README.md

---
license: cc-by-4.0
tags:
- single-cell
- rna-seq
- scRNA-seq
- h5ad
- anndata
- genomics
- benchmark
- shape-analysis
- synthetic
size_categories:
- 1M<n<10M
task_categories:
- tabular-classification
- tabular-regression
pretty_name: scShapeBench
configs:
  - config_name: scRNAseq
    data_files:
      - split: train
        path: data/scRNAseq/*.h5ad
  - config_name: sample
    data_files:
      - split: train
        path: data/sample/*.h5ad
  - config_name: synthetic
    data_files:
      - split: train
        path: data/synthetic/*.npz
  - config_name: annotations
    data_files:
      - split: train
        path: labels/annotations.parquet
  - config_name: labels
    data_files:
      - split: train
        path: labels/labels.parquet
---

# scShapeBench

A benchmark dataset for single-cell shape analysis with four configurations: **real-world scRNA-seq** data, **synthetic** data, **annotator labels**, and **aggregated labels**.


## Dataset Summary

scShapeBench is a curated collection of datasets assembled for benchmarking computational methods in single-cell shape analysis. It is organized into four configurations:

### scRNAseq
Real-world single-cell gene expression datasets. Each dataset is stored as an individual AnnData file with precomputed PCA embeddings and Leiden clustering.

- **Total cells:** 2,547,517
- **Total datasets:** 102
- **Total size:** ~116 GB
- **Format:** AnnData (`.h5ad`)

### synthetic
Synthetically generated single-cell data for controlled benchmarking.

- **Format:** NumPy compressed array (`.npz`) + per-sample metadata (`.json`)

### annotations
Per-dataset shape labels from 9 independent annotators. Each annotator assigned one or more shape categories to each dataset they reviewed.

- **Total datasets labeled:** 97
- **Annotators:** 9
- **Shape categories:** `archetypal`, `multi_branch`, `simple_traj`, `clusters`
- **Format:** Parquet, long format (one row per dataset–annotator pair)

### labels
Aggregated shape labels derived from the 9 annotator labels using three strategies described in the paper.

- **Total datasets:** 97
- **Aggregations:** `majority`, `soft`, `confidence_weighted`, `union`
- **Format:** Parquet, long format (one row per dataset–aggregation pair); values are floats in [0, 1] per shape category

## Dataset Structure

```
scShapeBench/
├── data/
│   ├── scRNAseq/
│   │   ├── SCD-0001.h5ad
│   │   ├── SCD-0002.h5ad
│   │   ├── ...
│   │   └── SCD-0112.h5ad
│   └── synthetic/
│       ├── sample_00000.npz
│       ├── sample_00000.json
│       ├── ...
├── labels/
│   ├── annotations.parquet  # Per-annotator shape labels (9 annotators)
│   └── labels.parquet       # Aggregated labels (majority, soft, confidence_weighted, union)
├── cell_metadata.csv    # Combined cell-level metadata (2.5M rows)
├── gene_metadata.csv    # Combined gene-level metadata
├── dataset_index.csv    # Per-file summary: dimensions, size
├── croissant.json       # Croissant 1.1 metadata
└── README.md
```

### File Naming

**scRNAseq:** Files are named `SCD-XXXX.h5ad` where XXXX is a zero-padded index. The numbering is not contiguous (e.g., SCD-0031, SCD-0032, SCD-0034–0036 are absent).

**synthetic:** Files are named `sample_XXXXX.npz` / `sample_XXXXX.json` with a zero-padded 5-digit index.

### Per-File Contents

Each `.h5ad` file contains:

| Component | Description |
|-----------|-------------|
| `X` | Gene expression matrix (cells × genes), log-normalized |
| `obs` | Cell metadata: `n_genes`, `leiden` (cluster assignment) |
| `var` | Gene metadata: `gene_ids`, `feature_types`, `genome`, `n_cells`, `highly_variable`, etc. |
| `obsm['X_pca']` | Precomputed PCA embeddings |
| `uns` | Clustering and HVG parameters |

### Dataset Index

The `dataset_index.csv` file provides per-file summary statistics:

| Column | Description |
|--------|-------------|
| `file_id` | Dataset identifier (e.g., SCD-0001) |
| `filename` | Filename |
| `n_cells` | Number of cells |
| `n_genes` | Number of genes |
| `file_size_bytes` | File size in bytes |

Dataset sizes range from 1,163 cells (SCD-0006) to 434,340 cells (SCD-0037).

