Spatial profiling defines three immunophenotypes and a perineural invasion-associated immune-evasion niche in gallbladder cancer

Abstract

Background & Aims

Gallbladder cancer (GBC) has a high risk of postoperative recurrence, and immunotherapy-based adjuvant approaches remain unstandardised. We aimed to define spatial immunophenotypes in GBC and evaluate their associations with clinical outcomes and immune-evasion mechanisms.

Methods

We used multiplex immunohistochemistry to quantify 16 immune-cell subsets across the tumour core, tumour margin and peritumour regions in 86 treatment-naïve GBCs, and integrated nerve bundle-centred spatial analysis with transcriptomic profiling (n = 9) to investigate perineural invasion (PNI)-associated immune-evasion mechanisms.

Results

Unsupervised clustering and a CD8/α-SMA dual-marker approach defined three spatial immunophenotypes: immune-inflamed, immune-excluded, and immune-desert. The immune-inflamed phenotype featured high intratumoural immune infiltration with elevated PD-L1 expression (median combined positive score: 13.85 vs. 0.41 vs. 0.25, p <0.001), was associated with better prognosis (overall survival [OS]: hazard ratio [HR] 0.25, 95% CI 0.12–0.54, p <0.001; recurrence-free survival [RFS]: HR 0.41, 95% CI 0.21–0.79, p = 0.008), and improved outcomes with adjuvant chemoimmunotherapy vs. chemotherapy alone (OS: HR 0.10, 95% CI 0.02–0.48, p <0.001; RFS: HR 0.29, 95% CI 0.12–0.70, p = 0.003). The immune-excluded phenotype was stroma-enriched (α-SMA, p <0.001), with immune cells largely restricted to the tumour margin or peritumour regions; the immune-desert phenotype showed sparse immune infiltration. PNI was associated with immunosuppressive chronic inflammation, characterised by enrichment of PD-L1⁺ myeloid cells and exhausted CD8⁺ T cells near invaded nerve bundles (cell density, p <0.001). Transcriptomic analyses suggested enhanced tumour neurotropism in PNI-positive tumours (gene-set enrichment, p <0.05), together with activation of IL-6/IGF-1 related chronic inflammatory signalling and lipid-metabolic rewiring involving LEP, CD36, and ADIPOQ, collectively shaping a local immune-evasion niche (FDR <0.05).

Conclusions

We defined three spatial immunophenotypes in GBC with distinct immune features and clinical outcomes. PNI was associated with the formation of a local immune-evasion niche.

Impact and implications

GBC lacks validated biomarkers to guide postoperative immunotherapy, whereas spatial immune architecture provides a clinically meaningful indicator of tumour immune heterogeneity. By defining three spatial immunophenotypes, we delineate their distinct clinical outcomes and immune-evasion features. Moreover, we demonstrate an association between PNI and the formation of a local immune-evasion niche. These findings may inform clinicians and clinical trialists seeking to optimise postoperative risk stratification and adjuvant treatment strategies, and researchers investigating PNI-related molecular mechanisms. Clinically, integrating spatial immunophenotyping with PNI assessment into routine pathology workflows may improve prognostic stratification and inform adjuvant treatment decisions; however, prospective validation is required before broad clinical adoption.

Graphical abstract

Highlights

• •Spatial immunophenotypes in GBC stratify patients into three subgroups.
• •Immune-inflamed, immune-excluded, and immune-desert GBCs exhibit distinct prognoses and immune features.
• •PD-L1⁺ myeloid cells and exhausted CD8⁺ T cells accumulate near invaded nerves.
• •PNI is associated with immunosuppressive inflammation and lipid-metabolic rewiring.

Introduction

Gallbladder cancer (GBC) is a highly aggressive biliary tract malignancy with an overall poor prognosis.¹ The 5-year survival rate for locally advanced GBC is approximately 30%,² and even after curative resection, the early postoperative recurrence rate remains around 30%.³ Thus, improving prognostic stratification and optimising adjuvant treatment strategies for resectable GBC remains a key challenge in clinical management.

Recent trials such as TOPAZ-1⁴ and KEYNOTE-966⁵ have shown that immunotherapy-based systemic regimens improve survival in patients with unresectable biliary tract cancer. However, evidence supporting immunotherapy as adjuvant treatment for resectable biliary tract cancer remains limited. This is largely due to the lack of reliable biomarkers to identify likely responders, limiting the use of immunotherapy in patients at high risk of recurrence.

The density of tumour-infiltrating lymphocytes is commonly used as an indicator of tumour immune contexture and has been associated with clinical outcomes. However, immune status inferred solely from immune-cell density may yield paradoxical associations. For example, the prognostic significance of FOXP3⁺ regulatory T cell (Treg) infiltration varies across malignancies.^6,7 These findings suggest that the clinical relevance of immune cells is not determined by abundance alone, but may also depend on spatial localisation and cellular interactions within the tumour microenvironment. Recent studies further suggest that distinct patterns of immune-cell spatial organisation are associated with distinct clinical outcomes and immune-evasion mechanisms.[8], [9], [10] Therefore, comprehensive profiling of immune-cell spatial distribution, abundance and phenotypes is warranted in GBC.

To address this gap, we applied multiplex immunohistochemistry (mIHC) to systematically characterise heterogeneity in the spatial distribution and abundance of 16 immune-cell subsets across three anatomical compartments of GBC (tumour core, tumour margin, and peritumour). We then assessed associations between spatial immune features and clinical outcomes, as well as key clinicopathological characteristics, including perineural invasion (PNI). Our findings may provide preliminary evidence to support mechanistic studies and clinical translation of spatial immunophenotypes and PNI in GBC.

