Computational modeling of cellular influence delineates functionally relevant and distinct cellular neighborhoods in primary and metastatic pancreatic ductal adenocarcinoma

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal cancer, with liver metastases significantly worsening outcomes. However, features of the tumor microenvironment (TME) that are distinct between primary and metastatic sites remain poorly defined. Cellular neighborhoods within the TME are recognized as functional units that influence tumor behavior. Conventional spatial methods, which assign equal weights to all cells in a region, fail to capture the nuances of cellular interactions. To address this, we report here the development of Functional Cellular Neighborhood (FunCN) quantification, which integrates both the proportion and proximity of surrounding cells. Applying FunCN to PDAC imaging mass cytometry data, we identified neutrophil-enriched interactions in liver metastases compared to primary tumors, correlating with elevated VISTA expression by tumor cells. Additionally, FunCN clusters around CD8⁺ T cells in pancreas and liver were associated with higher TIGIT and LAG3, respectively. These findings demonstrate the importance of spatial immune landscapes in PDAC and identify potential therapeutic opportunities.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is among the most deadliest malignancies (1), largely due to its propensity to metastasize early to distant organs and its poor response to existing treatments (2). More than 50% of PDAC patients present with liver metastases at diagnosis, and these are associated with particularly poor prognosis (3). Thus, understanding mechanisms that enable tumor cells to spread and proliferate at distant metastatic sites, including the liver, is crucial for the identification of novel therapeutic targets that may improve patient outcomes.

Recent studies have shown that primary and metastatic tumors are heterogeneous and that the tumor microenvironment (TME) plays a key role in modulating cancer progression and metastasis. The TME is defined not only by its cellular components but also by interactions between tumor cells and non-tumor cells, including immune cells and fibroblasts, that can support or suppress tumor growth (4). In fact, the distance between immune and tumor cells directly reflects both the effectiveness of immune cells in eliminating tumor cells and the tumor’s capacity to influence immune cells (5). For example, a recent study demonstrated that tumors with PD-1⁺ T cells located within approximately 10 μm of PD-L1⁺ tumor cells exhibited significantly longer progression-free survival following PD-1 blockade therapy (6), highlighting the importance of assessing spatial interactions when evaluating therapeutic outcomes. Previous studies have integrated cellular proximity and influence by taking into consideration distances between cells and the cell density within specific regions (7–9). However, these methods impose rigid spatial thresholds, either by fixing the number of nearest neighbors or by categorizing interactions using predefined ranges (e.g., 10–20 μm, 20–30 μm), which may overlook biologically relevant cells just beyond these cutoffs. Moreover, they assign equal weights to all included cells, regardless of their varying distance from the reference cell. This uniform weighting limits the resolution and accuracy of spatial interaction analysis within the TME.

To address these limitations and achieve a more comprehensive characterization of the TME, we developed a method, which we called Functional Cellular Neighborhoods (FunCN), that quantifies cellular interactions by considering both the proximity and frequency of neighboring cells. Specifically, FunCN uses a spatial kernel-based approach to model the decrease in a cell’s influence on other cells as they get farther apart. We applied this approach to investigate the TME of primary and metastatic PDAC and identified distinct neighborhoods of CD8⁺ T cells and tumor cells. We also found cellular interactions that were unique to the pancreas and liver. Overall, our study provides insight into the cellular interactions within primary and metastatic tumors that may guide future therapeutic strategies.

Materials and Methods

Human samples

Primary and metastatic tumors were obtained from 12 patients with metastatic PDAC through the Johns Hopkins Research Autopsy Program. Written informed consent was obtained from all research subjects, and the research protocol was approved by the institutional review board of the Johns Hopkins School of Medicine (protocol number NA_00036610). Tissue collection was carried out in compliance with the 1996 Declaration of Helsinki.

Collected tissue specimens were fixed in formalin and embedded in paraffin. Tissue sections were then stained with hematoxylin and eosin and reviewed by board certified anatomic pathologists (J.E.H. and L.D.W.). Regions containing tumor cells were then selected to construct tissue microarrays (TMAs) with each core measuring approximately 1 mm in diameter. Patient characteristics and the number of cores analyzed from each patient are shown in Supplementary Table S1.

Imaging mass cytometry (IMC) staining and IMC processing

IMC staining on TMAs was performed as previously described (10). Briefly, the TMA slides were baked, deparaffinized in xylene, and rehydrated in an alcohol gradient. Heat-mediated antigen retrieval was performed at 95 °C for 30 minutes using Antigen Retrieval Agent pH9 (Agilent S2367). The slides were then blocked with 3% BSA for 45 min at room temperature, followed by overnight staining at 4 °C with the antibody cocktail. Information on antibodies used for IMC is provided in Supplementary Table S2. After staining, images were acquired using the Hyperion Imaging System (Standard BioTools) at the Johns Hopkins Mass Cytometry Facility. IMC provides high-dimensional spatial profiling of tissue slides with a cell resolution at 1 μm. Images were visualized using MCD Viewer (Standard BioTools). For data analysis, image segmentation was carried out using CellProfiler (11) (version 3.1.9), ilastik (12) (1.4.1rc2-gpu), and HistoCAT (13) (version 1.76).

IMC data pre-processing

The IMC data was pre-processed with the removal of benign cells, followed by log-normalization of the expression data (log2(expr + 1)). We performed batch effect correction using the Harmony (14) R package (version 1.2.3), accounting for patient and tissue site differences. The adjusted expression values were subsequently shifted to ensure non-negative values by subtracting the minimum expression level for each marker.

Cell-type annotation

Cell clustering.

We performed clustering using a Self-Organizing Map (SOM) using FlowSOM (15) R package (version 2.12.0), which organized the high-dimensional IMC data into a 35×35 grid, resulting in 1,225 nodes. . Cell-type specific markers (CD3, CD8, FOXP3, CD4, granzyme B (GZMB), CD45RA, CD57, DCSIGN, CD15, CD86, CD163, CD206, CD68, HLA-DR, CD74, VISTA, COL1A1, smooth muscle actin and vimentin (SMA/VIM), pan-keratin (CK), and Ki-67) from the panel (Supplementary Fig. S1A) were selected for clustering. The cell expression values within each node were summarized by their median expression of the cells in that node, and outliers were controlled by applying a median absolute deviation (MAD) threshold of 2 median absolute deviations (MADs). Subsequent metaclustering of the SOM nodes using Consensus Clustering was performed with ConsensusClusterPlus R package (version 1.68) to identify distinct 50 clusters using Euclidean distance

Visualization and annotation.

