Cancer tissue is not a uniform mass of identical cells. It is a complex ecosystem where cancer cells, immune cells, structural cells, and proteins are arranged in specific spatial patterns. This organization -- or tumor microenvironment (TME) -- influences how aggressively cancer grows, how well it evades the immune system, and how it responds to treatment.
For decades, researchers have studied cancer either by analyzing gene expression (which genes are active and at what levels) or by examining tissue images under a microscope. Both approaches can predict patient outcomes, but they have been largely disconnected -- researchers could not directly link the spatial arrangement of cells in an image to the molecular activity happening in those same cells.
Spatial transcriptomics (ST) is a new technology that bridges this gap. It measures gene expression at thousands of specific locations across a tissue section, while simultaneously capturing the tissue image. This allows researchers to ask: which genes are active in specific regions of the tumor, and does gene activity correlate with how cells are organized in space?
This study is the first to link topological features extracted from cancer histology images to spatial transcriptomic data, identifying genes whose expression correlates with the physical arrangement of cells in breast and prostate cancer tissue.
Topological data analysis (TDA) is a branch of mathematics that studies the shape of data. In this context, it is applied to histopathology images to quantify how cells are organized in space -- not just individual cell shapes, but the patterns formed by many cells interacting at a tissue-wide level.
The specific TDA technique used is called persistent homology. Imagine each cell as a point on a map. As you conceptually grow circles around each cell, they begin to overlap and form connected shapes -- triangles, loops, rings. The birth and death (persistence) of these shapes at different scales creates a mathematical fingerprint of the tissue's spatial organization.
This approach captures two types of features: 0-dimensional features (clusters of cells) and 1-dimensional features (loops or ring-like structures in cell arrangements). Both types of features describe different aspects of how cells group and connect across the tissue section.
A key advantage of topological features over standard deep learning features is that they are not task-specific. A neural network trained to recognize cats cannot usefully describe tissue architecture, but topology is a general mathematical language that can describe any spatial arrangement of data points. This also makes topological features directly comparable across different cancer types.
The researchers analyzed six 10x Genomics Visium spatial transcriptomic slides: three breast cancer slides and three prostate cancer slides (two cancerous and one normal prostate). Each slide contained 2,000 to 4,000 measurement spots, each capturing both gene expression data and a corresponding tissue image patch.
From the tissue images, 1,400 image topological features (ITFs) were extracted for each spot using TDA methods applied to 350x350 pixel image patches centered on each measurement location. The top 150 most variable genes were selected per slide for the correlation analysis.
The core method was calculating Pearson correlation coefficients between each gene's expression level and each ITF value across all spots on a slide. Genes that strongly correlate with topological features are called topology-associated genes (TAGs) -- their activity goes up or down in concert with changes in cellular spatial organization.
TAGs were grouped into clusters using unsupervised hierarchical clustering, then subjected to functional enrichment analysis to identify which biological pathways or gene ontology terms were overrepresented. This reveals what biological processes are associated with specific tissue architecture patterns.
The most consistent finding across both cancer types was that extracellular matrix (ECM) genes are among the strongest topology-associated genes. Five ECM-related gene ontology terms appeared in both breast and prostate cancer samples: extracellular space, extracellular exosome, extracellular vesicle, extracellular organelle, and extracellular membrane-bounded organelle.
The ECM is the structural scaffold that surrounds and supports cells in tissue. It provides physical structure, regulates cell behavior, and mediates communication between cells. In cancer, the ECM is known to promote tumor growth, enable cancer cell migration, and create barriers to drug delivery. The finding that ECM gene expression correlates with tissue topology makes biological sense: ECM proteins literally shape how cells are arranged.
Specifically, Collagen Type I Trimer genes showed significant enrichment in both cancer types. Collagen is the most abundant structural protein in the ECM. Cancer-associated fibroblasts produce collagen, which creates tracks along which cancer cells can migrate. This collagen scaffold directly determines the spatial layout of cells -- connecting gene activity to visible tissue architecture.
Extracellular vesicles and exosomes -- tiny membrane-enclosed particles secreted by cells -- also consistently appeared as TAGs. These structures mediate communication between cells and have been shown to promote the spread of prostate cancer. Their association with topological features suggests that regions with specific spatial organization may be more active in cell-to-cell signaling.
