RFantibody is a structure-guided antibody and nanobody design pipeline built around an antibody-tuned version of RFdiffusion. It generates new CDR-mediated binding geometries against a chosen epitope, designs compatible loop sequences, and then filters the resulting complexes with an antibody-tuned RoseTTAFold2 model.
The method was introduced as a route to de novo, epitope-specific antibody design from structural input rather than immunization or random library discovery. In the published workflow, the computational design stage is followed by experimental screening, because the model can generate structurally accurate binders but still requires broad sampling to recover the strongest candidates.
ProteinIQ runs RFantibody on hosted GPU infrastructure, so the full three-stage pipeline can be launched from the browser without installing RosettaCommons dependencies locally.
The online interface accepts one target structure and one antibody or nanobody framework. The target can be uploaded as a PDB/ENT file or fetched from the RCSB. The framework must already be prepared in RFantibody's HLT format.
| Input | Description |
|---|---|
Target antigen structure | Antigen PDB or ENT structure, uploaded directly or fetched from the RCSB by PDB ID. Large targets should usually be truncated around the intended epitope because runtime scales poorly with target size. |
HLT antibody framework | Antibody or nanobody framework PDB in RFantibody HLT format. This is not a generic antibody structure file; it must use RFantibody chain conventions and CDR loop remarks. |
| Requirement | Description |
|---|---|
| Heavy chain | Must use chain ID H. For nanobodies, the single variable domain is treated as the heavy chain framework. |
| Light chain | If present, must use chain ID L. Nanobody frameworks omit this chain. |
| Chain order | Protein chains should appear in H then L order. In full HLT complexes, target chains are represented as T. |
| Loop remarks | The file must end with REMARK PDBinfo-LABEL annotations that mark the absolute residue indices of the CDR loops, for example REMARK PDBinfo-LABEL: 32 H1. |
| Accepted examples | native example frameworks include the nanobody file h-NbBCII10.pdb and the antibody Fv file hu-4D5-8_Fv.pdb. |
Generic antibody PDB files usually do not include those loop remarks or chain conventions. In the native RFantibody workflow, such structures are converted to HLT format before design.
| Setting | Description |
|---|---|
Backbone designs | Number of backbone dock designs to generate in the RFdiffusion stage (1-10, default 10). |
Sequences per backbone | Number of ProteinMPNN sequence variants to sample for each designed backbone (1-8, default 1). |
Sequence temperature | ProteinMPNN sampling temperature (0.1-1.0, default 0.1). Lower values bias toward more conservative loop sequences. |
RF2 recycles | Number of RoseTTAFold2 recycling iterations used during filtering (1-20, default 10). |
Deterministic mode | Requests deterministic sampling where supported, useful for reproducible pilot runs. |
| Setting | Description |
|---|---|
Design H1/H2/H3 | Toggles for heavy-chain CDR loop design. Enabled by default, matching H1:,H2:,H3: behavior. |
H1/H2/H3 length | Optional heavy-chain CDR length or length range, such as 7 or 5-8. Leave blank to design that loop at the native framework length. |
Design L1/L2/L3 | Toggles for light-chain CDR loop design. Enabled by default; native skips these loops for nanobody frameworks without chain L. |
L1/L2/L3 length | Optional light-chain CDR length or length range for antibody frameworks. Leave blank to design the loop at the native framework length. |
At least one loop must be enabled. To keep a loop fixed, turn off its design toggle. A blank length on an enabled loop does not skip the loop; it uses the native length syntax such as H1:.