To facilitate robust cell type annotation, we generated a heatmap visualization of the protein expression for each cluster. The expression data was normalized using min-max scaling to a [0, 1] range, based on the 33rd and 99th percentiles for heatmap visualization.

$$x_{vis}=\min \left(\max \left(0,a+\frac{\left(x-Q_a\left(x\right)\right)\left(b-a\right)}{Q_b\left(x\right)-Q_a\left(x\right)}\right),1\right),$$

where $Q_a\left(x\right)$ and $Q_b\left(x\right)$ are the a^th and b^th quantile values of the expression data $x$ before normalization. We used $a=0.33,b=0.99$ to highlight positive outliers of protein expressions compared to a symmetric mapping and consequently identify distinct cell types. This visualization improvement is especially beneficial in cases where data sparsity and a skewed distribution could lead to spurious normalization effects. For instance, setting $a=0.01,b=0.99$ for a highly skewed protein expression distribution results in lower quantiles ($x<0.5$) of expression values to map to ($x_{vis}>0.5$), resulting in a distorted and misleading visualizations of the expression values. Since we primarily use overexpression of proteins to identify cellular phenotypes, the more robust mapping from $x\to {\mathrm{x}}_{\mathrm{v}\mathrm{i}\mathrm{s}}$ for $x<0.5$ obtained using $a=0.33$, allowing for a more accurate cell type annotation. The normalized values were summarized within each cluster, and hierarchical clustering was applied.

Subtype annotation and statistical analysis of cellular abundance.

After annotation, subtypes of CD4⁺ T cells, other myeloid cells (excluding the neutrophil cell type) and stromal cells were identified using the same clustering method but with a 10×10 grid. The heatmap was then generated using ComplexHeatmap (16) R package (version 2.20) in R to visualize the clustering result. The abundance of each cell type was visualized using bar plots generated using ggplot2 package (version 3.5.1) in R for each tissue site, core, and patient. To evaluate differences in cell-type densities across tissue sites, an unpaired Wilcoxon rank-sum test was performed using the ggpubr package (version 0.6) in R. The resulting median values and standard deviations were then illustrated on scatter plots.

Functional cellular neighborhood (FunCN) quantification

We developed a method for quantifying functionally relevant cellular neighborhoods as observed by each individual cell in the spatial dataset. Previous approaches define a cell’s neighborhood either by including all cells within a given radius r from a reference cell (distance based cellular neighborhood [CN] (17,18)), or by selecting the k-nearest neighbors (KNN) of the reference cell (KNN-based CN (19)), and summarize the proportion of each cell type among the neighboring cells. However, these methods have limitations. First, they do not account for the relative distance of neighboring cells from the reference cell, treating nearby and farther cells equally within the boundary. Second, both methods place a definite boundary on the cellular neighborhood and hence do not allow us to model the influence of cells situated beyond a radius r or beyond the k-th nearest neighbor, respectively. For example, a large tertiary lymphoid structure (TLS) at a moderate distance from a cell may have a non-trivial influence on the cell’s function, despite being outside the boundaries described in the prior art. To overcome these limitations, our method for Functional Cellular Neighborhood (FunCN) quantification uses a spatial kernel-based approach to model the influence of all other surrounding cells weighted by their proximity to the reference cell. To quantify the FunCN of the i-th cell $C_i$, we computed a sum of kernel-weighted influences from all neighboring cells:

$$FunCN\left(C_i\right)={\sum }_{j\ne i}K\left(x_i,x_j\right)\cdot type\left(C_j\right),$$

where $type\left(C_j\right)$ is the cell-type of the j-th cell $C_j(j\ne i)$. $x_i$ and $x_j$ are the coordinates of the i-th and j-th cells, and $K(x_i,x_j)$ is a spatial kernel function which decreases as the distance between $x_i$ and $x_j$ increases. In this paper, we define $K$ as a 2-D Gaussian kernel

$$K\left(x_i,x_j\right)\doteq \exp \left(\frac{-{‖x_i-x_j‖}^2}{2{\sigma }^2}\right),$$

where ${‖\cdot ‖}^2$ is the Euclidean distance, and $\sigma$ is the width of the kernel. The resulting FunCN of each cell is a weight vector of length equal to the number of distinct cell types $N_{celltype}$, with nonnegative weights summing to 1, thereby representing the relative proportional contribution of each surrounding cell type (Supplementary Fig. S2A). This approach provides a continuous distance-based representation of spatial influence that extends beyond fixed boundaries. For example, when neighboring cells are equidistant and equal number, their FunCN score remains identical. In contrast, in a fixed-neighbor method (e.g., KNN with k = 10), cells located at the same distance may be inconsistently included or excluded depending on the cutoff, leading to arbitrary discontinuities in neighborhood composition. In another case, when comparing cells located at different radial distances from a reference cell, FunCN values progressively decrease as the radius increases, effectively reflecting the diminishing influence of more distant neighbors. (Supplementary Fig. S2B) This distance-weighted quantification captures continuous spatial effects that conventional fixed-neighbor or fixed-radius methods cannot achieve, as those approaches assign equal weights to all cells within the boundary.

The bandwidth parameter ($\sigma$) controls the spatial scale of neighborhood influence in FunCN quantification. A smaller $\sigma$ focuses the analysis on the immediate local microenvironment – capturing highly localized, short-range cellular interactions – whereas increasing $\sigma$ broadens the scope to encompass a wider region of the tissue, effectively integrating influences from more distant cells and grouping together spatially distinct microenvironments. In this study, the bandwidth was chosen to best capture the local microenvironment by placing greater emphasis on nearby cells, aligning with the critical distance range associated with improved patient survival due to enhanced T-cell infiltration within a 10 μm range of tumor (6,20). We used the spatstat (21) R package (version 3.3.1) to evaluate the weights and subsequently the weighted influence of each cell type for every cell in the dataset (see github repository https://github.com/DeshpandeLab/Spatial_Influence).