## Usage

### scRNAseq config

```python
import scanpy as sc
import pandas as pd

# Load a single dataset
adata = sc.read_h5ad("data/scRNAseq/SCD-0001.h5ad")
print(adata)

# Browse available datasets
index = pd.read_csv("dataset_index.csv")
print(index.sort_values("n_cells", ascending=False).head())

# Load cell metadata across all datasets
cell_meta = pd.read_csv("cell_metadata.csv")
print(cell_meta.groupby("file_id").size())
```

### synthetic config

```python
import numpy as np
import json

# Load a single synthetic sample
data = np.load("data/synthetic/sample_00000.npz")
meta = json.load(open("data/synthetic/sample_00000.json"))
```

### annotations config

```python
import pandas as pd

annotations = pd.read_parquet("labels/annotations.parquet")
# columns: dataset_id, annotator_id, archetypal, multi_branch, simple_traj, clusters

# Fraction of annotators who labeled a dataset as multi_branch
agreement = annotations.groupby("dataset_id")["multi_branch"].mean()
```

### labels config

```python
import pandas as pd

labels = pd.read_parquet("labels/labels.parquet")
# columns: dataset_id, aggregation, archetypal, multi_branch, simple_traj, clusters

# Get majority-vote labels (binary)
majority = labels[labels["aggregation"] == "majority"].set_index("dataset_id")

# Get union labels (any annotator assigned the class)
union = labels[labels["aggregation"] == "union"].set_index("dataset_id")
```

## Sample Data

The full scRNAseq configuration is ~116 GB. A representative sample (~2.6 GB) is available at [`data/sample/`](data/sample/) and accessible as the `sample` config:

```python
from datasets import load_dataset
ds = load_dataset("scShape-Benchmark/scShapeBench", "sample")
```

It contains **9 `.h5ad` files** selected to span the range of dataset sizes present in the full benchmark:

| File | Cells | Size |
|------|------:|-----:|
| `SCD-0006.h5ad` | 1,163 | 61 MB |
| `SCD-0014.h5ad` | 1,886 | 57 MB |
| `SCD-0015.h5ad` | 1,973 | 57 MB |
| `SCD-0010.h5ad` | 8,686 | 56 MB |
| `SCD-0002.h5ad` | 8,368 | 227 MB |
| `SCD-0052.h5ad` | 9,669 | 211 MB |
| `SCD-0062.h5ad` | 7,976 | 223 MB |
| `SCD-0071.h5ad` | 18,890 | 858 MB |
| `SCD-0074.h5ad` | 21,225 | 823 MB |

**Selection criteria:** Files were chosen to cover three size tiers — small (<80 MB, 4 files), medium (200–230 MB, 3 files), and large (>800 MB, 2 files) — so reviewers can inspect data at each scale without downloading the full corpus. Within each tier, files were selected by ascending file size from the full `dataset_index.csv`. All lightweight metadata files (`labels/`, `dataset_index.csv`, `cell_metadata.csv`, `gene_metadata.csv`) are available in full regardless of which scRNAseq files are downloaded.

## Croissant Metadata

This dataset includes a `croissant.json` file conforming to the [Croissant 1.1](https://docs.mlcommons.org/croissant/) metadata standard. This enables programmatic discovery and loading of dataset metadata through compatible tools.

## Citation

<!-- Add citation information here when available -->

## License

This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. You are free to share and adapt the material for any purpose, provided appropriate credit is given.