Patients and methods

Study cohort and selection criteria

This retrospective study included 86 patients with primary GBC who underwent surgical resection at Sun Yat-Sen Memorial Hospital between June 2017 and June 2024. The study was conducted in accordance with protocols approved by the Medical Ethics Committee, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, China. Inclusion criteria were as follows: availability of formalin-fixed, paraffin-embedded (FFPE) tumour specimens; postoperative histopathological confirmation of primary GBC; no history of preoperative anticancer therapy; and receipt of standard adjuvant treatment and regular follow-up after surgery. Patients were excluded if they refused to provide informed consent, had autoimmune disease, had a history of other malignancies, or had severe systemic comorbidities involving the heart or lungs. Overall survival (OS) was defined as the interval from the date of surgery to death from any cause or the last follow-up. Recurrence-free survival (RFS) was defined as the interval from the date of surgery to the first documented recurrence or the last follow-up. The last follow-up for all patients was conducted in March 2025.

The external validation cohort consisted of 149 cases included in a commercially available tissue microarray (catalogue no. YP-Gal1201) purchased from Shanghai Yingbeilan Biotechnology.

Fresh-frozen human GBC specimens used for RNA-sequencing (RNA-seq) profiling were collected as follows: PNI-positive GBC (n = 5) and PNI-negative GBC (n = 4). All specimens were confirmed by pathological diagnosis.

Dual and multiplex immunohistochemistry

Dual and multiplex immunohistochemistry were performed using a chromogenic double-staining kit (RCF040RM) and a seven-colour, six-plex immunohistochemistry kit (RC0086Plus-67RM) (both from Shanghai Huilan Biotechnology, Shanghai, China). Staining was performed according to the manufacturer’s instructions and our previously described methods.¹¹ Briefly, after deparaffinisation and rehydration, tissue sections underwent iterative cycles of antigen retrieval, quenching of endogenous peroxidase activity, serum blocking, primary and secondary antibody incubation, and incubation with chromogenic/fluorogenic substrates, followed by DAPI counterstaining and mounting with an antifade medium. The emission peaks of the Tyramide signal amplification (TSA) fluorophores were 480, 520, 570, 620, 690, and 780 nm. Details of the primary antibodies are listed in Table S1.

Region of interest (ROI) segmentation

Whole-slide images of H&E staining sections were acquired using an automated scanner (3D HISTECH). The tumour and non-tumour invasive borders were identified through a double-blind assessment performed by two pathologists (XXH and GH). Based on the invasive borders, the tissue regions were delineated, as illustrated in Fig. S1, with definitions from prior studies. The different tumour regions were defined as follows: tumour margin (M): a 1,000-μm wide area extending 500 μm on either side of the invasive border; tumour core (C): the tumour region excluding the tumour margin; and peritumour (P): the non-tumour region outside the tumour margin.

mIHC imaging and analysis

The Vectra Polaris automated quantitative pathology imaging system (Akoya Biosciences) was utilized to scan slides and obtain fluorescent whole-slide images. Using InForm (version 2.6, Akoya Biosciences, Marlborough, Massachusetts, United States), panoramic image analysis of three regions—tumour core, tumour margin, and peritumour—was conducted based on the tissue segmentation established in the previous step (Fig. S1). The analysis workflow included autofluorescence removal, spectral unmixing, cell segmentation, and phenotypic identification, followed by further co-expression and spatial distance analyses using the Phenoptr package in R software (R Foundation for Statistical Computing, Vienna, Austria). This procedure was performed by oncologists trained in the field (XH, XY, QXZ), who adjusted software parameters based on experimental controls to optimise detection metrics. As a quality control measure, all analysis data underwent a secondary review by two pathologists (XXH, GH) before export. All assessments ensured a double-blind approach to the maximum feasible extent.

Cellular neighbourhood analyses

Using CytoMAP (v1.4.21) and a previously described workflow,¹² we identified and clustered cellular neighbourhoods (CNs). Single-cell segmentation outputs with corresponding (x, y) coordinates were imported into CytoMAP, and per-cell neighbourhood composition features were computed to generate a neighbourhood feature matrix. This matrix was then clustered by k-means to define three cellular neighbourhoods (CN1–CN3), and cells were annotated according to their assigned CN class.

CD8 and α-SMA dual-marker model for spatial immunophenotyping

Intratumour CD8⁺ T-cell density was dichotomised at the cohort median of 342 cells/mm² into CD8-low (<342 cells/mm²) and CD8-high (≥342 cells/mm²). For CD8-high tumours, the ratio of parenchymal to stromal CD8⁺ T-cell density was calculated; cases with a ratio ≥0.6 were classified as immune-inflamed, and those with a ratio <0.6 as immune-excluded. For CD8-low tumours, the abundance of alpha-smooth muscle actin (α-SMA) was quantified as the percentage of the intratumour area and dichotomised at the cohort median of 14%; cases with α-SMA <14% were classified as immune-desert, and those with α-SMA ≥14% as immune-excluded. Using this algorithm, the CD8 and α-SMA dual-marker model stratified tumours into three spatial immunophenotypes: immune-desert, immune-excluded, and immune-inflamed.