Comparison of FunCN, KNN and fixed-radius approaches

FunCN was compared to the conventional KNN and fixed-radius approaches to assess differences in how spatial proximity is incorporated into cellular influence quantification. To compute the KNN for cellular influences, the FNN R package (version 1.1.4.1) was used to identify KNNs for each cell based on spatial coordinates, and the relative proportions of distinct cell types among these neighbors were calculated. For the fixed-radius approach, pairwise Euclidean distances between cells were computed using the spatstat R package, and the proportions of cell types within a specified radius were determined. Spatial density of cell counts was visualized using the spatstat R package. Comparison of FunCN and KNN method were visualized using dot plots, violin plots and histograms using ggplot2.

Cellular Neighborhood Clustering and Downstream Analysis

Cells were grouped by clustering their FunCN vectors using self-organizing map (SOM) clustering, which grouped cells with similar local microenvironment. To characterize the microenvironment of each CN cluster, we visualized the clusters using dot plot where dot size represented the median cellular influences of each cell type within a cluster and dot color indicated the scaled median value across CNs. Functional marker expression was further evaluated within each clustered CN using radial plots which displayed the scaled median expression of each marker across CNs, relating variations in local microenvironments to functional cell states. To examine trends in functional marker expression across CN clusters, line plots were generated showing absolute median expression values with 95% confidence intervals (calculated as median ± 2 × standard error) for each site. Core 43, which contained approximately 50% immune cell mixture, was excluded from the downstream analysis.

Results

Influence-weighted CNs incorporate proximity and frequency of neighboring cells to enhance TME characterization

We report here FunCN, a method for quantifying the CNs using a spatial kernel-based method to model cellular influence in the TME. In our approach, we applied a 2D spatial kernel to model how cellular influence on surrounding cells decays with increasing distance. Spatial kernels (mathematical functions) have been previously used to model influence of cells or biological patterns in spatially resolved transcriptomics (22). Extending beyond regional-scale analyses, we built on this concept to characterize FunCN for each individual cell, represented as a vector capturing the accumulated influence by all cell types within the tissue.

Applying this method to map the cellular architecture of PDAC across different sites, we utilized IMC with a 35-antibody panel to generate high-dimensional spatial data. A total of 267,239 cells from 64 tissue cores were analyzed, collected from 12 unique patients. This included liver metastasis samples from 10 patients and pancreatic primary tumor samples from 8 patients, with matched primary and metastatic tumors collected from 6 patients. After image processing and cell segmentation, cell types were annotated by clustering cells based on lineage marker expression (Fig. 1A).

Figure 1. Overview of spatial profiling and cellular influence analysis workflow.

A, Metastatic PDAC samples from the liver (n = 10) and pancreas (n = 8) spatial data were acquired using imaging mass cytometry, followed by data processing and cell type phenotyping. The schematic shown was created using BioRender. B, Schematic workflow illustrating the quantification of cellular influences using the FunCN method. Spatial objects were transformed and weighted using a spatial kernel, emphasizing cells that are in proximity. This approach captures variations in cellular influence quantification based on spatial organization despite the same cell counts within a region. C, Downstream analyses included cellular neighborhood classification and evaluation of functional states.

Following cell-type annotation, the FunCN model was applied. Similar to KNN-based or distance-based CNs, we normalized the FunCN to represent the proportional influence exerted by each distinct cell type. However, unlike KNN or distance-based CN’s, FunCN does not create an arbitrary boundary for the cellular neighborhood and instead, considers the potential influence from all cells in the tissue, weighted by a rapidly decreasing spatial kernel. By implicitly incorporating both the proximity and local abundance of cell types, FunCN can reveal subtle differences in spatial organizations through distinct patterns of proportional influence, even when overall cellular composition is identical (Fig. 1B). By associating the strongest influence from the nearest neighbors, this approach provides a high-resolution, spatially resolved representation of cell–cell interactions, revealing how variations in local arrangement shape cellular influence.

The kernel function and width can be set to model either physical interactions between adjacent cells or diffusion of a secreted cytokine (23,24) within a broader tissue region. The kernel bandwidth parameter determines the rate of decay of influence with distance. For instance, a smaller bandwidth places greater emphasis on the nearest cells and drastically decreases the influence as the distance increases. In contrast, a larger bandwidth results in a more gradual decrease in the influence over a broader radius, distributing weights more evenly across a larger area.

We then performed downstream analyses to identify groups of cells that shared similar local microenvironments. By leveraging spatial features derived from FunCN quantification, we clustered cells into distinct CNs, each representing a unique configuration of neighboring cell types and spatial proximity. Within each CN, we further examined cell state heterogeneity to understand how cellular behavior is shaped by spatial interactions within the TME (Fig. 1C).

Site-specific cellular compositions of primary and metastatic PDAC

To investigate differences in the cellular landscape between primary and metastatic PDAC, we examined site-specific cellular compositions. SOM clustering of protein expression profile was used to define cellular phenotypes as well as cells that could not be definitively assigned a cell type (“unassigned”, UA). Further refinement of the clustering identified regulatory T cells (Tregs), six distinct myeloid subtypes (M_I–M_VI) and seven stromal subtypes (Str_I–Str_VII), each characterized by specific marker expression patterns (Fig. 2A).

Figure 2. Cell type identification, and distribution across pancreatic and liver sites.

A, Cell-type phenotyping: Scaled expression profile of cell type signatures for each cluster is reflected as heatmap. B, Proportional representation of cell types across sites and the relative abundance of each site shown as a fraction of the total cell population. C, Cell types with significantly different densities (measured as the number of cells per mm² area) between pancreas and liver. Individual points represent per-sample values and error bars represent standard deviation. Differences were assessed using an unpaired Wilcoxon test. D, Validation of cell type annotations. Representative cell mask images (left) from pancreatic and liver samples were compared to matched MCD images displaying relevant cell type markers. Scale bar: 100 μm.

