Single-cell profiling uncovers a hypoxia-vascular-immune axis underlying poor immunotherapy response in acral versus cutaneous melanoma Authors: Qian Dong, Yiming Zhang, Fuchu He, Aihua Sun Published: 2026 ABSTRACT BackgroundAcral melanoma (AM), accounting for ~50% of melanomas in Asian populations, exhibits far poorer immunotherapy response (anti-PD-1 ORR < 20%) than cutaneous melanoma (CM, ~43.7%). While the tumor microenvironment (TME) plays a pivotal role in immunotherapy resistance, previous studies mainly focused on immune cells, neglecting stromal components like pericytes—creating a critical knowledge gap in understanding AM’s therapeutic refractoriness.MethodsWe integrated multiple single-cell transcriptomic datasets from 31 acral melanoma, 13 cutaneous melanoma and 35 normal skin samples. Comprehensive bioinformatics analyses including cell type annotation, differential expression, gene functional enrichment, and cell-cell interaction analysis, were performed to delineate the tumor microenvironment differences between acral and cutaneous melanoma. Key findings were validated using spatial transcriptomics, bulk RNA-seq datasets, and functional assays including qPCR and western blot.ResultsAM had a more immunosuppressive TME, enriched in collagen-secreting RGS5+/COL3A1+ pericytes that co-localized with KDR+/PODXL+ endothelial cells, associated with vascular abnormalities and hypoxia. These collagen-secreting RGS5+/COL3A1+ pericytes (enriched in AM) were significantly negatively correlated with the infiltration ratio of CD8+ tumor-infiltrating lymphocytes (TILs); moreover, in immunotherapy cohorts, high expression of the collagen-secreting pericyte marker RGS5 was associated with poorer response to immune checkpoint blockade (ICB) treatment. We also identified a hypoxia-related NECTIN2-TIGIT immunosuppressive axis: hypoxia upregulated NECTIN2 on multiple cells, interacting with TIGIT on T cells to potentially exacerbate dysfunction.ConclusionsThis study provides insights into vascular-stromal features associated with immunotherapy resistance in AM, highlighting RGS5+/COL3A1+ pericytes and the NECTIN2-TIGIT axis as targets, and guides development of effective combination immunotherapies for AM.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12967-026-08201-2. Background Melanoma is a highly lethal and heterogeneous skin cancer, with subtypes exhibiting distinct clinical presentations and prognoses [1]. Among these, cutaneous melanoma (CM) dominates in Caucasians (~90%), whereas acral melanoma (AM) constitutes approximately 50% of melanomas in Asian populations [2]. Notably, AM harbors molecular features that directly exacerbate its therapeutic challenges: compared to CM, it has lower tumor mutational burden (TMB), more extensive copy number variations (CNVs) [3, 4], and a far lower BRAF mutation rate (<20% vs ~50% in CM). This low BRAF mutation rate restricts the use of BRAF-targeted therapies in AM [5]. Meanwhile, though immunotherapy has revolutionized melanoma treatment [6], anti-PD-1 monotherapy achieved an objective response rate (ORR) of only < 20% in AM, far below the 43.7% ORR in CM [6–8], leaving AM with a severe “dual dilemma” of limited targeted and immunotherapeutic options. Resolving AM’s immunotherapy resistance requires a deeper understanding of how AM differs from CM. Single-cell technologies has recently advanced our understanding of this question: MPZ+ subgroup contributes to major differences in tumor cell composition between AM and CM [9], and AM exhibits a more pronounced immunosuppressive TME than CM, characterized by reduced infiltration of effector CD8+ T cells, natural killer (NK) cells and γδ T cells, along with enhanced CD8+ T cells exhaustion [10, 11]. However, existing studies have predominantly focused on tumor-intrinsic features or immune cells, leaving the heterogeneity of stromal components, such as fibroblasts, endothelial cells, pericytes, remains largely underexplored. This gap is critical, as existing evidence has clearly demonstrated that cancer-associated fibroblasts (CAFs) and abnormal vasculature not only regulate tumor growth and metastasis but also directly contribute to immunotherapy resistance by impairing immune cell infiltration and function [12, 13]. Unfortunately, prior single-cell analyses of AM have been limited by small sample sizes (<10 AM cases), hindering a comprehensive assessment of stromal cells and their contribution to TME dysfunction. To address these limitations, we integrated multiple public single-cell RNA-sequencing datasets (31 AM, 13 CM, and 35 normal skin samples) supplemented with spatial transcriptomics to validate cell co-localization, thus providing a comprehensive melanoma atlas that enabled us to systematically profile TME heterogeneity. Distinct from the previous studies, our study specifically targeted an underexplored question: how subtype-specific stromal TME regulate immune suppression in AM. We identified a collagen-secreting pericyte subset enriched in AM, which correlates with abnormal vasculature, hypoxia, and immune exclusion. We further uncovered a hypoxia-related NECTIN2-TIGIT immunosuppressive axis. These findings provide valuable insights into the potential mechanisms underlying AM’s poor immunotherapy response and lay a foundation for developing novel combination therapeutic strategies targeting stromal-immune interactions. Methods Data collection To comprehensively investigate tumor microenvironmental differences between AM and CM, we systematically collected and integrated published single-cell RNA-seq datasets from the GEO database and public resources. The following predefined inclusion/exclusion criteria to filter datasets and individual samples: (1) Clear anatomical annotation: Only samples with unambiguous annotation of acral or non-acral skin sites were included. (2) Primary tumor restriction: For tumor samples, only primary cutaneous lesions were retained; lymph node and distant metastatic samples were excluded. (3) Treatment‑naïve status: Samples from patients who had received immunotherapy prior to surgery were excluded. Specifically, GSE189889 and GSE215121 contained AM samples, whereas GSE215121 and GSE277165 included CM samples. Melanoma samples from a pan-cancer single-cell study [14] of tumor vasculature (DOI: 10.5281/zenodo.10651059) were also incorporated to better characterize the stromal microenvironment. Normal skin data were obtained from GSE165816, GSE130973, GSE179162, and GSE202352. scRNA-seq data quality control and analysis The scRNA-seq data were processed and analyzed using the R package Seurat (v 4.4.0). Low-quality cells and potential doublets were filtered out using the following criteria: 200 < nFeature_RNA <5000, 500 < nCount_RNA <40,000, and mitochondrial content < 20%. After quality control, a total of 309,951 high-quality single cells were retained for downstream analyses. Next, the data were normalized, batch effects removed, and clustered. Transcript counts were log-normalized using the “NormalizeData” function, and the top 3000 highly variable genes were identified using the “FindVariableGenes” function. Data scaling was then performed with the “ScaleData” function and used for PCA. Data integration and batch effect correction across datasets and samples were performed using the “RunHarmony” function from R package harmony (v 1.2.3), with the following parameters: theta = 2, lambda = 1, and sigma = 0.05. Then clustering and distribution of single cells was mapped using uniform manifold approximation and projection (UMAP). Cell clusters were annotated mainly based on the expression pattern of canonical markers for major cell types and well-known or newly identified markers for cell subsets, supported by literature and the CellMarker 2.0 database. For single-cell subclustering analysis, a similar procedure was applied. Besides, differentially expressed genes (DEGs) were identified using the “FindAllMarker” function with parameters set to min.pct = 0.25 and logfc.threshold > 0.25. Tissue similarity evaluation of cell types The Bhattacharyya distance was calculated to evaluate the similarity between AM and CM for each cell type, with reference to the open-source GitHub code (https://github.com/Chuxj/CRC-atlas). Tissue preference evaluation of cell subsets To assess the subtype preference of cell subsets in melanoma, the ratio of observed to expected (Ro/e) values was calculated using the STARTRAC package, where higher Ro/e values indicate a stronger