Proteo-R1

Proteo-R1 webserver overview

Proteo-R1 0.1.0 designs antibody complementarity-determining regions (CDRs) in antibody-antigen complexes. CIF mode prepares the complex and can generate CDR reasoning before diffusion. Prepared YAML mode runs an existing design file with optional CDR conditioning, framework structure inpainting or structure-only prediction. Results include designed PDB structures, sequences, confidence artifacts and supporting files.

The default workflow uses raw diffusion without framework structure inpainting. These designs are exploratory candidates. The authors' published-quality target requires inpainting with a matching canonical design YAML and reference structure assets; selecting inpainting alone does not establish experimental quality or reproduce a benchmark.

Pricing

Each Proteo-R1 run costs 20 credits flat, for any supported sample count. The quote is shown before submission.

Design samples	Credits per job
1	20
4	20
16	20

Each job processes one design record and at most one CDR conditioning sample. The 20-credit price covers the complete run, not each generated design. Sample count, input mode, inpainting, reasoning controls, recycling steps, sampling steps and optional matrices do not change the price, although they can affect runtime and output size.

Inputs

Each input accepts one file, up to 25 MiB. Prepared YAML mode requires no CIF upload.

Input	Formats	Requirements
Antibody-antigen structure	.cif, .mmcif	Required in CIF mode; protein atoms, chain annotations and residue IDs must support the CDR design task.
Prepared design (YAML)	.yaml, .yml	Required in Prepared YAML mode and for inpainting; also accepted in CIF mode to supply the design used after preparation.
Processed MSA assets (optional)	.npz, .zip	Required when the YAML uses processed MSAs; filenames must match the record, such as 8q7o_A__C_0.npz.
Reference structure NPZ (for inpainting)	.npz	Required for inpainting, with the same token and chain order as the canonical spec-mask YAML; rejected when inpainting is off.
Prepared CDR hidden states (optional)	.zip	Exactly one record_SAMPLE_N.safetensors file matching the record ID; CIF mode requires Generate CDR reasoning to be off.

ZIPs must contain unique, unencrypted files directly at their root, without folders or links, and expand to at most 256 MiB. An MSA ZIP accepts 1 to 4,096 matching NPZ files. A CDR-state ZIP accepts exactly one conditioning file.

Matching prepared assets

Prepared YAML retains the supplied sequences, masks and constraints. The record ID defaults to the input filename stem: the CIF filename in CIF mode or the YAML filename in Prepared YAML mode. MSA and CDR-state filenames must use that ID. The reference NPZ is stored under the selected record ID without changing its contents.

For CIF chain mapping, entry_heavy_light_antigen identifies the chains. 8q7o_A__C selects heavy chain A, no light chain and antigen chain C; complex_H_L_AB selects heavy H, light L and antigen chains A and B. Without this naming convention, preparation uses its H/L defaults.

Framework inpainting requires both a canonical spec-mask YAML and its matching reference structure NPZ. Processed MSA assets are required only when the YAML uses them; a native msa: empty design can run without an MSA upload, including during inpainting.

Generated or supplied CDR hidden states automatically enable conditioning during design. Without either, generation is unconditioned. The pinned checkpoint cannot perform direct text conditioning without hidden states.

Settings

Starting point

Parameter	Type	Default	Description
Starting point	enum	cif	Antibody-antigen CIF (cif) prepares a complex; Prepared YAML (prepared) runs a supplied native design.
Antigen design points	string	conditional	Required in CIF mode; comma-separated chain/residue tuples such as [C,4], [C,1], [C,71], with positive integer residue IDs matching the complex.
Generate CDR reasoning	boolean	true	CIF mode only; generates CDR hidden states, and must be off when supplying prepared hidden states.

Generation options

Parameter	Type	Default	Description
Framework structure inpainting	boolean	false	Requires the matching canonical spec-mask YAML and reference structure NPZ.
Predict structure only	boolean	false	Preserves the supplied sequence while predicting its structure.
Use epitope conditioning	boolean	true	Enables the model's epitope guidance.

Job settings

Parameter	Type	Default	Description
Record ID	string	optional	Defaults to the input filename stem; an override must start with a letter or digit and contain only letters, digits, dots, underscores or hyphens.
Random seed	integer	2025	Range: 0 to 2,147,483,647; applies to reasoning and generation, but repeated CIF-mode runs can still differ.
Design samples	integer	1	Range: 1 to 16 diffusion samples; the complete run costs 20 credits regardless of sample count.

