Skip to main content

​
Design De Novo Antibodies

RFantibody, from the IPD’s Baker Lab, designs antibodies, VHHs, and scFvs starting from a target structure and a desired epitope. This is the first method demonstrated to design de novo antibodies with atomic-level accuracy, as validated by cryo-EM. If you want to design CDRs of an antibody/nanobody and you have a known target structure, we recommend trying RFantibody!

​
Experimental Validation

The authors use cryo-EM to demonstrate the folding of the generated structures into Igs at the intended binding site. The authors show a validated influenza-targeting VHH (78 nM, RMSD 1.45Å to design), along with scFvs to TcdB (72 nM, all 6 CDRs accurate) and a PHOX2B peptide-MHC complex. Affinity maturation via OrthoRep improved initial designs by ~100x.

​
Methods

The RFantibody pipeline has three steps:
  1. RFdiffusion (Structure Generation): Finetuned RFdiffusion model which generates structure without amino acids assigned to the newly designed CDR loops (you’ll notice poly-Gs in your RFantibody outputs)
  2. ProteinMPNN (Sequence Design): Design of CDR sequences with ProteinMPNN to find sequences which fold into the RFantibody generated structures
  3. RF2 (Design Validation): A fine-tuned RoseTTAFold2 to evaluate if the designed sequences fold into their initial structures
On Tamarind, the entire pipeline can be run with our RFantibody tool—just provide an input target, framework, and epitope residues.

​
Filtering Methods

The authors recommend filtering by: RF2 pAE < 10, RMSD (design vs. RF2 predicted) < 2Å, and optionally Rosetta ddG < -20. These serve as minimums, and there isn’t a strong guarantee that sequences conforming to these standards are binders. The filtering step remains the most significant bottleneck to this pipeline. AlphaFold3 shows promise as a stronger filter (ipTM > 0.6), but more validation is needed.

​
Scale

The smallest number of designs where the authors found hits was a set of 95 VHHs. However, a more realistic number of designs to test is in the 10k range, with experimental success rates of 0-2% per target. This is primarily because RF2 is not a strong filter for distinguishing binders from non-binders. For the time being, yeast display or other high-throughput screening methods remain the most reliable path to identifying de novo binders.

​
Target Site Selection

A site candidate for binding should have at least 3 hydrophobic residues. Binding to charged polar sites remains challenging, as do sites with nearby glycans (which often become ordered upon binding, costing energy). For unstructured loops, see this paper for peptide-targeting strategies.

​
Hotspot Selection

RFantibody is highly sensitive to specified epitope/hotspots—more so than vanilla RFdiffusion. Where RFdiffusion tends to generate long helices when given poor hotspots, RFantibody will typically produce undocked antibodies instead. Tips for hotspot selection:
  • Select residues where the C-beta distance to CDR residues would be <8Å
  • Include a mix of hydrophobic residues when possible
  • Start with 2-5 hotspots and adjust based on results
Always run a pilot batch of 10-20 designs before generating thousands. If most outputs show undocked antibodies, try different hotspot combinations. Once you have settings that give reasonable docks, Tamarind can support tens of thousands of in-silico designs.

​
Truncating Your Target

RFdiffusion and RF2 scale as O(N²) with residue count, so truncating large targets significantly reduces compute time. Preserve ~10Å of protein around your epitope on all sides, maintain secondary structure elements (don’t cut mid-helix), and minimize chain breaks. For multi-domain proteins, cut at flexible linkers between domains. PyMOL works well for preparing truncated structures.

​
Nanobody Docking Behavior

Nanobody outputs often show side-on binding docks—this is expected behavior, not a bug. Natural nanobodies frequently bind this way with framework-mediated contacts. If you need a more traditional end-on approach angle, use an scFv/antibody framework instead, or adjust hotspots to favor different docking geometries.

​
CDR Length Selection

You can specify length ranges for each CDR, or use “auto” to preserve the length from your input framework. Suggested ranges based on natural distributions: H1 (7-10), H2 (6-8), H3 (5-13), L1 (8-13), L2 (7), L3 (9-11). Consider longer H3 loops when targeting deep hydrophobic pockets or epitopes requiring extended reach.

​
Framework Options

Two validated frameworks are available: h-NbBCII10 (nanobody/VHH) and hu-4D5-8_Fv (scFv). You can also upload a custom Chothia-annotated PDB if you have a preferred therapeutic framework—the tool will automatically detect CDR regions, or you can manually select them. Try RFantibody