Statistical and bioinformatic analyses

Statistical analyses were performed in R software (version 4.4.2; R Foundation for Statistical Computing). Survival curves were estimated using the Kaplan–Meier method and compared with the log-rank test. Continuous variables were compared using the Wilcoxon rank-sum test, Wilcoxon matched-pairs signed-rank test, or Kruskal–Wallis test, as appropriate. Associations between categorical variables were assessed using the Χ² test or Fisher’s exact test, and agreement between classification methods was evaluated using Cohen’s κ (kappa) coefficient. Univariable and multivariable survival analyses were conducted with Cox proportional hazards regression models. Spearman’s rank correlation coefficient was used to assess associations between continuous variables. Hierarchical clustering and heatmaps were generated with the pheatmap package. Statistical significance was defined as follows: p <0.05; ∗p <0.01; and ∗∗p <0.001.

Bulk RNA-seq data were processed and normalised using standard workflows. Immune-cell fractions were inferred with the CIBERSORT algorithm. Single-sample Gene Set Enrichment Analysis (ssGSEA) was performed with the GSVA package using Hallmark gene sets from the Molecular Signatures Database. Differentially expressed genes (DEGs) between groups were identified with the DESeq2 package, applying thresholds of p <0.05 and fold change ≥1.5. Functional enrichment of DEGs was evaluated using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses implemented in the clusterProfiler package. Protein–protein interaction (PPI) networks were constructed using the STRING database and visualised in Cytoscape.

Results

Spatial distribution of immune-cell subsets in GBC

To systematically map the spatial immune landscape of GBC, we quantified immune-cell densities (cells/mm²) spanning broad lymphoid and myeloid lineages across the tumour core, tumour margin, and peritumour regions (Fig. 1, Fig. 2A and B; Fig. S1), with subset definitions provided in Table S2.

Fig. 1.

Workflow for spatial profiling of the immune microenvironment in GBC.

FFPE serial sections from resected tumours were analysed by H&E and mIHC, with regions defined as tumour core (C), tumour margin (M), and peritumour area (P). Single-cell segmentation and phenotyping were performed across mIHC panels, followed by computational analyses. Immune-cell subsets and the markers used for lineage, functional state, and auxiliary annotation are summarised. CN, cellular neighbourhood; FFPE, formalin-fixed, paraffin-embedded; GBC, gallbladder cancer; mDC, myeloid dendritic cell; mIHC, multiplex immunohistochemistry; NK, natural killer cell; NKT, natural killer T cell; PCA, principal component analysis; pDC, plasmacytoid dendritic cell; PNI, perineural invasion; PPI, protein–protein interaction; RNA-seq, RNA-sequencing; ssGSEA, single-sample Gene Set Enrichment Analysis; Treg, regulatory T cell; α-SMA, alpha-smooth muscle actin.

Fig. 2.

Spatial distribution of immune-cell subsets in GBC.

(A) Representative images of four mIHC panels. (B) Representative H&E and matched mIHC images of the tumour core, tumour margin, and peritumour regions. (C) Heatmap of differential immune-cell infiltration across three regions. (D) Correlation heatmap of immune-cell subset infiltration across three regions (Spearman’s correlation coefficients). (E) Ridgeline plots showing immune-cell subsets enriched in each region (Kruskal–Wallis test). (Scale bar is shown in the upper left of the image, ∗p <0.05; ∗∗p <0.01; ∗∗∗p <0.001; ns, not significant.) GBC, gallbladder cancer; mIHC, multiplex immunohistochemistry.

Heatmaps and ridgeline plots demonstrated pronounced spatial heterogeneity in immune infiltration (Fig. 2C–E). CD123⁺ plasmacytoid dendritic cells (pDCs), CD208⁺ myeloid dendritic cells (mDCs), and CD4⁺AhR⁺ Th22 cells were significantly enriched in the tumour core. In contrast, CD4⁺ T cells, CD3^-CD56⁺ natural killer (NK) cells, and CD20⁺CD79a⁺ B cells showed significantly higher infiltration in the peritumour region. The tumour margin was preferentially enriched for several immunosuppressive and myeloid subsets, including CD4⁺FOXP3⁺ regulatory T cells (Tregs), CD20^-CD79a⁺ plasma cells, CD68⁺ macrophages (CD163⁺, M2-like; CD163^-, tending toward M1-like), and CD66b⁺ neutrophils. By comparison, CD3⁺ T cells, CD3⁺CD56⁺ natural killer T (NKT) cells, CD8⁺ T cells, and tryptase⁺ mast cells showed no significant differences across the three regions (Fig. S2). Correlation heatmaps further indicated broadly positive associations among immune-cell subset densities within the tumour core (Fig. 2D).

Overall, immune-cell subsets in the GBC tumour microenvironment exhibited clear region-specific infiltration patterns, underscoring the importance of spatial context for interpreting and characterising the tumour immune microenvironment.

Unsupervised clustering of spatial distribution profiles of immune-cell subsets defines spatial immunophenotypes in GBC

Based on immune-cell subset densities across the tumour core, tumour margin, and peritumour regions, we performed unsupervised clustering within each region to classify patients as immune-high or immune-low (Fig. 3A, upper panel), and then integrated the three region-level classifications to define three spatial immunophenotypes (immune-inflamed, immune-excluded, and immune-desert; Fig. 3A, lower panel). Representative multiplex immunohistochemistry images illustrated the characteristic spatial immune-infiltration patterns of these immunophenotypes (Fig. 3B).

Fig. 3.

Unsupervised clustering identified three spatial immunophenotypes in GBC with distinct clinical outcomes.

