Prompteina

Prompteina webserver overview

Prompteina designs all-atom protein structures from a plain English description. A prompt can specify a fold or CATH architecture, the number of helices, strands and sheets, disulfide bonds and salt bridges, size, compactness and symmetry, surface properties, a protein family such as a lysozyme, and for a binder the interface. Each design is returned as a PDB file, with the sequence and side chains built together with the backbone.

Two tasks are available, each with its own released model. Monomer designs a single chain. Protein binder designs a binder against a target structure you provide. A run takes one prompt written in the form or a table of prompts, one per row, with an optional length range and hotspot residues for each row.

The model was described by Lee et al. (2026) and builds on Proteina-Complexa. It reads the prompt through two frozen text encoders, Qwen3-Embedding-8B and Qwen3-8B. The authors report nine binders from 90 designs tested against GM2A, a protein with no earlier reported de novo binders.

Prompteina is in beta on ProteinIQ and runs the design step only. Refolding and scoring designs, the authors' agent for writing prompts, search guided by a folding model, the built-in benchmark targets by name and mmCIF targets are not offered. Other design models on ProteinIQ include Proteina-Complexa, RFD4-Proteina and BindCraft.

Pricing

Runs cost 42 credits per minute of measured runtime when the longest design has 300 residues or fewer, and 52 credits per minute above that, because longer designs need a larger GPU. For a binder, the residues of the target selection count towards the length.

Completed runs are charged in proportion to elapsed time, rounded up to a whole credit, and unused reserved credits are returned. Charged runtime is capped at 3,300 seconds, which is 2,310 credits on the smaller GPU and 2,860 on the larger. Set a spending limit before submission.

The quote shown before submission is an estimate of that charge, starting at 60 credits. It grows with the number of designs, their length and Sampling steps, and a Guidance weight other than 1 raises it. The charge follows the measured runtime.

The first table gives credits for a monomer prompt written in the form, with one fixed chain length, 400 sampling steps and a Guidance weight of 1. Columns give the chain length in residues.

Designs per prompt	100	200	300	400	500	600
1 to 8	60	64	72	102	119	139
16	81	90	105	156	189	230
50	185	217	269	424	541	684

One to eight designs of the same length cost the same. The step between 300 and 400 residues is the move to the larger GPU. With the defaults, 50 designs with lengths drawn between 100 and 500 residues, the quote is 486 credits and the run took about nine minutes in our tests.

The second table gives credits for a binder against a target selection of 115 residues, with a written prompt, 400 sampling steps and a Guidance weight of 1. Columns give the binder length in residues.

Designs per prompt	80	120	185	200	300	485
1 to 8	64	67	72	91	104	139
16	89	94	105	133	160	230

The step between 185 and 200 residues is the larger GPU again, because the target and the binder together pass 300 residues.

Sampling steps and guidance scale the quote. Eight designs of 200 residues cost 45, 52, 64 and 90 credits at 100, 200, 400 and 800 steps, and 90 instead of 64 credits at 400 steps with a Guidance weight of 2.

A prompt table runs every row, so the estimate adds the designs of all rows and uses a row's own nres, nres_min and nres_max where they are present. The prompt text, Hotspot residues and Random seed do not change the quote.

Inputs

A run takes one prompt or one prompt table, and a target structure for binder design. Prompts chooses between the first two.

Input	Used when	Accepted	Limit
Prompt	Prompts is Write one prompt	Plain text	One prompt per run
Prompt table	Prompts is Upload a prompt table	CSV (.csv), uploaded or pasted	Up to 1 MiB
Target structure	Task is Protein binder	PDB (.pdb), uploaded or fetched from RCSB	Up to 50 MiB, with protein atoms

Account upload limits may be lower. The longest design of a run, its chain plus the target selection for a binder, can have at most 600 residues, the most the GPUs used for Prompteina hold. A longer request is refused before it starts.

Prompts

The prompts shipped with the model describe the protein itself. A monomer example reads "A compact mainly alpha protein with an up-down bundle architecture formed by four helices." A binder example reads "A compact mainly alpha binder with a four-helix up-down bundle, a densely packed core and a flat hydrophobic protein-protein interface surrounded by polar contacts."

A binder prompt describes the binder and its interface and leaves the target out, because the target comes from the structure.

Prompt table

A table has a header row and one prompt per row, in the model's own format.

Column	Required	Meaning
description	Yes	The prompt for the row.
accession	No	A name that is added to the row's file names and viewer labels.
nres	No	A fixed length in residues for the row's designs.
nres_min, nres_max	No	A length range for the row. Both are needed; one alone is ignored and the length settings apply.
hotspot_residues	No	Binder only. Target residues for the row, such as A37 A39 A49 A98. When the column is present it replaces Hotspot residues for every row, and an empty cell means none for that row.

A row's nres is used first, then nres_min with nres_max, then the length settings. The number of designs is the number of rows times Designs per prompt.

The header must read description exactly. A table without that column is not refused: every prompt is read as empty text and the designs are generated without a prompt. A spreadsheet that saves "CSV UTF-8" adds a byte order mark to the first column name, so a table that starts with description loses its prompts in the same way; save plain CSV instead. A table with a header and no rows is refused.

Target structure

The binder task needs a PDB file of the target. Residues are named as the file numbers them: the chain letter and the residue number it uses. The model reads the first model of the file and ignores residues outside the selected ranges, including waters and ions. A file with a gap in its numbering needs the selection split at the gap, such as A1-50,A60-115.

