← All tools
Antibody design · Protein Design

RFantibody

(Work in Progress)

Design antibody or nanobody binders against a target structure.

Runs the whole RFantibody pipeline against a target and an HLT framework: antibody-finetuned RFdiffusion docks a backbone and rebuilds the chosen CDR loops, ProteinMPNN designs their sequences, and antibody-finetuned RF2 predicts the complex for filtering. The Antibody task designs a paired heavy/light framework; Nanobody designs a VHH.

No jobs before this; it will run immediately.

Ready
The HLT-formatted scaffold whose CDR loops and dock are designed. The two bundled frameworks are the ones used in the RFantibody preprint; pick the one that matches the task. "Custom" converts the Chothia-numbered antibody PDB below into HLT format instead.
The antigen structure. Crop it to the region around the epitope: RFdiffusion and RF2 both scale as O(N^2) in residue count.
Comma-separated chain IDs to keep from the target PDB, for example A or A,B. Everything else in the file is dropped before the run.
Read only when the framework is "Custom": a Chothia-numbered antibody structure, such as one downloaded from SAbDab. It is converted to HLT format using the chain IDs below.
Chain ID of the heavy chain in the antibody PDB.
Chain ID of the light chain in the antibody PDB.
Comma-separated target residues that define the epitope, written as a chain ID then a residue number: A305,A456. These use the target PDB's own chain IDs and numbering. RFantibody is sensitive to the choice; pilot a few designs first.
Comma-separated loops: hcdr1, hcdr2, hcdr3, and in the Antibody task lcdr1, lcdr2, lcdr3. A loop left out keeps the framework's own sequence and structure through both the diffusion and the sequence-design stages.
"auto" lets the model keep the framework loop's own length, or give a fixed length (7) or a range sampled per design (5-13).
"auto" lets the model keep the framework loop's own length, or give a fixed length (7) or a range sampled per design (5-13).
"auto" lets the model keep the framework loop's own length, or give a fixed length (7) or a range sampled per design (5-13).
"auto" lets the model keep the framework loop's own length, or give a fixed length (7) or a range sampled per design (5-13).
"auto" lets the model keep the framework loop's own length, or give a fixed length (7) or a range sampled per design (5-13).
"auto" lets the model keep the framework loop's own length, or give a fixed length (7) or a range sampled per design (5-13).
1-indexed absolute residue numbers in the HLT framework (heavy chain first, then light), as 26-32 or 26,27,28. Used only when "Select CDR indices" is on; an empty box leaves that loop's existing annotation alone.
1-indexed absolute residue numbers in the HLT framework (heavy chain first, then light), as 26-32 or 26,27,28. Used only when "Select CDR indices" is on; an empty box leaves that loop's existing annotation alone.
1-indexed absolute residue numbers in the HLT framework (heavy chain first, then light), as 26-32 or 26,27,28. Used only when "Select CDR indices" is on; an empty box leaves that loop's existing annotation alone.
1-indexed absolute residue numbers in the HLT framework (heavy chain first, then light), as 26-32 or 26,27,28. Used only when "Select CDR indices" is on; an empty box leaves that loop's existing annotation alone.
1-indexed absolute residue numbers in the HLT framework (heavy chain first, then light), as 26-32 or 26,27,28. Used only when "Select CDR indices" is on; an empty box leaves that loop's existing annotation alone.
1-indexed absolute residue numbers in the HLT framework (heavy chain first, then light), as 26-32 or 26,27,28. Used only when "Select CDR indices" is on; an empty box leaves that loop's existing annotation alone.
How many docked backbones the diffusion stage generates.
ProteinMPNN sampling temperature; higher is more diverse and less confident.
How many CDR sequences ProteinMPNN designs onto each backbone.
One-letter codes ProteinMPNN may not place. X is always omitted; C is omitted by default to avoid unpaired cysteines.
Gaussian noise added to backbone coordinates before sequence design.
Size of ProteinMPNN's k-nearest-neighbour graph.
Recycling iterations in the RF2 filtering stage. RFantibody's own examples use 10; fewer is faster and less confident.
Proportion of the hotspot residues revealed to RF2 when it predicts the designed complex. Showing all of them makes the prediction less independent of the design.
Ready