README.md

metadata

license: apache-2.0
task_categories:
  - biology
  - genomics
  - tabular-classification
  - tabular-regression
tags:
  - single-cell
  - scRNA-seq
  - transcriptomics
  - perturbation
  - batch-integration
  - cell-type-annotation
  - anndata
size_categories:
  - 100K<n<1M

🧬 ScDiVa Benchmark Datasets

📄 arXiv Paper | 💻 Model Repository | 🌐 Project Page

📖 Dataset Description

This repository contains the pre-processed benchmark datasets used to evaluate ScDiVa (Single-cell Deep Variational Analysis). All datasets are stored in .h5ad (AnnData) format, compatible with scanpy and anndata.

These datasets cover four major downstream tasks in single-cell genomics:

1. Batch Integration
2. Cell Type Annotation (Fine-tuning & Zero-shot)
3. Gene Perturbation Prediction
4. Gene Regulatory Network (GRN) Reconstruction

---

📂 File Organization

The datasets are categorized by their primary evaluation task in the ScDiVa paper.

1. Batch Integration & Atlas Building

Datasets containing multiple batches with technical effects to be removed.

Filename	Description	Cells
pbmc12k_processed.h5ad	PBMC dataset with 2 batches.	~12k
immune_processed.h5ad	Human Immune Atlas (large-scale integration).	~300k
bmmc_processed.h5ad	Bone Marrow Mononuclear Cells (NeurIPS 2021).	~90k
perirhinal_processed.h5ad	Mouse Brain Perirhinal Cortex.	~23k
covid19_processed.h5ad	PBMC from COVID-19 patients.	~45k

2. Cell Type Annotation (Fine-tuning)

Datasets used for supervised training and evaluation of cell type classifiers.

Filename	Description
hpancreas_processed.h5ad	Human Pancreas dataset (State-of-the-art benchmark).
ms_processed.h5ad	Multiple Sclerosis dataset.
myeloid_processed.h5ad	Myeloid cells (Set A).
myeloid_b_processed.h5ad	Myeloid cells (Set B).

3. Zero-shot Annotation

Datasets used to test the model's generalization capability without task-specific training.

Filename	Description
Pancrm.h5ad	Human Pancreas (distinct from hPancreas).
PBMC.h5ad	Standard PBMC benchmark.
PBMC_368K.h5ad	Large-scale PBMC dataset.
HumanPBMC.h5ad	Additional human PBMC variant.
Cell_Lines.h5ad	Cell line benchmarking data.
DC.h5ad	Dendritic Cells.
MCA.h5ad	Mouse Cell Atlas subset.
immune_processed.h5ad	Human Immune Atlas (Used for large-scale zero-shot evaluation).

4. Gene Perturbation Prediction

Datasets for causal inference tasks (predicting gene expression after CRISPR perturbation).

Filename	Description
adamson_processed.h5ad	Single-gene CRISPRi perturbations (Adamson et al., 2016).
norman.h5ad	Combinatorial (double-gene) CRISPRa perturbations (Norman et al., 2019).

5. Reconstruction & GRN Inference

Datasets used to evaluate the model's ability to reconstruct gene expression (Section 4.2) and infer regulatory networks (Section 4.6).

Filename	Description
zheng68k_processed.h5ad	Mouse hematopoietic stem and progenitor cell scRNA-seq dataset (68k cells).
hpancreas_processed.h5ad	Human Pancreas dataset.
pbmc12k_processed.h5ad	PBMC dataset with 2 batches.
immune_processed.h5ad	Human Immune Atlas(Primary dataset for GRN inference).

---

🚀 How to Use

Since these are .h5ad files, we recommend using huggingface_hub to download the specific file you need, and then loading it with scanpy.

Prerequisites

pip install scanpy huggingface_hub

Loading a Dataset

import scanpy as sc
from huggingface_hub import hf_hub_download

# Example: Load the PBMC12k dataset for Batch Integration
file_path = hf_hub_download(
    repo_id="warming666/ScDiVa",
    filename="pbmc12k_processed.h5ad",
    repo_type="dataset"
)

adata = sc.read_h5ad(file_path)

print(adata)
# AnnData object with n_obs × n_vars = 11990 × 2000 ...

---

⚠️ Data Processing Notes

• Normalization: Most datasets labeled _processed have typically undergone library size normalization (e.g., target sum 10,000) and log1p transformation. Please check the adata.X values or adata.uns for specific processing details.
• Splits: Train/Test splits, if applicable, are usually stored in adata.obs['split'] or indicated by boolean masks (e.g., train_mask).
• Gene Vocabulary: To work with the pre-trained ScDiVa model, ensure the gene symbols in adata.var_names match the model's vocabulary (41,818 genes).

---

📄 Citation

If you use these processed datasets in your work, please cite the ScDiVa paper:

@article{wang2026scdiva,
  title={ScDiva: Masked Discrete Diffusion for Joint Modeling of Single-Cell Identity and Expression},
  author={Wang, Mingxuan and Chen, Cheng and Jiang, Gaoyang and Ren, Zijia and Zhao, Chuangxin and Shi, Lu and Ma, Yanbiao},
  journal={arXiv preprint arXiv:2602.03477},
  year={2026}
}