| Setting | Description |
|---|---|
Hotspot residues | Comma-separated antigen residues such as B146,B170,B177. These residues bias the model toward a particular epitope. Bad hotspot definitions can produce undocked designs, so small pilot runs are usually preferable before scaling up. |
| Setting | Description |
|---|---|
Diffusion timesteps | RFdiffusion denoising steps passed as --diffuser-t (default 50). |
Final diffusion step | Final RFdiffusion step passed as --final-step (default 1). |
Return RFdiffusion trajectory | Returns native RFdiffusion trajectory quiver files. Enabled by default, matching native trajectory output. |
Omitted amino acids | ProteinMPNN --omit-aas value (default CX). |
ProteinMPNN noise | Optional ProteinMPNN --augment-eps coordinate noise. Leave blank to use the command-line default. |
RF2 hotspot visibility | Fraction of hotspot residues shown to RF2 during filtering (default 0.1). |
RF2 random seed | Optional RF2 seed. Leave blank for stochastic RF2, or enable deterministic mode to use seed 42. |
| Output | Description |
|---|---|
| Designed complexes | Final antibody-antigen or nanobody-antigen complexes in PDB format. |
| RFdiffusion score file | .sc targeting table returned for hotspot-guided runs. mindist is the minimum hotspot-to-designed-loop Cβ distance, and averagemin is the mean nearest distance across hotspots; both are reported in Å. |
| Final score file | .sc score table exported from the final RF2 quiver output. It includes interaction PAE, mean PAE, mean pLDDT, and Cα RMSD metrics in RFantibody's order. |
| Stage quiver files | RFdiffusion, ProteinMPNN, and RF2 quiver files for downstream inspection. |
| RFdiffusion trajectory | Xt-1 and pX0 trajectory quiver files when trajectory output is on. |
| File table | Downloadable list of all generated design files. |
RFantibody runs three design stages in sequence.
The first stage places a chosen framework against the target and redesigns the selected CDR loops. Rather than generating an arbitrary binder scaffold from scratch, this stage treats the framework as fixed context and focuses the generative step on loop-mediated docking and backbone creation around the target epitope.
Once loop backbones have been generated, ProteinMPNN samples amino acid sequences compatible with each designed structure. In RFantibody, this step is used to populate the redesigned loops while preserving the broader framework context.
The final stage predicts the structures of the designed sequences in complex with the target and scores whether the intended interface appears self-consistent. This filtering stage is important because antibody design campaigns generally require broad sampling, and only a subset of computational designs retain the desired dock after structure prediction.
RFantibody output is best treated as a candidate set rather than a final answer. Structural inspection still matters: plausible designs should maintain a well-packed interface, avoid obvious clashes, and keep the framework geometry reasonable outside the redesigned loops.
RFantibody's guidance recommends filtering designs with strong RF2 confidence and low disagreement between the design model and the RF2-predicted complex. In practice, RF2 pAE below roughly 10 Å and design-versus-prediction Cα RMSD below roughly 2 Å are used as minimal screening criteria before downstream experimental work or more detailed physics-based evaluation. For hotspot-guided runs, smaller mindist and averagemin values indicate that the generated CDR geometry landed closer to the requested epitope. Candidate numbers record generation order; they are not a score-based ranking.
Exploratory antibody CDR co-design for antibody-antigen complexes using Proteo-R1 reasoning and raw diffusion. The standard online workflow does not include the framework structure-inpainting assets required for the published-quality target.
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 uses generative diffusion models to design protein, peptide, nanobody, and Fab binders against protein and small-molecule targets.
AI-powered antibody CDR design using equivariant diffusion models. Generates optimized complementarity-determining region (CDR) sequences and structures for antibodies targeting specific antigens. Supports single CDR, multi-CDR co-design, and fixed-backbone sequence design modes.
PepMimic designs short peptides that mimic the binding interface of a known protein binder on its target. From a reference protein complex, a latent diffusion model generates peptide candidates constrained to the target interface, and each candidate is scored by interface-mimicry against the reference binder.
Design protein binders against a target structure with NVIDIA BioNeMo's Proteina-Complexa generative pipeline.
Design linear peptide binders for target proteins using a target sequence-conditioned masked language model. PepMLM generates peptide sequences optimized to bind specific protein targets based on ESM-2 protein language modeling.
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.
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.
All-atom generative AI for designing protein binders. Specify target binding sites and generate diverse binding proteins with fine-grained control over interaction parameters.
