Cell-NET Dataset

A large-scale multimodal dataset for single-cell analysis.

Overview

Cell-NET is a high-resolution multimodal single-cell dataset comprising 14.4 million single-cell samples. It bridges microscopic imaging, gene expression, spatial transcriptomics, and natural language annotations, enabling a bottom-up, interpretable approach to modeling biological systems.

🔹 Key Features

• Multi-scale imaging: Each single-cell instance has three hierarchical image patches:
  • Cell-level (64×64 px) – Captures fine-grained morphology.
  • Tissue-level (200 µm window) – Provides microenvironment context.
  • Whole-slide level – Offers global histopathological insights.
• Transcriptomics Data: RNA-seq profiles capturing over 36,000 genes per cell.
• Spatial transcriptomics integration: Mapping gene expression to tissue structure.
• Cell-to-cell communication (CCC) graphs: Infers interactions between neighboring cells.
• Vision-Language Annotations: Each cell has descriptive captions generated to summarize morphology, tissue context, and pathology.

---

📥 Download

🔗 Dataset Link

---

🖼️ Figures

Dataset Structure

Figure 1: Cell-NET integrates imaging, transcriptomics, and structured metadata for multimodal single-cell analysis.

📑 Data Structure

The Cell-NET dataset is stored in an HDF5 file, containing structured single-cell data across multiple modalities.

/Cell_0001  
├── expression          # Gene expression profile (36,601 genes)
├── images             # Multi-scale image patches
│   ├── cell_patch     # Cropped cell-level image (uint8)
│   ├── tissue_patch   # Larger tissue-level image (uint8)
│   └── WSI_patch      # Whole-slide image patch (uint8)
├── attributes.json     # Structured metadata (JSON)
├── caption.txt         # Natural language description (str)
└── CCC                 # Cell-Cell Communication (sparse matrix)

🔬 Multimodal Components

1️⃣ Multi-Level Imaging

• Cell-Level (64×64 pixels): Captures nuclear morphology and fine-grained details.
• Tissue-Level (200 µm window): Provides neighborhood context of the cell.
• WSI-Level: Whole-slide view to integrate local and global tissue information.

2️⃣ Gene Expression Profiles

• Derived from 10x Visium HD & Xenium platforms.
• Gene list file includes all detected genes per cell.
• Expression matrix provides log-normalized read counts.

3️⃣ Cell-Cell Communication (CCC)

• Computed using CellChat and COMMOT.
• Models ligand-receptor interactions within a 200 µm radius.
• Supports spatial organization and cellular interaction analysis.

4️⃣ Language Captioning

Each cell receives an automatically generated caption describing its morphological characteristics, local environment, and potential pathology.
Example:

"The microscopic image reveals pleomorphic cells with enlarged, hyperchromatic nuclei, prominent nucleoli, eosinophilic cytoplasm, and areas of dense and sparse cellularity. Keratinization and potential intercellular bridges suggest squamous differentiation. Disorganized tissue architecture and irregular cell clustering indicate a likely squamous cell carcinoma, typically found in the head and neck region."

📑 Labeled Attributes

Each cell in Cell-NET is annotated with structured metadata, allowing researchers to analyze multimodal relationships between morphology, spatial context, and transcriptomics.

Attribute	Example Value
source	"Human"
tissue	"Lymph node"
cell_type	"T-cell"
cell_disease_state	"Cancer"
tissue_disease_state	"Cancer"
Position_in_tissue	"(1945, 345)"
Position_in_WSI	"(893021, 398472)"
cell_diameter	"9.0 µm"
st_technology	"Visium HD"

---

📊 Applications

Cell-NET enables various downstream tasks, including:

✅ 1. Cell Type Classification

• Benchmark for single-cell vision models.
• Supports few-shot learning approaches.

✅ 2. Gene Expression Prediction

• Infers gene profiles from histological images.
• Bridges imaging and transcriptomics.

✅ 3. Spatial Domain Segmentation

• Identifies functional tissue regions.
• Useful for tumor microenvironment mapping.

✅ 4. Cell-Cell Interaction Modeling

• Analyzes spatially-resolved signaling networks.
• Supports multi-modal biological modeling.

---

📌 Citation

If you use Cell-NET, please cite:

@article{cellnet2025,
  author  = {},
  title   = {Cell-NET13M: Redefining Microscopic Insights with a Multimodal Single-Cell Spatial Transcriptomics Dataset},
  journal = {Under Review},
  year    = {2025}
}