kaichenxu/scCAFM-data

scCAFM Tutorial Data

This repository contains the prepared datasets used by the executable tutorials for scCAFM, a causality-aware single-cell RNA-seq foundation model. The collection supports five workflows:

1. inference of cell-specific gene regulatory networks (GRNs) in developing mouse pancreas;
2. pooled-GRN inference and ChIP-seq-based benchmarking in human and mouse embryonic stem cells;
3. validation of predicted regulatory edges with CRISPRi Perturb-seq in K562 cells;
4. cell-type annotation through full and partial supervised fine-tuning of EFM cell-state embeddings;
5. gene perturbation prediction using Norman CRISPRa data and contextual EFM gene embeddings.

The GRN and perturbation files are tutorial-ready derivatives of public research datasets. The PBMC annotation file is a byte-for-byte copy of the user-provided CTA input; its original archive accession and reuse terms are not recorded in the file. They are not replacements for the primary archives. Users should cite the relevant original study and repository record listed below.

Related resources:

• scCAFM model
• scCAFM source and tutorials
• scCAFM Space

Repository structure

scCAFM-data/
├── cell_specific_grns/
│   └── mPancreas.h5ad
├── chipseq_grn_recovery/
│   ├── hESC.h5ad
│   ├── mESC.h5ad
│   ├── hESC-ChIP-seq.csv
│   └── mESC-ChIP-seq.csv
├── perturbseq_edge_validation/
│   └── K562.h5ad
├── cell_type_annotation/
│   └── pbmc.h5ad
├── gene_perturbation_prediction/
│   ├── norman.h5ad
│   ├── go_prior.json
│   ├── go_edges.csv
│   └── split_manifest.json
├── LICENSES.md
└── MANIFEST.tsv

Dataset inventory

File	Contents	Dimensions
cell_specific_grns/mPancreas.h5ad	Processed E15.5 mouse pancreatic endocrinogenesis data with cell-population annotations, spliced/unspliced layers, PCA and UMAP coordinates	3,696 cells × 27,998 genes
chipseq_grn_recovery/hESC.h5ad	BEELINE-derived human embryonic stem-cell expression matrix	758 cells × 17,735 genes
chipseq_grn_recovery/mESC.h5ad	BEELINE-derived mouse embryonic stem-cell expression matrix	421 cells × 18,385 genes
chipseq_grn_recovery/hESC-ChIP-seq.csv	Directed human ChIP-seq reference edges (Gene1 regulator, Gene2 target)	441,991 edges
chipseq_grn_recovery/mESC-ChIP-seq.csv	Directed mouse ChIP-seq reference edges (Gene1 regulator, Gene2 target)	985,654 edges
perturbseq_edge_validation/K562.h5ad	QC-filtered raw-count K562 essential-scale CRISPRi Perturb-seq data	162,751 cells × 8,563 genes
cell_type_annotation/pbmc.h5ad	Human PBMC raw counts, 11 cell types, and the supplied training/test split	68,579 cells × 32,738 genes
gene_perturbation_prediction/norman.h5ad	Unchanged supplied Norman K562 CRISPRa source with normalized expression, raw counts and GEARS DE rankings	91,205 cells × 5,045 genes

SHA-256 checksums and exact byte sizes are provided in MANIFEST.tsv.

Data formats

AnnData files

The .h5ad files follow the AnnData convention:

• rows (obs) are cells;
• columns (var) are measured genes;
• X stores the expression matrix;
• additional annotations, layers and embeddings are retained where available.

Important fields include:

• mPancreas.h5ad
  • obs["clusters"]: eight detailed cell populations;
  • obs["clusters_coarse"]: five broader populations;
  • obs["species"]: mouse;
  • layers["spliced"] and layers["unspliced"];
  • obsm["X_pca"] and obsm["X_umap"].
• hESC.h5ad and mESC.h5ad
  • obs["species"];
  • obs["disease"], set to normal.
• K562.h5ad
  • X: raw counts after the recorded filtering and perturbation-QC steps;
  • obs["gene"]: perturbed gene or non-targeting;
  • obs["gene_id"]: Ensembl gene identifier for the perturbation;
  • obs["sgID_AB"]: paired guide identifiers;
  • obs["species"], obs["disease"], obs["cell_line"], obs["cell_type"] and obs["tissue"];
  • var["gene_name"]: gene symbol;
  • uns["basic_filter"] and uns["perturbation_qc"]: recorded preparation parameters and cell counts.

