eheiss/scbFM_data

scbFM data

Data used by scbFM, the code repository for Single-Cell Foundation Models for Bulk Transcriptomics: Evaluation and Adaptation.

Set SCBFM_ROOT_DIR to the directory containing datasets/, output/, and other/:

export SCBFM_ROOT_DIR=/path/to/experiment-workspace

Install the Hugging Face client and download the hosted files directly into the expected directory structure:

python -m pip install --upgrade huggingface_hub
hf download eheiss/scbFM_data --repo-type dataset --local-dir datasets

To download one file, add its repository path after the dataset name. For example:

hf download eheiss/scbFM_data TCGA/tcga.h5ad \
  --repo-type dataset \
  --local-dir datasets

1. Harmonized data for the controlled benchmark

These are the ready-to-use inputs for controlled pretraining, pre-adaptation, the nine bulkRNA-seq downstream tasks, and their direct-expression baselines. Files marked as hosted can be downloaded directly from this repository.

Required structure

datasets/
|-- DepMap/
|   `-- depmap.h5ad
|-- DiSignAtlas/
|   `-- disignatlas.h5ad
|-- GDSC/
|   |-- drug_features.npz
|   |-- drug_response_prediction_IC50.csv
|   |-- drug_response_prediction_IC50_drug_resp_6e86893658933c75_cv_folds.csv
|   `-- gdsc.h5ad
|-- SurvBoard/
|   `-- data_reproduced/
|       |-- TCGA/
|       |   `-- *_data_complete_modalities_preprocessed.csv
|       `-- splits/
|           `-- TCGA/
|               `-- ...
|-- TCGA/
|   `-- tcga.h5ad
|-- bulk/
|   |-- preadapt_bulk_RAW.h5ad
|   `-- pretraining_bulk_RAW.h5ad
|-- pseudo_bulk/
|   `-- pseudo_bulk_RAW.h5ad
`-- sc/
    `-- pretraining_sc_RAW.h5ad

Files and download sources

File or directory	Purpose	Download source
bulk/pretraining_bulk_RAW.h5ad	Controlled bulkRNA-seq pretraining corpus	Available in this repository
bulk/preadapt_bulk_RAW.h5ad	Controlled bulkRNA-seq pre-adaptation corpus	Available in this repository
sc/pretraining_sc_RAW.h5ad	Controlled scRNA-seq pretraining corpus	Available in this repository
TCGA/tcga.h5ad	Cancer classification, pan-cancer survival, and binary vital-status prediction	Available in this repository
DiSignAtlas/disignatlas.h5ad	Disease classification	Available in this repository
DepMap/depmap.h5ad	Gene-essentiality prediction	Available in this repository
GDSC/gdsc.h5ad	Harmonized drug-response expression matrix	Available in this repository
GDSC/drug_response_prediction_IC50.csv	Drug-response targets and sample-drug pairs	Available in this repository
GDSC/drug_features.npz	Molecular drug representations	Available in this repository
GDSC/*_cv_folds.csv	Canonical drug-response cross-validation assignments	Available in this repository
pseudo_bulk/pseudo_bulk_RAW.h5ad	Cell-type deconvolution	Available in this repository
SurvBoard/data_reproduced/	Cohort-specific survival matrices and predefined splits	Available in this repository

2. Primary data for reproducing preprocessing

The primary files below are needed only to regenerate the harmonized matrices and controlled corpora using the scripts in scbFM/data/.