(A) Unsupervised clustering based on immune-cell subset densities in the tumour core, tumour margin, and peritumour regions classified patients as immune-high or immune-low within each region (upper), and integration of the three region-level classifications defined three spatial immunophenotypes: immune-inflamed, immune-excluded, and immune-desert (lower). (B) Representative mIHC images of tumours with each spatial immunophenotype. (C) Kaplan–Meier curves for OS and RFS by spatial immunophenotype, with multivariable Cox regression forest plots (log-rank test, univariable and multivariable Cox proportional hazards regression). (D) Kaplan–Meier curves for OS and RFS comparing adjuvant chemoimmunotherapy vs. adjuvant chemotherapy alone within each spatial immunophenotype subgroup (log-rank test). (E) Summary heatmap of immune-cell subset enrichment patterns across the three spatial immunophenotypes. Scale bar is shown in the upper left of the image. GBC, gallbladder cancer; mIHC, multiplex immunohistochemistry; OS, overall survival; RFS, recurrence-free survival.

Survival analyses demonstrated significant differences in postoperative OS and RFS across the three spatial immunophenotypes (Fig. 3C). In multivariable Cox regression, spatial immunophenotype remained an independent prognostic factor for both OS and RFS, and PNI was an independent adverse factor for RFS (Fig. 3C; Table S6–S7). In treatment-stratified analyses (chemotherapy alone vs. chemoimmunotherapy), a significant OS and RFS benefit from chemoimmunotherapy was observed only in the immune-inflamed subgroup (Fig. 3D; Table S3–S5).

Integrative analyses of spatial enrichment (Fig. 3E; Fig. S3A–F) showed that CD3⁺ T cells, CD3⁺CD8⁺ T cells, and tryptase⁺ mast cells displayed phenotype-dependent regional enrichment patterns in both the overall cohort and within each immunophenotype. By contrast, CD3⁺CD4⁺AhR⁺ Th22 cells, CD3⁺CD56⁺ NKT cells, CD123⁺ pDCs, and CD208⁺ mDCs exhibited relatively subtype-specific spatial enrichment differences in the immune-excluded phenotype, whereas the remaining subsets showed broadly consistent spatial patterns across subgroups.

CD8 combined with α-SMA stratifies spatial immunophenotypes in GBC

In clinical translation, biomarkers must balance accessibility, cost, and standardised workflows. Accordingly, we sought a simplified alternative to the unsupervised clustering approach while preserving stratification performance.

Prior studies have shown that the spatial distribution of intratumoural CD8⁺ T cells can stratify spatial immunophenotypes.^13,14

Interestingly, we observed immunophenotype-dependent CD8⁺ T-cell spatial enrichment in GBC (Fig. 3E). We then performed k-means clustering on the spatial-neighbourhood composition feature matrix derived from tumour ROIs, defining three CN classes (CN1–CN3) (Fig. 4A–D). Correlation profiling of CN densities against immune-cell subsets indicated that CN1 represented a lymphocyte-enriched neighbourhood, showing positive associations with CD8⁺ and CD4⁺ T cells, NK/NKT cells, dendritic cells (mDC and pDC) and mast cells (Fig. 4D). In contrast, CN2 was a myeloid-dominant neighbourhood, characterised by enrichment of macrophages and neutrophils and relative depletion of lymphoid populations. CN3 displayed a B cell/plasma cell-enriched neighbourhood signature, correlating positively with B cells and plasma cells. Importantly, CN1 abundance correlated strongly with intratumoural CD8⁺ T-cell density (R = 0.911, p <0.001); however, ROIs deviating from this relationship tended to exhibit higher α-SMA abundance (Fig. 4E). Principal component analysis (PCA) of CN1–CN3 densities across tumour ROIs identified two dominant axes (PC1 and PC2) (Fig. 4F). PC1 primarily captured a CN1–CN2 compositional gradient, separating CN2-high (myeloid-dominant; negative PC1) from CN1-high (lymphocyte-enriched; positive PC1) ROIs; accordingly, ROIs shifted towards the CN1/positive-PC1 direction showed higher intratumoural CD8⁺ T-cell density. PC2 largely reflected variation in CN3 abundance, distinguishing ROIs with higher levels of the B cell/plasma cell-enriched neighbourhood (CN3). Notably, ROIs with higher α-SMA abundance (% stroma) tended to be displaced away from the CN1/positive-PC1 direction, implicating stromal activation as a key determinant of heterogeneity in CD8⁺ T-cell spatial infiltration.

Fig. 4.

A simplified CD8/α-SMA dual-marker model stratifies spatial immunophenotypes in GBC.