Reasoning settings

These controls apply only in CIF mode with Generate CDR reasoning enabled.

Parameter	Type	Default	Description
Maximum reasoning tokens	integer	2048	Range: 1 to 8,192; very small budgets can truncate generated CDR JSON.
Reasoning temperature	number	0.7	Range: 0 to 2; controls reasoning-model sampling temperature.
Reasoning top-p	number	0.8	Range: 0 to 1; controls nucleus sampling.
Reasoning top-k	integer	20	Range: 0 to 200; controls the token-candidate sampling limit.

Diffusion settings

Parameter	Type	Default	Description
Recycling steps	integer	3	Range: 1 to 20; number of structure-refinement recycling steps.
Sampling steps	integer	200	Range: 1 to 1,000; discrete sequence diffusion requires exactly 200, while continuous noise and structure-only prediction allow other counts.
Step scale	number	1.638	Range: 0 to 10; diffusion step-size scale.
Gamma 0	number	0.605	Range: 0 to 10; stochasticity parameter in the diffusion schedule.
Gamma minimum	number	1.107	Range: 0 to 10; threshold for the schedule's stochasticity term.
Noise scale	number	0.901	Range: 0 to 10; injected-noise scale.
Rho	number	8	Range: 0 to 20; noise-schedule parameter.
Generation temperature	number	1	Range: 0 to 5; sequence-sampling temperature.
Noise type	enum	discrete_absorb	Choices: discrete_absorb, discrete_uniform or continuous.

Returned artifacts

Parameter	Type	Default	Description
Return full PAE matrix	boolean	false	Includes full predicted aligned error NPZ artifacts when produced.
Return full PDE matrix	boolean	false	Includes full predicted distance error NPZ artifacts when produced.
Skip CDR prediction JSON	boolean	false	Suppresses the CDR prediction JSON; leaving it off retains that output.

Outputs

The results include an ordered structure collection, a summary table and downloadable supporting artifacts. Files are retained when produced by the selected mode, with their original scientific values and ordering.

Output	Contents
Ranked PDB structures	Designed complexes named like record_model_0.pdb, using native zero-based ranks.
Sequence table	Native .seq tab-separated file with Rank, Sequence and amino acid recovery fields when available; presented as TSV.
CDR prediction JSON	Chain-separated CDR sequences, regions, sequence metrics and per-CDR amino acid recovery when available.
Confidence JSON	confidence_score, ptm, iptm, complex_plddt, complex_pde and chain-pair metrics when produced.
pLDDT NPZ	Per-residue local-confidence arrays.
Optional PAE/PDE NPZ	Full predicted aligned error or predicted distance error matrices when requested.
Preparation and model artifacts	Generated YAML, masked JSON, extracted/x-mask CIF, processed structure/MSA files, CDR hidden states and Protenix feature dumps when produced.
Stage logs	Complete standard output and error for preparation and design, including failed stages when available.
Run provenance	Source and model revisions, runtime assets, sampling settings and whether inpainting was used.

Understanding results

confidence_score is the model's ranking score when finite values are available. ptm describes predicted overall structural confidence, iptm describes confidence in interchain arrangement, and pLDDT describes local structural confidence. PAE and PDE describe predicted alignment and distance errors. Amino acid recovery fields compare designed residues with the reference sequence; they do not measure binding affinity.

Unavailable confidence values appear blank in the summary, with a warning when no usable confidence scores are returned. The original JSON and NPZ files are retained. A rank or a valid PDB file alone does not establish design quality.

These outputs are candidate designs for downstream assessment. Structure review can identify clashes and unrealistic loops; sequence review can flag unusual motifs. Confidence metrics, binding assessment and developability checks support triage, but do not replace experimental validation.

Native limitations

• CIF preparation requires .cif or .mmcif metadata; PDB input is not accepted or converted automatically.
• The pinned source cannot preprocess FASTA, convert raw CSV MSAs or generate MSAs automatically. Prepared designs require native YAML and processed MSA NPZ assets, or intentional use of msa: empty.
• PDB is the supported structure-output format; native model and confidence NPZ artifacts are retained alongside it.
• Direct text conditioning without hidden states is unavailable, and a fixed seed does not guarantee identical CIF-mode results.