DenMark: A Bayesian Hierarchical Model for Identifying Cell-Density Correlated Genes from Spatial Transcriptomics

DenMark is a novel Bayesian hierarchical model that leverages Hilbert space Gaussian process approximations to jointly model cell locations and gene expression in single-cell-resolution spatial transcriptomics, enabling the robust identification and uncertainty quantification of genes whose expression correlates with local cell density across diverse tissues and platforms.

The Gist

Imagine you are walking through a bustling city. You notice that in some neighborhoods, the streets are packed with people (high density), while in others, it's quiet and sparse (low density).

Now, imagine that every person in this city is wearing a glowing shirt that changes color or brightness depending on how crowded the street is. In the crowded squares, some people's shirts might glow bright red, while in the quiet parks, they might glow blue.

The Problem:
Scientists have a new, super-powerful camera (called Spatial Transcriptomics) that can take a photo of every single cell in a piece of tissue (like a slice of your brain or a tumor) and see exactly where they are and what "shirt" (gene) they are wearing.

However, most existing computer tools are like photographers who only care about where the glowing shirts are. They ask, "Is this red shirt here? Is that blue shirt there?" They are great at finding patterns of location, but they miss a crucial question: "Does the brightness of the shirt change because of how many people are standing nearby?"

If you have a crowd of 100 people, does the average person's shirt glow brighter? Or dimmer? Or stay the same? Current tools struggle to answer this because they treat the crowd size as a fixed background, rather than something that actively influences the people.

The Solution: DenMark
The authors of this paper created a new tool called DenMark (Density-dependent Marked point process). Think of DenMark as a smart detective that doesn't just look at the map; it investigates the relationship between the crowd and the glowing shirts.

Here is how DenMark works, using simple analogies:

1. The "Grid" Game (Simplifying the Chaos)

The tissue is full of millions of individual cells. Trying to track every single one at once is like trying to count every grain of sand on a beach while a storm is blowing.

• What DenMark does: It draws a giant invisible grid over the tissue, like a chessboard. It groups the cells into these little squares (grids).
• The Magic: Instead of looking at individual cells, it looks at the "average vibe" of each square. How many people are in this square? How bright is the average shirt here?

2. The "Shared Secret" (The Core Innovation)

This is the most important part. DenMark assumes that the "crowd" and the "shirts" are connected by a secret handshake.

• The Old Way: Imagine the crowd and the shirts are two separate movies playing on different screens.
• The DenMark Way: DenMark realizes they are actually scenes from the same movie. It uses a mathematical trick (called a Bayesian Hierarchical Model) to say: "Okay, let's assume the brightness of the shirts is partly caused by the crowd size, and partly caused by something unique to that specific gene."

It separates the signal into two parts:

1. The Crowd Effect: How much does the gene change just because there are more neighbors?
2. The Unique Effect: How much does the gene change for reasons other than the crowd?

3. The "Speedy Calculator" (HSGP)

Usually, doing this kind of math on millions of cells is like trying to solve a Rubik's cube while running a marathon. It takes forever.

• The Fix: DenMark uses a "shortcut" called a Hilbert Space Gaussian Process. Imagine instead of calculating the exact distance between every single person in the city, you use a smart map that estimates the distances based on a few key landmarks. It's an approximation, but a very smart one that keeps the results accurate while making the computer run 100 times faster.

What Did They Find?

The authors tested DenMark in two real-world scenarios:

• The Brain (Mouse): They looked at brain cells called astrocytes. They found that certain genes (like Aqp4) act like social butterflies: the more astrocytes are packed together, the more these genes "glow" (express themselves). This helps explain how brain cells organize themselves to handle stress or injury.
• The Tumor (Breast Cancer): They looked at a battle zone between cancer cells and immune cells. They discovered a fascinating "tug-of-war."
  • In areas packed with cancer cells, certain genes turn on.
  • In areas packed with immune cells, those same genes turn off.
  • This reveals a hidden language between the tumor and the immune system that previous tools missed.

Why Does This Matter?

Before DenMark, scientists were like people looking at a crowd and guessing why people were wearing certain clothes. They might see a pattern, but they couldn't prove if the crowd caused the clothing choice.

DenMark gives us the proof. It allows scientists to say with confidence: "Yes, the density of cells in this specific neighborhood is directly causing these genes to turn on or off." This is a huge step forward for understanding how tissues are built, how diseases like cancer grow, and how we might design better drugs to target these specific neighborhoods.

In short: DenMark is the tool that finally connects the dots between where cells are and what they are doing, revealing that the "neighborhood" matters just as much as the "individual."

Technical Summary

1. Problem Statement

Recent advances in single-cell-resolution spatial transcriptomics (ST) allow for the profiling of gene expression while preserving precise cellular locations. A critical biological phenomenon is the relationship between local cell density (the abundance of cells per unit area) and gene expression. Changes in cell density influence cell-cell communication, signaling diffusion, and resource competition, often driving coordinated transcriptional responses.

However, existing computational frameworks for ST data primarily focus on identifying Spatially Variable Genes (SVGs)—genes with heterogeneous expression patterns across space—rather than Density-Correlated Genes (DCGs). Current methods often treat cell locations as fixed covariates or model expression conditional on spatial coordinates, failing to explicitly model the stochastic variation in cell density itself. Consequently, there is a lack of rigorous statistical tools to quantify the correlation between local cell density and gene expression with proper uncertainty quantification.

2. Methodology: The DenMark Framework

The authors propose DenMark (Density-dependent Marked point process), a unified Bayesian hierarchical framework designed to jointly model cell locations and gene expression.

Core Statistical Model

DenMark treats the data as a Marked Point Process (MPP) where:

• The Point Pattern: Represents the spatial locations of cells (cell density).
• The Marks: Represent the gene expression counts associated with each cell.