enrichment of a given cluster within a specific subtype [15]. Gene functional enrichment analysis We conducted single-cell enrichment analysis for each cell using R package AUCell with default parameters (v 1.28.0) [16], and GSEA analysis was performed using the R package clusterProfiler with default parameters (v 4.14.6) [17]. Reference genesets were download at MsigDB (https://www.gsea-msigdb.org/gsea/msigdb/human/collections.jsp). Cell-cell interaction network analysis We assessed potential cell-cell interactions based on CellChat (v 1.6.1) pipeline, with default parameters and threshold [18]. CellChat objects were first created separately for AM and CM samples using the “createCellChat” function. Subsequently, a comparative analysis between AM and CM CellChat objects was conducted to identify significant differences in signaling pathways and ligand-receptor interactions across cell types. The CellChatDB.human dataset from the CellChatDB database was used as the reference for intercellular communication. Spatial transcriptome analysis To validate the findings from single-cell transcriptomic analysis, we collected a publicly available spatial transcriptomics dataset of cutaneous melanoma published by Love et al. (10.5061/dryad.ksn02v7b1 [19]. The spatial transcriptomic data were analyzed and annotated using Seurat, following the single-cell analysis workflow described above. The spatial localization information (X/Y coordinates) of the spots was obtained from the dataset provided by the authors at the link mentioned above. To define the cellular neighborhood of a given spot, we calculated the Euclidean distance between its coordinates and those of all other spots. The 20 spatially nearest spots (excluding the spot itself) were then selected as its neighbors. For each pericyte spot, the composition of its neighborhood was summarized by calculating the proportion of each cell type among the 20 neighboring spots. Proportions were computed as the count of neighbors annotated as a given cell type divided by the total number of neighbors (20). To compare the neighborhood composition between the two pericyte subsets (COL3A1+ vs. MHY11+), we performed two-sided Wilcoxon rank-sum tests for each cell type proportion across all pericyte spots. Bulk RNA-seq validation Bulk RNA-seq data containing both AM and CM samples were obtained from the GSE190113 dataset. Using the R package BayesPrism with default parameters (v 2.2.2) [20], we calculated the relative proportion of cell types of bulk RNA-seq samples by incorporating single-cell annotations as prior information, and compared the proportions of the cell subsets of interest between AM and CM. Survival analysis For the Kaplan-Meier analysis, we grouped patients collected from Gide et al.’s cohort based on the expression of genes using the “surv_cutpoint” function in the survminer R package (v.0.5.0). The “ggsurvplot” function was then applied for survival curve visualization and comparison. Cell culture A375 and SK-MEL-28 cells, received as gifts from the Xiangya Hospital, Hospital for Skin Diseases (Institute of Dermatology), were authenticated by STR profiling and confirmed to be free of mycoplasma contamination prior to use. The cells were maintained in Dulbecco’s modified Eagle’s medium (DMEM, Gibco, C11995500BT) supplemented with 10% fetal bovine serum (FBS, Gibco, 10,099–141) and 1% penicillin/streptomycin (Gibco, 15,140–122) at 37 °C under 5% CO2. For hypoxic culture, cells were maintained in a hypoxia chamber with 1% O2, 5% CO2, and 94% N2 placed in a 37 °C incubator for the indicated times. Real-time quantitative PCR (qPCR) Total RNA was extracted using FreeZol reagent (Vazyme, R711-02). Subsequently, 1 μg of RNA was quantified by NanoDrop and reverse transcribed using ReverTra AceTM qPCR RT Master Mix (Toyobo, FSQ-201). The qPCR analysis was performed with specific oligonucleotides as detailed below, employing Taq Pro Universal SYBR mix (Vazyme, Q712-03) on the Bio-Rad CFX96TM Real-time System. F-NECTIN2 (5’-3’): GAGCACAGCCCACTCAAGAC, R-NECTIN2 (5’-3’): AAGGTGGGCAGCTCATGGTA; F-PGK1 (5’-3’): GAACAAGGTTAAAGCCGAGCC, R-PGK1 (5’-3’): GTGGCAGATTGACTCCTACCA; F-ACTB (5’-3’): CACCATTGGCAATGAGCGGTTC, R-ACTB (5’-3’): AGGTCTTTGCGGATGTCCACGT. Western blot