Skip to content

C3 nanoparticle scaffold design

A bounded RFdiffusion v1 → I3 RPXdock → tied ProteinMPNN → AF3 workflow. It generates a C3 scaffold around a fixed antigen stub, docks its asymmetric unit into an icosahedral cage, and refolds the selected trimer.

Inputs and dependencies

  • files/input_pdbs/dn5_stump.pdb: fixed motif, residues 1–7 of chains A/B/C.
  • files/input_pdbs/dn5b_stalk.pdb: antigen stalk in the same reference frame.
  • RFdiffusion v1 weights: configure RFDIFFUSION_WEIGHTS_DIR.
  • Separately obtained AF3 weights: configure ALPHAFOLD3_WEIGHTS_DIR.
  • Apptainer or Docker and a supported GPU for RFdiffusion/AF3; RPXdock runs on CPU.

From the repository root, in the installed Chaperone environment:

chaperone install rfdiffusion1
chaperone install rpxdock
chaperone install ligandmpnn
chaperone install alphafold3
python examples/nanoparticle-design/pipeline.py --test --rundir=/absolute/path/nanoparticle-smoke

Use chaperone list-available to inspect available image names. Add --local inside an existing GPU allocation; otherwise the engine submits to Slurm using your site settings. No IPD partition or private container is required. The RPXdock recipe builds pinned public RPXdock/willutil sources and bundles their small_ilv_h scoring tables. Its first build compiles native extensions.

Steps and acceptance criteria

  1. Generate one 157-residue-per-chain C3 scaffold, keeping the seven-residue motif.
  2. Align the stalk through the RFdiffusion motif mapping. Reject more than eight stalk CAs within 3 Å of designed scaffold CAs (motif residues excluded).
  3. Require C3 CA agreement within 2 Å; orient the axis to z, preserve the asymmetric unit's radial displacement, residue identities, and motif metadata.
  4. Dock into I3. RFdiffusion's undesigned residues are glycine placeholders, so this example sets ignored_aas="": RPX scores N/CA/C-derived backbone frames for them instead of excluding glycine. These frames use the upstream Rosetta-derived mean centroid offset; output residue identities and atoms are retained. The historical score threshold has not been calibrated for this placeholder-inclusive scoring protocol. Emit the highest-scoring model as a true C3 subgroup, with its corresponding score and model index. Require RPX score ≥70.
  5. Design two sequences (one with --test) with corresponding residues tied across all three chains and antigen motif identities fixed. MPNN returns a packed all-atom structure for each sequence (output_mode="structures", one sidechain pack per design).
  6. Refold with one AF3 sample; compare against the docked trimer. Require backbone RMSD ≤2 Å and mean pLDDT ≥90.

The run directory contains normal engine manifests, structures, score tables, and results.csv. A one-design run may yield no passing candidates. An early filter rejection means later stages have not been exercised. Increase scale only after checking the small run's outputs and resource use.

Scope

CPU regression tests cover rigid-transform invariance, aligned clash counting, C3 geometry, motif/lineage preservation, model-score matching, and rejection of invalid docking outputs. CLI tests resolve the bounded configuration without running the neural models. A native CPU integration smoke on the upstream C3_1na0-1_1.pdb.gz fixture completed an I3 search (beam size 100) and Chaperone compilation: one 357-CA trimer, three mapped motif residues, retained parent lineage, and RPX score 0.03 (below the selection threshold). This used Python 3.12, NumPy 1.26.4, and the pinned RPXdock/willutil sources. The container recipe itself still needs a build/run check on the deployment system.

A subsequent CoreWeave H200 check completed public RFdiffusion v1 (86507b6, Base checkpoint, 50 steps) for the full 471-residue C3 scaffold. Its 157-residue asymmetric unit had zero stalk clashes and C3 CA RMSD 0.000638 Å. A native CPU I3 RPXdock search (beam 100, placeholder-inclusive scoring) produced a 471-CA trimer with 21 mapped motif residues and score 0.14, below the example gate. The RFD1 stack used PyTorch 2.4.0+cu124, DGL 2.4.0+cu124, NumPy 1.26.4, and e3nn 0.5.1.

The same candidate then completed tied ProteinMPNN on CUDA, one sidechain pack, and one AF3 sample (seed 42) on the H200. Packed and predicted structures each contain 3,825 finite-coordinate atoms. All three 157-residue sequences are identical; all 21 fixed motif residue identities and all 84 existing input motif atom coordinates were preserved exactly through MPNN. The public runner compiled AF3 pLDDT 77.11 and docked-trimer backbone RMSD 4.007 Å. AF3 model inference took 47.5 seconds after featurisation.

This was an execution diagnostic: it continued past the RPX gate to exercise packing and refolding. It fails both the RPX and AF3 selection criteria, so no accepted nanoparticle design is claimed. Native public RFD1/RPX installations, the public LigandMPNN image, and an existing AF3 image with the maintained Chaperone runner were used. Fresh image builds remain a separate validation step. Ordinary engine artifacts include rpx.csv, diagnostic.csv, structures, and logs for this run.

AF3 evaluates the trimer, not assembly of the 60-subunit cage. The source recipe's N-terminus check, Rosetta symmetry wiggle, and full-cage antigen/contact checks are not implemented here. These require additional validation before selecting experimental nanoparticles; a higher RPX score is not a substitute. The bundled small scoring tables and historical threshold are example settings, not a calibrated assembly-quality predictor.

Adapted from the baker-laboratory/pipelines nanoparticle example (main snapshot 95ef0b6), based on Cyrus Haas's nanoparticle workflow and the RFdiffusion work. The two input structures retain their source coordinates. The public helpers correct the predecessor's unaligned clash calculation, recentering of the monomer, motif loss, and rank/model-index confusion.