Spatial Proximity of Tumor-Immune Interactions Predicts Patient Outcome in Hepatocellular Carcinoma

Abstract

Background and Aims:

The fitness and viability of a tumor ecosystem are influenced by the spatial organization of its cells. We aimed to study the structure, architecture, and cell-cell dynamics of the heterogenous liver cancer tumor microenvironment using spatially resolved multiplexed imaging.

Approach and Results:

We performed co-detection by indexing (CODEX) multiplexed immunofluorescence imaging on 68 hepatocellular carcinoma (HCC) biopsies from Thai patients (TIGER-LC) as a discovery cohort, and then validated the results in an additional 190 HCC biopsies from Chinese patients (LCI). We segmented and annotated 117,270 and 465,632 cells from the TIGER-LC and LCI cohorts, respectively. We observed four groups of TIGER-LC patients (IC1, IC2, IC3, IC4) with distinct tumor-immune cellular interaction patterns. In addition, patients from IC2 and IC4 had much better overall survival than those from IC1 and IC3. Noticeably, tumor and CD8⁺ T cell interactions were strongly enriched in IC2, the group with the best patient outcomes. Close proximity between tumor and CD8⁺ T cells was a strong predictor of patient outcome in both the TIGER-LC and the LCI cohorts. Bulk transcriptomic data from 51 of the 68 HCC cases was utilized to determine tumor-specific gene expression features of our classified subtypes. Moreover, we observed that the presence of immune spatial neighborhoods in HCC as a measure of overall immune infiltration is linked to better patient prognosis.

Conclusions:

Highly multiplexed imaging analysis of liver cancer reveals tumor-immune cellular heterogeneity within spatial contexts, such as tumor and CD8⁺ T cell interactions, which may predict patient survival.

Graphical Abstract

Introduction

The number of diagnoses of primary liver cancer continues to rise rapidly, with a low 5-year survival rate at 18% (1). Without clear molecular drivers, the heterogeneity in hepatocellular carcinoma (HCC) poses difficulty in identifying key early features of disease progression and defining new therapeutic strategies (2). Advancement in spatial pathobiology imaging techniques like co-detection by imaging (CODEX) (3) has enabled the simultaneous measure of >50 markers across millions of cells with robust ability to understand the tumor ecosystem’s spatial organization. Identifying co-localization and neighborhood patterns of cellular interactions from spatial data has started to uncover communication networks and niches within tumors (4). Computational analysis of multiplexed imaging data has revealed the spatial information of cells can inform tumor-immune dynamics in human colorectal (5), breast (6, 7), bladder (8), lung (9), skin (10), and pan-cancer analysis (11). Spatially resolved technologies have begun to be leveraged to study HCC, encouraging that our ability to probe the spatial architecture may yield novel insights to biomarkers, antitumor immunity, and the development of targeted treatments.

The HCC tumor microenvironment (TME) has been characterized by spatial transcriptomics (12, 13), single-cell in situ hybridization (RNAscope) to study the spatial heterogeneity of apoptotic cell death in HCC (14), and deep learning to identify histological subtypes from H&E staining based on the spatial organization of tumor cells and lymphocytes (15). Detailed whole-slide CODEX imaging of 15 HCC patients revealed regulatory T cell accumulation in tumors, thought to be an unfavorable indicator for patient outcome (16, 17). In addition, immune compartment communication networks in HCC patients were defined by imaging mass cytometry, revealing key features of tumor associated macrophages (TAMs) (18, 19). For example, study of 14 HCC patients showed close spatial proximity of Arg1^hi macrophages to CD8⁺ T cells was associated with poorer effector T cell functions, a feature more prevalent in non-responding tumors to immunotherapy (18). This highlights the crucial need to clearly characterize spatial immune cell interactions with larger HCC cohorts. Still, little is known about which spatial immune interaction dynamics may predict better patient outcomes or drive tumor aggressiveness in liver cancer.

