MINGL Quantifies Borders, Gradients, and Heterogeneity in Multicellular Tissue Organization

MINGL is a probabilistic framework that transforms discrete spatial-omics neighborhood annotations into continuous measures of tissue architecture, enabling the quantification of cellular gradients, interface heterogeneity, and hierarchical interactions across diverse biological contexts.

The Gist

Imagine your body is a massive, bustling city made of millions of tiny citizens (cells). In this city, people don't just live in random spots; they form neighborhoods, districts, and regions. Some neighborhoods are like busy marketplaces (immune cells), others are quiet residential zones (epithelial cells), and some are industrial areas (stromal cells).

For a long time, scientists studying these "cities" had a very rigid way of looking at them. They would draw hard lines on a map and say, "You live in Neighborhood A," or "You live in Neighborhood B." If a cell was standing right on the fence between two neighborhoods, the old methods forced it to pick a side. They treated the city as a collection of separate, solid blocks.

The Problem:
Real life isn't like that. Cities have fuzzy borders, transition zones, and gradients. A cell on the edge of a tumor isn't just "tumor" or "healthy"; it's a mix of both, acting as a bridge. Furthermore, sometimes a neighborhood changes its character slowly over a distance (a gradient), and other times it changes abruptly (a sharp border). Old tools missed these nuances, treating uncertainty as noise rather than a signal.

The Solution: MINGL
The authors of this paper created a new tool called MINGL (Mixture-based Identification of Neighborhood Gradients with Likelihood estimates). Think of MINGL not as a map drawer, but as a probability detective.

Instead of forcing a cell to say, "I am 100% in Neighborhood A," MINGL asks, "How much do you feel like you belong to Neighborhood A, and how much to Neighborhood B?" It gives every cell a "membership card" with percentages.

Here is what MINGL can do that the old tools couldn't, explained through simple analogies:

1. Finding the "Border Patrol" (Identifying Borders)

The Old Way: If a cell was on the border, it was forced into one neighborhood, hiding the fact that it was actually interacting with the next one.
The MINGL Way: MINGL identifies cells that have a foot in two camps. It calls these "Border Cells."

• Analogy: Imagine a town square where the bakery meets the park. The old map said, "You are either in the bakery or the park." MINGL says, "Ah, this person is standing right on the edge, smelling the bread and watching the ducks. They are a 'Border Cell'!"
• Why it matters: In cancer, these border cells are crucial. The paper found that at the edge of a tumor, specific immune cells and tumor cells mix together. These are the "hot zones" where the body tries to fight the cancer or where the cancer tries to invade. MINGL found these zones clearly, showing us where the real action happens.

2. Drawing the "City Interaction Map" (Hierarchical Networks)

The Old Way: Scientists knew Neighborhood A existed and Neighborhood B existed, but they didn't know how they talked to each other.
The MINGL Way: By looking at all the border cells, MINGL draws a network map. It shows which neighborhoods are neighbors and how strongly they interact.

• Analogy: Think of a subway map. The old way just listed the stations. MINGL draws the lines connecting them and shows how thick the lines are (how many people are traveling between them).
• Discovery: In the intestine, MINGL found that "Plasma Cell" neighborhoods act like a central hub, connecting immune cells, gut lining cells, and structural cells. It's like finding a major train station that connects the whole city, even though the station itself is small.

3. Measuring the "Steepness of the Hill" (Gradients)

The Old Way: Scientists assumed transitions between neighborhoods were either instant (a wall) or didn't exist.
The MINGL Way: MINGL measures the gradient. Is the transition a steep cliff or a gentle slope?

• Analogy: Imagine walking from a snowy mountain peak to a warm beach.
  • Steep Transition: You step off a cliff and instantly hit the water. (This is like the boundary between the inner and outer parts of an immune follicle in the intestine).
  • Gradual Transition: You walk down a gentle hill where the snow slowly turns to grass, then to sand. (This is like the transition between different layers of the gut lining).
• Why it matters: Diseases often change the "steepness" of these hills. MINGL can measure if a disease is making a smooth transition suddenly become a jagged cliff, which tells us how the tissue is breaking down.

