hoodscanR: profiling single-cell neighborhoods in spatial transcriptomics data

The paper introduces hoodscanR, a Bioconductor package that overcomes limitations in existing methods by enabling accurate identification of cellular neighborhoods and generation of cell-specific profiles in spatial transcriptomics data, as demonstrated through its application to breast and lung cancer datasets to reveal subtle transcriptional changes and disease mechanisms.

The Gist

Imagine you are looking at a bustling city from a drone. In the past, scientists could only take a "smoothie" of the whole city, blending all the people, buildings, and traffic into one giant cup to see what was inside. Later, they got better and could look at individual people (cells) and see what they were doing. But there was a catch: they lost the map. They knew who was there, but not where they were standing or who they were talking to.

Spatial transcriptomics is the new technology that gives us the map back. It tells us exactly where every cell is in the tissue. But now, we have a new problem: we have a massive amount of data, but we don't have a good way to understand the "neighborhoods."

Enter hoodscanR. Think of it as a super-smart urban planner for your biological city.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Hard Line" Mistake

Imagine a street corner where a bakery, a park, and a coffee shop meet.

• Old tools were like a strict zoning officer who would say: "This corner is 100% a Bakery" or "This corner is 100% a Park." They forced every spot into one single box.
• The Reality: In biology, cells are messy. A cell might be standing right on the edge of a tumor, surrounded by immune cells and blood vessels. It belongs to all those neighborhoods at once. Old tools missed this nuance.

2. The Solution: The "Soft Focus" Lens

hoodscanR is different. Instead of drawing a hard line, it uses a soft focus lens.

• It looks at a specific cell and asks: "Who are your neighbors?"
• It calculates a probability score. It might say, "This cell is 60% in the 'Tumor Neighborhood,' 30% in the 'Immune Neighborhood,' and 10% in the 'Stromal Neighborhood.'"
• The Analogy: It's like realizing you aren't just a "student" or just a "musician." You might be 70% student, 20% musician, and 10% baker. hoodscanR captures that mix, which is crucial for understanding how cells actually behave.

3. How It Works (The Magic Recipe)

The software does three main things:

1. The Radar Scan: It uses a fast algorithm (like a high-speed radar) to find the closest neighbors to every single cell in the tissue.
2. The Neighborhood Profile: It creates a "ID card" for every cell. This ID card doesn't just say "I am a Tumor Cell." It says, "I am a Tumor Cell, and I am currently hanging out with 50 B-cells and 20 Macrophages."
3. The Detective Work: Once it knows who is hanging out with whom, it can spot patterns.
   • Example: It found that in lung cancer, when tumor cells hang out near specific immune cells (B-cells), they start acting differently. They might become more aggressive or more resistant to drugs.

4. Why This Matters: The "Third Place" Effect

In sociology, a "third place" is a social surrounding separate from the two usual social environments of home and the workplace (like a coffee shop or park). In cancer, the Tumor Microenvironment (TME) is that third place.

• The Discovery: The paper shows that the "third place" (the neighborhood) changes the personality of the cells.
• Real World Example: In the study, they looked at breast cancer. They found that cells near hormone receptors acted differently than cells far away. In the brain, neurons in one specific "neighborhood" had different genes turned on than neurons in a neighboring "neighborhood," even though they were the same type of neuron.
• The Takeaway: You can't understand a cell just by looking at its DNA. You have to look at its address and its neighbors.

5. Speed and Compatibility

The authors built hoodscanR to be fast (like a sports car) and compatible (like a universal charger).

• It works with data from many different high-tech microscopes (like 10X Genomics and Nanostring).
• It is so fast that it can process huge datasets (hundreds of thousands of cells) in seconds, whereas other tools might take hours or crash.

Summary

hoodscanR is a new tool that helps scientists stop looking at cells as isolated islands and start seeing them as part of a complex, shifting community. By understanding the "neighborhood" a cell lives in, doctors and researchers can better understand why cancer grows, how it resists treatment, and where to aim new therapies to break up the bad neighborhoods and build better ones.

In short: It turns a blurry, black-and-white map of a city into a high-definition, color-coded guide that shows exactly who is talking to whom, and why it matters.

Technical Summary

1. Problem Statement

Spatial transcriptomics (ST) technologies (e.g., 10X Genomics Xenium, Nanostring CosMx) have revolutionized the field by preserving the spatial coordinates of cells, offering insights into tissue microenvironments (TME) that bulk and single-cell RNA-seq miss. However, existing analytical tools face critical limitations:

• Lack of Partial Membership: Most current methods (e.g., Giotto, Squidpy, SpaGCN) assign cells to discrete, exclusive spatial domains or neighborhoods. They fail to capture the nuance where a cell resides in a mixed environment (e.g., a tumor cell surrounded by both stromal and immune cells).
• Absence of Cell-Level Profiles: Existing tools often generate domain-level annotations but lack the capability to provide detailed, probabilistic neighborhood profiles for individual cells.
• Limited Downstream Utility: Many tools focus on clustering or co-localization but do not easily facilitate downstream analyses, such as differential expression (DE) based on specific neighborhood contexts or uncertainty quantification.
• Computational Efficiency: Many state-of-the-art methods struggle with the scale and resolution of modern high-throughput ST datasets (hundreds of thousands of cells).

2. Methodology

The authors developed hoodscanR, a Bioconductor R package designed to address these gaps through a probabilistic, cell-centric approach.

