A unified spatial transcriptome profiling of ten mouse organs

This paper presents a unified, high-quality spatial transcriptomic dataset generated via the Stereo-seq platform across ten mouse organs, providing matched histological images, cell-type annotations, and multi-resolution expression matrices to serve as a standardized resource for developing and benchmarking deep learning models in spatial biology.

The Gist

Imagine you have a giant, incredibly detailed map of a bustling city. Usually, scientists can either look at the city from a satellite (seeing the whole layout but missing the details) or walk down the street and talk to one person at a time (seeing the details but losing the big picture).

This paper is about creating a super-map that does both at the same time, but for the inside of a mouse's body.

Here is the story of what the researchers did, explained simply:

1. The Problem: The "Missing Puzzle Pieces"

In the world of biology, scientists are trying to understand how organs work by looking at two things at once:

• The Blueprint: Which genes are active in a specific spot? (The "who is doing what" list).
• The Photo: What does the tissue actually look like under a microscope? (The "where are they" picture).

The problem is that most existing maps are either low-quality, messy, or only cover one tiny part of the body. It's like trying to build a massive AI robot to understand cities, but you only have a few blurry photos of a single street corner. The robot can't learn well because it doesn't have enough high-quality data.

2. The Solution: The "Grand Tour" of the Mouse

The team at BGI Research decided to take a "Grand Tour" of a mouse. They didn't just look at one organ; they mapped 10 different organs (the brain, kidneys, lungs, skin, heart, etc.) all at once.

They used a high-tech camera system called Stereo-seq. Think of this camera as a super-powered microscope that can take a photo of a tissue slice and instantly tag every single gene in every single cell with a GPS coordinate.

• The Scale: They took 23 different "slices" (like cutting a loaf of bread) from 21 different chips.
• The Variety: They used different staining methods (like taking photos in color vs. black and white) to make sure the data works for everyone.

3. Two Ways to Look at the Data: "The Pixel" vs. "The Neighborhood"

The researchers created two versions of their map for each organ, like offering a map in two different zoom levels:

• The "Cell-Bin" (The Single-Pixel View): This is the high-definition mode. It looks at one single cell at a time. It's like zooming in so close you can see the face of every person in the crowd. This is great for identifying exactly who is there (e.g., "That's a specific type of immune cell").
• The "Bin-50" (The Neighborhood View): Sometimes, the photo isn't clear enough to see individual faces. So, they grouped cells into little 25x25 micrometer squares (like a city block). This is the "neighborhood" view. It's slightly less detailed but still very useful for seeing the general layout of the city.

4. Why This Matters: Training the "Digital Brain"

Why go through all this trouble? Because scientists are building AI models (digital brains) to understand biology.

• The Analogy: Imagine you are teaching a child to recognize animals. If you only show them one blurry picture of a dog, they might think a cat is a dog. But if you show them 1,000 high-quality, clear photos of dogs, cats, and birds from different angles, they learn instantly.
• The Result: This dataset is a "textbook" of 1,000 high-quality photos. It allows AI models to learn how genes and tissues interact in 3D space. This helps researchers discover new diseases, understand how organs develop, and test new drugs faster.

5. The Proof: It Actually Works

The team didn't just dump the data; they proved it was good.

• Consistency: They looked at two slices of the same brain and found the "neighborhoods" matched perfectly.
• Accuracy: They checked if the map matched known landmarks. For example, in the testis, the map correctly showed that sperm cells are in the center and stem cells are on the outside, just like a real textbook says.
• The "Cell-Bin" Advantage: They showed that the high-definition "Cell-Bin" view could spot tiny, rare cell groups that the "Neighborhood" view missed. It's like the difference between seeing a crowd from a helicopter vs. walking through it; the helicopter sees the shape, but walking through lets you see the individuals.

The Bottom Line

This paper is a gift to the scientific community. It's a massive, standardized, high-quality library of mouse organ maps. Just as Google Maps changed how we navigate the world, this dataset changes how scientists navigate the complex world inside our bodies, making it easier to build better AI tools to cure diseases.

Technical Summary

1. Problem Statement

Spatial transcriptomics (ST) has emerged as a transformative technology for linking tissue architecture with gene expression. However, the development of robust deep learning models for ST (e.g., for automated cell annotation or multimodal analysis) is currently hindered by a lack of large-scale, high-quality, and unified training datasets.

• Data Scarcity & Heterogeneity: Existing public datasets are often fragmented, generated across different platforms, protocols, and processing strategies, leading to technical heterogeneity and poor cross-dataset comparability.
• Resolution Trade-offs: There is a need to systematically evaluate the trade-offs between different spatial resolutions (e.g., single-cell vs. aggregated bins) to determine the optimal data format for specific analytical tasks.
• Standardization: The field lacks a standardized resource encompassing diverse tissue types with matched histological images and expression matrices to benchmark new algorithms.