4. Spotting the "Chameleon Neighborhoods" (Heterogeneity)

The Old Way: If two patients had the same "Tumor Neighborhood," scientists assumed they were identical.
The MINGL Way: MINGL looks at the internal composition of the neighborhood. It asks, "Is this neighborhood in Patient A made of the same ingredients as in Patient B?"

• Analogy: Imagine two bakeries both labeled "Bakery." The old way assumed they sold the same bread. MINGL walks inside and says, "Bakery A sells mostly sourdough, but Bakery B has switched to mostly donuts and has very few bakers left."
• Discovery: In Barrett's Esophagus (a pre-cancerous condition), MINGL found that while some neighborhoods stayed stable (like a "Plasma Cell" neighborhood that acts like a loyal anchor), others completely changed their recipe as the disease progressed. This helps doctors understand which patients are changing their tissue structure the most.

5. The "Goldilocks" Cluster Selector (Finding the Right Number)

The Old Way: When scientists tried to group cells, they had to guess: "Should I make 5 groups? 10? 50?" It was a game of trial and error.
The MINGL Way: MINGL calculates a "Goldilocks score." It balances two things:

1. Fit: Are the groups distinct enough?
2. Confidence: Are we sure a cell belongs in its group?

• Analogy: It's like trying to sort a pile of mixed Legos. If you make too few piles, you mix red and blue bricks together (bad fit). If you make too many piles, you separate every single brick by a tiny scratch (too much noise). MINGL automatically finds the "just right" number of piles that makes the most biological sense.

The Big Picture

MINGL changes the way we see the body. It stops treating tissue like a mosaic of solid, separate tiles and starts treating it like a living, breathing landscape with hills, valleys, borders, and bridges.

By accepting that cells can belong to multiple places at once, MINGL reveals the hidden conversations happening at the edges of our tissues. This helps us understand how healthy bodies stay organized and how diseases like cancer break those rules, opening new doors for better treatments and diagnostics.

Technical Summary

1. Problem Statement

Current spatial-omics analysis methods typically rely on discrete clustering algorithms to assign cells to a single "cellular neighborhood" (CN) or organizational unit. While effective for identifying conserved compositions, this approach has significant limitations:

• Loss of Continuity: It ignores natural biological gradients, transitional zones, and heterogeneity within tissues.
• Border Neglect: Cells at the interfaces between organizational units are forced into a single category, obscuring critical "border cells" that mediate signaling and regulation.
• Homogeneity Assumption: It treats organizational units as internally homogeneous, missing subtle compositional changes driven by disease or individual variation.
• Arbitrary Parameter Selection: Determining the optimal number of clusters (neighborhoods) often relies on manual "guess-and-check" workflows, leading to user bias and inconsistent results across datasets.
• Lack of Quantitative Metrics: There are no standard metrics to quantify the steepness of transitions between states or the degree of architectural heterogeneity across conditions.

2. Methodology: MINGL Framework

The authors introduce MINGL (Mixture-based Identification of Neighborhood Gradients with Likelihood estimates), a probabilistic framework that converts discrete neighborhood annotations into continuous measures of tissue architecture.

Core Algorithmic Steps:

1. Feature Construction: For each cell, a feature vector is constructed based on the composition of its $k$-nearest neighbors (kNN) using cell-type labels (or higher-level labels like CNs/Communities).
2. Centroid Definition: For each pre-defined organizational unit (e.g., a specific CN), a centroid is defined as a multivariate Gaussian distribution. The mean and standard deviation of the feature vectors for cells in that unit define the distribution.
3. Probabilistic Assignment (GMM): Instead of hard assignment, MINGL uses a Gaussian Mixture Model (GMM) with diagonal covariance. It calculates the likelihood of each cell belonging to every organizational unit, resulting in a vector of membership probabilities for each cell.
4. Derived Metrics:
   • Border Cells: Defined as cells with membership probability $>0.25$ in two or more organizational units.
   • Interaction Networks: Constructed by counting cells with co-membership in pairs of units, creating weighted graphs of organizational interactions.
   • Gradient Quantification: Uses log-ratio scores of probabilities between two units, binned and analyzed via derivatives to calculate transition steepness (curvature) and local gradient magnitude.
   • Heterogeneity ($\Delta$-probability): Calculates the difference between condition-specific probabilities and global probabilities to measure structural reorganization.
   • Optimal Cluster Selection: Uses a harmonic composite score balancing fit (log-likelihood) and confidence (assignment probability) to identify a stable range of cluster numbers, avoiding over-splitting or over-merging.