• Core Algorithm:

  • Approximate Nearest Neighbor (ANN): Uses k-dimensional trees to efficiently identify the $k$-nearest neighbors for every cell based on spatial coordinates.
  • Probabilistic Neighborhood Assignment: Instead of hard clustering, hoodscanR calculates a neighborhood probability distribution for each cell. It uses a SoftMax function enhanced by a hyperparameter $\tau$ (tau) to weigh the influence of neighboring cells based on their distance and annotation (e.g., cell type).
    • Formula: $p_{h_j}(x; \tau) = \frac{\sum \mathbb{1}_{h_j}(x_i) \cdot \exp(-d(x, x_i)/\tau)}{\sum \exp(-d(x, x_i)/\tau)}$
    • This allows a single cell to have partial membership in multiple neighborhoods (e.g., 60% "Tumor" neighborhood, 30% "B-cell" neighborhood).
  • Hyperparameters:
    • $k$: Number of nearest neighbors.
    • $\tau$: Controls the spatial scale (local vs. global interactions). Smaller $\tau$ emphasizes immediate neighbors; larger $\tau$ incorporates broader context.

• Key Metrics & Downstream Analysis:

  • Perplexity: A metric derived from Shannon entropy to quantify the diversity/complexity of a cell's neighborhood. High perplexity indicates a heterogeneous environment.
  • Permutation Testing: An empirical test to calculate p-values for perplexity, identifying regions with statistically significant cellular complexity.
  • Unsupervised Clustering: Cells are clustered based on their neighborhood probability vectors (using K-means) to define "neighborhood-based spatial domains."
  • Differential Expression (DE): Enables DE analysis by grouping cells based on their neighborhood profiles (e.g., comparing tumor cells in "macrophage-rich" vs. "stroma-rich" neighborhoods).

• Infrastructure: Built on the SpatialExperiment class, ensuring seamless integration with the Bioconductor ecosystem (Seurat, scran, scuttle).

3. Key Contributions

1. Partial Membership Modeling: hoodscanR is the first tool to provide cell-level partial membership across multiple neighborhoods, capturing the continuous and mixed nature of tissue microenvironments.
2. Cell-Level Neighborhood Profiles: Generates a comprehensive probability matrix for every cell, enabling granular downstream analyses that were previously impossible with discrete domain assignments.
3. Computational Efficiency: Demonstrates superior speed compared to existing tools (e.g., ~21x faster than Banksy), making it scalable for large datasets (e.g., >500k cells).
4. Flexibility in Annotation: Supports not only cell-type annotations but also gene-expression-based neighborhoods (e.g., grouping cells by hormone receptor status like ER/PR/AR).
5. Uncertainty Quantification: Introduces perplexity and permutation testing to statistically validate the complexity of cellular neighborhoods.

4. Results

The authors validated hoodscanR using multiple datasets: 10X Xenium (Breast Cancer), Nanostring CosMx (NSCLC), MERFISH (Mouse Colon), STARmap (Mouse Cortex), and Visium HD (Mouse Brain).

• Benchmarking (Spatial Domain Identification):

  • Compared against 7 state-of-the-art tools (Seurat, Banksy, BayesSpace, MERINGUE, SpaGCN, Stagate, Utag) across 16 datasets.
  • Performance: hoodscanR achieved the highest composite score (mean of ARI, NMI, Purity, Homogeneity) across all datasets.
  • Speed: It was approximately 21-fold faster than the second-fastest method (Banksy).
  • Robustness: Maintained high accuracy (NMI > 0.8) even when cell-type annotations were downgraded from high-resolution to low-resolution.

• Case Study: Breast Cancer (Xenium):

  • Successfully identified DCIS cells residing in mixed neighborhoods of myoepithelial and tumor cells, assigning probabilistic scores (e.g., 61% DCIS, 31% Myoepithelial) that matched the visual spatial distribution of neighbors.

• Case Study: NSCLC (CosMx):

  • Identified Tertiary Lymphoid Structures (TLS) as distinct neighborhoods containing B cells, cDC1, and stromal cells.
  • Transcriptional Changes: Performed DE analysis on tumor cells from "macrophage-rich" vs. "stroma-rich" neighborhoods. Found significant upregulation of collagen-related genes (e.g., COL1A1, COL11A1) in macrophage-associated tumor cells, linking spatial context to drug resistance mechanisms.

• Case Study: Mouse Brain (Visium HD):

  • Applied gene-expression-based neighborhoods (using markers like Neurod6, Calb2).
  • Revealed that neurons in different neighborhoods (e.g., Neurod6+ vs. Calb2+) exhibited distinct transcriptional profiles related to synaptic plasticity and calcium signaling, demonstrating the tool's ability to detect subtle spatial transcriptional variations.

5. Significance

• Biological Insight: hoodscanR shifts the paradigm from discrete spatial domains to continuous, probabilistic neighborhood profiling. This allows researchers to uncover how specific microenvironmental compositions (e.g., immune infiltration) directly influence the transcriptional state of individual cells.
• Clinical Relevance: By identifying specific neighborhood contexts (like TLS or collagen-rich tumor microenvironments), the tool aids in understanding therapy resistance and identifying novel therapeutic targets in cancer.
• Scalability: Its speed and compatibility with Bioconductor make it a practical solution for the rapidly growing volume of high-resolution spatial data.
• Versatility: The ability to define neighborhoods based on any user-defined annotation (cell type, gene expression, ligand-receptor pairs) makes it a flexible framework for hypothesis generation across diverse biological contexts, from oncology to neurobiology.

In conclusion, hoodscanR fills a critical methodological gap in spatial transcriptomics by providing a robust, efficient, and probabilistic framework for characterizing cellular neighborhoods, thereby enabling a deeper understanding of tissue architecture and disease mechanisms.