Beyond the shared ECM signature, each cancer type had unique topology-associated gene sets. In breast cancer, the ZAG-PIP complex consistently appeared across multiple slides. This complex of zinc alpha2-glycoprotein (ZAG) and prolactin-inducible protein (PIP) plays roles in cell cycle regulation, immune response, and is connected to androgen signaling -- notably relevant as both proteins are also studied in prostate cancer.
In prostate cancer, the dominant topology signature was smooth muscle contraction genes. This reflects the anatomy of the prostate, which contains abundant smooth muscle tissue. Both 0-dimensional (cluster) and 1-dimensional (ring) topological features correlated with smooth muscle gene expression across all prostate slides, including the normal prostate tissue sample.
In the prostate acinar cell carcinoma sample, there was also enrichment for IgA immunoglobulin complex genes. The prostate normally secretes IgA, and changes in IgA expression may indicate altered immune function in prostate cancer. IgA-producing B-cells have been identified as markers of an immunosuppressive cell population in prostate cancer -- suggesting that specific spatial arrangements of immune cells are associated with IgA-mediated immune suppression.
The presence of GI dysmotility gene enrichment in prostate cancer samples is an unexpected finding that may reflect shared signaling pathways between prostate smooth muscle and gastrointestinal smooth muscle biology, or could indicate that certain spatial patterns in prostate tissue are associated with abnormal muscle signaling pathways.
In the FFPE human breast cancer slide, the researchers went deeper to examine whether specific topological features could predict immune cell activity. Some regions of the tumor showed denser immune infiltration -- clusters of immune cells attacking the tumor. These regions have distinct spatial arrangements measurable by ITFs.
Using a deep learning tool called DEGAS, the researchers estimated T-cell abundance across the entire spatial transcriptomic array. They confirmed that regions with high T-cell enrichment also expressed high levels of CD8 (a cytotoxic T-cell marker), TCR (T-cell receptor genes), and MHC-I/HLA genes (which cancer cells use to present antigens to T-cells). These signals converged on specific tumor-peripheral regions.
Critically, a specific topological feature called ITF121 showed a similar spatial distribution to these immune signaling markers. This suggests that a measurable change in how cells are physically arranged in the tissue corresponds to areas where CD8 T-cells are actively engaging with cancer cells through MHC-I antigen presentation.
Machine learning models (LightGBM) trained on just 12 image topological features could predict the expression of 19 immune-related genes with Pearson correlations of 0.49-0.58. This means that tissue architecture alone -- without measuring gene expression directly -- carries significant information about the molecular immune state of the tumor.
The study demonstrates that tissue architecture -- how cells are arranged in space -- is not merely a consequence of cancer, but is actively associated with specific molecular programs. ECM genes, immune signaling genes, and structural genes all show coordinated activity in regions with specific spatial organization.
For diagnostic pathology, this finding suggests that quantifying tissue topology from standard hematoxylin and eosin (H&E) stained slides could provide molecular information without requiring expensive genomic assays. A pathologist reviewing a slide cannot see which genes are active, but topological image features measurable by computers might carry that information implicitly.
From a therapeutic perspective, the consistent enrichment of ECM-related genes in both cancers underscores the TME as a treatment target. Drugs that modulate collagen networks, extracellular vesicle signaling, or ECM composition could potentially alter the spatial organization of tumors in ways that make them more accessible to immune cells or therapeutic drugs.
The ability to predict immune gene expression from image topology also has implications for immunotherapy response prediction. If the spatial arrangement of cells in a tumor can indicate whether cytotoxic T-cells are actively engaging with cancer cells, topology-based features could help identify patients most likely to benefit from checkpoint inhibitor therapy.
This study establishes the first direct link between topological features from cancer histology images and spatial transcriptomic gene expression. TAGs were found in every analyzed slide, across both breast and prostate cancers, confirming the general validity of the approach.
The most robust finding is the universal association between ECM and collagen genes and tissue topology -- a biologically compelling result given that ECM proteins physically determine how cells are spatially organized. This connection between molecular activity and structural organization is measurable and quantifiable.
Cancer-specific signatures -- ZAG-PIP in breast cancer and smooth muscle/IgA in prostate cancer -- suggest that topology-based analysis can reveal biologically meaningful, cancer-type-specific molecular programs. These could serve as biomarkers for diagnosis, prognosis, or therapy selection.
Topology-based image analysis offers a scalable, task-agnostic framework for extracting biological meaning from routine pathology images. As spatial transcriptomic datasets grow larger, methods that bridge imaging and molecular data will become increasingly important tools for understanding and treating cancer.