(A) K-means clustering of the cell neighbourhood composition matrix identified three neighbourhood classes (CN1, CN2, and CN3). (B) Stacked bar plots show the CN1, CN2, and CN3 proportions within tumour ROIs for each patient. (C) Box plots compare overall cell densities across CN subtypes (Kruskal–Wallis test). (D) Bubble plot shows correlation patterns between CN subtypes and intratumoural densities of immune-cell subsets (Spearman’s correlation). (E) Scatter plot shows the correlation between CN1 density and CD8⁺ T-cell density; point size indicates the α-SMA⁺ area fraction within each ROI (Pearson correlation). (F) Principal component analysis of CN densities, with point size indicating α-SMA⁺ area fraction and colour indicating intratumoural CD8⁺ T-cell density; arrows denote CN1–CN3 contributions to PC1 and PC2. (G) Schematic of the CD8/α-SMA-based spatial immunophenotyping strategy, integrating intratumoural CD8⁺ T-cell density, α-SMA abundance, and the parenchyma-to-stroma CD8⁺ T-cell density ratio to classify tumours as immune-desert, immune-excluded, or immune-inflamed. (H) Cohen’s κ assessing agreement between spatial immunophenotypes defined by unsupervised clustering (spatial immunophenotype ^a) and by the CD8/α-SMA model (spatial immunophenotype ^b). (I) Kaplan–Meier curves for OS and RFS stratified by CD8/α-SMA-defined spatial immunophenotypes (log-rank test). (J) Representative dual-IHC images (left) and corresponding OS Kaplan–Meier curves (right) showing CD8/α-SMA-based stratification in the external validation cohort (log-rank test). (Scale bar is shown in the upper left of the image, ∗p <0.05; ∗∗p <0.01; ∗∗∗p <0.001; ns, not significant.) α-SMA, alpha-smooth muscle actin; CN, cellular neighbourhood; GBC, gallbladder cancer; IHC, immunohistochemistry; OS, overall survival; RFS, recurrence-free survival; ROI, region of interest.

Building on these observations, we integrated intratumoural CD8⁺ T-cell density, α-SMA abundance, and the parenchyma-to-stroma CD8⁺ T-cell density ratio to construct a CD8/α-SMA–based three-class decision tree that stratified patients into immune-desert, immune-excluded, and immune-inflamed spatial immunophenotypes (Patient and methods; Fig. 4G). Spatial immunophenotype (Fig. 4H^b) (CD8/α-SMA–based) showed strong agreement with spatial immunophenotype (Fig. 4H^a) (unsupervised clustering–defined) (κ = 0.913, p <0.001; Fig. 4H) and reproduced OS and RFS stratification (Fig. 4I). In an external validation cohort, dual-colour immunohistochemistry (IHC) for CD8 and α-SMA, applying the same decision rules, likewise identified the three immunophenotypes and significantly separated postoperative OS (Fig. 4J), supporting the feasibility and scalability of this dual-marker model for clinical implementation.

PNI is associated with immunosuppressive chronic inflammation and poor prognosis in GBC

Fig. 5A shows an integrated heatmap grouped by the three spatial immunophenotypes, highlighting distinct distributions of clinicopathological variables and tumour microenvironment components; notably, the immune-inflamed phenotype displayed the highest PD-L1 scores across the combined positive score (CPS), tumour cell score (TC), and immune-cell score (IC) (Fig. 5B). The Sankey plot summarises the relationship between PNI status and spatial immunophenotypes, showing no significant association (χ² = 2.45, p = 0.294; Fig. 5C). Stratified analyses further indicated that, compared with PNI-negative tumours, PNI-positive tumours exhibited broadly increased infiltration densities of multiple immune-cell subsets within the tumour core (Fig. 5D), a pattern not consistently observed when stratifying by microvascular invasion (MVI), differentiation, or T/N/M stage (Fig. S4A–E), suggesting that PNI may more closely capture an intratumoural chronic inflammatory milieu. We then assessed the association between PNI and PD-L1 expression and found that PD-L1 IC scores were significantly higher in PNI-positive tumours, whereas CPS and TC did not differ markedly (Fig. 5E). This pattern was consistent across the immune-inflamed, immune-excluded, and immune-desert subgroups (Fig. S5A), indicating that the PD-L1 increase is mainly driven by enhanced PD-L1 expression in immune cells rather than tumour cells. To quantify cell–cell spatial proximity, we applied a previously reported distance-based effective score (Fig. 5G and H). Using this metric, PNI positivity was specifically associated with higher effective scores between PD-1⁺CD8⁺ T cells and PD-L1⁺CD68⁺CD163⁺ M2-like macrophages or PD-L1⁺CD123⁺ pDCs (Fig. 5F–J), and these associations persisted within each spatial immunophenotype (Fig. S5B). Survival analyses demonstrated shorter recurrence-free survival in PNI-positive patients (Fig. 5K), and PNI positivity was significantly associated with higher N stage (Fig. 5L) and elevated vimentin expression (Fig. 5M). Taken together, these findings associate PNI with increased PD-L1 expression in immune cells, closer spatial apposition of CD8⁺ T cells to immunosuppressive myeloid cells, and poorer clinical outcomes.

Fig. 5.

PNI is associated with immunosuppressive chronic inflammation and poor prognosis in GBC.

(A) Heatmap summarising the distribution of variables stratified by spatial immunophenotype (immune-inflamed, immune-excluded, and immune-desert). (B) Box plots comparing PD-L1 CPS, TC, and IC scores across spatial immunophenotypes (Kruskal–Wallis test). (C) Sankey plot showing the relationship between PNI status and spatial immunophenotype (Χ² test). (D) Heatmap and violin plots comparing immune-cell infiltration densities between PNI⁺ and PNI⁻ tumours (Wilcoxon rank-sum test). (E) Box plots comparing PD-L1 CPS, TC, and IC scores between PNI⁺ and PNI⁻ tumours (Wilcoxon rank-sum test). (F) Representative mIHC images of the panel used for distance-based effective score analysis. (G, H) Definition and schematic of the effective score for spatial proximity analysis. (I, J) Box plots comparing effective scores between the indicated cell pairs across three each spatial immunophenotypes and PNI status (Kruskal–Wallis test, Wilcoxon rank-sum test). (K) Kaplan–Meier curve of RFS stratified by PNI status (log-rank test). (L) Stacked bar plot showing the association between PNI status and N stage (Fisher’s exact test). (M) Stacked bar plot showing the association between PNI status and vimentin expression (Χ² test). (Scale bar is shown in the upper left of the image, ∗p <0.05; ∗∗p <0.01; ∗∗∗p <0.001; ns, not significant.) CPS, combined positive score; GBC, gallbladder cancer; IC, immune-cell score; PNI, perineural invasion; TC, tumour cell score.

