Endometrial cancer is now classified into four molecular subtypes by the TCGA framework: POLE ultramutated (best prognosis), MSI-H (microsatellite instability-high, good response to immunotherapy), CN-L (copy number-low, diverse outcomes), and CN-H (copy number-high, worst prognosis). These subtypes have meaningfully improved treatment planning, but they still miss a critical dimension: the spatial organization of immune and stromal cells surrounding the tumor.
Traditional molecular analysis provides an average signal across the entire tumor, masking the complex spatial relationships between cell types. For example, whether T cells are present in the tumor center versus the invasive margin has enormous implications for immune killing - and this spatial information is invisible to bulk RNA sequencing. Additionally, molecular subtypes alone do not fully explain why some patients within the same subtype recur while others remain disease-free. The tumor microenvironment (TME) - all the immune, stromal, and vascular cells surrounding the cancer cells - is a major source of this unexplained heterogeneity.
This study used Imaging Mass Cytometry (IMC) - a technology that simultaneously measures up to 40+ proteins at single-cell resolution while preserving their spatial location within the tissue - to profile the tumor microenvironment of 40 endometrial cancer patients across all four molecular subtypes. By mapping which cells are where and how they interact spatially, the study aimed to discover new prognostic features and build a machine learning model to predict disease recurrence.
Formalin-fixed paraffin-embedded tumor specimens were sectioned from 40 early-stage endometrioid endometrial carcinoma patients, with 2 tissue regions of interest (ROIs) collected per patient, yielding 80 ROIs total. An optimized panel of 41 markers (39 proteins plus DNA) was measured by IMC, covering epithelial, endothelial, stromal, and immune cell lineages, as well as functional markers including Ki-67 (proliferation), HIF1a (hypoxia), CXCR4 (immune chemotaxis), and the immune checkpoints PD-1 and PD-L1.
Cell segmentation was performed using the pre-trained TissueNet algorithm, which delineates individual cell boundaries. Batch correction across samples used the Harmony algorithm. Cell type clustering was performed with the Rphenograph algorithm, identifying 10 main cell populations including T lymphocytes, B lymphocytes, NK cells, myeloid cells, endothelial cells, epithelial cells, fibroblasts, stromal cells, CD90+ cells, and other lineages. In total, 951,754 cells were analyzed across the 80 ROIs.
Three types of spatial features were extracted per ROI: cell population frequencies, cell-cell interaction patterns (using permutation tests to identify which cell type pairs were significantly close to each other), and functional marker expression levels. Cellular neighborhoods (CNs) were defined by identifying the 20 nearest spatial neighbors around each cell, then clustering these neighborhood patterns into 10 CN categories. A random forest classifier was trained on these features with a 70/30 training-validation split, repeated 1,000 times, and evaluated by AUC.
The four molecular subtypes showed dramatically different cellular compositions. POLE-type tumors had the highest overall cell density (significantly greater than CNL and MSI-H subtypes). Compared to CN-L, both POLE and MSI-H subtypes showed significantly higher proportions of active CD4+ and cytotoxic CD8+ T cells - consistent with these subtypes' better response to immunotherapy. CN-L tumors were characterized by higher proportions of epithelial and fibroblastic cells and lower immune infiltration, with enhanced epithelial-mesenchymal transition (EMT) - the biological process by which cancer cells gain invasive properties.
Six coordinated cellular programs (NMF1-6) were identified using non-negative matrix factorization (NMF) analysis. The most clinically significant was NMF6, described as a tumor progression program, which was enriched for CD90+CD105+ endothelial cells, M2 macrophages (immunosuppressive type), and regulatory T cells (Tregs) - a combination that collectively suppresses anti-tumor immunity and promotes tumor growth. CN-L tumors showed higher NMF1 and NMF5 (epithelial and fibroblastic programs) and lower NMF3 (lymphocyte program) and NMF6.
Among the 10 cellular neighborhood types identified, CN5 (activated T cell-enriched neighborhoods) was significantly more common in POLE and MSI-H subtypes than in CN-L, consistent with the better prognosis of these subtypes. The study also found that CD90+ cell density correlated with patient age, while myeloid cell-enriched neighborhoods and NMF6 expression were associated with larger tumor size (diameter greater than 2 cm).