RFantibody is a structure-guided antibody and nanobody design pipeline built around an antibody-tuned version of RFdiffusion. It generates new CDR-mediated binding geometries against a chosen epitope, designs compatible loop sequences, and then filters the resulting complexes with an antibody-tuned RoseTTAFold2 model.
The method was introduced as a route to de novo, epitope-specific antibody design from structural input rather than immunization or random library discovery. In the published workflow, the computational design stage is followed by experimental screening, because the model can generate structurally accurate binders but still requires broad sampling to recover the strongest candidates.
ProteinIQ runs RFantibody on hosted GPU infrastructure, so the full three-stage pipeline can be launched from the browser without installing RosettaCommons dependencies locally.
The online interface accepts one target structure and one antibody or nanobody framework. The target can be uploaded as a PDB/ENT file or fetched from the RCSB. The framework must already be prepared in RFantibody's HLT format.
| Input | Description |
|---|---|
Target antigen structure | Antigen PDB or ENT structure, uploaded directly or fetched from the RCSB by PDB ID. Large targets should usually be truncated around the intended epitope because runtime scales poorly with target size. |
HLT antibody framework | Antibody or nanobody framework PDB in RFantibody HLT format. This is not a generic antibody structure file; it must use RFantibody chain conventions and CDR loop remarks. |
| Requirement | Description |
|---|---|
| Heavy chain | Must use chain ID H. For nanobodies, the single variable domain is treated as the heavy chain framework. |
| Light chain | If present, must use chain ID L. Nanobody frameworks omit this chain. |
| Chain order | Protein chains should appear in H then L order. In full HLT complexes, target chains are represented as T. |
| Loop remarks | The file must end with REMARK PDBinfo-LABEL annotations that mark the absolute residue indices of the CDR loops, for example REMARK PDBinfo-LABEL: 32 H1. |
| Accepted examples | native example frameworks include the nanobody file h-NbBCII10.pdb and the antibody Fv file hu-4D5-8_Fv.pdb. |
Generic antibody PDB files usually do not include those loop remarks or chain conventions. In the native RFantibody workflow, such structures are converted to HLT format before design.
| Setting | Description |
|---|---|
Backbone designs | Number of backbone dock designs to generate in the RFdiffusion stage (1-10, default 10). |
Sequences per backbone | Number of ProteinMPNN sequence variants to sample for each designed backbone (1-8, default 1). |
Sequence temperature | ProteinMPNN sampling temperature (0.1-1.0, default 0.1). Lower values bias toward more conservative loop sequences. |
RF2 recycles | Number of RoseTTAFold2 recycling iterations used during filtering (1-20, default 10). |
Deterministic mode | Requests deterministic sampling where supported, useful for reproducible pilot runs. |
| Setting | Description |
|---|---|
Design H1/H2/H3 | Toggles for heavy-chain CDR loop design. Enabled by default, matching H1:,H2:,H3: behavior. |
H1/H2/H3 length | Optional heavy-chain CDR length or length range, such as 7 or 5-8. Leave blank to design that loop at the native framework length. |
Design L1/L2/L3 | Toggles for light-chain CDR loop design. Enabled by default; native skips these loops for nanobody frameworks without chain L. |
L1/L2/L3 length | Optional light-chain CDR length or length range for antibody frameworks. Leave blank to design the loop at the native framework length. |
At least one loop must be enabled. To keep a loop fixed, turn off its design toggle. A blank length on an enabled loop does not skip the loop; it uses the native length syntax such as H1:.