PBMC annotation data

cell_type_annotation/pbmc.h5ad preserves raw counts in X; labels in obs["cell_type"]; 20,579 training and 48,000 test cells in obs["split"]; and assay, tissue and disease annotations. Gene symbols are in var_names, with Ensembl identifiers in var["gene_ids"]. The tutorial adds species="human" in memory and constructs a reproducible stratified 70:30 training/test split across all 68,579 cells, independently of the supplied original split.

ChIP-seq edge tables

Both CSV files contain:

Column	Definition
Gene1	Regulator/transcription factor
Gene2	Putative target gene supported by the BEELINE ChIP-seq reference

The ChIP-seq networks are experimental reference networks, not absolute biological ground truth.

Download and load

Download the complete snapshot:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="kaichenxu/scCAFM-data",
    repo_type="dataset",
    local_dir="tutorial_data",
)

Load the files:

import anndata as ad
import pandas as pd

pancreas = ad.read_h5ad(
    "tutorial_data/cell_specific_grns/mPancreas.h5ad"
)

hesc_chip = pd.read_csv(
    "tutorial_data/chipseq_grn_recovery/hESC-ChIP-seq.csv"
)

For reproducible analyses, pin the dataset repository to a commit revision when calling snapshot_download.

Provenance and preparation

Mouse pancreatic endocrinogenesis

mPancreas.h5ad is derived from the processed E15.5 pancreatic endocrinogenesis dataset distributed through scVelo. The underlying experiment is available from GEO under GSE132188 and was described by Bastidas-Ponce et al. The publication copy corrects an erroneous obs["species"] value from human to mouse; the expression matrix, layers, annotations and embeddings are otherwise unchanged during this packaging step.

Embryonic stem-cell expression and ChIP-seq references

The hESC and mESC materials are reformatted from the experimental expression data and ChIP-seq reference networks distributed with BEELINE:

• BEELINE data record: 10.5281/zenodo.3378975
• BEELINE study: 10.1038/s41592-019-0690-6
• hESC source experiment: GSE75748
• mESC source experiment: GSE98664

The expression matrices were placed into cell-by-gene AnnData files and supplied with the species and condition fields required by the scCAFM tutorials. The directed ChIP-seq edge tables retain the BEELINE regulator-target representation.

K562 Perturb-seq

K562.h5ad is derived from the K562 essential-scale CRISPRi Perturb-seq experiment sampled at day 6 in Replogle et al.:

• processed-data record: 10.25452/figshare.plus.20029387.v1
• study: 10.1016/j.cell.2022.05.013
• raw sequencing archive: GSE146194

The preparation started from the raw single-cell count object. The stored provenance records 310,385 original cells, 192,648 cells after perturbation-level QC and 162,751 retained cells after cell-level QC. Basic filters require at least 200 detected genes per cell and at least three cells per gene. The stored perturbation-QC metadata records a minimum of more than 25 filtered cells, a maximum knockdown percentage of −0.3 (inclusive), more than 50 differentially expressed genes, and a 10th-percentile control threshold for cell-effect filtering.

PBMC cell-type annotation

The source is the user-provided Sunlab file /data1021/xukaichen/data/CTA/pbmc.h5ad. It is copied without modification and its checksum is verified against that source. The file does not record an original archive or licence in uns. No original accession or redistribution rights are asserted here; add the verified primary citation and reuse terms before publicly publishing this component. The supplied train/test split is preserved, and validation is derived only from training cells.

Intended use

The collection is intended for:

• running the curated scCAFM tutorials;
• testing scCAFM data loading and preprocessing;
• reproducing the tutorial-level GRN inference and validation examples;
• educational exploration of single-cell GRN and cell-type annotation workflows.