Required structure

datasets/
|-- ARCHS4/
|   |-- human_gene_v2.latest.h5
|   `-- archs4.h5ad                         # generated
|-- DepMap/
|   |-- gene_essentiality_data.h5ad
|   `-- depmap.h5ad                         # generated
|-- DiSignAtlas/
|   |-- disease_annot_data.h5ad
|   `-- disignatlas.h5ad                    # generated
|-- GDSC/
|   |-- drug_response_expr_data.csv
|   |-- drug_response_prediction_IC50.csv
|   |-- drug_features.npz                   # generated
|   `-- gdsc.h5ad                           # generated
|-- GTEx/
|   |-- GTEx_Analysis_2025-08-22_v11_RNASeQCv2.4.3_gene_reads.parquet
|   `-- gtex.h5ad                           # generated
|-- SurvBoard/
|   `-- data_reproduced/
|       |-- TCGA/
|       `-- splits/
|-- TCGA/
|   |-- TCGA_cancer_data.h5ad
|   `-- tcga.h5ad                           # generated
|-- bulk/
|   |-- pretraining_bulk_RAW.h5ad           # generated
|   `-- preadapt_bulk_RAW.h5ad              # generated
|-- pseudo_bulk/
|   |-- source_cell_chunks/                 # downloaded from CELLxGENE Census
|   |-- source_cell_chunk_manifest.csv
|   |-- pseudo_bulk_sampling_plan.csv
|   `-- pseudo_bulk_RAW.h5ad                # generated
`-- sc/
    |-- census_RAW_chunks/                  # downloaded from CELLxGENE Census
    |-- sampling_plan_RAW.csv
    `-- pretraining_sc_RAW.h5ad             # generated

Files and download sources

Primary resource	Expected source file or directory	Download source
ARCHS4	ARCHS4/human_gene_v2.latest.h5	ARCHS4 portal
GTEx	GTEx/GTEx_Analysis_2025-08-22_v11_RNASeQCv2.4.3_gene_reads.parquet	GTEx portal
TCGA	TCGA/TCGA_cancer_data.h5ad	BulkFormer resource page
DiSignAtlas	DiSignAtlas/disease_annot_data.h5ad	BulkFormer resource page
DepMap	DepMap/gene_essentiality_data.h5ad	BulkFormer resource page
GDSC	GDSC/drug_response_expr_data.csv and GDSC/drug_response_prediction_IC50.csv	BulkFormer resource page
CELLxGENE Census	sc/census_RAW_chunks/ and pseudo_bulk/source_cell_chunks/	CELLxGENE Census API via scripts in scbFM/data/
SurvBoard	SurvBoard/data_reproduced/	Available in this repository

The Census scripts create resumable sampling plans, manifests, and chunks. Keep those generated files together when transferring an interrupted download to an offline compute cluster.

3. Data for complementary analyses

The complementary analyses translate the controlled findings to scGPT and BulkFormer, investigate bulkRNA-seq pretraining scale, and test retention of single-cell capabilities. The scale analysis reuses bulk/pretraining_bulk_RAW.h5ad and needs no additional dataset.

Required structure

datasets/
|-- DepMap/
|   `-- gene_essentiality_data.h5ad
|-- DiSignAtlas/
|   `-- disease_annot_data.h5ad
|-- GDSC/
|   `-- drug_response_expr_data.csv
|-- TCGA/
|   `-- TCGA_cancer_data.h5ad
`-- scgpt_single_cell/
    |-- covid/
    |   |-- batch_covid_subsampled_train.h5ad
    |   `-- batch_covid_subsampled_test.h5ad
    `-- lung/
        |-- sample_proc_lung_train.h5ad
        `-- sample_proc_lung_test.h5ad

Files and download sources

File or directory	Complementary analysis	Download source
TCGA/TCGA_cancer_data.h5ad	Full-vocabulary BulkFormer evaluation on classification and survival	See Section 2
DiSignAtlas/disease_annot_data.h5ad	Full-vocabulary BulkFormer disease classification	See Section 2
DepMap/gene_essentiality_data.h5ad	Full-vocabulary BulkFormer gene-essentiality evaluation	See Section 2
GDSC/drug_response_expr_data.csv	Full-vocabulary BulkFormer drug-response evaluation	See Section 2
scgpt_single_cell/covid/*.h5ad	Cell-type annotation and zero-shot batch integration	Official scGPT COVID-19 files
scgpt_single_cell/lung/*.h5ad	Cell-type annotation and zero-shot batch integration	Official scGPT Lung-Kim files