Herein, we sought to further understand the immune and microenvironmental spatial landscape of HCC tumors. Our study examines the cell level protein staining of HCC formalin-fixed, paraffin-embedded (FFPE) tissue tumor samples with 16 markers. The antibody panel was selected to distinguish subsets of immune cell lineages. We show that CODEX produces highly specific staining of tumor cells with mixtures of stromal and infiltrating immune interactions. Our analysis demonstrates the utility of examining the spatial organization and structures of specific cellular immune neighborhoods in liver cancer as their spatial proximity to tumor cells influences patient outcome. The major findings of our study are 1) highly multiplexed imaging of tumor-immune interactions can be useful to predict HCC patient survival 2) spatial proximity of CD8⁺ T cells and tumor cell interactions are indicators of good prognosis and 3) immune infiltrating spatial neighborhoods in liver cancer are linked to better patient outcomes.

Patients and Methods

Cohorts and clinical specimens

Following CODEX staining, a subset of Thailand Initiative in Genomics and Expression Research for Liver Cancer (TIGER-LC) and Liver Cancer Institute (LCI) tumor cores were excluded from further analysis after selecting highest quality samples from the tissue microarray (TMA). For analysis, TIGER-LC had n=58 tumor cores and n=58 matching nontumor cores. LCI had n=132 tumor cores. All patients in both cohorts gave informed consent with approval by each Institutional Review Board. Clinical characteristics for selected patients in TIGER-LC and LCI TMAs can be found in Table S1. (Note: only n=51 TIGER-LC patients had complete survival information. Clinical information was not present for every sample).

Computational Imaging Quantification

Nuclear segmentation and pixel fluorescence intensities for the 16-plex panel combined with DAPI as the nuclear stain were quantified with HALO Image Analysis Platform from Indica Labs using the High Plex Fluorescence module version 4.0.4. The expression positivity threshold for each marker was manually tuned and the HALO TMA module was run to apply the settings to all cores.

Clustering and UMAP Generation

We identified unique cellular clusters from the CODEX staining of 16 markers utilizing the expression profiles of all single-cells in the TIGER-LC cohort (n = 117,270). Seurat clustering in R was performed on single-cells with greater than two expressed markers (setting: min.features=2). Data was scaled and dimensional reduction with PCA was performed. The first 10 PCs were selected for UMAP analysis based on the elbow point method. Clustering settings: resolution=0.2, Louvain clustering algorithm. The remaining default parameters were used for the functions FindNeighbors and FindClusters. For both cohorts, the broader categories of tumor (E-cadherin, HNF4alpha, beta-catenin) and immune (CD45) were classified. Next, a second round of clustering only on the immune cells was used to refine the cell-type labels to include T cells, B cells, and macrophages. In total, we identified 14 unique cell types. Cells expressing only DAPI and a subset of CD44⁺ cells which fell close to our tumor cell population in the UMAP space were labeled Unclassified.

The LCI Cohort (n=465,632 cells) was annotated in a similar manner. We scaled the data, reduced the dimensionality with PCA, and annotated the clusters in the UMAP space. We identified 15 unique cell types. Cell type assignments and visualizations of our LCI validation cohort can be found in Fig. S6.

Cell-Cell Interactions and Voronoi Diagrams

Voronoi diagrams with the x and y coordinates of each cell’s centroid were used to build spatial networks of each patient’s tumor. We used the Giotto R package (20) to implement the Voronoi tessellation (Delaunay triangulation) method to connect single-cells. We generated the cell-cell interactions using the function createSpatialNetwork (settings: minimum_k = 0, k=10, maximum_distance_delaunay = 100). For our analysis, neighboring cells with a distance less than 40μm (100 pixels) separating their centroids are counted as a cell-cell interaction. One micron on a tumor core was equivalent to 2.5 pixel units.

Immune Cell-Cell Interaction Clustering

A matrix of immune cell-cell interaction frequencies for each patient was calculated as the total number of cell-cell interactions of type X divided by the total cell-cell interactions for that sample. This strategy was taken because for each individual core some samples have whole tissue while others have a smaller tissue region. Clustering was performed with the ComplexHeatmap package in R. The Pearson correlation coefficient (complete linkage method) was used to hierarchically cluster the patients (matrix columns) based on immune cell-cell interaction composition frequency. The clustering method for each cell type (matrix rows) was Ward.D2. Cell-cell interactions (matrix rows) were included for analysis if >10% of samples overall had at least 10 interaction counts for that interaction. Thus, we focused on immune cell phenotypes consistently found across all tumors.