The seven stromal subtypes displayed distinct phenotypic characteristics. Str_I (COL1A1⁺SMA/VIM^+/lo) was akin to myofibroblastic cancer-associated fibroblasts (myCAF), while Str_II (COL1A1^hiCD74⁺SMA/VIM⁺HLA-DR⁺) akin to antigen-presenting cancer-associated fibroblasts (apCAF). Str_III (COL1A1^hiCD74^hi) represented an extracellular matrix (ECM)-rich stromal population, and Str_IV (SMA/VIM^hi) resembled a canonical myofibroblastic CAF subtype. The remaining subtypes Str_V–Str_VII, formed a podoplanin^hi stromal population with distinct combinatorial marker profiles: Str_V co-expressed high expression of CD74, COL1A1, HLA-DR, and SMA/VIM; Str_VI demonstrated moderate COL1A1 and CD74 expression; and Str_VII additionally high HLA-DR expression. The six myeloid subtypes demonstrated combinations of myeloid and activation markers. M_I cells (CD163^loCD206^lo) and M_II cells (HLA-DR^hiCD68^hiCD163^hiCD206^hiCD86⁺) were the most abundant across both tissue sites. M_III and M_IV and M_VI subtypes were defined by DC-SIGN expression, with M_III co-expressing CD86, M_IV co-expressing CD163, and M_VI co-expression CD206, CD163 and HLA-DR. M_V represented a CD68^hi population with mild CD86 and HLA-DR expression. Collectively, these profiles indicate the presence of phenotypically diverse myeloid populations (Supplementary Fig. S1B–D). Among the resolved populations, we also observed an immune cluster that co-expressed markers of multiple immune lineages (CD4, CD8, DC-SIGN, CD86, CD206, and HLA-DR). Histological examination and raw images confirmed that this region contained a highly dense mixture of immune cell types, including dendritic cells (DCs), CD8⁺ T cells, and CD4⁺ T cells. A subset of UA cells exhibited low expression across all markers. Based on histologic review, UA cells were mostly poorly differentiated tumor cells or other mesenchymal cells (Supplementary Fig. S1E).

Comparison of cell-type proportions between primary pancreatic tumors and liver metastases revealed distinct site-specific differences (Fig. 2B–C). M_I cells (CD163^loCD206^lo) were more abundant in the pancreas whereas Tregs, Str_II (akin to apCAF), M_IV (CD163^hiDC-SIGN⁺), and neutrophils were more enriched in the liver. When comparing the relative abundance of immune cells, both pancreatic primary tumors and liver metastatic lesions exhibited high proportions of myeloid cells and CD15⁺ neutrophils, with M_I and M_II subtypes being particularly enriched (Supplementary Fig. S1F–G). These findings highlighted distinct immune and stromal compositions in primary pancreatic tumors versus metastatic tumors in the liver, warranting an in-depth evaluation of site-specific spatial interactions and functional implications. Cell-type annotation was validated through the examination of cell masks colored by cell type at the pixel level (Fig. 2D) and histological review of tissue cores.

FunCN reveals enhanced resolution of spatial interactions in the TME

To validate our method, we compared the combined influence of CD8⁺ T and CD4⁺ T cells on tumor cells in core 51 (pancreas) with the spatial density plot of these two immune cell types in the same region. We observed high levels of combined CD8⁺ T-cell and CD4⁺ T-cell influence on tumor cells located near areas where CD8⁺ T cells and CD4⁺ T cells were densely concentrated (Fig. 3A). We then explored how our FunCN approach detected distance-based spatial patterns compared to the conventional KNN method. Specifically, when quantifying cellular influence from the 10 nearest neighbors in core 51, the FunCN method produced a wider, continuous range of values, whereas the KNN method provided only 0.1-step discrete increments (Fig. 3B). The broader spectrum from the FunCN approach was especially pronounced at mid-range values, suggesting that in the regions of very low or very high cell frequencies, the frequency of cells was the main driver of cellular influence, whereas in the mid-range, distance-based weighting played a larger role.

Figure 3. Validation of the FunCN approach for quantifying spatial cellular influences.

A, Visualization of CD8⁺ T and CD4⁺ T-cell influences on tumor cells within a representative core (Core 51). The top panel shows a spatial density plot representing T cell counts, while the bottom panel presents cellular influence values of tumors derived from the FunCN kernel weighting. B, Comparison of cellular influences from CD8⁺ T cells, myeloid cells, and stromal cells across all cells in the representative core, using the FunCN approach versus the KNN-based approach (k=10). For each discrete KNN spatial weight, continuous spatial weights were quantified using the FunCN. C, Violin plots showing cellular influence scores computed by FunCN and KNN methods for CD8⁺ T cells, myeloid cells, and stromal cells when these populations appeared as the first to fourth nearest neighbors (N1–N4). D, Spatial organization of tumor cells showing the largest differences in spatial weights between the FunCN and KNN methods. Tumor microenvironment (TME) of a tumor cell within a ±50 μm radius are visualized, highlighting the increased emphasis on cellular influences when cells are in proximity.

Additionally, unlike the KNN approach, which calculated cellular influence based on a fixed number of neighbors, the FunCN approach applied a continuous distance-weighted kernel that decays with increasing distance from the reference cell (Supplementary Fig. S2A–B). As a result, FunCN showed a gradual decrease in CD8⁺ T-cell, myeloid-cell, and stromal-cell influences as the neighbor order increased (with N1 being the closest and N4 being the furthest neighbor), whereas the KNN method, which treats all neighbors equally regardless of distance, resulted in uniform influence values across neighbor orders (Fig. 3C). To evaluate the effect of bandwidth on spatial influence, we examined $\sigma$ values of 5, 10, 30 and 50, using representative core 51. Smaller $\sigma$ value placed stronger emphasis on immediate neighbors, yielding a highly localized view of the microenvironment, while larger $\sigma$ values produced progressively broader and more diffused influence maps (Supplementary Fig. S2C). In the case of the KNN method, when k is 10 or higher, it compensates for sparse local regions by expanding its boundary to include additional neighbors, resulting in inconsistent neighborhood sizes and influence estimates that do not scale with distance. Increasing k or the neighborhood radius (or $\sigma$ in FunCN) further broadens the zone of influence, causing the effects of rare but spatially proximal cells to be overshadowed by more abundant cell types. By tuning $\sigma$, FunCN allows the spatial scale of analysis to be flexibly adjusted according to the research question, ranging from fine-scale characterization of the local microenvironment to broader assessment of tissue-level organization. In this study, we set $\sigma$ to be 10 to examine the local microenvironment, ensuring that the influence of rare yet spatially close cells were retained without over-emphasizing only the immediately adjacent ones.