2. Methodology

The authors generated a comprehensive dataset using the Stereo-seq platform, known for its subcellular resolution (500 nm) and broad transcriptome coverage.

• Sample Collection:
  • Subjects: Healthy 12-week-old C57BL/6J mice.
  • Organs: 10 distinct organs (brain, kidney, lung, thymus, large intestine, skin, spleen, ovary, testis, uterus).
  • Scale: 23 tissue sections processed across 21 chips.
• Experimental Workflow:
  • Preparation: Fresh-frozen tissues were embedded in OCT, cryosectioned (10 µm), and mounted on Stereo-seq chips.
  • Staining: Each section was paired with either ssDNA (fluorescence) or H&E (brightfield) staining to enable cell segmentation.
  • Library Prep & Sequencing: Utilized Stereo-seq V1.2 and V1.3 kits. mRNA was captured by spatially barcoded probes, reverse-transcribed, and sequenced on the MGI DNBSEQ-Tx platform.
• Data Processing Pipeline:
  • Raw Data: Processed using the SAW (v8.1.0) pipeline. Reads were aligned to the mm10 mouse genome (STAR), and UMIs were collapsed.
  • Matrix Generation:
    • Cell-bin: For high-quality images allowing reliable cell segmentation, single-cell resolution expression matrices were generated using the CellBin package.
    • Bin-50: For sections where single-cell segmentation was unreliable or as a comparative standard, 50 × 50 pixel (25 µm × 25 µm) square bins were aggregated.
  • Annotation: Cell types were annotated using cell2location (via the STOmics SDAS repository), leveraging three independent reference single-cell atlases (Mouse Cell Atlas, Tabula Muris Senis, and original cell2location data).
  • Validation: Annotations were validated via cross-slice consistency, concordance with canonical marker genes, and comparison against known anatomical atlases (e.g., Allen Mouse Brain Atlas).

3. Key Contributions

• Unified Dataset Creation: The paper presents the first large-scale, unified ST dataset covering 10 major mouse organs with matched histological images, generated under a standardized protocol.
• Dual-Resolution Resource: The dataset provides both single-cell resolution (cell-bin) and intermediate resolution (bin-50) matrices for the same samples, allowing direct comparison of resolution impacts.
• Benchmarking Standard: By including diverse tissue types and staining methods (ssDNA/H&E), the dataset serves as a robust benchmark for developing and testing deep learning models in spatial transcriptomics.
• Open Access: All raw and processed data (images, GEF matrices, annotations) are publicly available via the STOmics DB and Zenodo.

4. Key Results

• Data Quality & Statistics:
  • Bin-50: Consistently captured >700 genes and >1,000 UMIs per spot across all tissues.
  • Cell-bin: Over 76% of sections allowed for accurate single-cell segmentation. These matrices showed a median of >120 genes and >170 UMIs per cell.
  • Scalability: Larger bins (200 µm × 200 µm) captured significantly higher gene counts (>4,500 genes) and UMIs (>16,000), demonstrating the trade-off between spatial resolution and transcriptome depth.
• Annotation Accuracy:
  • Spatial Coherence: Annotations accurately recapitulated known histological structures (e.g., distinct layers in the brain, centripetal distribution of germ cells in the testis, and S1 proximal tubules in the kidney).
  • Consistency: High concordance was observed between biological replicates (different slices of the same organ) and between different staining methods (ssDNA vs. H&E).
  • Marker Validation: Canonical markers (e.g., Meis2 in brain, Rag1 in thymus, Spp2 in kidney, Slc4a1 in spleen) showed precise spatial co-localization with annotated cell types.
• Resolution Comparison (Cell-bin vs. Bin-50):
  • Global Level: Both resolutions yielded highly consistent cell type proportions.
  • Local/Structural Level: Cell-bin resolution demonstrated superior performance.
    • It preserved intact hollow structures (e.g., tubules, follicles) better than bin-50.
    • It identified rare or specific cell populations missed by bin-50 (e.g., macrophages in the testis, endothelial/epithelial cells in the ovary, and diverse subtypes in the kidney).
    • Bin-50 often failed to resolve fine-grained cellular heterogeneity within complex tissue architectures.

5. Significance

• Accelerating AI Development: This dataset provides the "ground truth" data necessary to train and validate next-generation deep learning models for spatial transcriptomics, addressing the critical bottleneck of data scarcity.
• Methodological Insight: The study definitively demonstrates that while bin-50 data is useful for broad tissue mapping, single-cell resolution (cell-bin) is essential for resolving fine-grained tissue architecture, rare cell types, and complex cellular interactions.
• Reproducibility: By standardizing the workflow across 10 organs and providing matched images and matrices, the authors set a new standard for reproducibility in spatial omics research.
• Resource Availability: The public availability of this multi-organ, multi-resolution dataset enables the global research community to perform multimodal analysis and develop tools that are generalizable across different tissue types.