Gene Set Enrichment Analysis using Transcriptomic Data

Gene set enrichment analysis (GSEA) was conducted using GSEA v4.2.3. There were 25,162 initial genes from the TIGER-LC cohort (GSE76297) from n=51 HCC patients with matching transcriptomic and CODEX data (21). We performed differential gene expression analysis of tumor compared to nontumor tissue using a paired t-test. The p-values were adjusted with FDR correction. The remaining 16,672 genes (FDR<0.05) were used as inputs for GSEA. The HALLMARK gene matrix from c2.v7.5.1 of the Molecular Signatures Database (MSigDB) was utilized. GSEA was run with default settings: 1000 permutations, gene set size filters (min=15, max=500). Pathways with nominal p-value <0.05 were plotted.

Immune Neighborhood Analysis

We utilized the spatial latent Dirichlet allocation (Spatial-LDA) topic modeling to identify cellular microenvironments in our CODEX data (22). The tumor-immune boundary compartmentalization has previously been studied with this method which uses cell type coordinates and the marker intensities to model smooth transitions between microenvironments (7, 22). We used default settings for training and fitting a Spatial-LDA model (training size fraction=0.9, difference penalty=250, neighborhood radius=100, number of iterations=3), identifying five TME topics. Index cells for input had at least five nearest neighbors by Voronoi network analysis. To examine the neighborhood analysis feature extraction, we first trained the model on the highest-quality samples from the TIGER-LC cohort by setting a threshold minimum of 500 index cells. This led to the inclusion of n=31 cores. For the LCI cohort, we used all n=192 cores. We excluded the marker CD31 from the model because we focused our analysis on interactions between tumor-and immune cells. Each cell is given a weight, indicating how much a TME resembles the given topic. We calculated the mean of all cellular weights for each patient for Topic 4 as the metric for survival analysis.

Statistical Analysis

We used R version 4.1.0. GraphPad Prism version 8 was used for all statistical tests. Student’s t test was used for two group comparisons. For multiple comparisons (i.e. gene expression across subgroups) we performed one-way ANOVA. We corrected for multiple testing using FDR correction. The log-rank test p-value in GraphPad Prism was used for survival analysis when comparing two groups (i.e. low vs. high T cell interactions). The log-rank test for trend p-value in GraphPad Prism was used for tertile comparison when splitting groups into low, medium, and high. Criteria for statistical significance was p < 0.05.

Complete details of the materials and methods can be found in the Supporting Information.

Results

Multiplexed CODEX imaging reveals the spatial tumor-immune landscape in HCC

We determined the spatial landscape of HCC by designing a 16-plex target antibody panel for CODEX staining of the tumor ecosystem to delineate the spatial organization of the tumor, immune, and stromal cells (Fig. 1A). We used the TIGER-LC HCC tumor samples (n=58) and nontumor samples (n=58) as a discovery cohort and the LCI tumor samples (n=132) as a validation cohort. Staining was performed on TMAs of 1.0-millimeter (mm) cores from archival HCC FFPE tissue (Fig. 1B, C). The marker panel included markers specific to lymphocytes (T and B cells), myeloid cells (macrophages and dendritic cells), endothelial cells, epithelial cells, and the nucleus (Fig. S1A). Histology was reviewed for the alignment of consecutive tissue slices with immune-dense or tumor regions, indicating the accuracy and robustness of the CODEX staining (Fig. 1D).

Figure 1. CODEX workflow and staining.

(A) Workflow of liver tumor TMA generation and quality control, CODEX staining, image feature extraction, and computational analysis.

(B) Histopathology of the representative HCC tumor shown in panel C with CODEX staining.

(C) Selected TIGER-LC HCC tissue core imaged with a 16-plex CODEX marker panel. The following markers are shown (DAPI, blue; CD31, red; CD8, green; CD4, cyan; CD45, pink; CD163, white; CD68, orange, E-cadherin, yellow).

(D) Higher magnification of the CODEX staining for two regions from the selected tissue core for additional markers (staining not shown here Ki67, CD44 see Fig. S1D).

We performed single-cell segmentation and feature extraction per marker using the HALO^® Image Analysis Platform. A total number of 117,270 cells were obtained from the TIGER-LC HCC cohort, with 36/58 cores having > 1,000 extracted cells. We further performed UMAP analysis of the cells and annotated them based on known markers unique to T cells (CD4, CD8, CD3E, CD45, CD45RA, CD45RO), B cells (CD20), tumor-associated macrophages (TAMs) (CD163, CD68), dendritic cells (DCs) (CD11c), and tumor-associated endothelial cells (TECs) (CD31) (Fig. 2A, B). Malignant cells classified as tumor were determined with high level of E-cadherin, HNF4alpha, and beta-catenin. We observed cells with multiple marker positivity, e.g., TAM-CD163⁺CD68⁺ (Fig. S1B, C). However, we also noticed a small group of CD20+ cells with background signal from HNF4a staining, so we termed the overall population as B cell-CD20⁺HNF4a⁺ cells. With the resolved cell types, we determined the cellular composition of the tumor. We found an overall proportion of 44.4% of tumor and 35.74% of immune cells in the TIGER-LC HCC cohort, with a specific immune composition of CD8⁺ T cells (5.57%, n=6,534), CD4⁺ T cells (3.75%, n=4,397), B cell-CD20⁺HNF4a⁺ (3.3%, n=3,872), TAM-CD163⁺CD68⁺ (3.86%, n=4,528), and TAM-CD68⁺ (3.27%, n=3,838) (Fig. S2A). We further mapped the annotated cells back to their spatial coordinate locations in the tissue. We observed well-aligned spatial structures including the tumor-immune border and the vasculature, giving confidence in the phenotype determination (Fig. 2C, Fig. S1D, E, F, G and Fig. S2B).

Figure 2. Cell type classification.

(A) UMAP representation of the CODEX protein expression space for 117,270 annotated cells from tumor tissue of TIGER-LC HCC patients.

(B) Dotplot showing the 16 CODEX markers and their corresponding cell-type assignments in HCC tumor tissue.

(C) Spatial mapping of the assigned cell-types back onto the tumor coordinates for a representative tissue core.

(D) UMAP of the CODEX expression of 16,336 annotated cells from nontumor tissue of hepatocellular carcinoma patients.

(E) Dotplot showing the 16 CODEX markers and their corresponding cell-type assignment in HCC nontumor tissue.

We performed similar analysis for the TIGER-LC HCC non-tumor tissues (Fig. 2D, E). Noticeably, the nontumor nuclei area (e.g., epithelial) was smaller compared to the tumors (Fig. S2C, H). The number of Ki67⁺ and TEC-CD31⁺ cells was increased in the tumor compared with the nontumor tissues, indicating high proliferation rate and vasculature in the tumor (Fig. S2D, E). We did not include this group of Ki67⁺ cells (proliferation status) for further analysis of tumor immune interactions. In addition, we found immune cells had closer proximity with tumor cells in the tumor tissues than with epithelial cells in the non-tumor tissues, suggesting elevated immune infiltration in the tumor (Fig. S2F, G).