The model decomposes spatial variation into two components:

1. Shared Spatial Component: A latent field shared between cell density and gene expression, capturing the correlation driven by local crowding.
2. Gene-Specific Component: A residual spatial field unique to the gene, capturing expression patterns independent of density.

Mathematical Formulation:

• Cell Density ($\lambda_1$): Modeled as a Log-Gaussian Cox Process (LGCP). The log-density is a linear combination of an intercept and a latent Gaussian Process (GP), $\omega_1(g)$.
  $$\log(\lambda_1(g)) = \beta_1 + a_{11}\omega_1(g)$$
• Gene Expression Intensity ($\lambda_2$): Modeled conditional on the observed cell count. The log-intensity includes a shared term linked to $\omega_1(g)$ and a gene-specific term linked to an independent GP, $\omega_2(g)$.
  $$\log(\lambda_2(g)) = \beta_2 + a_{21}\omega_1(g) + a_{22}\omega_2(g)$$
  • Here, $a_{21}$ quantifies the strength and direction of the density-expression correlation.
  • $a_{22}$ captures gene-specific spatial variation.

Approximation and Scalability:

• Grid Discretization: To handle the intractable likelihood integrals of continuous point processes, the tissue is discretized into a rectangular grid. Cell counts and aggregated gene expression are summed within each grid cell.
• Hilbert Space Gaussian Process (HSGP): Standard GPs scale cubically with the number of grid points ($O(N^3)$), making them computationally prohibitive for high-resolution ST data. DenMark employs a low-rank HSGP approximation (Solin & Särkkä, 2020). This approximates the GP using a finite set of basis functions (eigenfunctions of the Laplacian operator), reducing computational complexity to linear or near-linear scaling while preserving spatial structure.
• Inference: Parameters are estimated via Hamiltonian Monte Carlo (HMC) using the Stan probabilistic programming language.

Model Comparison:
To determine if a gene is a DCG, DenMark compares the full model (where $a_{21} \neq 0$) against a baseline nested model (where $a_{21} = 0$, implying independence) using the Watanabe-Akaike Information Criterion (WAIC).

3. Key Contributions

1. Novel Statistical Framework: DenMark is the first unified Bayesian framework to explicitly model the joint stochastic process of cell locations and gene expression to infer density-expression correlations.
2. Scalability: By integrating HSGP approximation, the method enables the analysis of large, single-cell-resolution datasets (e.g., 10x Xenium, MERFISH) that would be computationally infeasible with exact GP methods.
3. Uncertainty Quantification: As a Bayesian method, DenMark provides full posterior distributions for the correlation parameter ($\rho$), allowing for rigorous uncertainty quantification rather than just point estimates.
4. Interpretability: The model explicitly separates density-driven effects from gene-specific spatial patterns, offering a clear biological interpretation of how much of a gene's spatial variation is driven by cell crowding.

4. Results

Simulation Studies

• Accuracy: DenMark accurately recovers true spatial correlation structures and parameter values (including the correlation coefficient $\rho$) in simulated data.
• Efficiency vs. Accuracy: The HSGP approximation achieves a balance between computational speed and accuracy. Increasing the number of basis functions improves accuracy but increases computation time; a moderate number of basis functions (e.g., 625) was found to be sufficient.
• Grid Sensitivity: Finer grid resolutions yield better approximations of the underlying continuous process, with bias decreasing as grid resolution increases.

Application 1: Mouse Brain (MERFISH)

• Dataset: MERSCOPE data from mouse brain tissue (78,329 cells, 483 genes).
• Validation: DenMark successfully identified known density-correlated genes (e.g., Aqp4, Cxcl12, Cxcr4) involved in astrocyte migration and brain homeostasis.
• Discovery: It identified novel DCGs (e.g., Gprc5b, Adcyap1r1) that showed strong density correlation but were not necessarily the top Spatially Variable Genes (SVGs) detected by standard tools like SPARK-X.
• Biological Insight: Gene Set Enrichment Analysis (GSEA) revealed that DCGs were enriched for nervous system development and protein binding, while suppressed in trace-amine receptor activity (suggesting these pathways are active in low-density regions).

Application 2: Breast Cancer (10x Xenium)

• Dataset: High-resolution data from human breast tumors (167,780 cells).
• Microenvironment Analysis: DenMark analyzed distinct regions of interest (ROIs): Invasive Tumor, DCIS-1, and DCIS-2.
• Antagonistic Interplay: The model revealed an antagonistic relationship between tumor and immune cells. Genes positively correlated with tumor cell density (e.g., CDH1, CLDN4) were often negatively correlated with immune cell density (e.g., TRAC, PTPRC), and vice versa.
• Distinct Signals: There was limited overlap between DCGs identified by DenMark and SVGs identified by SPARK-X, confirming that density-correlation captures a distinct biological signal from general spatial heterogeneity.
• Functional Enrichment: DCGs in invasive tumors were enriched for immune activation, myeloid leukocyte activation, and Ras signaling pathways.

5. Significance

DenMark addresses a critical gap in spatial transcriptomics analysis by shifting the focus from "where genes are expressed" to "how gene expression relates to cellular crowding."

• Biological Mechanism: It provides a statistical mechanism to link tissue architecture (density) directly to molecular state (transcription), which is crucial for understanding phenomena like tumor-immune evasion, developmental patterning, and tissue remodeling.
• Tooling: It offers a scalable, open-source tool (available on GitHub) that allows researchers to move beyond descriptive spatial maps to quantitative inference of density-expression relationships.
• Future Directions: The framework lays the groundwork for more complex models, such as spatially varying coregionalization or multi-gene interactions, further advancing the quantitative understanding of tissue biology.