Multi-omics analysis reveals an immune-evasive niche surrounding perineural invasion in GBC

Based on the discrimination by H&E and CK19/S100B dual-IHC, we annotated nerve bundles in tumours with PNI (PNI⁺) and without PNI (PNI⁻), and defined two concentric rings (inner and outer) centred on each nerve bundle, generating four nerve bundle-related ROIs: PNI⁺ inner, PNI⁺ outer, PNI⁻ inner, and PNI⁻ outer (Fig. 6A). Representative mIHC images for this nerve bundle-centred spatial analysis are shown in Fig. 6B. Quantification demonstrated higher overall immune-cell infiltration densities in PNI⁺ inner and PNI⁺ outer regions than in PNI⁻ inner and PNI⁻ outer regions (Fig. 6C). CD68⁺CD163⁺ M2 macrophages and CD123⁺ pDCs exhibited a clear gradient distribution, peaking in PNI⁺ inner, decreasing in PNI⁺ outer, and remaining markedly lower in PNI⁻ inner and PNI⁻ outer (Fig. 6D). In contrast, total CD3⁺CD8⁺ T-cell densities were comparable between PNI⁺ inner and PNI⁺ outer regions (Fig. 6D). Using the CD8⁺ T-cell functional panel (Fig. 6B, panel D and 6E), we observed a significant increase in both the absolute abundance of exhausted T cells (Tex) and their proportion among CD3⁺CD8⁺ T cells in PNI⁺ inner regions, accompanied by a relative reduction in effector T cells (Teff), indicating a predominantly exhausted CD8⁺ T-cell state adjacent to PNI⁺ nerves. Consistently, PD-L1⁺ M2 macrophages, PD-L1⁺ pDCs, and PD-L1⁺ neutrophils were significantly enriched in PNI⁺ inner regions (Fig. 6F). Collectively, these data indicate that the niche surrounding invaded nerves is characterised by PD-L1⁺ myeloid-cell enrichment and CD8⁺ T-cell exhaustion.

Fig. 6.

Perineural spatial analysis identifies an immune-evasion niche surrounding invaded nerve bundles.

(A) Representative H&E, CK19/S100B, and mIHC images of PNI⁺ and PNI⁻ nerve bundles (left) and nerve-centred ROI definition (right): PNI⁺ inner, PNI⁺ outer, PNI⁻ inner, and PNI⁻ outer. (B) Representative images of mIHC panels used for perineural analysis. (C) Heatmap of immune-cell subset densities across the four ROIs. (D) Box plots of CD68⁺CD163⁺ M2 macrophages, CD123⁺ pDCs, and CD3⁺CD8⁺ T cells across the four ROIs (Kruskal–Wallis test). (E) CD8⁺ T-cell states (precursor-exhausted T cells (Tpex), exhausted effector T cells (Tex), effector T cells (Teff)): representative images (left), densities across ROIs (top right), and Tpex/total CD3⁺CD8⁺ T-cell proportions (bottom right) (Kruskal–Wallis test). (F) Violin plots of PD-L1⁺ immune-cell densities across ROIs (Kruskal–Wallis test). (Scale bar is shown in the upper left of the image, ∗p <0.05; ∗∗p <0.01; ∗∗∗p <0.001; ns, not significant.) mIHC, multiplex immunohistochemistry; PNI, perineural invasion; ROI, region of interest; Teff, effector T cells; Tex, exhausted T cells.

At the transcriptomic level, we performed bulk RNA-seq to compare PNI⁺ and PNI⁻ tumours. Representative CK19/S100b dual-IHC images for the two groups are shown in Fig. 7A. Fig. 7B summarises CIBERSORT-inferred immune-cell fractions and changes in immune-related gene-set scores derived from the transcriptomic data. Using tumour type–specific PNI gene sets curated across multiple cancers in a prior study,¹⁵ we observed a conserved PNI-associated transcriptional programme in PNI⁺ GBC (Fig. 7C). Among canonical PNI mechanisms, the neurotrophic signalling gene sets—reflecting a neural-like, neurotropic propensity—showed the most prominent increase (Fig. 7D), suggesting that tumour cells acquire neurotrophic and neurally directed traits that facilitate PNI. Pathway enrichment of genes upregulated in PNI⁺ tumours (Fig. 7E and F) indicated heightened activity in neural-like programmes, neuroinvasive tropism, and cell–cell interaction pathways. Conversely, downregulated genes were predominantly enriched in epithelial barrier and antimicrobial defence pathways (Fig. S6), consistent with impaired barrier function and attenuated innate antimicrobial immunity. PPI analysis further highlighted IGF1, IL6, NTRK2, and NRXN1 as neuroinflammatory hub molecules, together with lipid metabolism–related factors (LEP, CD36, and ADIPOQ) as central network nodes (Fig. 7G).

Fig. 7.

Transcriptomic profiling reveals a PNI-associated immune-evasion niche.

