold, or metric weight changes the source policy. Preserve empty native defaults until a result gives you a specific reason to intervene.

Ignoring the 73 to 76 residue warning

The authors report a BoltzGen memorization issue for some protein targets at binder lengths 73 to 76, where diversity can collapse toward ubiquitin-like designs. Avoid treating candidates in this range as independent evidence; check sequence similarity and consider alternative lengths.^[1]

BoltzGen limitations

• Generated confidence and affinity values are predictions, not experimental measurements.
• Selectivity, off-target binding, developability, and biological function are not established by the final rank.^[1]
• The supplied target conformation may omit induced fit, alternative states, partners, or post-translational context.
• ProteinIQ campaigns are capped at 100 generated candidates per job, below the scale recommended in the source documentation.^[2][3]
• Large or heavily constrained tasks can still fail to produce enough acceptable designs.
• The preprint's prospective results are task-specific and do not provide a universal success rate for new targets.^[1]

BoltzGen alternatives and related tools

Alternatives to BoltzGen

The right comparison depends on the experimental product. A peptide, miniprotein, nanobody, Fab, and redesign campaign have different synthesis, expression, screening, and developability constraints even when all aim at the same target.

Tools to use with BoltzGen

A practical BoltzGen checklist

• The protocol matches the molecule that will be synthesized or expressed.
• The target structure, biological assembly, chains, ligands, and state are intentional.
• All mmCIF chain and residue controls use canonical identifiers.
• The first run uses a small campaign and mostly native defaults.
• Binder length or scaffold choice has an experimental rationale.
• Advanced constraints encode known requirements rather than guesses.
• The shortlist is supported by confidence, refolding, interface, and sequence evidence.
• Candidates are diverse in both sequence and structure.
• Native CSV, structure, FASTA, and provenance files are archived.
• Binding, specificity, developability, and function will be tested independently.

BoltzGen FAQs

Can I use BoltzGen without installing it?

Yes. The ProteinIQ webserver runs BoltzGen v0.3.2 on hosted GPU infrastructure and returns the ranked structures, sequences, metrics, reports, and provenance files.^[3]

Does BoltzGen need a target structure?

The hosted peptide, protein, redesign, nanobody, and Fab protocols require a protein structure. Protein-small molecule design instead accepts a ligand as SMILES or a CCD code.

How many BoltzGen designs should I generate?

Use a small trial to validate the task, then increase sampling. ProteinIQ accepts 1 to 100 candidates per job. The source authors recommend much larger pools for discovery campaigns, so multiple independent hosted jobs are more appropriate than treating one small run as exhaustive.^[2][3]

What is a good BoltzGen score?

There is no universal cutoff. Compare native ranks only within the same campaign, require agreement across confidence, refolding, interface, and sequence metrics, inspect the structures, and validate the selected molecules experimentally.

Does a high iPTM mean the design will bind?

No. A high predicted interface TM score supports the model's confidence in the proposed interface, but it does not establish affinity, specificity, expression, or activity.

Can BoltzGen design cyclic peptides?

Yes. The peptide protocol supports backbone cyclization, and direct-binder constraints also support specified disulfide and staple bonds where the design chemistry requires them.^[1][2]

Can I upload my own nanobody framework?

Yes. Select Upload my framework, provide the structure and framework chain, and specify the canonical CDR residue ranges that BoltzGen may redesign. ProteinIQ does not infer CDR boundaries or antibody numbering.

Are the FASTA files native BoltzGen outputs?

BoltzGen's native pipeline stores full-chain sequences in its metric tables rather than emitting FASTA. ProteinIQ creates one lossless FASTA projection per retained design and also returns the unchanged native CSV files.

Sources▼

1. BoltzGen: Toward Universal Binder Design bioRxiv · 2025. https://doi.org/10.1101/2025.11.20.689494
2. BoltzGen source code and documentation GitHub, HannesStark/boltzgen · August 26, 2026. https://github.com/HannesStark/boltzgen/tree/a3149cf18eeb58648d1abbb27539bd73f746cdda
3. Use BoltzGen Online ProteinIQ · August 26, 2026. https://proteiniq.io/app/boltzgen
4. 8JJS: Human K-Ras G12D (GDP-bound) in complex with cyclic peptide inhibitor AP10343 RCSB Protein Data Bank · 2023 · August 26, 2026. https://www.rcsb.org/structure/8JJS