Settings

Prompt

Parameter	Type	Default	Description
Task (task)	enum	monomer	monomer or binder. Each task uses its own released model.
Prompts (prompt_source)	enum	write	write takes one prompt typed in the form. table takes a prompt table.

Design

Parameter	Type	Default	Description
Length, minimum (monomer_length_min)	integer	100	Monomer. Shortest chain to sample, from 1 to 600. A table row's own length takes precedence.
Length, maximum (monomer_length_max)	integer	500	Monomer. Longest chain to sample, from 1 to 600 and at least the minimum. Each design draws its own length in the range.
Target residues (target_input)	string	Required	Binder. Chain and residue ranges as numbered in the PDB file, such as A1-115, joined by commas, such as A96-174,A306-446. Every part is a range, and every residue number in it must exist in the file.
Hotspot residues (hotspot_residues)	string	Optional	Binder. Target residues the binder should contact, written as chain and number such as A37 A39 A49 A98, separated by spaces or commas. Single residues only.
Binder length, minimum and maximum (binder_length_min, binder_length_max)	integer	Required	Binder. Shortest and longest binder, from 1 to 600, with the minimum no greater than the maximum. The target selection and the binder together can be up to 600 residues.
Designs per prompt (designs_per_prompt)	integer	50 for monomer, 1 for binder	Designs generated for each prompt, at least 1.
Guidance weight (guidance_w)	number	1	At least 0. 1 turns classifier free guidance off. The authors recommend 2 for monomer design and 1 for binder design.

Sampling

Parameter	Type	Default	Description
Sampling steps (nsteps)	integer	400	Denoising steps per design, at least 1.
Random seed (seed)	integer	5	From 0 to 4294967295.
Job name	string	Optional	A label for the run.

Outputs

Viewer displays the designed structures and Files lists every download.

Download	Contents
job_0_n_<n>_id_<k>.pdb	One designed structure, in a folder of the same name. <n> is the number of residues in the file, which for a binder includes the target, and <k> counts the designs of that length from 0. An accession from a prompt table is added to the name.
rewards_text_monomer_generate_0.csv, rewards_text_binder_generate_0.csv	One row per design: pdb_path, pdb_index, aatype, total_reward and metadata_tag.
timing_0.csv	job_id, total_time in seconds, which includes loading the model, and nsamples.
config.yaml, hydra.yaml, overrides.yaml	The resolved configuration and command line of the run.
generate.log	The model's own log.
input/prompts.csv	The table the model read. For a written prompt, the one-row table ProteinIQ wrote from it.
records/run.log	The command, the model's log and the measured resource use of the run.
records/provenance.json	Settings, command, input SHA-256 digests, resource measurements, the source commit and the model revisions.
records/NOTICE.txt, records/NVIDIA-OPEN-MODEL-LICENSE.txt	The notice of the source code and the license of the model weights.

Understanding results

A finished run returns structures and no verdict. The default monomer job returns 50 PDB files with lengths scattered between 100 and 500 residues, and nothing in them says which fold. The total_reward column of the rewards table is NaN for every row because no reward model is used, and the other files record settings, not quality. Each design is a candidate, and the checks that rank candidates come afterwards.

The usual check is a refold. Refold each designed sequence with ESMFold2 or AlphaFold2, then compare the result with the design by RMSD and read its pLDDT; a design that its own sequence does not fold back into is unlikely to hold.

The authors score a binder by refolding it and reading the interface. On ProteinIQ, predict the complex with Boltz-2 and read ipTM and ipSAE at the interface, which the ipSAE tool calculates. The authors also redesign the sequence of a backbone with ProteinMPNN before refolding.

A prompt is a request, and the output does not report whether the model followed it. Measure the property you asked for: DSSP assigns the secondary structure of a design, so its helices and strands can be counted, and the file name carries its length. If every design looks alike and ignores the prompt, check the header of a prompt table, because a column not named exactly description is read as empty text.

A monomer prompt draws a length for each design between Length, minimum and Length, maximum, so 50 designs of one prompt span the range, while all designs of one binder prompt share one length. The lengths do not depend on Random seed: a different seed gives different structures at the same lengths, and the same prompt, settings and seed return the same designs.

The number in a file name is a counter, not a rank. Design 8 is not worse than design 1, and the counter restarts for each length. The viewer numbers designs from 1 while the file name counts from 0.

A binder's n includes the target, so a file of 188 residues from a 115 residue target holds a binder of 73. Hotspots ask the model to place the binder against those residues; they are not a measured contact, so check the interface in the refolded complex.

The model's own quick start is a table of three monomer prompts with one design each at a Guidance weight of 2, quoted at 90 credits. The default job with a single written prompt is quoted at 486 credits, and setting both length fields to 100 returns 50 designs of 100 residues for 185.

A binder against a 115 residue target selection, A1-115, with Hotspot residues A37 A39 A49 A98, a binder length of 80 to 100 and 8 designs of one written prompt, is quoted at 65 credits. All eight files carry the same n, a value from 195 to 215 that sums the target and the binder.

Sources

• Lee J, Li Y, Qiang B, Lau G, Ettinger A, Smith NA, Donovan-Maiye R, Baker D, Hajishirzi H, Koh PW. "Designing Protein Structures from Natural Language." arXiv (2026). arXiv:2610.13917
• Prompteina source code, Apache-2.0. github.com/jleechung/prompteina
• Prompteina model weights, NVIDIA Open Model License. huggingface.co/jleechung/prompteina