Cells cultured in 6-cm dishes were lysed with SDT lysis buffer (0.1 M Tris-HCl, 0.1 M DTT, 4% SDS, and a protease inhibitor cocktail; Sigma, S8830-20TAB). Equal amounts of protein were then resuspended in SDS sample buffer and denatured at 95 °C. The proteins were separated by 10% SDS-PAGE and transferred onto 0.45-µm nitrocellulose membranes. The blots were blocked with 5% non-fat milk and probed with primary antibodies (Anti-Nectin-2, Proteintech, 84,460–5-RR; anti-HIF1α, Proteintech, 66,730–1-Ig; anti-GAPDH, HRP-60004). After secondary antibody (HRP-conjugated Goat Anti-Mouse IgG(H+L), Proteintech, SA00001-1; HRP-conjugated Goat Anti-Rabbit IgG(H+L), Proteintech, SA00001-2) incubation, proteins were detected using an ECL chemiluminescent detection kit (NCM, P10300) and visualized on a GE ImageQuant 800 imaging system. Statistical analysis All data analyses were conducted using the R software (v.4.4.2). Differences between two groups were assessed using either the Student’s t-test or Wilcoxon rank-sum test, depending on data distribution. Correlations between continuous variables were assessed using Spearman’s rank correlation test. Statistical significance was defined as a P value less than 0.05. *for p < 0.05, ** for p < 0.01, *** for p < 0.001, **** for p < 0.0001. Results Comprehensive single-cell atlas construction and basal tumor microenvironment feature analysis of AM and CM To systematically dissect the similarities and differences in cellular composition and molecular characteristics between AM and CM, we collected published single-cell RNA-sequencing data of AM and CM samples [10, 11, 14, 21], as well as normal acral and non-acral skin tissues [22–25], comprising a total of 82 samples from 75 donors (Fig. 1A, Fig. S1A and Table S1). After quality control, a total of 309,951 cells were obtained and clustered by cell types and pathological status rather than batch, indicating minimal batch effects (Fig. 1B). According to different gene expression patterns and characteristic markers, we identified 9 major cell clusters using UMAP based clustering, including melanocyte cells (MLANA), T/NK cells (CD2), B/Plasma cells (CD79A), monocytes/macrophages/dendritic cells (CD68), granulocytes/mast cells (TPSB2), keratinocytes (KRT5), fibroblasts (DCN), endothelial cells (PECAM1), and pericytes/smooth muscle cells (SMCs) (ACTA2) (Fig. 1B–C).Fig. 1Single-cell transcriptomic landscapes of AM, CM, and normal skin tissues. A Workflow of this study. B UMAP plot showing the nine major cell types, dataset sources, and tissue origins of cells (n = 309,951 cells). C Bubble diagram showing the expression of marker genes for different cell types. Bubble size indicates expression proportion, while the color gradient from blue to red signifies higher expression levels. D Heatmap showing differentially expressed genes (DEGs) between the two melanoma subtypes and their corresponding normal tissues. UMAP, uniform manifold approximation and projection; NK, natural killer; DC, dendritic cell; Mela, melanocyte; Mye, myeloid, including monocyte, macrophage, DC. Fibro, fibroblast; Endo, endothelial cell; Mural, mural cell, including pericyte and smooth muscle cell (SMC) Single-cell transcriptomic landscapes of AM, CM, and normal skin tissues. A Workflow of this study. B UMAP plot showing the nine major cell types, dataset sources, and tissue origins of cells (n = 309,951 cells). C Bubble diagram showing the expression of marker genes for different cell types. Bubble size indicates expression proportion, while the color gradient from blue to red signifies higher expression levels. D Heatmap showing differentially expressed genes (DEGs) between the two melanoma subtypes and their corresponding normal tissues. UMAP, uniform manifold approximation and projection; NK, natural killer; DC, dendritic cell; Mela, melanocyte; Mye, myeloid, including monocyte, macrophage, DC. Fibro, fibroblast; Endo, endothelial cell; Mural, mural cell, including pericyte and smooth muscle cell (SMC) Cell proportion analysis revealed that immune cells, including T/NK cells, B/Plasma cells, and myeloid cells, accounted for a significantly higher proportion in tumor tissues (Fig. S1B, C), consistent with the activation of immune responses within the