Flexible-peptide binder design¶
Two rounds of RFD3 → MPNN → AF3 selection for the peptide sequence
TVGSDGPPSFRSLSG. The peptide sequence is fixed; its conformation is not.
The initial CIF is constructed from sequence and has placeholder coordinates.
select_fixed_atoms=False allows RFD3 to generate peptide coordinates together
with a 60–140-residue binder. This is not a fixed-structure target example.
Run¶
Install the foundry, ligandmpnn, and alphafold3 containers and their checkpoints;
AF3 weights require separate access. Configure the model paths following the
installation guide. From the repository root in the Chaperone environment:
Append --local on an allocated GPU; otherwise use your Slurm site configuration.
Input CIF/JSON files are generated under the selected run directory. No private
structure or scratch path is needed. The normal default is four RFD3 designs,
two MPNN sequences per design per round, and one AF3 sample. --test reduces each
model to one sample. Motif sequence identities remain fixed during MPNN.
Filters and outputs¶
Round one requires AF3 ipTM ≥0.7, minimum interface PAE ≤2 Å, and backbone RMSD
≤3 Å. Only passing candidates enter the second MPNN/AF3 round, which requires
ipTM ≥0.85, minimum interface PAE ≤1.1 Å, and backbone RMSD ≤2 Å. RMSD uses the
RFD3 structure as reference. Outputs are normal engine artifacts and
results.csv; zero passing rows are a valid scientific outcome and must not be
reported as successful execution of stages skipped after filtering.
This port preserves the predecessor's filters; they have not been calibrated
against experimental peptide binding. Input generation and bounded CLI
configuration are CPU-tested. Bounded GPU execution is described below; binder quality is not established. Adapted from baker-laboratory/pipelines, snapshot 95ef0b6, with
corrected description of peptide coordinate conditioning and smaller defaults.
MPNN returns side-chain-packed structures, with one packing sample per designed sequence; refolding consumes those sequences and structure context.
GPU execution check¶
One CoreWeave H200 diagnostic completed RFD3 followed by two rounds of packed LigandMPNN, AF3, and backbone/motif RMSD calculation. MPNN explicitly requested CUDA. Both packed outputs matched their sampled FASTA sequences and preserved all 15 fixed peptide residues, with zero displacement of observed fixed heavy-atom coordinates from each round's input after compilation. Coordinates and reported RMSDs were finite. AF3 also retained the target peptide sequence.
The diagnostic bypassed quality selection to exercise both rounds. It failed all three original gates in each round:
| Metric | Round one result / requirement | Round two result / requirement |
|---|---|---|
| ipTM | 0.33 / ≥0.70 | 0.32 / ≥0.85 |
| Minimum interface PAE | 6.40 / ≤2.00 | 6.52 / ≤1.10 |
| Backbone RMSD | 8.51 Å / ≤3.00 Å | 5.87 Å / ≤2.00 Å |
The normal pipeline would stop after round one. This demonstrates execution and fixed-sequence/packing behavior, not a successful peptide binder. The run used native AtomWorks 2 RFD3, the public MPNN container, and an existing AF3 image with Chaperone's public inference runner; it was not a fresh image build test.