(A) Representative CK19/S100B dual-IHC images of PNI⁻ (upper) and PNI⁺ (lower) GBC samples. (B) Heatmaps comparing CIBERSORT-inferred immune-cell abundances (left) and enrichment scores of immune-related gene sets (right) between PNI⁻ and PNI⁺ tumours. (C) Heatmap comparing enrichment scores of pan-cancer PNI gene sets between PNI⁻ and PNI⁺ tumours. (D) Box plots comparing enrichment scores of PNI-related pathways between PNI⁻ and PNI⁺ tumours (Kruskal–Wallis test). (E) GO biological process enrichment of genes upregulated in PNI⁺vs. PNI⁻ tumours (bubble plot; dot size, gene count; colour, adjusted p value). (F) KEGG pathway enrichment of genes upregulated in PNI⁺vs. PNI⁻ tumours (bubble plot; dot size, gene count; colour, adjusted p value). (G) PPI network of genes upregulated in PNI⁺ tumours; node size and colour indicate connectivity. (Scale bar is shown in the upper left of the image, ∗p <0.05; ∗∗p < 0.01; ∗∗∗p <0.001; ns, not significant.) GBC, gallbladder cancer; GO, Gene Ontology; IHC, immunohistochemistry; KEGG, Kyoto Encyclopedia of Genes and Genomes; PNI, perineural invasion; PPI, protein–protein interaction.

In summary, multi-omics analyses indicate that PNI-positive GBC exhibits a neural-like invasive phenotype, enhanced neuro-chronic inflammatory signalling, and lipid-metabolic rewiring, and forms an immune-evasion niche around invaded nerve bundles.

Discussion

Based on spatial distribution data from 16 immune-cell subsets across the tumour core, tumour margin, and peritumour regions, we classified patients into three spatial immunophenotypes—immune-inflamed, immune-excluded, and immune-desert—using unsupervised clustering. These immunophenotypes stratified postoperative RFS and OS; only the immune-inflamed subgroup showed clinical benefit from adjuvant chemoimmunotherapy. Given the cost and limited accessibility of multi-marker assays in routine practice, we developed a simplified CD8/α-SMA dual-marker model that showed high concordance with the multi-marker unsupervised classification, supporting its clinical translatability. Moreover, our study indicates that PNI is associated with immunosuppressive chronic inflammation, tumour neurotropism, and lipid-metabolic reprogramming, and is characterised by an immune-evasion niche. Collectively, these findings provide preliminary biological evidence to inform mechanistic studies and clinical translation of PNI.

Spatial distribution maps showed that neutrophils, Tregs, macrophages, and plasma cells were consistently enriched at the tumour margin across the three spatial immunophenotypes. A recent study¹⁶ reported that preferential accumulation of neutrophils at the invasive front of GBC promotes the establishment of an invasive niche at the tumour margin. Another study¹⁷ reported markedly increased macrophage infiltration in invasive GBC. These findings suggest that, beyond neutrophils, mechanistic and functional dissection of Treg, macrophage, and plasma-cell enrichment at the tumour margin may further elucidate the molecular basis of GBC aggressiveness.

In this study, we classified GBC into three spatial immunophenotypes—immune-inflamed, immune-excluded, and immune-desert—thereby validating a previously proposed classification framework in gallbladder cancer¹⁸ and further demonstrating that this stratification has prognostic value for clinical outcomes, consistent with findings reported in other malignancies.^13,14,19 Distinct spatial immunophenotypes may reflect different intrinsic immune-evasion mechanisms. The immune-evasion features we observed across the three spatial immunophenotypes are consistent with the immunosuppressive microenvironment described in the molecular subtyping study of GBC by Koshiol et al.²⁰

Regarding the mechanisms underlying the immune-excluded phenotype, the inflammatory stroma subtype within the STIM classification proposed by Sia et al.²¹ in intrahepatic cholangiocarcinoma and the mesenchymal subtype described by Llovet et al.²² in extrahepatic cholangiocarcinoma provide relevant clues. A canonical mechanism of stroma-enriched tumour is fibroblast activation and extracellular matrix deposition, which creates a physical barrier that restricts immune effector-cell infiltration and limits the penetration of immunotherapeutic agents into the tumour parenchyma. This is consistent with our results showing that the immune-excluded subtype exhibited significantly higher α-SMA and collagen IV abundance, was associated with poorer prognosis, and did not derive benefit from adjuvant chemoimmunotherapy. Interestingly, we observed significant accumulation of tryptase⁺ mast cells and CD3⁺CD8⁺ T cells at the tumour margin exclusively in the immune-excluded subtype. In PCA based on neighbourhood clustering, α-SMA abundance emerged as the major factor driving differences in CD8⁺ T cell neighbourhood distribution, further supporting the stromal-immune exclusion effect reported by Bengsch et al.²³ in hepatocellular carcinoma and Wen et al.²⁴ in GBC. Moreover, prior studies[25], [26], [27] suggest that mast-cell infiltration can drive fibroblast activation and stromal remodelling through the release of proteases and fibroblast-activating mediators. Notably, our results showed that mast cells were enriched at the tumour margin in the immune-excluded subtype, providing a structural basis for this mechanism.

These findings underscore that future therapeutic strategies for the immune-excluded subtype should prioritise targeting the extracellular matrix in combination with immunotherapy. In addition, local cell–cell interaction networks centred on mast cells may represent novel therapeutic targets for modulating fibroblast activation and ECM remodelling, warranting focused investigation in future mechanistic and translational studies.