| Setting | Description |
|---|---|
Hotspot residues | Comma-separated antigen residues such as B146,B170,B177. These residues bias the model toward a particular epitope. Bad hotspot definitions can produce undocked designs, so small pilot runs are usually preferable before scaling up. |
| Setting | Description |
|---|---|
Diffusion timesteps | RFdiffusion denoising steps passed as --diffuser-t (default 50). |
Final diffusion step | Final RFdiffusion step passed as --final-step (default 1). |
Return RFdiffusion trajectory | Returns native RFdiffusion trajectory quiver files. Enabled by default, matching native trajectory output. |
Omitted amino acids | ProteinMPNN --omit-aas value (default CX). |
ProteinMPNN noise | Optional ProteinMPNN --augment-eps coordinate noise. Leave blank to use the command-line default. |
RF2 hotspot visibility | Fraction of hotspot residues shown to RF2 during filtering (default 0.1). |
RF2 random seed | Optional RF2 seed. Leave blank for stochastic RF2, or enable deterministic mode to use seed 42. |
| Output | Description |
|---|---|
| Designed complexes | Final antibody-antigen or nanobody-antigen complexes in PDB format. |
| RFdiffusion score file | .sc targeting table returned for hotspot-guided runs. mindist is the minimum hotspot-to-designed-loop Cβ distance, and averagemin is the mean nearest distance across hotspots; both are reported in Å. |
| Final score file | .sc score table exported from the final RF2 quiver output. It includes interaction PAE, mean PAE, mean pLDDT, and Cα RMSD metrics in RFantibody's order. |
| Stage quiver files | RFdiffusion, ProteinMPNN, and RF2 quiver files for downstream inspection. |
| RFdiffusion trajectory | Xt-1 and pX0 trajectory quiver files when trajectory output is on. |
| File table | Downloadable list of all generated design files. |
RFantibody runs three design stages in sequence.
The first stage places a chosen framework against the target and redesigns the selected CDR loops. Rather than generating an arbitrary binder scaffold from scratch, this stage treats the framework as fixed context and focuses the generative step on loop-mediated docking and backbone creation around the target epitope.
Once loop backbones have been generated, ProteinMPNN samples amino acid sequences compatible with each designed structure. In RFantibody, this step is used to populate the redesigned loops while preserving the broader framework context.
The final stage predicts the structures of the designed sequences in complex with the target and scores whether the intended interface appears self-consistent. This filtering stage is important because antibody design campaigns generally require broad sampling, and only a subset of computational designs retain the desired dock after structure prediction.
RFantibody output is best treated as a candidate set rather than a final answer. Structural inspection still matters: plausible designs should maintain a well-packed interface, avoid obvious clashes, and keep the framework geometry reasonable outside the redesigned loops.
RFantibody's guidance recommends filtering designs with strong RF2 confidence and low disagreement between the design model and the RF2-predicted complex. In practice, RF2 pAE below roughly 10 Å and design-versus-prediction Cα RMSD below roughly 2 Å are used as minimal screening criteria before downstream experimental work or more detailed physics-based evaluation. For hotspot-guided runs, smaller mindist and averagemin values indicate that the generated CDR geometry landed closer to the requested epitope. Candidate numbers record generation order; they are not a score-based ranking.
Exploratory antibody CDR co-design for antibody-antigen complexes using Proteo-R1 reasoning and raw diffusion. The standard online workflow does not include the framework structure-inpainting assets required for the published-quality target.
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 uses generative diffusion models to design protein, peptide, nanobody, and Fab binders against protein and small-molecule targets.
AI-powered antibody CDR design using equivariant diffusion models. Generates optimized complementarity-determining region (CDR) sequences and structures for antibodies targeting specific antigens. Supports single CDR, multi-CDR co-design, and fixed-backbone sequence design modes.
PepMimic designs short peptides that mimic the binding interface of a known protein binder on its target. From a reference protein complex, a latent diffusion model generates peptide candidates constrained to the target interface, and each candidate is scored by interface-mimicry against the reference binder.
Design protein binders against a target structure with NVIDIA BioNeMo's Proteina-Complexa generative pipeline.
Design linear peptide binders for target proteins using a target sequence-conditioned masked language model. PepMLM generates peptide sequences optimized to bind specific protein targets based on ESM-2 protein language modeling.
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.
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.
All-atom generative AI for designing protein binders. Specify target binding sites and generate diverse binding proteins with fine-grained control over interaction parameters.
