🧬 biology

SEMA: A Framework of Spatial-Aware Enrichment for Multi-pathway Analysis

SEMA is a novel framework that integrates spatial context into pathway enrichment analysis, outperforming existing methods to reveal critical spatially resolved biological insights, such as metabolic activities near tertiary lymphoid structures and fibroblast-mediated immune regulation, across diverse tissue and cancer datasets.

Original authors: Xiaokai Fan

Published 2026-07-20

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaokai Fan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are looking at a massive, bustling city from a drone. If you just count the people in each building, you get a list of numbers: "100 people here, 50 there." That's like traditional gene sequencing, which tells scientists which genes are active in a sample but treats every cell like it's floating in a vacuum, ignoring its neighbors. But cities aren't just lists of buildings; they are defined by how those buildings interact. A coffee shop next to a library creates a different vibe than a coffee shop next to a construction site. In biology, this "city" is your body's tissue, and the "buildings" are cells.

Recently, scientists developed a new way to take pictures of these cellular cities, called spatial transcriptomics. Instead of just listing the genes, this technology maps exactly where each gene is turned on within the tissue's layout. However, there's a catch: most tools used to analyze these maps still act like they are reading a list of words without understanding the sentences. They look at a single cell and ask, "What is this cell doing?" but they miss the bigger picture of how that cell is chatting with the cells right next to it. This paper introduces a new tool to fix that blind spot, helping researchers understand not just who is in the room, but how the room is arranged and what that arrangement means for the city's health.

The New Mapmaker: Meet SEMA

Meet SEMA (Spatial-Aware Enrichment for Multi-pathway Analysis). Think of SEMA as a super-smart neighborhood watch that doesn't just look at one house, but checks out the whole block to understand what's really going on.

In the old way of doing things, if a scientist wanted to know if a specific group of genes (a "pathway") was active in a tissue, they would look at a single spot and calculate a score. It was like judging a party by looking at one person in the corner. If that person was dancing, you might think the whole party was wild. But what if that person was just dancing alone while everyone else was asleep? The old methods often made this mistake, missing the fact that biology happens in neighborhoods, not in isolation.

SEMA changes the game by using a "neighborhood rule." When it calculates the score for a specific cell, it doesn't just look at that cell's genes. It also looks at the genes of the cells immediately surrounding it, giving a little weight to the neighbors' activities. It's like realizing that if your neighbor is baking cookies, the smell (the signal) is probably coming from their house, even if your own kitchen is empty. By blending a cell's own activity with its neighbors' activity, SEMA creates a much clearer, smoother map of where biological processes are actually happening.

Putting SEMA to the Test

The authors didn't just build this tool; they put it through the wringer to see if it actually works better than the existing methods. They tested it in three very different "cities":

The Organized City (The Brain): They looked at the mouse brain and the human prefrontal cortex. These areas have very distinct, organized neighborhoods (like the visual cortex or the hippocampus). When they tried to find these specific regions using gene activity, SEMA was much better at drawing the boundaries than the other tools. It could tell the difference between two adjacent brain regions with much higher accuracy, essentially saying, "Yes, this is definitely the library, and that is definitely the park," while other tools were a bit fuzzy.
The Tiny Village (The Lymph Node): They zoomed in on tiny, complex structures called germinal centers inside lymph nodes. These are small, crowded areas where immune cells gather to fight infections. Because these areas are so small and packed, it's hard for old tools to spot them. SEMA, however, spotted them clearly, matching up almost perfectly with the actual cell counts found by other advanced methods.
The Hidden Fortresses (Tertiary Lymphoid Structures): In tumors, the body sometimes builds "fortresses" called tertiary lymphoid structures (TLS) to fight cancer. These are hard to find because they look a bit like the surrounding tumor tissue. SEMA was the best at identifying these structures, finding them more accurately and using less computer power (memory and time) than the competition.

What SEMA Discovered in Cancer

Once the authors were confident SEMA was the best mapmaker, they used it to explore 93 breast cancer samples and 123 samples from various other cancers. Here is what they found:

The Metabolic Connection:
They discovered a strong spatial link between "functional" TLS (the active, fighting versions of those immune fortresses) and a specific metabolic process called arachidonic acid metabolism. It's as if they found that every time the immune fortress was fully operational, the surrounding neighborhood was buzzing with a specific type of chemical energy. This suggests that this metabolic pathway might be fueling the immune fight, a clue that wasn't obvious before.