Tumor-immune spatial interactions classify HCC subtypes

We determined the spatial interaction dynamics of tumor and immune cells via Voronoi networks based on the Delaunay triangulation of individual cells. Specifically, network-connected adjacent cells within a 40μm (100 pixel) radius were used to define a single cell-cell interaction (Fig. 3A and Fig. S3A, B). The mean cell to cell interaction distance was 15.8 μm (39.5 pixels), s.d.= 6.7 μm (16.7 pixels) (Fig. 3B). As an example, tumor and infiltrating CD8⁺ T cell interactions connected by the Voronoi network are indicated in Fig. 3C. We further performed hierarchical clustering analysis of the tumor-immune interactions across all the HCC samples in the TIGER-LC cohort. We found four immune class (IC) subtypes based on the interaction activities (Fig. 3D). Noticeably, patients in IC2 and IC4 had much better overall survival compared with those in IC1 and IC3 (Fig. 3E). We found that IC2, the best survival group, was enriched in CD8⁺ T cell to tumor cell interactions and CD4⁺ T cell to tumor cell interactions. When evaluating the interactions among immune cells, we observed strong CD4⁺ and CD8⁺ T cell interactions in this group of patients. The IC4 cluster was enriched in tumor cells interacting with both B cell-CD20⁺HNF4a⁺ and TAM-CD163⁺CD68⁺ cells, although these two cell types did not seem overlapping. In IC1, elevated interactions between tumor cells and TAM-CD68⁺ cells was observed. The IC3 cluster was enriched in tumor to DC interactions, but largely lacked infiltrating T cells. Gene set enrichment analysis (GSEA) on the bulk transcriptomic data demonstrated that IC1 and IC3 were enriched in cell cycle related pathways (Fig. 3F and Fig. S4A). In addition, we performed PCA analysis of the tumors using the bulk transcriptomic data based on the most variable genes. We found IC1 and IC3 are similar in the PC space while distinct from IC2 and IC4 (by PERMANOVA test IC1 vs IC3 p=0.244) (Fig. 3G). Collectively, these results suggest that the spatial interaction profiles of tumor and immune cells can reflect the intrinsic tumor biology.

Figure 3. Patient clustering by spatially defined subtypes of tumor-immune interactions.

(A) Diagram of cell-cell interaction calculations using Voronoi tessellation

(B) Density distribution of all cell-cell interaction distances calculated per tumor sample. The red dotted line represents the mean distance between interacting cells.

(C) Representative image of a tumor core with tumor-immune interactions for CD8⁺ T-cells highlighted.

(D) Hierarchical clustering of HCC patients (n=58) stratified into four immune cluster (IC) groups. Clustering is based on the percentage of six tumor-immune interactions on the top half of the panel. The bottom annotation (HC1, HC2, HC3) represents our former transcriptome signature (21).

(E) Overall survival of patients (n=51) with HCC from different immune subtypes. The p-values on the figure comparing the survival of two groups are reported from the log-rank (Mantel-Cox) test output from GraphPad Prism. The permutation p-value for the survival analysis was 0.027 (27/1000). Figure S3C shows the n=7 patients which did not have long-term survival data or matching transcriptome data and are not included in the survival plot.

(F) Gene Set Enrichment Analysis (GSEA) of the four IC groups, colored by normalized enrichment score (NES). The comparison was IC1/IC3 vs IC2/IC4. Red indicates positively enriched. Blue indicates negatively enriched.

(G) Principal component analysis (PCA) of the top 1164 variable genes across the four IC groups by one-way ANOVA (adj. p <0.05). Centroids are plotted for each of the four groups as the mean of PC1 and PC2 for all points in the group.

We also compared the IC group classification (i.e., IC1, IC2, IC3, IC4) in the TIGER-LC HCC cohorts with previous identified molecular subtypes of HCC (23–25) using a nearest template prediction algorithm (26). We found HC1 determined in our previous study (21) of this cohort are enriched in IC3, while HC2 and HC3 are enriched in IC2 and IC4 (Fig. 3D). A Chi-Square test of independence between the IC groups and other HCC signatures revealed relevance to former classifications: Hoshida subclasses (p=0.05), Sia’s Immune Class (p=0.06), and Chiang’s HCC subclasses (p=0.12) (Fig. S5C). Comparisons between the classifications using Fisher’s Exact Test showed that: IC4 was enriched in Hoshida’s S3 (odds ratio (OR) 8.2, p=0.012); IC1 was enriched with Hoshida’s S1 (OR 4.41, p=0.05); IC3 was enriched with Chiang’s Proliferation class (OR 3.96, p=0.05); and IC4 was enriched with Chiang’s CTNNB1 class (OR 9.3, p=0.003). In addition, the CD8⁺ T cell enriched group IC2 are overrepresented in Sia’s Immune Class (9/21 cores, OR 2.64, p=0.09). Taken together, these analyses suggest that the spatial dynamics of tumor-immune interactions can stratify HCC patients into subgroups with distinct clinical outcomes.

