Explore antibody CDR designs with Proteo-R1 reasoning and raw diffusion. Learn more
Input
Inputs
Upload one antibody-antigen complex in mmCIF or CIF format with the chains and residue IDs needed for the masked CDR design task.
Settings
Output
Configure inputs to begin
Set options on the left, then click “Submit job”.
What is Proteo-R1?
Proteo-R1 is a reasoning-guided antibody CDR design method for antibody-antigen complexes. It uses a language-model reasoning stage to identify CDR design decisions, then a conditional diffusion model to generate designed antibody structures and sequence artifacts.
The standard online workflow runs Proteo-R1's raw-diffusion mode without framework structure inpainting. Treat these results as exploratory candidates. The Proteo-R1 authors state that structure inpainting is required for the published-quality target; it needs a matching canonical YAML, processed MSA files, and ground-truth structure assets that cannot be derived safely from a CIF upload alone.
The online tool is intended for antibody-antigen complexes where the antigen residue positions of interest are already known. It accepts one mmCIF or CIF structure file and a set of antigen design-point tuples such as [C,4], [C,1], [C,71].
How to use Proteo-R1 online
Upload one antibody-antigen mmCIF or CIF file, enter the antigen residue tuples that should guide CDR redesign, provide the chain-mapping record ID, and run the job. Proteo-R1 returns ranked designed PDB structures, designed sequence tables, CDR prediction JSON, confidence artifacts, provenance, and intermediate files used during the reasoning and generation stages.
Inputs
| Input | Description |
|---|---|
Antibody-antigen structure | One .cif or .mmcif file containing the antibody-antigen complex. Chain IDs and residue IDs must match the design-point tuples. |
Antigen design points | Comma-separated chain/residue tuples, for example [C,4], [C,1], [C,71]. These positions guide the CDR reasoning step toward the antigen surface of interest. |
Record ID | Required chain mapping in entry_heavy_light_antigen form. For example, 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. |
PDB input is not accepted for this workflow. Proteo-R1's CDR preparation step expects mmCIF/CIF structure metadata, so the tool requires a CIF-family upload instead of converting PDB files automatically.
Settings
| Setting | Description |
|---|---|
Design samples | Number of diffusion samples to generate. More samples broaden the search but increase runtime. Billing scales linearly at 250 credits per sample. |
Random seed | Seed used by the reasoning and generation stages. The default is 2025. |
Maximum reasoning tokens | Token budget for the CDR reasoning step. The default is 2048. Very small values can truncate the generated CDR JSON. |
Reasoning temperature, Reasoning top-p, Reasoning top-k | Sampling controls for the reasoning model. Defaults follow the Proteo-R1 command-line settings. |
Recycling steps | Structure refinement recycling steps. The default is 3. |
Sampling steps | Diffusion sampling steps passed directly to Proteo-R1. The source default is 200. |
Noise type | Diffusion noise mode. The default is discrete_absorb. |
Return full PAE matrix | Returns full predicted aligned error NPZ artifacts when produced. |
Return full PDE matrix | Returns full predicted distance error NPZ artifacts when produced. |
Skip CDR prediction JSON | Disables the CDR prediction JSON artifact. Leave off for normal runs. |
Outputs
| Output | Description |
|---|---|
| Ranked PDB structures | Designed antibody-antigen structures named like record_model_0.pdb. |
| Sequence table | .seq table with designed sequence and CDR amino acid recovery fields when available. |
| CDR prediction JSON | Chain-separated CDR predictions, CDR regions, sequence metrics, and per-CDR amino acid recovery where available. |
| Confidence JSON | Confidence metrics such as confidence_score, ptm, iptm, complex_plddt, complex_pde, and chain-pair metrics. Some values may be blank when the model does not return finite scores for a run. |
| pLDDT NPZ | Per-residue pLDDT arrays for generated structures. |
| Optional PAE/PDE NPZ | Full error matrices when requested. |
| Preparation artifacts | YAML, masked JSON, x-mask CIF, hidden-state safetensors, and Protenix feature dump files used by the run. |
| Run provenance | Exact source, submodule, model revisions, runtime assets, sampling settings, and whether structure inpainting was used. |
How Proteo-R1 works
Proteo-R1 runs in two main stages.