We also examined the distribution of CD8⁺ T-cell influence on tumor cells based on the range of distances between them in core 51 (Supplementary Fig. S2D). When tumor cells were arranged by their average distance to CD8⁺ T cells within a 30 μm radius with varying number of CD8⁺ T cells in this range, the FunCN method ($\sigma$ = 10) demonstrated a gradual decline in CD8⁺ T-cell influence with increasing distance. In contrast, the KNN (k = 10) and Radius (r = 30) methods produced influence scores that were primarily driven by the proportion of CD8⁺ T cells within their respective cutoffs rather than spatial distance.

To examine the spatial organization of tissue specimens where the two methods would differ the most, we selected the top five tumor cells with the largest difference in CD8⁺ T-cell influence between the FunCN and KNN methods. A ±50 μm window was set around each reference tumor cell. The results showed that CD8⁺ T cells in close proximity to the tumor cells were assigned higher influence values, with their influence considered more significant in our method (Fig. 3D). This highlighted the ability of our method to capture interactions between tumor and immune cells.

Using our FunCN quantification method, we identified distinct spatial interactions in the primary pancreatic and metastatic liver TME (Supplementary Fig. S3A–B). Tumor cells at both sites showed limited cellular influence from other cell types, with only minimal contributions from M_I and Str_I. However, comparing the pancreas and liver, there was a greater influence of Str_II, Tregs, and neutrophils on tumor cells in the liver than in the pancreas, indicating distinct tumor–immune/stromal interactions in each of the sites. In the liver metastatic TME, stromal subtypes had more spatial influence from T cells. Neutrophil influence on CD8⁺ T cells was more enriched in the liver, whereas influence from other myeloid cells on CD8⁺ T cells was more enriched in the pancreas. These site-specific enrichment of cellular influences, particularly those involving tumor cells and CD8⁺ T cells, suggested that the same cell types could have different functions depending on the tissue microenvironment and the neighboring cell types.

Delineating distinct CNs around CD8⁺ T cells

We next conducted a more integrated analysis of the spatial influences to understand distinct CNs surrounding CD8⁺ T cells and tumor cells. To this end, CD8⁺ T cells were clustered based on their surrounding cell influences (FunCN scores), yielding eight distinct CNs via SOM clustering. These neighborhoods, derived from different patients and tissue sites, each showed unique cellular influences of multiple cell types, depicted by dot size indicating magnitude and color representing relative differences across CNs (Fig. 4A). Two CNs, CN1 and CN2, represented CD8⁺ T-cell neighborhoods enriched for tumor cells. Stromal cells appeared predominantly in CN2–6, with enrichment in CN3 (Str_I abundant) and CN6 (Str_II abundant). CN5 comprised CD8⁺ T cells closely surrounded by M_I myeloid cells and Str_I stromal cells. Three CNs (CN4, CN7, and CN8) were enriched with T-cell and myeloid-cell subtypes at varying levels: CN4 was mixed with CD8⁺ T cells, CD4⁺ T cells, and Tregs; CN7 was primarily CD8⁺ T cells; and CN8 was mixed with CD8⁺ T and M_II type myeloid cells.

Figure 4. Cellular neighborhood (CN) characterization of CD8⁺ T cells.

A, CNs are depicted in a dot plot indicating the cellular influences comprising each neighborhood. Dot size represents the median cellular influence, while color indicates the z-score of differences across distinct CNs. A bar plot shows the sample fraction of each CN. B, Stacked bar plot illustrating the abundance of CNs across different sites. C, Representative images of CD8⁺ T cells in different CNs within a single core, demonstrating varying dominant cellular influences experienced by CD8⁺ T cells in that core. D, For each CN, one CD8⁺ T cell was sampled from each representative core to validate its TME, highlighting relevant cellular influences as defined in the dot plot. E, Analysis of CD8⁺ T cell functional marker expression, shown as radial plots indicating the relative z-scores of marker differences across CNs. F, Line plot of median expression levels for functional markers with a shaded 95% confidence interval. The dashed line indicates the median expression of all CD8⁺ T cells in each site, with all CNs combined.

Among these CNs surrounding CD8⁺ T cells, a large proportion of CD8⁺ T cells was found in proximity to neighborhoods enriched with CD8⁺ T-cells themselves (CN7), stromal cells (CN3, CN6), and M_I-type myeloid cells (CN5). When comparing the two TME sites, we observed a predominance of CN7 over CN4 in the pancreas and the opposite for liver (Fig. 4B), indicating greater influence from CD4⁺ T cells and Tregs on CD8⁺ T cells within the liver TME.

We further validated the CD8⁺ T cell–associated CNs by visual inspection. In each core, CD8⁺ T cells were classified into different CNs based on spatial composition of their surrounding cells, with representative neighboring cell types summarized next to the spatial map (Fig. 4C). To further confirm the spatial organization of CNs within CD8⁺ T cells, one cell per CN was sampled from each of two representative tissue cores 51 (Pancreas) and 26 (Liver), with X and Y coordinates zoomed within ±50 μm around the reference cells (Fig. 4D). Across all CNs, CD8⁺ T cells exhibited a higher frequency and closer interactions with the specific cell types that characterize each neighborhood, consistent with the cellular compositions of the CNs.

We next explored the functional states of CD8⁺ T cells in relation to the CNs (Fig. 4E). CD8⁺ T cells in CN3 (stromal-dominant; driven by Str_I [COL1A1⁺ SMA/VIM^+/lo “myofibroblastic”-like] and minor M_I [CD163ˡᵒCD206ˡᵒ; less activated macrophages] influences), exhibited lower expression of markers that denote activation (GZMB, ICOS, HLA-DR) in both the pancreas and liver. This decreased expression of activation markers was also accompanied by lower expression of inhibitory markers (PTPN22, TIM3, and TIGIT). Taken together, our findings suggested that Str_I and M_I myeloid cells could diminish the activation of CD8⁺ T cells and impair their effector functions.

In contrast, CD8⁺ T cells in CN8 (enriched for M_II myeloid cells [CD68^hi, HLA-DR^hi, CD86⁺, CD163^hi, CD206^hi macrophages] and CD8⁺ T cells) exhibited higher expression of PTPN22, TIM3, TIGIT, LAG3, and HLA-DR compared to other CNs in both pancreas and liver, consistent with a state of immune exhaustion (Fig. 4E). However, CD8⁺ T cells in CN8 in the liver also expressed higher levels of GZMB compared to the pancreas, suggesting the presence of differential regulation by the tissue sites despite the similar cellular neighborhoods. We posit that other immunosuppressive signals in the microenvironment could inhibit overall T cell function despite elevated GZMB expression in the liver (25).

We also compared the functional marker expression level for each CN across both tissue sites. A line plot of median expression levels, with shaded areas representing 95% confidence intervals, illustrated trends in expression changes across CNs and highlighted differences between sites (Fig. 4F). Both sites exhibited remarkably similar trends for several functional markers, highlighting the critical importance of the surrounding CN in shaping the functional states of CD8⁺ T cells. For example, the expression trends for TIM3, HLA-DR, PTPN22, GZMB, and to a lesser degree LAG3 and TIGIT, were shifting together across the CNs regardless of the tissue site. Furthermore, CD8⁺ T cells in the liver, which were in close proximity to tumor cells with minimal stromal barriers, exhibited exceptionally high ICOS expression. Immune checkpoint marker expression also varied by site, with TIGIT levels elevated in the pancreas and LAG3 expression more prominent in the liver, indicating distinct immune exhaustion landscapes.

CNs around tumor cells and their association with tumor cell expression of functional state markers

We applied the same approach to cluster tumors based on their neighborhoods. Prior to CN classification, we first categorized tumors as either bulk or boundary tumor types, using a tumor cell influence cutoff of 0.7 (tumors with greater than and equal to 0.7 influence were classified as bulk, and those with less than 0.7, where 30% of the influence was attributed to other cell types, were classified as boundary tumors). We examined the boundary tumor cells because cellular influences from other cell types originate at the tumor boundary and propagate inward, allowing us to compare these influences and relate them meaningfully to the overall TME. This approach successfully distinguished boundary tumors, which exhibited greater interaction with other cell types, from bulk tumors, whose cellular influence primarily came from other tumor cells. (Fig. 5A–B). Pancreas and liver sites had similar proportions of bulk and boundary tumor types (Fig. 5C).

Figure 5. Functional characterization of tumor cells across sites.

A, Representative IMC images from the pancreas (core 18) and liver (core 25) showing the tumor marker CK. Scale bar: 200 μm. B, Tumor cells were classified into boundary (BD) and bulk (BK) tumors based on a tumor influence cutoff of 0.7, with values below the cutoff classified as boundary tumors and values ≥0.7 classified as bulk tumors. C, Bar plot illustrating the abundance of BD and BK tumor types across different sites. D-E, Violin plots comparing the pancreas and liver for BD and BK tumor types. The overlaid boxplots indicate the median (center line) and interquartile range (IQR) (25^th-75^th percentiles), with whiskers extending to 1.5*IQR. D shows significant differences in mean cellular influences, and E shows differences in mean expression of tumor markers CK and CD86. Statistical significance was assessed with an unpaired Wilcoxon test. F, Cell mask images from matched primary and metastatic tumor sites (PA6, PA11) (top) and tumor cells colored according to CK expression levels (bottom). G, Spearman correlation between mean cellular influence (mean per core) and mean tumor marker expression, with rho values shown as color gradients in the heatmap. Significant p-values are indicated (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001).

Based on their FunCN scores, we identified 12 unique CNs for boundary tumors (Supplementary Fig. S4A). The majority of tumor cells were surrounded by other tumor cells, often with stromal cells. This predominant tumor-associated neighborhood, CN2, exhibited the highest overall proportion among all CNs (Supplementary Fig. S4B). A similar neighborhood but with additional UA cells (mostly poorly differentiated tumor cells) was identified as CN1. Across CNs, stromal cells represented the most abundant non-tumor neighbors. Str_I (COL1A1⁺ SMA/VIM^+/lo, myCAF-like) in particular, exhibited varying magnitudes of influence in several CNs, with the strongest association observed in CN3. Meanwhile, other stromal subtypes were enriched in distinct CNs: CN7 (Str_II; COL1A1^hiCD74⁺HLA-DR⁺SMA/VIM⁺, apCAF-like), CN4 (Str_III; COL1A1^hiCD74^hi), and CN8 (Str IV; SMA/VIM^hi). As for immune cells, M_I subtype myeloid cells (CD163^loCD206^lo) exerted influence in CN9, CN10, and CN5, augmented by contributions from M_II (HLA-DR^hiCD68^hiCD163^hiCD206^hiCD86⁺) in CN10 and M_III (DC-SIGN⁺CD86⁺) in CN5. Other immune-enriched neighborhoods included CN6 (NK cell-enriched), CN11 (immune-mixed), and CN12 (neutrophil-enriched).

To examine site-specific differences in their local microenvironment, we compared the distribution of boundary tumor cells across CNs between the pancreas and liver. The proportion of boundary tumor cells belonging to CN3 (Str_I-predominant neighborhood) was significantly higher in the pancreas, whereas the proportion in CN7 (Str_II-predominant neighborhood) was significantly higher in the liver (Supplementary Fig. S4C). In addition, M_I-abundant neighborhood, CN9 was significantly more represented in the pancreas than the liver. Although not statistically significant, neutrophil influence was particularly high in CN12 within several liver cores (Supplementary Fig. S4C). Consistent with these CN-based observations, tumor cells overall had stronger cellular influences from Tregs and Str_II (apCAF-like) in the liver, but more influences from M_I and Str_I (myCAF-like) cells in the pancreas (Fig. 5D).

We next evaluated the expression of functional markers on the tumor cells across the CNs: CK for differentiation, PD-L1 and VISTA for immunosuppression, CD86 for costimulation, and Ki-67 for proliferation (Supplementary Fig. S4D–E). Tumor cells predominantly surrounded by Str_I cells (CN3), and M_I cells (CN9) exhibited lower levels of these markers than those in other CNs. In the pancreas, CK expression showed a marked decline from bulk tumors to boundary tumors belonging to CN2 (tumor-predominant with surrounding myCAF-like stromal cells) and CN1 (UA; mostly poorly differentiated tumor cells). A similarly pronounced lower level in CK expression was also observed in myeloid-enriched neighborhoods (CN9 and CN10) in the pancreas. CN12 (neutrophil-enriched) and CN5 (M_III-enriched; DC-SIGN⁺CD86⁺) were associated with high tumor expression of immunosuppressive molecules PD-L1 in the pancreas and VISTA in both Pancreas and the liver (above median with better confidence; less shaded area), implicating the role of myeloid in promoting tumor-mediated immunosuppression.

The overall difference of CK level between bulk and boundary tumors could largely be attributed to CN2 and CN3 (Str_I; myCAF-like), which contained the highest proportion of boundary tumor cells among all CNs. Bulk tumors in the pancreas exhibited significantly higher CK levels compared to boundary tumors in both the pancreas and liver as well as to bulk tumors in the liver (Fig. 5E). This reduction suggests loss of epithelial phenotype and increased epithelial-to-mesenchymal transition (EMT) features, particularly in tumor regions adjacent to stromal area. In contrast, CK expression did not differ significantly between bulk and boundary regions within the liver. These observations align with prior work (26) showing that EMT markers enriched at tumor-stromal interfaces in primary PDAC and metastatic lesions being composed largely of mesenchymal-like clones that have already undergone EMT, lacking the same interface-driven transition patterns seen in primary tumors. Furthermore, CD86, which plays active role in regulating the activity of T cells (27), showed similar expression pattern as CK.

Differences in CK levels across tumor cells from different sites and spatial organizations were also evident in the spatial visualization of tumors (Fig. 5F). In the pancreas, tumors cells exhibited a more clustered formation, creating dense aggregations with overall higher CK expression (appearing red). Metastatic tumor cells in the liver, on the other hand, demonstrated a more diffuse and less compartmentalized shape with lower CK levels (more yellow overall). As expected from our analysis, in the pancreas, tumor cells with elevated CK were more abundant, particularly in clustered regions, whereas CK levels were relatively lower at boundary zones at the surface with stromal and myeloid cells. As for the immune cell in the TME, the liver microenvironment appeared to recruit more Tregs and neutrophils within the stromal areas, suggesting a more immunosuppressive TME that may contribute to immune evasion and reduced responsiveness to immune checkpoint therapies.

Evaluating site-specific differences in the immune–tumor interactions and functional states

Building on these findings, we sought to further dissect site-specific immune regulation. To do this, we used Spearman correlation analysis to evaluate the pairwise relationships between cellular influences and functional states of tumor cells, using core-level averages of cellular influences and functional marker expression (Fig. 5G). The spatial influences of two cell types, CD8⁺ T cells and M_I (CD163ˡᵒCD206ˡᵒ) cells, showed opposing trends in relation to tumor functional marker levels in the pancreas and liver. In the pancreas, higher CD8⁺ T-cell influence inversely correlated with tumor expression of immunoregulatory markers, resulting in lower CD86, PD-L1, and VISTA. In contrast, in the liver, greater CD8⁺ T-cell influence correlated with increased expression of these markers. M_I exhibited similar relationships, though to a lesser extent. Other site-specific myeloid-cell influences on tumor cells were also observed. Only in the liver, neutrophil influence correlated positively with elevated VISTA levels, while M_III cells (DC-SIGN⁺CD86⁺) were significantly associated with increased expression of both VISTA and PD-L1. These varying immune-tumor relationships between the two sites suggested the presence of distinct immune regulatory mechanisms in each organ, influencing tumor immune evasion strategies differently. Namely, neutrophil/myeloid-driven enhancement of immunosuppressive environment was particularly notable in the liver.

To further investigate immunosuppression by VISTA, specifically as it related to neutrophil influence, we first classified the tumor regions into four inflammatory types based on the presence or absence of CD8⁺ T-cell influence and PD-L1 expression, following the framework by Martin et al. (28): Type I (“CD8⁺ T⁺ PD-L1⁺”), Type II (“CD8⁺ T⁻ PD-L1^∓”), Type III (“CD8⁺ T⁻ PD-L1⁺”), and Type IV (“CD8⁺ T⁺ PD-L1⁻”). CD8⁺ T-cell status was defined using FunCN scores, with CD8⁺ T⁺ and CD8⁺ T⁻ tumors based on whether their per-region CD8⁺ T-cell influence exceeded or fell below the median value. PD-L1 status was similarly assigned based on median PD-L1 expression across tumor regions. These classifications provide a framework for TME archetypes, with Type I tumors harboring PD1/L1-inhibited T cell states, Type II representing immune ignorance, Type III exhibiting tumor-derived immune suppression, and Type IV suggesting the presence of alternative immunosuppressive pathways. We verified that tumor cells classified as “PD-L1⁺” exhibited higher PD-L1 expression (Supplementary Fig. S5A). A higher proportion of CD8⁺ T⁺ PD-L1⁻ tumor cores was observed in the pancreas, whereas CD8⁺ T⁺ PD-L1⁺ tumor cores were more prevalent in the liver (Supplementary Fig. S5B). We further observed that VISTA expression varied across these tumor types in both pancreatic and liver sites (Supplementary Fig. S5C). Based on this classification, we observed that, specifically in the liver, neutrophil influence significantly correlated with VISTA expression in Type III TME (Spearman R=0.74, p=0.031). In the pancreas, although a significant correlation was observed in Type IV tumors (Spearman R=0.79, p=0.0098), the overall neutrophil influence was minimal (Supplementary Fig. S5D). Using MiniCAD Design File (MCD) images, we further confirmed that tumor regions classified as “CD8⁺ T⁺” exhibited a higher concentration of CD8⁺ T cells in close proximity to the tumor cells (Supplementary Fig. S5E). This analysis supports the role of neutrophils in shaping an immunosuppressive environment in the liver tumors, particularly in Type III “CD8⁺ T⁻ PD-L1⁺” tumors, where immune evasion is mediated by tumor-intrinsic expression of immunosuppressive molecules. VISTA is a known negative regulator of T-cell activity and has been implicated in immune evasion mechanism (29,30). Clinically, high VISTA expression by tumor cells has been associated with worse overall survival in PDAC (30). Together, although further clinical studies are needed to validate these observations, our finding highlights the potential impact of VISTA-mediated immunosuppression and implicates the potential need for VISTA-targeted therapies, especially in liver metastatic and neutrophil-enriched diseases, to improve patient outcomes.