Spatial interactions of CD8⁺ T cells and tumor cells are associated with patient outcomes

We noticed that IC2 was largely driven by T cell presence, with CD8⁺ T cell to tumor interactions showing increased frequency compared with other groups (Fig. 4A). In addition, IC2 had elevated CD4⁺ to CD8⁺ T cell interactions (Fig. S4B). Analysis of bulk transcriptomic data indicated that IC2 had a higher cytotoxicity score based on gene expression of 15 cytotoxicity-related markers (Fig. S4C). To further determine the role of CD8⁺ T cells in HCC, we split the patients into three groups based on the interaction frequency of CD8⁺ T cells and tumor cells. We found that in the TIGER-LC cohort, CD8⁺ T cell and tumor interactions are positively linked to patient outcome, where patients with high level of interactions had much better overall survival (Fig. 4B). These findings were further validated in 132 HCC patients from the LCI cohort, which had 465,632 total cells (Fig. 4C and Fig. S6). We didn’t observe the relationship in the TIGER-LC HCC adjacent nontumor tissue samples, suggesting that the link is tumor specific (Fig. S4D, E, F, G). A univariate and multivariate cox proportional hazards regression analysis revealed the CD8⁺ T cells-tumor spatial interaction from our survival analysis (Fig. S5A) are independent of age and gender in both cohorts (Table S6). Finally, Fig. S5D shows the improvement when using spatial information in survival analysis compared to composition alone.

Figure 4. CD8⁺ T cell interactions with tumor cells drives better patient outcomes.

(A) Frequency boxplot of CD8⁺ T cell to tumor cell-cell interactions of total tumor-immune interactions: IC1 (n=10), IC2 (n=21), IC3 (n=10), IC4 (n=17). The line inside the boxplot represents the median value, with the box spanning the first to third quartiles. Statistical differences between two groups on boxplots in subpanels are shown by Student’s t test in GraphPad Prism.

(B-C) Survival analysis for low (n=17), medium (n=18), and high (n=16) CD8⁺ T cell to tumor interaction groups for the TIGER-LC cohort (B) and LCI cohort with groups low (n=45), medium (n=44), and high (n=43) (C). The p-value on the survival figure is based on the log-rank test for trend in GraphPad Prism.

(D) The percentage of TIGER-LC CD8⁺ T cells which have TAM-CD68⁺ as nearest neighbors comparing groups IC1 (n=10) vs IC2 (n=21).

(E) The percentage of LCI CD8⁺ T cells which have TAM-CD68⁺ as nearest neighbors comparing groups low CD8⁺ T cell-tumor interactions (n=66) versus high CD8⁺ T cell-tumor interactions (n=66) split into equal sized groups based on the percent interaction.

(F) Heatmap showing tumor-immune and immune-immune median nearest neighbor distances in μm. Darker color indicates closer proximity of the cells.

We observed that some of the patients in IC1, the poor outcome group, also had high level of CD8⁺ T cell and tumor cell interactions (Fig. 3D). Noticeably, TAM-CD68⁺ to tumor interactions were universally elevated in IC1 (Fig. 3D). Further analysis demonstrated that IC1 had a higher percentage of CD8⁺ T cells whose nearest neighbors were TAM-CD68⁺ compared with IC2 (elevated CD8⁺ T cell and tumor cell interactions with better survival), suggesting close spatial proximity of TAM-CD68⁺ and T cells instead of a random distribution of the cells (Fig. 4D). Here, the percentage was calculated as the proportion of CD8⁺ T cells with at least one or more TAM-CD68⁺ cell as nearest neighbors within 40μm. We observed consistent results in the LCI cohort (Fig. 4E). Higher levels of TAM-CD68⁺ to tumor cell interactions were linked to poorer clinical outcomes of HCC patients (Fig. S5B). Collectively, these results suggest that TAM-CD68⁺ cells might have an immune suppression role in the tumor.