3. Key Contributions

• Unified Discrete/Continuous Representation: MINGL treats tissue organization as a probabilistic landscape, allowing for the simultaneous analysis of distinct units and the continuous transitions between them.
• Quantification of Borders: It explicitly identifies and characterizes "border cells" as biologically meaningful regions of mixing rather than noise.
• Steepness Metrics: Introduces mathematical metrics to quantify whether transitions between tissue states are sharp (step-like) or gradual (gradient-like).
• Automated Cluster Resolution: Provides a data-driven, unbiased method to recommend an optimal starting number of clusters for spatial analysis.
• Modality Agnostic: The framework operates on cell-type labels and spatial coordinates, making it applicable to various spatial-omics platforms (e.g., CODEX, Visium).

4. Key Results

The authors validated MINGL across three datasets: Human Melanoma, Healthy Human Intestine, and Barrett's Esophagus Progression.

A. Border Identification & Enrichment

• Melanoma: Identified border cells enriched at the tumor-normal interface. These borders were highly enriched for specific immune cells (CD4+ T cells, macrophages) and tumor subtypes, suggesting active coordination at the interface.
• Intestine: Successfully identified borders at anatomical scales (e.g., Mucosa vs. Muscularis Mucosa) and cellular scales. Border cells were enriched for transit-amplifying cells and dendritic cells, consistent with known biological feedback loops.

B. Hierarchical Interaction Networks

• Constructed interaction graphs showing that organizational units interact in specific hierarchies.
• Plasma Cell Enriched communities emerged as critical "hubs" connecting stromal, epithelial, and immune compartments across multiple biological scales (tissue units, communities, and neighborhoods).

C. Gradient Steepness & Transitions

• Immune Follicles: Quantified the transition from Inner to Outer Follicle. The analysis revealed a steep, hyperbolic transition (high curvature), indicating a sharp boundary between B-cell zones.
• Epithelial-Immune Interface: The transition between Plasma Cell Enriched and Secretory Epithelial communities showed a gradual, linear gradient (low curvature), indicating a more diffuse mixing zone.

D. Heterogeneity in Disease (Barrett's Esophagus)

• MINGL detected that Tumor states exhibited the highest structural heterogeneity (largest $\Delta$-probability) compared to normal tissue.
• It identified that while some neighborhoods (e.g., Plasma Cell Enriched) remained stable across disease states, others (e.g., Mature Intestinal/Immune) underwent significant compositional remodeling (e.g., shift from CK7+ to Ki67+ epithelial cells).
• It revealed patient-specific heterogeneity, identifying outliers where spatial organization diverged significantly from the cohort average.

E. Optimal Cluster Number

• Applied to the intestine dataset, the composite score method suggested N=17 clusters.
• This resolution successfully separated key structures (e.g., Inner/Outer Follicle, Plasma Cell Enriched) without the artificial splitting seen at higher resolutions (N=28) or the merging seen at lower resolutions (N=6).

5. Significance

MINGL represents a paradigm shift in spatial biology analysis by moving from descriptive (categorizing cells) to quantitative (measuring architecture).

• Biological Insight: It reveals that tissue borders are active regulatory zones enriched with specific cell types, not just geometric boundaries.
• Clinical Utility: By quantifying architectural heterogeneity and transition steepness, MINGL offers potential biomarkers for disease progression (e.g., distinguishing aggressive tumor states) and therapeutic response.
• Standardization: The framework reduces the subjectivity in spatial analysis workflows, providing a reproducible, scalable, and mathematically grounded approach to understanding tissue organization across diverse biological contexts.

The tool is implemented as a Python package (scverse ecosystem) and is available on GitHub, facilitating broad adoption across the spatial-omics community.