cfy2yue's picture
Upload scPerturb normalized payload
766e785 verified
|
Raw
History Blame Contribute Delete
1.64 kB
# scPerturb-derived model-ready H5AD
This directory contains 252 human cell-level derivatives prepared from the
scPerturb collection: 23 gene-perturbation units and 229 chemical-perturbation
units. It does not contain the upstream raw downloads.
| modality | units | cells | reference | effect |
| --- | ---: | ---: | ---: | ---: |
| genepert | 23 | 3,468,800 | 218,422 | 3,250,378 |
| chempert | 229 | 1,306,002 | 179,106 | 1,126,896 |
For every file, `X` is per-cell CP10K natural `log1p`, confirmed raw UMI counts
are in `layers["counts"]`, and the independently trained 30-dimensional scVI
representation is in `obsm["X_scVI"]`. Control/reference selection must use
`obs["is_reference"]`; vehicle/DMSO must not be reinterpreted as untreated or
time zero.
The on-disk directory name `scPeturb` is a retained historical spelling. The
machine-readable source label and center paths use `scPerturb`; paths are
case-sensitive. Always obtain paths from `../manifest/TRAINING_UNITS.tsv`.
These are processed/model-ready derivatives, not raw source files. Read
`../README.en.md`, `../DATA_DICTIONARY.md`, and `../USAGE.md` before use. The
source release, original-study citations, licenses, and redistribution status
must be taken from the Stage 1 source-provenance table; do not infer a data
license from the scPerturb software repository.
The audited redistribution basis for these files is the scPerturb v1.4 dataset
record (`10.5281/zenodo.13350497`), which applies CC BY 4.0. Public release must
attribute scPerturb and each original study, link the license, and identify the
CellClip modifications described in `../PROCESSING.md`.