Hammerhead ribozyme rescaffolding

Design a 60-nucleotide RNA around six catalytic nucleotides from the Schistosoma hammerhead crystal structure, PDB 2OEU, and its active-site Mn²⁺. RFD3NA chooses the positions of the unindexed motif within one designed RNA chain. AF3 then predicts the RNA and ion, and the metrics compare the predicted motif with the crystal reference.

RFD3NA → motif/clash filter → AF3 → crystal motif RMSD
       → catalytic distances and angle → geometry filter

Inputs

The source directory is examples/ribozyme-hammerhead/. files/2oeu_motif.pdb contains A20–22, A36, B6–7 and Mn. files/2oeu_native.cif is the full native reference for optional controls. The unindexed groups are joined with commas, not /0: introducing a molecule break changes the connectivity requested of the model. Input JSON is generated under the run directory, leaving the bundled files unchanged.

Dependencies and bounded run

Install the foundry and alphafold3 images. Stage the public RFD3NA checkpoint (foundry install rfd3na --checkpoint-dir /absolute/path/checkpoints) and set FOUNDRY_CHECKPOINT_DIRS. Obtain AF3 model parameters under their applicable terms and set ALPHAFOLD3_WEIGHTS_DIR.

python examples/ribozyme-hammerhead/pipeline.py --test --rundir=/absolute/path/hammerhead-smoke

Append --local inside an allocated GPU session. Otherwise the engine submits jobs through your Slurm site settings. The default produces two 60-nt designs; --test requests one, followed by one AF3 sample if the first filter passes. The RFD3NA cloud backend is not currently enabled: use a configured cluster or local GPU with the staged checkpoint.

Measurements and selection

The generation filter requires motif insertion RMSD ≤1 Å, zero conjoined residues, and zero reported interresidue backbone/sidechain clashes. These checks do not establish complete phosphodiester connectivity.

After AF3, the example requires crystal-reference RNA motif RMSD ≤2 Å, C17 O2′–C1.1 P attack distance ≤4 Å, and O2′–P–O5′ angle ≥140°. General-acid/base distances and Mn–A9/Mn–G10.1 distances are reported for inspection. Mn is excluded from the RNA motif RMSD and its distances are not selection gates. These settings preserve the source example's computational criteria; they are not a validated predictor of cleavage activity.

Normal engine manifests, structures, metric tables, and results.csv are written to the run directory. An empty table means no design passed. If the first filter rejects the sole smoke design, AF3 has not been tested by that run.

Validation scope

CPU tests verify that every atom addressed by the catalytic measurements exists in the bundled motif and that the generated specification requests one 60-nt RNA molecule. A bounded NVIDIA H200 run using the public RFD3NA 1190 checkpoint completed 200 diffusion steps (seed 42), producing 60 RNA residues and one Mn. Motif insertion RMSD was 0.257 Å, with zero conjoined residues, zero reported backbone clashes, and one reported sidechain clash. That candidate fails the example’s clash gate. A diagnostic continued through AF3 and both metric blocks to verify execution, without changing the shipped selection gates. AF3 seed 42 / one sample completed with 1,285 finite-coordinate atoms, 60 RNA residues, and one Mn. All catalytic measurements were finite. Mean pLDDT was 82.13; crystal motif RMSD was 5.166 Å, attack distance 4.046 Å, and attack angle 73.12°. This candidate also fails the refold gates. There was no accepted design in this one-sample check.

Generation used the native AtomWorks 3 Foundry environment and checkpoint 1190; refolding used the existing AF3 image with this repository's maintained runner, not a newly built public container. AF3 model inference took 37.8 seconds after featurisation on the H200. The ordinary run artifacts include generation.csv, refolding.csv, structures, manifests, and logs under the validation run's rna directory. This verifies GPU model execution, output compilation, motif mapping, and scoring; it does not certify a fresh container installation.

Neither a successful model execution nor these geometry filters establish RNA catalytic activity, stability, or experimental function. Check O3′–P connectivity and native controls before a scientific pilot.

Adapted from baker-laboratory/pipelines, main snapshot 95ef0b6. Historical success rates from that source are not evidence for this public port.