Neighborhood analysis reveals immune infiltrated tumor microenvironment

Fig. 4F shows that tumor-immune interactions showed closest nearest neighbor proximity, followed by homotypic interactions (e.g. T cell-CD8⁺ to T cell-CD8⁺). The closest immune-immune heterotypic interaction was CD4⁺ T cell to CD8⁺ T cell at 17.4 μm. These observations prompted us to perform the neighborhood analysis described below to further investigate the spatial organization of the cells in HCC.

Neighborhood analysis was performed using a spatial-latent Dirichlet allocation (spatial LDA) model (22), which determines the local microenvironment based on the patterns in marker expression of adjacent cells, incorporating spatial localization. We identified five topics in the HCC samples from TIGER-LC cohort (Fig. 5A). Noticeably, Topic 4 aligned with the boundaries of immune infiltration, giving higher weights to immune markers (CD8, CD4, CD45, CD45RA, CD45RO, CD3E, CD68, CD163, CD44, CD20). In addition, patients with a higher average score of Topic 4 across all cells (indicating more immune infiltration) had a trend towards better survival outcomes (Fig. 5B). We found that the degree of immune infiltration assigned by the modelling aligned well with the CODEX staining and the manual cell-type assignment (Fig. 5C), tracing the tumor-immune boundary. We further examined this immune-infiltrated neighborhood signature in the LCI cohort (Fig. 5D–5F). We found Topic 4 in the LCI cohort was enriched in the same set of immune markers with similar trends as the TIGER-LC cohort, which is also associated with better prognosis (Fig. 5D,E). Univariate and multivariate cox proportional hazards regression analyses revealed that the immune neighborhood patterns are independently associated with overall survival in the LCI cohort with a similar trend but not statistically significant in the TIGER-LC cohort (Table S6). Taken together, our results demonstrate that features of spatial immune aggregates in liver cancer influence patient outcomes.

Figure 5. Immune infiltrated cellular neighborhood predicts better patient survival.

(A) Heatmap of the five modeled spatial-LDA tumor tissue microenvironments (topics) in the TIGER-LC Cohort. Blue indicates higher weight and red indicates lower weights for each marker.

(B) Overall survival in the TIGER-LC cohort separated by median value for the Topic 4 weight score across all cells.

(C) A representative TIGER-LC tumor core showing alignment of the immune-infiltrated regions in the CODEX staining, our cell type annotations, and predicted cellular neighborhood Topic 4. Blue indicates immune cells and grey indicates tumor cells in the spatial-LDA subplot. Darker blue intensity indicates cells which have a higher weight for Topic 4. Only immune cells are colored by weight.

(D) Heatmap of the modeled spatial-LDA topics in the LCI validation cohort.

(E) Overall survival in the LCI cohort separated by median value for the Topic 4 weight score across all cells.

(F) Alignment of the CODEX immune staining, cell assignment, and spatial-LDA Topic 4 boundaries. Both patients in panel C and F were assigned a higher immune-infiltrated neighborhood score and had better outcomes. The survival figure p-values are the log-rank test in GraphPad Prism.

Discussion

The findings of this study highlight key observations on the immune spatial architecture of HCC tumors. Crucial molecular details of HCC tumor biology have been captured by single-cell RNA sequencing, but the technology loses spatial resolution and restricts its utility to study positional cell-cell interactions (2, 27, 28). The advantage of utilizing a multiplexed immunofluorescence imaging technology, such as CODEX, is the visualization of intricate cell populations and their heterogenous spatial distributions in tumors. The spatial context of single cells in cancer ecosystems, such as compartmentalized immune communities or hot versus cold immune infiltration phenotypes, have shown relevance to understanding different immunotherapy responses and overall cancer outcomes (5, 7, 13, 15, 29).

Using CODEX multiplexed imaging, we captured a snapshot of the liver cancer tumor-immune spatial microenvironmental interactions and showed spatial dynamics of immune cells are linked to patient outcomes. Herein, we demonstrated survival subclass separation using tumor-immune interaction spatial features at one timepoint. This demonstrates the value of the rich spatial information contained in the localization and interactions within a tumor core. Using only the spatial interaction dynamics of CD8⁺ T cells with tumor cells demonstrated a strong patient survival trend. In addition, cores with greater T cell to tumor interactions showed higher cytotoxicity scores by transcriptional status.