The study's most novel finding was the identification of CD90+CD105+ double-positive endothelial cells as a key immunomodulatory population enriched in the CN-H subtype. Flow cytometry confirmed that CN-H tumors contained 23.2% of these cells versus 12.8% in POLE tumors (p less than 0.05). POLE tumors, which have the best prognosis, had the sparsest distribution of these cells across all subtypes.
Laboratory experiments validated the functional significance of this cell population. When CD90+CD105+ endothelial cells were co-cultured with THP-1 macrophage precursor cells, the macrophages preferentially polarized toward the M2 phenotype (immunosuppressive) rather than M1 (anti-tumor), as shown by elevated IL-10 (M2 cytokine) and unchanged TNF-alpha (M1 cytokine) levels. Transwell migration experiments confirmed that CD90+CD105+ endothelial cells actively recruit macrophages, and angiogenesis experiments showed they promote tumor blood vessel formation. Co-culture with the endometrial cancer cell line Ishikawa accelerated cancer cell proliferation.
Spatial interaction analysis in CN-H patients further revealed that interactions between CD90+CD105+ endothelial cells and immunosuppressive cells (M2 macrophages with high CXCR4 expression and Tregs) strongly predicted poor outcomes. Conversely, interactions between CD90-CD105- endothelial cells (normal endothelial cells) and activated CD4+ T cells were associated with better prognosis. This suggests that the balance between immunosuppressive and anti-tumor endothelial-immune interactions determines clinical outcomes within the same molecular subtype.
The random forest model was trained on three feature categories derived from IMC data: cell population frequencies, cell-cell interaction patterns, and functional marker expression. Feature importance analysis identified the top five features for two separate tasks: molecular subtype classification (distinguishing the four TCGA subtypes) and recurrence prediction (predicting which CN-H patients would recur). The molecular subtype model achieved an AUC of 0.923; the recurrence model achieved an AUC of 0.945.
The key features for recurrence prediction in CN-H patients included: IL-1beta expression on activated CD4+ T cells (associated with better prognosis, as this cytokine promotes immune activation), interactions between CD90-CD105- normal endothelial cells and activated CD4+ T cells (also protective), and interactions between inflammatory fibroblasts and NK cells (an independent risk factor for poor prognosis). The model's high performance despite the relatively small cohort suggests that spatial proteomic features encode strong prognostic signals.
This is one of the first studies to demonstrate that spatial proteomics data from routine clinical tissue samples - specifically FFPE (formalin-fixed paraffin-embedded) archival tissue, the standard format for clinical specimens - can be used to train machine learning models that predict recurrence with very high accuracy. If validated in larger cohorts, this approach could be applied to standard surgical specimens without requiring fresh tissue, making it practically feasible for clinical deployment.
This study establishes a new framework for understanding endometrial cancer heterogeneity: the spatial ecostructure - the organized spatial architecture of all cell types within and around the tumor - is a critical determinant of clinical outcomes that cannot be captured by molecular subtyping alone. Two patients with identical CN-H molecular subtypes can have dramatically different recurrence risks depending on the spatial relationships among their endothelial, macrophage, T cell, and fibroblast populations.
The identification of CD90+CD105+ endothelial cells as a tumor-promoting immunosuppressive population is a clinically actionable finding. These cells are enriched in the highest-risk CN-H subtype, promote M2 macrophage polarization and angiogenesis, and their spatial interactions with immunosuppressive cells predict poor prognosis. Anti-angiogenic therapies targeting the CD90/CD105 axis, or strategies to block the CD90+CD105+ endothelial cell-macrophage interaction, represent potential therapeutic directions worth investigating in CN-H endometrial cancer.
The study's limitations include its single-center retrospective design with 40 patients, which necessitates validation in larger multicenter cohorts. The cross-sectional design captures a single tumor snapshot rather than the dynamic evolution of the microenvironment over time. Future studies incorporating longitudinal multi-time-point sampling, organoid co-culture experiments, and humanized mouse models will be needed to establish causal mechanisms and confirm the clinical utility of spatial ecostructural modeling in endometrial cancer treatment planning.