The Secret Handshake:
They also looked for "ligand-receptor pairs," which are like secret handshakes cells use to talk to each other. While everyone knew about a famous handshake called CXCL13-CXCR5 in these immune fortresses, SEMA highlighted a different, equally important pair: CXCL13-CXCR3. This pair was found to be highly active right where the immune fortresses were strongest. The authors suggest this pair might be a key player in organizing the immune response, and patients with high levels of this signal tended to have better survival rates.

The Great Wall vs. The Mixed Crowd:
Perhaps the most surprising finding came from looking at how different cell types are arranged across many different cancers. The authors found that fibroblasts (a type of support cell) act as the "glue" or the central hub in the tumor city. They are usually the closest neighbors to the cancer cells.

However, the arrangement matters immensely. The team found two distinct patterns:

The Mixed Crowd: In some tumors, fibroblasts and cancer cells are all jumbled together in a chaotic mix.
The Segregated City: In other tumors, there is a clear, sharp boundary. The fibroblasts form a distinct ring or wall around the cancer cells, keeping them separate from the immune cells.

The paper suggests that this "segregated" pattern—where the fibroblasts form a clear boundary—is actually a good thing. Patients with this clear separation tended to have stronger immune activity and better survival outcomes. It seems that when the support cells form a neat, organized boundary, the immune system can do its job better. In contrast, when everything is mixed up, the immune system seems less effective.

Why This Matters

This paper doesn't claim to have cured cancer or solved the mystery of life. Instead, it offers a better way to read the map. By acknowledging that cells live in neighborhoods and that their location matters just as much as their identity, SEMA helps scientists see patterns that were previously hidden in the noise. It suggests that the physical layout of a tumor—how the cells are organized and separated—might be a crucial clue for understanding how the disease behaves and how patients might respond to treatment. It's a reminder that in biology, just like in a city, it's not just about who lives there, but how they are arranged.

Technical Summary: SEMA – A Framework of Spatial-Aware Enrichment for Multi-pathway Analysis

Problem Statement

Spatial transcriptomics (ST) technologies have evolved from bulk-level resolution to single-cell precision, enabling the characterization of tissue architecture, cell-cell interactions, and functional niches such as tertiary lymphoid structures (TLS). While gene set enrichment analysis (GSEA) is a standard tool for interpreting bulk and single-cell RNA-seq data, current methods for ST often neglect the explicit spatial context. Traditional approaches treat spots or cells as isolated data points, calculating enrichment scores based solely on local gene expression. This can lead to misinterpretations where biological functions are erroneously assigned to spatially disconnected locations. There is a critical gap in methods that integrate physical spatial proximity into pathway enrichment to accurately decode the spatially resolved data, particularly for understanding tissue functionality which emerges from physical cell interactions.

Methodology

The authors developed SEMA (Spatial-Aware Enrichment for Multi-pathway Analysis), a framework that integrates spatial neighborhood information into pathway scoring. The methodology proceeds through the following steps:

Raw Score Calculation: For a given gene set $G$ , a raw score ( $S_i^{raw}$ ) is calculated for each spot or cell $i$ as the average expression of genes present in both the gene set and the expression matrix.
Spatial Graph Construction: A k-nearest neighbor (KNN) graph is constructed using the Euclidean spatial coordinates of the spots/cells. An undirected adjacency matrix is formed where neighbors are defined by the $k$ closest cells.
Weighted Smoothing:
- A Gaussian kernel bandwidth ( $\sigma$ ) is automatically estimated based on the distances to the $k'$ -th nearest neighbor to control the decay of spatial influence.
- Gaussian kernel weights ( $w_{ij}$ ) are assigned to edges based on spatial distance ( $d_{ij}$ ): $w_{ij} = \exp(-d_{ij}^2 / 2\sigma^2)$ .
- These weights are row-normalized to create a weighted adjacency matrix $\tilde{W}$ .
Spatial-Aware Score Computation: The final spatial-aware score ( $S_i^{spatial}$ ) is a convex combination of the spot's own raw score and the weighted average of its neighbors' raw scores:
$S_i^{spatial} = (1 - \lambda) \cdot S_i^{raw} + \lambda \cdot \sum_j \tilde{W}_{ij} S_j^{raw}$
where $\lambda \in [0,1]$ is a spatial weight parameter.
Spatial Correlation Analysis: To investigate relationships between pathways or cell types, the authors employ bivariate Moran's I. This statistic uses the SEMA scores to measure global spatial correlation, identifying co-localization (positive correlation) or spatial exclusion (negative correlation) between two gene sets across the tissue.