It is not intended to serve as a new primary archive, a clinical resource or a comprehensive reprocessing of the source studies.

Limitations

• The files are processed tutorial derivatives; consult the primary repositories for raw data and full experimental metadata.
• ChIP-seq binding provides population-level regulatory evidence but does not prove that every edge is active in every cell.
• A Perturb-seq expression shift can be indirect or off-target and does not by itself prove a direct TF-target interaction.
• The workflows use different organisms, assays and processing histories; matrices should not be concatenated or compared without an explicit harmonization strategy.
• The hESC and mESC AnnData files contain only the metadata needed by the tutorial and do not reproduce every field from their primary archives.

Access, rights and attribution

This is a mixed-source collection, so no single software license applies to every data file. The repository therefore uses the Hugging Face other license designation. Per-source licenses and attribution requirements are documented in LICENSES.md. Redistribution through this repository does not replace the original terms, and users must follow the terms associated with each source dataset.

Citation

When using a file from this repository, cite the corresponding original dataset and study listed in Provenance and preparation, as well as the scCAFM model or paper associated with the analysis.

Suggested repository reference:

Xu, K. (2026). scCAFM Tutorial Data. Hugging Face Datasets. https://huggingface.co/datasets/kaichenxu/scCAFM-data

Norman gene perturbation prediction

The tutorial file is a byte-for-byte copy of the supplied /data1021/xukaichen/data/GPP/norman_perturb.h5ad, stored as gene_perturbation_prediction/norman.h5ad. The study describes K562 single/double CRISPRa perturbations: Norman et al. (2019), GSE133344.

Original source contents

The file contains 91,205 cells and 5,045 genes, normalized log expression in X and raw counts in layers["counts"]. Cell barcodes are in obs_names; obs retains condition, cell_type, dose_val, control and condition_name. Ensembl identifiers are in var_names and symbols in var["gene_name"]. Cell type is K562.

The source GEARS DE rankings and dropout caches are preserved, together with the recorded cell-type annotation correction. The file has no tutorial split assignments or control pairings. Provenance and checksums are recorded in split_manifest.json and MANIFEST.tsv.

The supplied source is already gene-filtered. Numerical reconstruction is consistent with median-library normalization before gene filtering and natural log1p, with inferred median library size 13,162. Exact upstream QC and gene selection are not recorded. This is not a copy of the complete primary raw archive or a claim of byte identity with the official GEARS file.

Preparation in the tutorial

The notebook uses source X without normalizing or log-transforming it again. It selects all supported intervention targets plus training-control HVGs in a 2,047-gene EFM panel, retains counts on the same axes, derives the required GEARS observation fields, and fixes one control input per response cell.

Outer training gene/combination fractions are 75%/75%; nested training fractions are 90%/90% for validation, using seed 1. All 7,353 controls belong to training. The split has 49,849 training, 10,754 validation and 28,754 test cells, with 138, 31 and 107 perturbed source labels, respectively. Original labels, including the two single-guide orientations, are retained.

Original source DE rankings are filtered to the selected panel; the first 20 retained genes define DE20. Validation DE20 selects checkpoints and test DE20 is used only for scoring. Rankings and outcome expression are not features for predicting perturbed responses.

External prior and model settings

go_prior.json records the 9,853-node perturbation universe and source hashes for GEARS datafiles 6153417, 6934320 and 6934319. go_edges.csv contains 202,314 directed edges retained by the reference top-21-per-target rule. The co-expression branch remains inactive.

The notebook compares Frozen EFM + GEARS (freeze_layers=12) and Partial Fine-tuning EFM + GEARS (freeze_layers=10, last two blocks trainable). SFM is frozen in both. The tutorial uses at most 10 epochs, decoder learning rate 0.0003, trainable EFM learning rate 0.000003, and a halving schedule every three epochs. Training stops after four epochs without validation DE20 MSE improvement and restores the best validation checkpoint. Fitting displays compact training-loss histories and a checkpoint summary; evaluation displays test metrics and inline figures.

The dataset archive contains inputs and provenance. Training histories, selected checkpoints and held-out results are generated by the accompanying tutorial.