A partition-based spatial entropy for co-occurrence analysis with broad application.

The paper introduces Regional Co-occurrence Entropy (RCE), a novel spatial entropy measure that quantifies context-dependent interactions between categorical points and their environments, demonstrating its broad utility across diverse fields such as spatial biology, urban geography, and ecology.

The Gist

Imagine you are a detective trying to solve a mystery, but instead of looking for clues in a single room, you are looking at an entire city, a forest, or even a microscopic slice of a brain. Your job is to figure out who is hanging out with whom, and where.

For a long time, scientists had tools to count how many people (or cells, or birds) were in a specific area. They could also tell if two types of things were near each other. But they struggled to answer a crucial question: "Does this specific pair of friends only hang out in the park, or do they hang out everywhere equally?"

This paper introduces a new mathematical tool called Regional Co-occurrence Entropy (RCE). Think of it as a "Social Radar" that doesn't just count people, but maps out where specific friendships are strongest and where they are missing.

Here is how it works, using simple analogies:

1. The Problem: The "Blind" Map

Imagine you have a map of a town divided into neighborhoods (Partitions). You see two types of people: Rich Folks (Intact roofs) and Struggling Folks (Damaged roofs).

• Old tools could tell you: "There are 50 Rich Folks in Neighborhood A."
• Old tools could also tell you: "Rich Folks and Struggling Folks are often neighbors."
• The Missing Piece: Old tools couldn't easily tell you if Rich and Struggling folks were mixing specifically in Neighborhood A, or if they were just randomly scattered everywhere. Maybe they only mix near the river, but not near the school.

2. The Solution: The "Social Radar" (RCE)

The authors created a new formula (RCE) that acts like a heat map for friendships.

• It looks at pairs of things (like Cell A and Cell B, or Bird X and Bird Y).
• It checks if they are standing close to each other (within a certain distance).
• It then asks: "Is this pair hanging out together more often in this specific neighborhood than in that one?"

If the answer is "Yes, they are only hanging out in the Park," the radar beeps loudly (low entropy). If they are hanging out randomly everywhere, the radar stays quiet (high entropy).

3. Real-World Detective Cases

The authors tested their "Social Radar" on three very different mysteries:

Case A: The Brain Detective (Alzheimer's Disease)

• The Scene: A tiny slice of a mouse brain with Alzheimer's.
• The Characters: Immune cells (Microglia) and Support cells (Astrocytes).
• The Mystery: We know these cells gather around "plaques" (gunk that builds up in the brain), but which specific types are working together?
• The RCE Find: The radar revealed that a specific "Protective" immune cell and a specific "Stressed" support cell were hugging each other tightly right next to the plaques, but ignoring each other everywhere else.
• Why it matters: This suggests these two specific cells are teaming up to fight the disease in that exact spot. It's like finding out that two specific firefighters only work together when the fire is in the kitchen, not the garage.

Case B: The Town Planner (Social Mixing)

• The Scene: A village in St. Lucia.
• The Characters: Houses with "Good Roofs" (Wealthy) and "Bad Roofs" (Less Wealthy).
• The Mystery: Do rich and poor neighbors mix well? Does a river or a canal separate them?
• The RCE Find: The radar showed that rich houses mostly sat next to other rich houses, and poor next to poor. This "clumping" happened everywhere in the village, not just in specific neighborhoods.
• Why it matters: The river didn't cause the segregation; the segregation was just a general pattern of the whole town. The tool helped prove that the river wasn't the "bad guy" keeping people apart.

Case C: The Bird Watcher (Nature)

• The Scene: A nature reserve in Florida.
• The Characters: 16 different species of birds.
• The Mystery: Do certain birds only hang out in grassy areas, while others prefer forests?
• The RCE Find: The radar spotted a specific pair of birds (Bachman's Sparrow and Common Ground Dove) that only hung out together in grassy fields. They ignored the forests.
• Why it matters: This tells ecologists that the grass itself is a special "meeting place" for these two birds, perhaps because they eat the same seeds or protect each other there.

The Big Takeaway

Before this tool, scientists had to guess if interactions were happening in specific places. Now, they have a mathematical magnifying glass that can instantly spot:

1. Who is interacting.
2. Where they are interacting.
3. How strong that interaction is compared to random chance.

It's like upgrading from a black-and-white photo of a crowd to a high-definition video that highlights exactly which people are whispering secrets to each other in different rooms of the house. This helps doctors, city planners, and biologists understand the hidden rules that govern how things organize themselves in space.

Technical Summary

1. Problem Statement

Despite the rapid accumulation of spatial data across fields like spatial biology, ecology, and geography, existing analytical methods struggle to simultaneously account for object characteristics (e.g., cell types, species, building types) and environmental context (spatial partitions).

• Limitations of Current Tools:
  • Standard Spatial Statistics: Often infer interactions based on annotations or gene expression without utilizing spatial coordinates effectively, or fail to distinguish between aggregated patterns and random distributions.
  • Batty's Entropy: Can detect if an object type is over/under-represented in specific partitions (e.g., tissue sub-regions) but cannot determine if objects of different types are physically close (interacting) or merely co-located in the same region by chance.
  • Leibovici's Entropy & Bayesian Co-occurrence: Capture co-occurrences between object types but ignore regional variations. They may falsely identify interactions between objects separated by physical barriers if they fall within the same broad category.
• The Gap: There is a need for a metric that answers not just "who" interacts, but "where" and "how varied" these interactions are across different environmental partitions.