The immune feature of the immune-desert phenotype is a widespread lack of effective immune-cell infiltration across all spatial compartments.¹⁸ Our results further support this notion: in the immune-desert phenotype, multiple immune-cell subsets were significantly reduced across the three regions, accompanied by markedly lower PD-L1 expression and stromal abundance, which were closely associated with poorer outcomes and limited benefit from adjuvant immunotherapy. Taken together, research and therapeutic strategies for the immune-desert subtype should focus on converting these prototypical immune-cold tumours into immune-hot tumours, thereby restoring sensitivity to immunotherapy.

The immune-inflamed phenotype is characterised by high intratumoural infiltration of diverse immune cells, elevated immune-checkpoint expression, and clinical benefit from immunotherapy, consistent with the immune features reported previously.^22,28 Previous studies have highlighted associations between histopathological features and heterogeneity of the tumour immune microenvironment.^24,29 Notably, we found that PNI was significantly associated with higher intratumoural infiltration across the majority of immune-cell subsets, indicating a nerve-associated inflammatory microenvironment within tumours. However, our nerve bundle-centred spatial analyses indicate that PNI is associated with an immune-evasion niche, characterised by enrichment of PD-L1⁺ myeloid cells and exhausted CD8⁺ T cells, consistent with immunosuppressive chronic inflammation. Clinically, within the immune-inflamed subgroup, GBC patients with PNI had significantly worse outcomes.

Although evidence in GBC remains limited, studies in other gastrointestinal tumours^30,31 support a close link between PNI, chronic inflammation, and immune evasion. Consistent with prior studies,[32], [33], [34] our transcriptomic analyses indicate that PNI is not merely an anatomical phenomenon but is accompanied by acquisition of a neural-like transcriptional signature in tumour cells, providing a mechanistic basis for tumour spread along nerve bundles. Building on this, tumour cells engage in close crosstalk with Schwann cells, promoting neurogenic chronic inflammation (e.g. IL6 and IGF1) alongside lipid-metabolic reprogramming (e.g. LEP, CD36, and ADIPOQ). Collectively, these processes establish a structural and metabolic foundation for an invasive, immunosuppressive perineural niche around invaded nerve bundles. Notably, recent evidence³⁵ linked PNI to resistance to anti-PD-1 therapy, reinforcing the translational rationale for targeting PNI to improve clinical outcomes and increase the likelihood of benefit from immunotherapy.

Our study has several limitations. First, the relative rarity of GBC constrained our cohort size, and the retrospective real-world design may introduce unavoidable selection and confounding biases; therefore, multi-centre prospective studies are warranted to validate the utility of spatial immunophenotypes for prognostic stratification and guiding adjuvant treatment decisions. Second, our conclusions should be interpreted with caution because they are primarily derived from integrative, correlative analyses of multiplex immunohistochemistry and bulk RNA-sequencing; rigorous in vitro and in vivo experiments will be needed to establish causality.

In conclusion, we defined three spatial immunophenotypes in GBC, each associated with distinct clinical outcomes and immune-evasion features. We also identified a close link between PNI and an immune-evasion niche. Collectively, these findings support the clinical evaluation of spatial immunophenotypes and PNI in GBC and may help inform therapeutic decision-making.

Abbreviations

CN, cellular neighbourhood; CPS, combined positive score; DEGs, differentially expressed genes; FFPE, formalin-fixed, paraffin-embedded; GBC, gallbladder cancer; GO, Gene Ontology; IC, immune-cell score; IHC, immunohistochemistry; KEGG, Kyoto Encyclopedia of Genes and Genomes; mDC, myeloid dendritic cell; mIHC, multiplex immunohistochemistry; Mø, macrophage; MVI, microvascular invasion; NK, natural killer cell; NKT, natural killer T cell; OS, overall survival; PC, plasma cell; PCA, principal component analysis; pDC, plasmacytoid dendritic cell; PNI, perineural invasion; PPI, protein–protein interaction; RFS, recurrence-free survival; RNA-seq, RNA-sequencing; ROI, region of interest; ssGSEA, single-sample Gene Set Enrichment Analysis; TC, tumour cell score; Teff, effector T cells; Tex, exhausted T cells; Treg, regulatory T cell; α-SMA, alpha-smooth muscle actin.

Authors’ contributions

Conceived this study: XH, ZYZ, WBL, YRP, JW. Performed the experiments: XH, XY, LQL, XXH. Analysed data: XH, XY, SCX. Queried the patient follow-up: SCX, QXZ. Prepared the figures and tables: XH, XY. Wrote the manuscript: XH, XY, SCX, HLY. Reviewed the manuscript: ZYZ, WBL, YRP. Provided the tissue samples: JKR, XXH. Performed the pathology and image evaluation: GH, XXH, XH, QXZ, XY. Provided the research facilities: JW, ZYZ, WBL, YRP, JW.

Data availability

The data can be obtained from the corresponding author on reasonable request.

Financial support

This work was supported by the National Natural Science Foundation of China (No. 81971463 and 81902413); the Science and Technology Program of Guangdong Province, China (No. 2025A1515010410, 2024A1515010619, and 2021A1515010109); Grant [2013] 163 from the Key Laboratory of Malignant Tumour Molecular Mechanism and Translational Medicine of Guangzhou Bureau of Science and Information Technology; Grant KLB09001 from the Key Laboratory of Malignant Tumour Gene Regulation and Target Therapy of Guangdong Higher Education Institutes.

Conflict of interest

The authors declare no conflicts of interest.

Please refer to the accompanying ICMJE disclosure forms for further details.

Supplementary data

The following are the Supplementary data to this article: