Boltz-2.1

Boltz-2.1 webserver overview

Boltz-2.1 predicts the 3D structure of complexes built from proteins, DNA, RNA and small molecules. Each job returns one or more sampled structures with confidence metrics, and can also request binding metrics for a ligand-protein or protein-protein complex.

ProteinIQ sends each job to the hosted Boltz API and publishes the finished structures, tables and records in the job's results. No MSA server or GPU settings are needed. For the self-hosted model with templates, constraints and affinity options, use Boltz-2.

Pricing

Boltz-2.1 is billed per sample, and each sample is priced by the size of the complex in tokens. A token is about one residue or nucleotide. Each SMILES ligand adds one token per heavy atom, and each CCD ligand adds 64 tokens. Copies count once per chain.

Complex size (tokens)	1 sample	5 samples	10 samples
Up to 256	51	254	507
257 to 512	102	507	1,013
513 to 1,024	203	1,013	2,026
1,025 to 2,048	406	2,026	4,051
More than 2,048	811	4,051	8,102

For example, a 300-residue protein with one aspirin ligand is 313 tokens and costs 102 credits for one sample. The total is rounded once for the whole job. Binding metrics, MSA use and the advanced sampling settings do not change the price.

The exact quote is calculated before submission. If Boltz estimates a noticeably higher cost when the job starts, the run is not started and the credits are refunded.

Inputs

A job needs at least one molecule. Inputs are combined into one complex in submission order, and ProteinIQ shows the chain ID assigned to each one.

Input	Accepted formats	Limits and behavior
Protein	FASTA text, .fasta, .fa or .txt file, or a UniProt sequence	Up to 10 inputs, files up to 10 MB, one sequence each. Single-letter amino acid codes, including X, U and O.
DNA	FASTA text, .fasta, .fa or .txt file	Up to 10 inputs, one sequence each, using A, C, G, T and N.
RNA	FASTA text, .fasta, .fa or .txt file	Up to 10 inputs, one sequence each, using A, C, G, U and N.
Ligand	One SMILES string, a .smi or .smiles file, or a PubChem compound	Up to 10 inputs, each with exactly one SMILES.
Ligand (CCD)	A Chemical Component Dictionary code such as ATP, NAD, HEM or SAH	Up to 10 inputs. Codes are 1 to 5 letters or digits.

Every input accepts 1 to 10 copies, and each copy becomes its own chain. Job name is an optional label for the saved run.

Settings

Prediction

Parameter	Type	Default	Description
Samples (num_samples)	integer	1	Structures sampled for the complex, from 1 to 10. Each sample is billed.
Binding metrics (binding)	enum	auto	auto, ligand_protein, protein_protein or none. See the binding rules below.
Binder chains (binder_chains)	string	optional	Protein chains treated as the binder, comma-separated. Shown for protein_protein; defaults to the last protein chain.

Binding metrics follow these rules:

• auto requests ligand-protein binding metrics when the complex has exactly one single-copy ligand and every other chain is a protein. Otherwise the prediction runs without them and the summary explains why.
• ligand_protein requires the same conditions and stops the job with an error when they are not met.
• protein_protein needs at least two protein chains. The binder chains must be protein chains and must leave at least one protein chain as the partner.
• none returns structures and confidence metrics only.

Advanced settings

Parameter	Type	Default	Description
Use MSA (use_msa)	boolean	true	Lets Boltz build multiple sequence alignments for protein chains. Off sends an empty MSA.
Recycling steps (recycling_steps)	integer	3	Model recycling iterations, from 1 to 10.
Sampling steps (sampling_steps)	integer	200	Diffusion sampling steps, from 50 to 500.
Step scale (step_scale)	number	1.638	Diffusion step scale, from 0.1 to 5. Lower values give more diverse samples.

Outputs

Viewer shows the predicted structures. Results lists one row per sample. Files holds every download.

Download	Contents
boltz21_sample_0.cif onwards	Predicted complex for each sample, numbered from 0.
boltz21_sample_0_pae.npz	Predicted aligned error for each sample, when Boltz returns it.
boltz21_samples.csv	The results table: sample number, best-sample flag, Boltz's confidence metrics and structure file.
boltz21_prediction.json	Boltz's full prediction record, including binding metrics when they were requested.

The predicted structures can be sent to other structure tools in workflows.

Understanding results

The results table uses Boltz's own metric names, with nested metrics flattened into columns joined by underscores. Best sample marks the sample Boltz reports as its best. Binding metrics describe the whole complex, so they appear in the run summary and the prediction record rather than per sample. Confidence metrics estimate how reliable the predicted structure is; they are not measured affinities.