| name | diffdock-docking |
| description | Run a pinned DiffDock release for diffusion-based protein-ligand pose prediction. Use for candidate pose generation from receptor coordinates and ligand structures, especially without a fixed pocket; do not interpret DiffDock confidence as affinity. |
| license | MIT |
DiffDock Docking
Gate and preflight
Follow $cx-compute-environment. Ask once before cloning/installing a pinned
commit, downloading weights, and using GPU compute. Name all network hosts and
never send private structures to a hosted service without approval.
Workflow
- Prepare inputs with
$cx-molecular-input-preparation; record receptor
assembly/state and every ligand microstate.
- Pin repository commit, environment lock, checkpoint checksum, inference
configuration, seed, samples per complex, and hardware.
- Smoke-test on the upstream example, then predict each complex into an isolated
output directory. Preserve all generated poses and confidence values.
- Convert outputs without changing atom order or stereochemistry; verify atom
mapping and clashes.
- Treat confidence as within-model pose ranking. Apply
$cx-docking-validation; use an independent affinity model or experiment for
affinity claims.
Boundaries
- Confidence is neither probability of binding nor binding affinity.
- Check training/template overlap for benchmark targets and ligands.
- For predicted receptors, report receptor uncertainty and avoid claiming
precision in low-confidence pockets.