The framework was benchmarked against established methods including AddModuleScore, AUCell, UCell, and ssGSEA.

Key Contributions and Results

1. Benchmarking and Performance

SEMA was rigorously evaluated across multiple datasets and tasks, consistently outperforming existing major methods in accuracy, specificity, and efficiency:

Anatomical Reconstruction: In mouse brain coronal sections and human dorsolateral prefrontal cortex (DLPFC) datasets, SEMA achieved higher Area Under the ROC Curve (AUC) and Precision-Recall AUC (PRAUC) in recovering signals for fine-scale regional marker gene sets compared to AddModuleScore, ssGSEA, UCell, and AUCell.
Microstructure Detection: In human lymph node data, SEMA demonstrated superior ability to detect signals in germinal centers (GCs) based on B cell, follicular dendritic cell (FDC), and T follicular helper (Tfh) markers. It showed stronger concordance with cell-type abundances inferred by deconvolution methods (cell2location and CARD) and achieved a higher AUC (0.707) in identifying pathologist-annotated lymphoid aggregates and GCs in 10x Xenium data.
TLS Identification: Using three distinct TLS signature gene sets on clear cell renal cell carcinoma (ccRCC) data, SEMA showed superior performance in distinguishing TLS-positive spots with higher AUC/PRAUC and lower false positive rates.
Efficiency: SEMA required significantly less computation time and memory than the compared methods, making it suitable for large-scale applications.

2. Biological Insights: Breast Cancer and Functional TLS

Applying SEMA to 93 breast cancer spatial transcriptomics datasets revealed:

Metabolic Correlation: A strong spatial correlation was identified between functional TLS (defined by markers like CXCL13, CD3D, MS4A1) and the arachidonic acid metabolism pathway. This finding was consistent across both KEGG and Reactome gene sets and validated in bulk RNA-seq data (4,148 samples).
Ligand-Receptor Pairs: Analysis of ligand-receptor pairs within functional TLS highlighted CXCL13-CXCR3 as a top-ranked pair, showing strong spatial co-localization with functional TLS signatures. This pair was associated with improved patient survival in TCGA-BRCA and METABRIC cohorts, suggesting a potentially underappreciated role in TLS activity beyond the canonical CXCL13-CXCR5 axis.

3. Pan-Cancer Cellular Organization

A pan-cancer analysis of 123 spatial transcriptomics samples (TNBC, OV, LC, HCC, ccRCC) uncovered general principles of cellular organization:

Fibroblast Hubs: Fibroblasts emerged as central spatial hubs, consistently showing strong spatial associations with immune cells (B cells, plasma cells, CD4+ T cells, macrophages) and serving as the closest cell type to epithelial/carcinoma cells.
Segregation vs. Mixing: Samples were stratified based on the spatial relationship (bivariate Moran's I) between fibroblasts and epithelial/carcinoma cells.
- Segregated State: Samples with sharper spatial segregation (clearer boundaries) between fibroblasts and tumor cells were associated with enhanced immune-related pathway expression (e.g., antibody production) and significantly improved patient survival across multiple cohorts.
- Mixed State: Samples with mixed spatial patterns correlated with different Cancer-Associated Fibroblast (CAF) subtypes. Specifically, segregated states were linked to antigen-presenting (apCAFs), matrix (mCAFs), dividing (dCAFs), and reticular-like (rCAFs) subtypes, whereas intermixed states correlated with metabolic CAFs (meCAFs).

Significance and Claims

The authors position SEMA as a powerful and efficient tool for uncovering spatially resolved pathway activities. The paper claims that by explicitly incorporating spatial context, SEMA avoids the misinterpretations inherent in methods that treat spatial data as isolated points.

The significance of the work lies in its ability to:

Decode Spatial Architecture: Provide a more faithful representation of biological function in tissues by accounting for the physical execution space of cellular functions.
Enable Large-Scale Analysis: Offer a computationally efficient solution for the rapidly expanding scale of spatial transcriptomic datasets.
Generate Novel Biological Hypotheses: Reveal specific spatial associations, such as the link between arachidonic acid metabolism and TLS, and the prognostic value of fibroblast-tumor cell segregation, which may inform strategies for remodeling the tumor microenvironment.

The authors acknowledge limitations, noting that the metabolic and ligand-receptor findings are based on gene expression and require future validation via spatial metabolomics and functional assays. Similarly, cell-type distributions are inferred from marker gene sets, and future integration of single-cell resolution data could further refine these interaction maps.