Atlas-scale spatially aware clustering with support for 3D and multimodal data using SpatialLeiden

This paper presents an extension of the SpatialLeiden algorithm that enables scalable, atlas-scale clustering of multi-sample, 3D, and multimodal spatial omics data through flexible neighbor-graph multiplexing, delivering superior modularity and stability on standard hardware while integrating with the scverse ecosystem.

The Gist

Imagine you have a massive library of books, but instead of words, the pages are filled with tiny dots representing cells in a body. Some of these books are just flat pictures (2D), some are thick stacks of pages that need to be read in order to see a 3D story (3D), and some books even have pictures, text, and audio tracks all at once (multimodal).

The problem? Trying to organize this library is a nightmare. If you just look at the text, you might miss the pictures. If you look at one page, you miss the story of the whole stack. And if you have hundreds of these books from different people, the "ink" might look slightly different on each one due to how they were printed (batch effects).

Enter "SpatialLeiden," a new super-organizer for biological data.

Here is how this paper explains it, using simple analogies:

1. The Problem: The "Messy Library"

Scientists use a technology called Spatial Omics to map where every cell is in a tissue and what it's doing. But as data gets bigger (like mapping an entire human brain) and more complex (adding 3D depth and different types of data like proteins and images), old tools break. They are like trying to organize a library with a single, rigid rulebook that doesn't work for 3D books or books with audio tracks.

2. The Solution: The "Smart Neighborhood Map"

The authors upgraded their tool, SpatialLeiden, to be a "universal librarian." Instead of forcing all data into one shape, it builds a flexible neighborhood map.

• The Analogy: Imagine you are organizing a massive city festival.
  • Old tools might only look at who is standing next to whom in a single line.
  • SpatialLeiden looks at the whole city. It knows that people standing next to each other in a 2D photo are neighbors, but it also knows that the person on the 2nd floor of a building is a neighbor to the person on the 3rd floor (3D). It even knows that the person wearing a red hat (RNA data) and the person holding a sign (histology image) are part of the same group, even if they look different.

3. Key Features (The "Superpowers")

A. The "Atlas-Scale" Power

The tool can handle massive datasets. The authors tested it on the Allen Brain Cell Atlas, which is like trying to organize 150 slices of a human brain all at once.

• The Result: It successfully grouped cells into "neighborhoods" (domains) that perfectly matched the known map of the brain. It did this so efficiently that it could run on a standard laptop, not just a supercomputer.

B. The "3D Stitching" Magic

When scientists slice tissue to look at it, they get flat pages. To see the 3D structure (like a tumor growing inside a lymph node), they need to stitch these pages together.

• The Analogy: Imagine trying to understand a 3D sculpture by looking at 2D X-rays. Old tools might treat each X-ray as a separate island. SpatialLeiden connects the islands. It realizes that a cell on Page 1 is physically connected to a cell on Page 2, creating a continuous 3D neighborhood. This revealed the true 3D structure of cancer tissue, showing how different cell types surround a tumor.

C. The "Multimodal" Mixer

Modern science often measures the same cell in different ways: its genes (RNA), its proteins, and its shape (images).

• The Analogy: Think of a cell as a person.
  • RNA is what they say.
  • Proteins are what they wear.
  • Images are what they look like.
  • Old tools might try to organize people based only on what they say. SpatialLeiden listens to what they say, looks at what they wear, and checks their appearance all at once. It creates a group based on the whole person.
  • Real-world win: In a study of colorectal cancer, using just the RNA data missed some details, and just the image data missed others. But when SpatialLeiden combined them, it perfectly identified dangerous "dysplastic" (pre-cancer) areas that neither method could find alone.

4. Why This Matters

• It's Flexible: It doesn't care how you prepared your data. You can use any method to clean up the data first, and SpatialLeiden will still work.
• It's Scalable: As technology improves and we get millions of data points, this tool won't crash. It scales linearly, meaning it gets slower at a predictable, manageable pace.
• It's Open: It's free software that fits into the standard tools scientists already use, making it easy for everyone to adopt.

The Bottom Line

This paper introduces a universal translator and organizer for the future of biology. Whether you are mapping a whole brain, studying a 3D tumor, or combining images with genetic data, SpatialLeiden helps scientists see the "big picture" clearly, turning a chaotic mess of data into a coherent map of life.

Technical Summary

1. Problem Statement

Spatially resolved transcriptomics (SRT) generates high-dimensional datasets that capture gene expression within the intact tissue context. While essential for constructing spatial atlases and analyzing tissue architecture, interpreting these datasets is challenging due to:

• Scalability: Existing clustering methods struggle to scale to atlas-level datasets (hundreds of samples, millions of cells).
• Complexity: Modern studies involve multi-sample integration, 3D volumetric data (from serial sections), and multimodal data (combining transcriptomics, histology, proteomics, etc.).
• Inconsistency: Batch effects, cross-slice variability, and the lack of unified frameworks for integrating diverse data modalities often lead to inconsistent domain alignment and biological interpretation.
• Limitations of Current Tools: Many specialized tools are either computationally inefficient, limited in modality count, or rigid in their preprocessing requirements.

2. Methodology: SpatialLeiden

The authors extend their existing SpatialLeiden algorithm to create a generalized framework capable of handling atlas-scale, 3D, and multimodal data. The core methodology relies on flexible neighbour-graph multiplexing on batch-corrected latent spaces.

• Core Mechanism: SpatialLeiden applies the Leiden community detection algorithm to a collection of neighbourhood graphs. It is agnostic to the specific graph creation method, allowing users to choose optimal integration strategies (e.g., scVI, Harmony, BBKNN) for batch correction.
• Multi-Sample & Atlas-Scale Integration:
  • Data from multiple samples are concatenated into a single AnnData object.
  • Batch correction is performed in the gene expression latent space before constructing the k-NN graph.
  • The algorithm integrates spatial neighbourhood graphs with gene expression graphs to identify coherent domains across samples.
• 3D Volumetric Clustering:
  • For serially sectioned data, the algorithm incorporates z-axis coordinates directly into the spatial neighbourhood graph.
  • This enables the detection of 3D cellular neighbourhoods spanning across slices, preserving inter-slice continuity.
  • Requires precise spatial alignment of serial sections prior to graph construction.
• Multimodal Integration:
  • Utilizes the MuData structure (from the mudata package) to handle multiple data layers (e.g., RNA, H&E histology, ATAC-seq).
  • Each modality is treated as an extendable adjacency element.
  • Users can assign custom weights to different layers (e.g., RNA vs. Histology) to control their influence on the final clustering.
  • Compatible with the scverse ecosystem (e.g., scanpy, squidpy) and tools like LazySlide for image processing.

3. Key Contributions

• Generalized Framework: A unified approach that handles single-sample, multi-sample, 3D, and multimodal data without requiring specialized, modality-specific algorithms.
• Scalability: Demonstrated linear scaling, capable of processing atlas-scale datasets (e.g., ~150 serial sections) on standard hardware (high-end notebooks).
• Preprocessing Agnosticism: By focusing solely on neighbour graphs and clustering, the method allows researchers to leverage custom preprocessing pipelines and future spatially-aware batch correction methods.
• Open-Source Implementation: The tool is released as a Python package (spatialleiden) compatible with the widely used scverse ecosystem, facilitating broad adoption.

4. Key Results

The authors validated SpatialLeiden across three distinct use cases:

• Atlas-Scale Multi-Sample Clustering (DLPFC & ABC Atlas):
  • DLPFC: Successfully integrated latent spaces using scVI, Harmony, and BBKNN, maintaining clustering accuracy (high Adjusted Rand Index and Normalized Mutual Information) across patients.
  • Allen Brain Cell (ABC) Atlas: Clustered ~150 serial MERFISH sections. The resulting domains aligned closely with the Common Coordinate Framework (CCF) parcellations. The method revealed smooth composition transitions across brain slices and refined multi-scale structures (e.g., isocortical layers) better than non-integrated clustering.
• 3D Reconstruction (Open-ST Metastatic Lymph Node):
  • Applied to aligned Open-ST data, SpatialLeiden identified 3D cellular neighbourhoods spanning multiple sections.
  • Domain composition remained stable across profiled sections, successfully delineating tumor microenvironment organization (e.g., distinct profiles for cancer-associated fibroblasts, macrophages, and immune cells).
• Multimodal Integration (Visium HD Colorectal Cancer):
  • Integrated matched H&E histology and RNA data.
  • Synergy: While histology alone refined smooth muscle clusters and RNA alone separated dysplasia from neoplasia, joint clustering simultaneously resolved both features, providing a more comprehensive tissue map than either modality alone.

5. Significance and Impact

• Biological Insight: The ability to perform robust, efficient clustering on atlas-scale and 3D datasets enables the construction of comprehensive spatial atlases and the discovery of complex tissue architectures that were previously difficult to resolve.
• Computational Efficiency: By demonstrating linear scaling and low memory footprint, SpatialLeiden lowers the barrier to entry for analyzing massive spatial omics datasets, removing the need for high-performance computing clusters for many use cases.
• Future-Proofing: The modular design (agnostic to preprocessing and extensible to new modalities) positions SpatialLeiden as a foundational tool for the evolving field of spatial biology, accommodating the exponential growth in data complexity and volume.
• Community Adoption: Integration with scverse and availability as open-source software ensures that the method can be easily adopted by the broader bioinformatics community, fostering reproducibility and standardization in spatial analysis.

Limitations Noted: The authors acknowledge that benchmarking remains challenging due to a lack of high-quality real-world multimodal datasets and the reliance on precise registration for 3D clustering, where misalignment can distort results. Additionally, manual weighting of modalities in multimodal integration introduces complexity.