2. Methodology: Regional Co-occurrence Entropy (RCE)

The authors propose a new spatial entropy measure called Regional Co-occurrence Entropy (RCE). It quantifies the randomness of co-occurrences between categorical objects (e.g., cell types $i_1, i_2$) within specific spatial partitions (e.g., tissue regions, neighborhoods).

• Core Definitions:
  • Space: A 2D area $T$ divided into $G$ partitions ($T_g$).
  • Points: $n$ points (objects) with categories $X$.
  • Co-occurrence: Defined as $m$ points (e.g., pairs for $m=2$) located within a distance $d$ of each other within the same partition.
  • Variable $Z$: Counts co-occurrences $z_r$ for specific category tuples.
• Calculation:
  1. Partition-based Adjacency: A matrix is constructed where entries represent the proportion of unordered $m$-tuples occurring in partition $g$ with distance $< d$.
  2. Absolute Entropy ($H_{RC}$): Calculated using Shannon's entropy formula on the distribution of co-occurrences across partitions.
  3. Relative RCE: Normalized to a range of $[0, 1]$.
     • Low RCE (near 0): Indicates specific co-occurrence pairs are over-represented in specific partitions (high signal of environment-dependent interaction).
     • High RCE (near 1): Indicates co-occurrences are uniformly distributed across partitions (random/no specific environmental driver).
  4. Decomposed RCE: The method allows breaking down the total entropy to identify which specific pairs (tuples) contribute most to the signal (lowest decomposed RCE values).
• Statistical Validation: Significance is assessed via permutation tests (shuffling category labels while keeping spatial coordinates fixed) to generate a null distribution.
• Implementation: An R package (RC.entropy) is provided. It handles edge cases (e.g., zero counts in partitions) by adding a pseudo-count to avoid division by zero in logarithmic calculations.

3. Key Contributions

• Novel Metric: Introduction of RCE, a unified framework combining partition structure and distance-based co-occurrence to detect environment-dependent associations.
• Versatility: The method is domain-agnostic, applicable from micro-scale (cellular) to macro-scale (geographic/ecological) data.
• Hypothesis Testing: Provides a rigorous statistical framework to test whether spatial diversity or interactions are driven by specific environmental partitions.
• Software Tool: Release of an open-source R package with vignettes for reproducibility.

4. Results & Applications

The authors validated RCE across three distinct domains:

A. Spatial Biology: Alzheimer's Disease (Mouse Brain)

• Data: Xenium in situ transcriptomics of 10,335 cells (microglia and astrocyte subtypes) mapped against amyloid-beta (A$\beta$) plaques.
• Partitions: "Plaque regions" (expanded 30µm around plaques) vs. "No-plaque regions."
• Findings:
  • Observed RCE (95.1%) was significantly lower than random permutations (98.6%), indicating non-random, plaque-enriched interactions.
  • Key Interactions: Identified strong enrichment of Protective DAMs (pDAM) and Astrocytes (Ac10) specifically within plaque regions.
  • Novel Insight: Confirmed pDAMs cluster with Ac10 astrocytes to form peri-plaque glial nets, while pathogenic DAMs do not migrate to plaques to the same extent. This refines the understanding of cell-to-cell activation mechanisms in AD.

B. Geography: Social Mixing in Dennery Village (St. Lucia)

• Data: Aerial drone imagery of 1,312 buildings classified by rooftop condition (Intact/High Wealth vs. Damaged/Low Wealth).
• Partitions: 6 districts defined by rivers and canals.
• Findings:
  • RCE was 92%, falling within the range of random permutations.
  • Conclusion: The null hypothesis was accepted; arbitrary geographic divisions (rivers/canals) do not drive social diversity.
  • Observation: Wealth segregation exists (intact-intact and damaged-damaged pairs are common), but this clumping is uniform across all districts, suggesting the drivers of social mixing are not the physical barriers used for partitioning.

C. Ecology: Bird Species in Disney Wilderness Preserve

• Data: 753 birds of 16 species across 90 sites.
• Partitions: 3 environments based on vegetation cover (Forest-dominant, Grass-dominant, Mixed).
• Findings:
  • Observed RCE was significantly lower than 99.7–99.9% of random permutations.
  • Key Interactions: Identified specific species pairs driven by vegetation. For example, Bachman's Sparrow (BACS) and Common Ground Dove (COGD) co-occur preferentially in grass-dominant sites, despite neither species being individually enriched in that habitat.
  • Conclusion: Vegetation cover is a strong driver of community composition and species associations.

5. Significance

• Scientific Impact: RCE bridges the gap between spatial statistics and information theory, offering a tool to detect context-dependent interactions that are invisible to standard co-occurrence or partition-based analyses.
• Practical Utility: It enables rapid hypothesis generation and testing in large spatial datasets without the heavy computational cost or complex parametrization required by model-based approaches (e.g., Gibbs point processes).
• Broad Applicability: The framework is applicable to diverse fields including:
  • Medical Imaging: Understanding tissue organization and disease mechanisms.
  • Ecology: Analyzing species distribution and habitat drivers.
  • Urban Planning/Geography: Studying social segregation and community dynamics.
  • Agriculture/Climate: Modeling landscape management and environmental change.

In summary, the paper presents a robust, flexible, and computationally efficient method to uncover how environmental contexts shape the spatial organization of complex systems, from cellular interactions to human settlements.