Lastly, while more limited in sample size, we sought to account for patient-specific variability and evaluated the site-specific differences only using 6 primary–metastatic matched patient samples in a paired fashion. There was notable inter-patient variability, not only in immune-cell influences on tumor cells but also in the expression of functional state markers by tumor cells (Supplementary Fig. S5F–G). We focused on patients exhibiting marked differences between primary tumors and metastases. For example, in patient PA11, the metastatic site showed a reduction in effector T cells (CD8⁺ T cells and CD4⁺ T cells) and NK cells, alongside an increased influence of Tregs. This shift was accompanied by a reduction in CK levels and an increase in CD86 levels in the metastatic site. In PA6, a substantial reduction in tumor functional markers (CK, PD-L1, CD86, and VISTA) was observed in metastasis. The influences from other cell types (T cell, myeloid, and stromal) on tumor cells was low, indicating fewer interactions between the tumor cells and their microenvironment (stromal and immune cells). In PA10, as discussed earlier, a marked increase in VISTA-associated cell types, such as M_III (DC-SIGN⁺CD86⁺) and neutrophils, was observed alongside elevated VISTA and PD-L1 levels in metastasis. Together, these findings highlight patient-specific immune adaptations, suggesting the need for personalized immunotherapeutic strategies.

Discussion

In this paper we present FunCN, a spatial kernel-based model for quantifying cellular influence and defining CNs. FunCN incorporates both the frequency and the proximity of distinct cell types in the tissue to the target cell. Compared to prior methods, FunCN established an enhanced characterization of the cellular organization and enabled a comprehensive mapping of the immune and stromal microenvironment of primary pancreatic tumors and liver metastases. FunCN is especially effective in quantifying the strong localized influence of low-abundance cell types such as lymphoid and myeloid cells in the TME. It is important to emphasize that although FunCN quantifies the influence to represent the spatial organization of the TME, the functional state of cells remains inherently complex. For instance, determining whether a higher measured cellular influence translates to a more biologically meaningful impact on tumor behavior is still an open question. Future work could focus on integrating functional marker levels with proximity and abundance metrics to more comprehensively elucidate patient outcomes. Establishing the threshold at which cellular influences begin to significantly impact cell behavior is also a critical area for further investigation. Furthermore, different cell types can affect tumor function in distinct ways, and their combined impact is challenging to fully disentangle. Our approach enables systematic quantification of cellular influences, providing a valuable tool for addressing these open questions. Its broad applicability across diverse datasets would support advanced analyses that integrate functional markers and spatial organization of the TME to better understand tumor behavior and improve outcome prediction.

Through FunCN-based spatial analysis, we identified key immune and stromal interactions that may shape tumor immunogenicity. In pancreatic primary tumors, reduced CK expression was observed in tumor areas adjacent to stromal cells, indicating a tumor–stromal interaction that modulates epithelial differentiation. In liver metastases, neutrophils and DC-SIGN⁺CD86⁺ myeloid cells (M_III subtype) were linked to increased VISTA expression, suggestive of a myeloid-driven immunosuppressive microenvironment. The strong influence of neutrophils in mediating tumor immunogenicity through VISTA, particularly in CD8⁺ T- PD-L1+ Type III tumors, suggests that immune checkpoint therapies relying on T-cell activation may be less effective in these tumors. These findings suggest the need for therapeutic strategies targeting VISTA and neutrophil-mediated immune suppression to enhance immune infiltration and improve treatment responses in liver metastases.

Furthermore, LAG3 expression was consistently elevated in CD8⁺ T cells at liver metastatic sites across most CNs. This finding aligns with a previous study (31) reporting an increased proportion of LAG3⁺CD8⁺ T cells in the liver. In contrast, at the pancreatic primary site, TIGIT expression was elevated in CD8⁺ T cells across the CNs (32) These results suggest the presence of tissue-specific drivers of distinct immunoregulatory axes. Similarly, our analysis uncovered opposing trends in CD8⁺ T-cell influence on tumor functional states between pancreatic and liver tumors. In pancreatic tumors, CD8⁺ T-cell influence was negatively correlated with functional markers (CD86, PD-L1, and VISTA), suggesting immune-mediated tumor suppression. However, in liver metastases, CD8⁺ T-cell influence was positively correlated with these markers. Together, these findings indicate site-specific immune regulatory mechanisms, e.g., driven by molecular pathways and/or parenchymal signals, beyond what could be identified by the current antibody panel and IMC alone. These data provide insights into why immune modulatory therapies that target different T-cell suppressive pathways have different success rates depending on the organ sites hosting the tumors.

By integrating high-dimensional IMC spatial profiling with our FunCN model, our study introduces FunCN, a high-resolution, distance-based spatial modeling method that quantifies cellular influence at the scale of local proximities. We then integrate these interactions to characterize the broader tumor microenvironment in pancreatic metastases. Unlike fixed-boundary neighborhood methods, FunCN weights neighboring cells by their spatial proximity, revealing patterns such as neutrophil-associated VISTA upregulation, reduced CK expression at tumor–stromal interface, and site-specific difference in CD8⁺ T cell–associated functional states. Together these findings highlight the complexity of the metastatic immune landscape. These findings reinforce the need for site-specific, multi-target immunotherapeutic strategies, particularly those addressing VISTA, TIGIT, LAG3, and neutrophil-driven immunosuppression, to enhance immune activation and improve treatment outcomes in PDAC metastases.

Synopsis:

The authors report a computational framework that enables high-resolution spatial analysis of immune interactions in the tumor microenvironment and apply it to PDAC samples, revealing distinct immune landscapes between primary and metastatic tumors that may inform immunotherapy strategies.

Data and code availability

Imaging mass cytometry data have been deposited at https://doi.org/10.5281/zenodo.15596960 and are publicly available as of the date of publication. All original code has been deposited at GitHub (https://github.com/DeshpandeLab/Spatial_Influence) and is publicly available. Any additional information required to reanalyze the data reported in this paper is available from the corresponding authors upon request.