| name | scgpt-single-cell |
| description | Run pinned scGPT checkpoints for single-cell embeddings, cell-type annotation, reference mapping, perturbation modeling, or fine-tuning. Use when a pretrained single-cell transformer is requested and donor/batch/feature provenance can be preserved. |
| license | MIT |
scGPT Single-Cell Analysis
Gate
Ask once before installing a pinned scGPT revision, downloading checkpoints or
reference assets, and using GPU compute. Never upload private human data without
explicit authorization and governance review. Follow $cx-compute-environment.
Workflow
- Define the independent biological unit, task, donors, tissues, batches,
perturbations, count layer, gene identifiers/release, labels, and deployment
population. Split by donor/sample/experiment before cell-level processing.
- Pin code/checkpoint/token dictionary, gene mapping, normalization/binning,
max sequence length, feature selection, pooling, fine-tuning configuration,
seeds, and hardware. Fit learned preprocessing on training data only.
- Smoke-test a small public example. Run into
artifacts/<run-id>/scgpt/;
retain AnnData schema, split IDs, preprocessing, embeddings/predictions,
checkpoint/config, learning curves, environment, logs, and excluded cells.
- Compare PCA/logistic, nearest-reference, and batch-aware baselines. Report
donor-level uncertainty, batch/tissue slices, calibration, OOD detection, and
nuisance predictability from embeddings.
- Record with
$science-provenance; review with $science-review.
Boundaries
- Cells from one sample are not independent replicates.
- Integration can erase biology; cell-type labels and perturbation predictions
remain model outputs, not experimental or clinical truth.