Analysis of multicellular anatomical structures from spatial omics data using sosta

This paper introduces *sosta*, an open-source Bioconductor package that enables the direct analysis of multicellular anatomical structures in spatial omics data, shifting the focus from single-cell arrangements to biologically relevant tissue-level organization.

The Gist

Imagine you are looking at a bustling city from a helicopter.

The Old Way (Current Methods):
Most scientists studying biological tissues right now are like people looking at the city through a telescope that only focuses on individual people. They count how many people are wearing red shirts, how many are running, and who is standing next to whom. They map out neighborhoods based on who is standing where. This is great for understanding the "cells" (the people), but it misses the bigger picture: the buildings, parks, and districts themselves.

In biology, these "buildings" are anatomical structures like crypts in the intestine or germinal centers in the tonsils. These structures are made of many cells working together, and their shape and size tell a huge story about health and disease. But current tools often ignore the shape of the building and only look at the people inside.

The New Way (This Paper's Solution):
The authors of this paper, Samuel, Helena, and Mark, have built a new tool called sosta (Spatial Omics Structure Analysis). Think of sosta as a tool that switches the camera from a "people counter" to a "city planner."

Instead of just counting cells, sosta looks at the density of cells to draw the outlines of the actual buildings (anatomical structures). Once it draws the building, it can measure:

• How wide the building is.
• Is it a perfect circle, or is it squashed and weird?
• How does the "traffic" (gene expression) change as you walk from the front door to the back wall?

Two Real-World Examples from the Paper

1. The Intestine: From a Smooth Tunnel to a Crumpled Mess
Imagine the healthy lining of your intestine as a series of perfect, smooth, finger-like tubes (called crypts). When cancer starts to form, these tubes get distorted, thick, and irregular.

• What the old tools did: They might have said, "There are more cancer cells here."
• What sosta did: It measured the width and shape of the tubes. It found that as the tissue moved from healthy to pre-cancer to full cancer, the tubes got wider and their shapes became more chaotic. It was like measuring the "crumbliness" of a piece of paper to see how damaged it is. This gave them a new way to spot the early stages of disease that just counting cells might miss.

2. The Tonsil: Finding the "Traffic Flow"
Inside your tonsils, there are special zones called "Germinal Centers" where immune cells (B-cells) mature. These zones have a specific layout: a "Light Zone" and a "Dark Zone." Cells move through them like cars on a highway, changing as they go.

• What the old tools did: They might have just said, "These cells are in the Light Zone."
• What sosta did: It drew the outline of the Germinal Center and created a virtual ruler from one side to the other. It then asked, "How does the activity of specific genes change as you move along this ruler?"
• The Result: They found 1,500+ genes that act like traffic lights, turning on and off in a specific pattern as cells move from the "Dark" to the "Light" zone. This revealed the hidden "instruction manual" of how immune cells mature, which was hard to see before.

Why This Matters (The "So What?")

The authors also warn about a statistical trap. If you measure 10 different buildings in one hospital room, those 10 buildings aren't 10 totally independent experiments; they are all part of the same room. If you don't account for that, you might think you found a pattern that isn't really there. sosta is smart enough to handle this, ensuring the results are real.

The Bottom Line

sosta is an open-source software package (like a free app for scientists) that helps researchers stop looking just at the "people" in the tissue and start looking at the "architecture."

By treating anatomical structures as the main unit of analysis, it bridges the gap between:

1. Microscopes: Which show us the cells.
2. Pathologists: Who have spent 100 years looking at tissue shapes to diagnose disease.

It allows scientists to ask questions like, "Is the building getting too wide?" or "Is the hallway getting too crowded?"—questions that are crucial for understanding how diseases like cancer develop, but which were previously very hard to answer with data.

Technical Summary

1. Problem Statement

Spatial omics technologies have revolutionized biology by enabling high-resolution quantification of molecular features while preserving spatial context. However, current computational analysis methods predominantly focus on single-cell arrangements (e.g., cell neighborhoods, spatial domains, or cell-cell interactions) or the detection of spatially variable features/gradients.

The authors identify a critical gap: biological function often emerges from multicellular anatomical structures (e.g., crypts, germinal centers, tissue layers), yet there is a lack of general frameworks to analyze data at this specific "structure level." Existing methods often fail to account for the heterogeneity inherent in these multicellular arrangements or to handle multi-sample, multi-condition datasets where nested correlations exist (e.g., multiple structures within a single patient sample).

2. Methodology

The authors introduce sosta (spatial omics structure analysis), an open-source Bioconductor R package designed to bridge classical histopathology and spatial omics. The methodology is built on a structure-based analysis workflow:

• Density-Based Reconstruction:

  • Instead of relying on concave hulls or manual segmentation, sosta reconstructs anatomical structures by estimating the intensity of a point pattern (e.g., centroids of specific cell types, transcript locations, or spatial domains).
  • It utilizes a Gaussian kernel density estimator (based on the spatstat package) to generate a density image.
  • Thresholding: Structures are defined by applying a cut-off threshold to the density image. The threshold is automatically estimated using the average of the first and second modes of the intensity distribution ($c = \frac{m_1 + m_2}{2}$), though manual adjustment is supported.
  • The resulting structures are stored as polygons (using the sf package).

• Quantification of Features:

  • Morphological Metrics: Once reconstructed, the package calculates geometric features for each structure, such as area, circularity, eccentricity, and specific metrics like "fibre width" (distance between inner and outer borders).
  • Cell-Based Features: It computes intersections between cells and structures and measures distances to structure borders, enabling the analysis of cell type composition dynamics relative to anatomical landmarks.

• Differential Feature Analysis:

  • To address the statistical challenges of multi-sample, multi-condition data, the authors advocate for Linear Mixed Effects (LME) modeling.
  • They demonstrate that ignoring the nested correlation structure (e.g., multiple structures within one slide, multiple slides per patient) leads to a significant increase in False Discovery Rates (FDR). LME models effectively retain FDR control by treating structures as non-independent experimental units.

• Gradient Analysis:

  • The framework allows for the definition of anatomically relevant axes (e.g., Light Zone to Dark Zone in germinal centers) to quantify gene expression gradients along these unified anatomical trajectories, rather than relying solely on discrete cell type labels.

3. Key Contributions

• New Paradigm: Introduces "structure-based analysis" as a distinct genre of spatial omics analysis, shifting the unit of analysis from individual cells to multicellular anatomical structures.
• Software Tool (sosta): Provides a modular, extensible, and automated Bioconductor package for reconstructing, characterizing, and comparing anatomical structures across scales.
• Statistical Framework: Highlights and implements strategies (LME models) to correctly handle the nested correlation structure inherent in spatial omics data, preventing inflated false discovery rates.
• Benchmarking: Demonstrates that density-based reconstruction (as implemented in sosta) outperforms concave hull approaches (like imcRtools, SPIAT, CellCharter) in accuracy, particularly in noisy settings, and offers superior automated parameter selection compared to GRIDGENE.

4. Results and Applications

The authors validated sosta using two distinct public datasets:

• Case Study 1: Colorectal Cancer (CRC) Transformation:

  • Data: Spatial transcriptomics of healthy colon (HC), tubulovillous adenomas (TVA), and colorectal cancer (CRC).
  • Findings:
    • Fibre Width: A metric representing crypt thickness increased progressively from HC to TVA to CRC, correlating strongly with inferred pseudo-time ($\rho=0.46$).
    • Shape Index: Reconstructed crypts showed a transition from regular shapes in healthy tissue to irregular shapes in premalignant regions ($\rho=0.22$).
    • Transition Detection: The method successfully identified "transition crypts" (areas of de novo transformation) by analyzing cell type frequencies adjacent to reconstructed structures, revealing candidates missed by manual annotation.

• Case Study 2: Human Tonsil Germinal Centers (GCs):

  • Data: Spatial transcriptomics of human tonsil tissue.
  • Findings:
    • The authors reconstructed GCs and defined anatomical axes from the Light Zone (LZ) to the Dark Zone (DZ).
    • Using the Maximal Information Coefficient (MIC), they identified 1,569 genes with spatially varying expression profiles along the LZ-DZ axis.
    • This approach recovered known markers of B-cell proliferation and maturation (e.g., MKI67, CIITA) without relying on pre-defined discrete cell type labels, demonstrating the ability to recover interpretable biological gradients.

5. Significance

• Bridging Disciplines: The work effectively bridges the gap between traditional histopathology (which relies on visualizing tissue architecture) and modern spatial omics (which provides molecular depth).
• Scalability: The approach is versatile, applicable to structures ranging from small clusters (e.g., tertiary lymphoid structures) to large tissue sections containing thousands of cells.
• Biological Discovery: By focusing on anatomical structures, researchers can uncover morphological changes and gene expression gradients that are biologically meaningful but may be obscured when analyzing cells in isolation.
• Community Resource: As an open-source Bioconductor package, sosta lowers the barrier for the community to perform rigorous, structure-aware spatial analysis, promoting reproducibility and standardization in the field.