| name | gnina-docking |
| description | Run reproducible local GNINA docking or CNN rescoring on protein-ligand systems. Use when GPU-assisted pose generation, refinement, or CNN reranking is wanted; keep CNN pose score, predicted affinity, and experimental affinity as distinct quantities. |
| license | MIT |
GNINA Docking
Gate and preflight
Follow $cx-compute-environment; ask once for the pinned GNINA binary/container,
model download, GPU compute, and any structure download. Require prepared inputs
from $cx-molecular-input-preparation. Prefer the official release binary or a
digest-pinned container over a source build.
Workflow
- Record GNINA release, binary checksum/container digest, CUDA/GPU, CNN model,
empirical scoring function,
cnn_scoring mode, box/autobox source, seed,
exhaustiveness, and number of modes.
- Smoke-test the binary and one control complex. Use an experimental reference
ligand for autoboxing only when it is legitimate for the evaluation split.
- Run docking; retain SDF properties, raw log, receptor, reference ligand, and
exact command under
artifacts/<run-id>/gnina/.
- Keep empirical energy, CNN pose score, and CNN affinity output in separate
columns. Do not select whichever score makes the result look best.
- Validate redocking and ranking with
$cx-docking-validation.
Boundaries
- CNN confidence/score is not experimental affinity and is not calibrated across
arbitrary targets or model versions.
- Blind whole-protein docking needs stronger controls and higher search effort;
label it exploratory.
- Flexible-sidechain docking requires a justified residue set and separate
validation; more flexibility does not guarantee realism.
