Antibody loop redesign with regular RFD3¶
This public example remodels one CDR-H3 loop in an existing antibody Fv–CD28 complex with the regular public RFD3 checkpoint. It then designs the loop sequence with ProteinMPNN and refolds the complex with AF3. It does not require the private antibody modelhub checkpoint or abMPNN-PoE used by the original example.
Inputs and scope¶
The bundled structure is cropped from PDB 8S6Z, as distributed with Foundry's binder-design tutorial. Chain A is the antibody heavy variable domain, B the light variable domain, and C CD28. Constant domains, glycans, waters, and other nonprotein components are omitted. This is a simplified structural example; glycan-dependent binding is outside its scope.
The source author residue labels include insertion codes. The bundled PDB uses
consecutive residue numbers within each chain; files/residue_mapping.json
records every source label. The eight-residue H3 segment corresponding to source
heavy-chain residues 95–102 is regenerated. All remaining coordinates and sequences
are supplied as fixed context. These labels describe this fixture, not a general
antibody numbering algorithm.
RFD3: fixed Fv/CD28 context + eight-residue H3 gap
→ ProteinMPNN: heavy chain only, fixed motif sequences preserved
→ AF3: one complex prediction
→ backbone RMSD and prediction confidence
RFD3 is a general model. Successful execution does not show equivalence to the antibody-tuned model, native binding, developability, or preservation of affinity. The script reports metrics without inventing an antibody acceptance threshold.
Run¶
Install Chaperone and configure the foundry, ligandmpnn, and alphafold3
runtimes and their model weights. AF3 weights require authorized access.
From the checkout root:
Add --local inside an existing GPU allocation. The default is already one
backbone, one sequence and one AF3 sample. Outputs and lineage are stored in the
run directory; results.csv contains the final metrics.
For another antibody, prepare its structure and RFD3 contig explicitly. Preserve framework and antigen residues as fixed motifs, and update chain selections if needed. Replacing only the structure path does not update the fixture's loop numbering.
Validation¶
Regression tests check the input crop, contiguous numbering, exact loop gap, fixed-framework sequence policy, packed output and smoke configuration.
A one-design test completed on a CoreWeave NVIDIA H200: regular RFD3 (200 steps, seed 42), ProteinMPNN with one side-chain pack, AF3 with one sample, and backbone RMSD. The packed structure contained 2,684 finite-coordinate atoms; all three chain sequences matched the sampled FASTA, and framework, light-chain and antigen sequences were unchanged. AF3 reported mean pLDDT 83.5 and complex backbone RMSD 6.23 Å. This is execution evidence, not evidence of a good antibody design.
The native packer moved some fixed atoms despite repack_everything=0.
After Chaperone's coordinate-preservation correction, all 2,628 observed heavy
atoms across the 340 fixed residues exactly matched the staged input PDB.
Sampled and packed sequences still matched. The sequence-only AF3 input was
unchanged, so its prediction was reused.
The test used the native Foundry AtomWorks 2 environment at revision 0999dce,
the public LigandMPNN image, and an existing AF3 image with Chaperone's current
public runner and licensed weights. It did not rebuild every release container.
The full example path executed without scientific gate bypasses; this reporting
example does not define an antibody success threshold.