Skip to content

data.md

Latest commit

 

History

History
36 lines (28 loc) · 3.23 KB

data.md

File metadata and controls

36 lines (28 loc) · 3.23 KB
Raw

Data

Everything needed to run the tutorials except the foundation-model checkpoints is in one Zenodo record: https://doi.org/10.5281/zenodo.23153537 (three .rar archives, about 650 MB). heterosc.zenodo.download_data() downloads, verifies the MD5 and extracts them.

Archive Contents
annotation raw/ original .h5ad (Pancreas, Skin, Liver, Lung, Myeloid) · processed/ .h5ad with cell_type, var['ensembl_id'], n_counts, obs['master_split'] · splits/ cell_id → split tables · annotation_split_summary.csv
drug_response response_table_merged_leave_drug_out.csv (cell_line_id, drug_id, canonical_smiles, auc, fold) · drug_fold_assignment.csv · drug_graph_features_merged.pkl · drug_fingerprint_merged.csv (MACCS 167 + ECFP4 2048) · cellline_expression.h5ad · gdsc_drugname_to_smiles_cache.json · raw/DepMap_Model.csv
gene_perturbation norman.zip, adamson.zip (GEARS PertData) · source .h5ad · gene-symbol→Ensembl tables · GEARS simulation split (*_simulation_3_0.75*.pkl) · GEARS GO resources

Archive MD5: annotation.rar 0cb08bcad13e54b0b42492b929bbd097 · drug_response.rar 43fec0e20632007264c238c7a04029f5 · gene_perturbation.rar 7b50c606d74eb7dddc499a882eb5ee74.

Opening .rar needs 7-Zip (7z), unrar, or pip install rarfile plus unrar. .pkl files should only be loaded from trusted sources.

Not in the record

  • Foundation-model checkpoints and embeddings. Download the checkpoints from the official releases (links in the README) and extract embeddings with notebooks 02a–d.
  • Raw GDSC dose-response files. Download from the GDSC website under its terms (scripts/download_data.py --gdsc data/gdsc). The processed response table in the record is enough for the tutorials.

Protocols encoded in the data

  • Annotation: cell types with fewer than 10 cells removed; stratified 70/15/15 train/validation/test, random seed 42, shared by all four models. Only cells with a valid embedding from all four models are used.
  • Drug response: 109,152 cell line–drug pairs, 223 drugs, 551 cell lines (the expression file holds 561 profiles; 551 have drug-response data), target = AUC from GDSC1 + GDSC2 (GDSC2 takes precedence for overlapping pairs). fold = 10-fold leave-drug-out: KFold over the sorted drug list, seed 42, 22–23 drugs per fold.
  • Perturbation: Norman and Adamson processed by GEARS; simulation split with seed 3 and train_gene_set_size = 0.75.

Embedding sizes

Expert Cell embedding Source
Geneformer V2-104M 768 CLS token, last hidden layer
scGPT whole-human 512 CLS embedding (L2-normalised)
CellFM (80M) 1536 CLS token
scFoundation 768 max-pool over encoder tokens

Sources and attribution

Please cite the original data providers: Cheng et al. 2021 (Myeloid), MacParland et al. 2018 (Liver), the CellFM benchmark data (Zeng et al. 2024), Norman et al. 2019, Adamson et al. 2016, Roohani et al. 2024 (GEARS), Yang et al. 2013 and Iorio et al. 2016 (GDSC), Barretina et al. 2012 and Ghandi et al. 2019 (CCLE). GDSC-derived values follow the GDSC data usage policy; DepMap data are generally CC BY 4.0.