Poorer survival HCC cores may face immunosuppressive features in the TME, preventing CD8⁺ T cell cytotoxicity and infiltration. The best survival group IC2 with the most CD8⁺ T cell-tumor interactions (and greatest cytotoxicity score) largely lacked TAM-CD68⁺ interactions. In contrast, some patients in IC1 with the most TAM-CD68⁺ cells had fewer CD8⁺ T cell-tumor interactions. Similarly, in the LCI cohort, patients with lower CD8⁺ T cell-tumor interactions (and poorer survival outcomes) had a greater degree of CD8⁺ T cells whose nearest neighbors were TAM-CD68⁺ cells. This supports previous findings that close proximity of specific subsets of TAMs to T cells can influence poorer effector T cell functions and cytotoxicity in HCC (18, 19). It is evident the immunological composition and spatial landscape of HCC is highly heterogenous. Developing immunotherapy treatment approaches which target the unique microenvironmental subtypes differently may be crucial.

Here, we focused on CD8⁺ T cells infiltrating at higher levels in patients with better outcomes. However, regulatory T cell (Tregs) infiltration in the TME [14, 15] has been previously implicated in unfavorable prognosis of HCC. Our current panel resolved tumor-immune interactions for lineages of CD4⁺ and CD8⁺ T cells but could not distinguish Tregs due to lack of FOXP3 staining. Future work should include additional cell markers for HCC with more functional states to improve resolution. For example, a comprehensive T cell multiplexed panel incorporating immune-exhausted checkpoints (PD-1, LAG-3, CTLA-4, TIM-3) and transcription factors (Tox, T-bet, Eomes) may help distinguish patients with immunosuppressive tumor ecosystems (30). Integrating single-cell spatial information with transcriptomics will help identify patient subsets which may respond best to therapies actively targeting cells in need of reinvigoration from exhausted states.

In both of our cohorts, spatial neighborhood modelling revealed the presence of an immune-infiltrated microenvironment which may predict patient survival. The modelling approach identifies neighborhoods (Topics) based on the spatial coordinates and close proximities of individual marker expressions. In this case, higher scores of Topic 4 overlapped well with CODEX imaging immune aggregates. The ability to separate patient survival based on immune clusters with CODEX demonstrates the utility of the method. Recent efforts have begun examining the influence of tertiary lymphoid structures/immune aggregates in cancer outcomes and immunotherapy responses (11). A trained clinical pathology classifier built on a well-defined multiplexed panel may be useful to stratify and monitor HCC patients.

Limitations of our study include the use of available antibodies, when studying markers not inventoried (pre-conjugated) or screened by Akoya Biosciences. Developing an antibody panel for CODEX can be costly and labor-intensive due to each custom antibody requiring individual conjugation with DNA oligonucleotides and validation (3). Our approach is also limited by using cell-based nuclear segmentation and manual tuning adjustments of markers. Liver tissue staining is often prone to high autofluorescence, requiring our careful attention to feature extraction and phenotyping. Recent advances in segmentation algorithms with deep learning will allow for more precise and scalable approaches (31).

Overall, our study highlights that utilizing spatial dynamics of cellular interactions within liver cancer can inform factors of better patient outcomes. Novel insights to spatial biomarkers will help the development of targeted treatments, especially for highly aggressive cancers. Future work in multiplexed imaging of liver cancer to identify how immune niches modulate or shift between microenvironment states upon immunotherapy will help further improve clinical outcomes.

Abbreviations

CODEX

CO-Detection by indEXing

FFPE

Formalin-Fixed Paraffin-Embedded

LCI

Liver Cancer Institute

TAMs

Tumor Associated Macrophages

TECs

Tumor Associated Endothelial Cells

TIGER-LC

Thailand Initiative in Genomics and Expression Research for Liver Cancer

TMA

Tissue Microarray

TME

Tumor Microenvironment

UMAP

Uniform Manifold Approximation and Projection for Dimension Reduction