First, the CDR preparation and reasoning stage identifies the CDR design task from the uploaded complex and design-point tuples. It creates the masked design representation, emits CDR hidden states, and records the intermediate files used by the generation model.
Second, the generation stage runs conditional diffusion to produce designed structures and sequences. Results are ranked by the confidence score when finite confidence values are returned. The output files preserve the model's native structure, sequence, CDR, confidence, and error artifacts so they can be inspected or downloaded together.
Advanced structure inpainting assets
The standard online workflow requires a CIF-family structure file, design-point tuples, and an explicit chain-mapping record ID. It runs raw diffusion. Proteo-R1's published-quality structure-inpainting mode depends on a prepared asset bundle rather than a few independent files.
Full inpainting needs matching assets with the same record IDs and the directory layout expected by Proteo-R1:
- Spec-mask YAML: A canonical YAML file describing the design case and masked residues.
- Processed MSA NPZ files: Precomputed MSA assets in Proteo-R1's processed format.
- Ground-truth structure NPZ files: Structure-format NPZ files prepared for the same records.
These assets are folder-shaped because the YAML, MSA files, and structure files must agree on record IDs and paths. Separate uploaded files can only work if the application reconstructs that directory layout before running Proteo-R1.
The current online tool does not claim the published inpainting quality target. Advanced inpainting asset upload can be added later as a separate expert workflow after validating it against a real prepared Proteo-R1 asset bundle.
Interpreting results
Proteo-R1 outputs should be treated as candidate designs for downstream triage. The ranked PDB files are the primary structures to inspect, while the .seq and CDR JSON files make it easier to compare how the model changed CDR regions.
Useful follow-up checks include:
- Structure inspection: Review the antibody-antigen interface for clashes, buried unsatisfied groups, or unrealistic loop placement.
- Sequence review: Compare redesigned CDR sequences against the starting antibody and flag unusual motifs before synthesis.
- Confidence artifacts: Use pLDDT and confidence JSON as screening signals, not experimental validation.
- Downstream scoring: Consider additional structure validation, binding assessment, and developability checks before selecting designs for experiments.
Related tools
RFantibody
Structure-based de novo antibody and nanobody design pipeline combining antibody-tuned RFdiffusion, ProteinMPNN sequence design, and antibody-tuned RoseTTAFold2 filtering.
DiffAb
AI-powered antibody CDR design using equivariant diffusion models. Generates complementarity-determining region (CDR) sequences and structures for antibody structures and antibody-antigen complexes. Supports single- and multi-CDR co-design, antibody optimization, fixed-backbone sequence design, and structure prediction.
Genie 3
Generate protein structures and scaffolds with Genie 3, an all-atom SE(3)-equivariant diffusion model. Genie 3 supports unconditional protein generation, motif scaffolding, and hotspot-targeted binder design.
BioPhi
Antibody humanization and humanness evaluation platform from Merck. Sapiens mode uses deep learning trained on the Observed Antibody Space (OAS) to humanize antibody sequences, while OASis mode evaluates humanness using 9-mer peptide search against human antibody databases.
BoltzGen
BoltzGen uses generative diffusion models to design protein, peptide, nanobody, and Fab binders against protein and small-molecule targets.
CoSiNE
Simulate antibody affinity maturation, steer heavy-chain evolution with SARS-CoV binding predictors, constrain mutations to CDR or framework regions, and score paired-antibody variants with CoSiNE likelihoods and optional somatic-hypermutation correction.
EvoDiff
EvoDiff is a diffusion-based protein sequence generation framework from Microsoft Research. ProteinIQ currently runs the EvoDiff-Seq OA_DM_38M model for unconditional protein generation, motif scaffolding, and user-sequence inpainting.
ODesign
All-atom generative AI for designing protein binders. Specify target binding sites and generate diverse binding proteins with fine-grained control over interaction parameters.
PocketFlow
PocketFlow is a structure-based molecular generative model that designs novel drug-like molecules within protein binding pockets. It uses autoregressive flow modeling with chemical knowledge to generate 100% chemically valid, highly drug-like compounds.
PocketXMol
PocketXMol is a pocket-interacting generative foundation model for small-molecule or peptide docking and design in protein binding pockets.