tumor. Differential expression analysis between the two melanoma subtypes and their respective normal tissue controls identified both shared and subtype-specific alterations (Fig. 1D). Functional enrichment analysis of genes upregulated in both AM and CM showed that melanoma cells upregulated pathways related to cell cycle, translation, and cytokine signaling; T/NK cells exhibited elevated interferon signaling and cytotoxicity, with increased expression of GZMA and GZMM; myeloid cells showed enhanced antigen presentation and interferon response; fibroblasts significantly upregulated collagen metabolism and degradation pathways and molecules like COL6A3, POSTN, MMP19; endothelial cells exhibited marked activation of vasculature development and inflammatory response; pericytes/SMCs showed upregulation of extracellular matrix (ECM) and TNF-α signaling pathways (Fig. S1D). These findings suggest that while AM and CM share some common features, they also display subtype-specific microenvironmental characteristics, which we explored further in subsequent analyses. AM microenvironment core traits hypoxia enhancement links to NECTIN2-TIGIT immunosuppressive axis activation Based on the constructed cellular atlas, we further investigated the key differences in the TME components and intercellular signals between AM and CM. Using the Bhattacharyya distance [15] to assess the transcriptional similarity of major cell types, the results showed that pericytes/SMCs exhibited the most significant discrepancy between AM and CM, followed by B/Plasma cells, endothelial cells, etc. (Fig. S2B). Gene set enrichment analysis (GSEA) revealed that pericytes/SMCsin AM were significantly enriched in assembly of collagen fibrils and ECM receptor interaction, while pericytes/SMCs in CM were enriched in oxidative phosphorylation and myoblast proliferation. Notably, the hypoxia pathway was significantly upregulated in all major cell types in AM compared to CM (Fig. 2A and Fig. S2A). Key glycolytic enzymes such as ALDOA and LDHA were also upregulated in AM (Fig. 2B), indicating that AM generally exhibits a more hypoxic TME.Fig. 2Differences in tumor microenvironment and cell-cell interaction between AM and CM. A Violin plots comparing hypoxia and glycolysis AUCell scores of major cell types between AM and CM. B Differential gene expression analysis showing genes upregulated in AM and CM across major cell types. Genes upregulated in AM are shown in red, those upregulated in CM are shown in blue, and glycolysis-related genes are highlighted in green. C Scatter plots illustrating the incoming and outcoming interaction strength of the different cell types in AM and CM (melanoma cells not shown). D Ranking of significant signaling pathways based on differences in overall information flow within the inferred networks between AM and CM. Signaling pathways colored red are enriched in AM and colored blue are enriched in CM. E Representative ligand-receptor pairs with significant differences between AM and CM. Dot color reflects communication probabilities and dot size represents computed p-values. F Violin plots showing the expression differences of selected receptor and ligand molecules between AM and CM across specific cell types. G Scatter plots showing that NECTIN2 expression is significantly positively correlated with hypoxia and glycolysis ssGSEA scores in TCGA-SKCM dataset (left) and Farshidfar et al.’s dataset (right). Data were analyzed using a Spearman’s correlation test. H Effect of hypoxia on NECTIN2 expression: A375 and SK-MEL-28 cells were cultured under normoxic or 1% O2 conditions for 24 h. Protein levels of NECTIN2 and HIF1α were detected by Western blot, with GAPDH as the loading control. I mRNA expression of NECTIN2 and PGK1 (a HIF1α target gene) evaluated by qPCR. Data are presented as mean ± SD; statistical significance was determined by Student’s t-test. *for p < 0.05, ** for p < 0.01, *** for p < 0.001, **** for p < 0.0001. cDC, conventional DC; pDC, plasmacytoid DC; Mono, monocyte; macro, macrophage; VasEndo, vascular endothelial cell; LymphEndo, lymphatic endothelial cell Differences in tumor microenvironment and cell-cell interaction between AM and CM. A Violin plots comparing hypoxia and glycolysis AUCell scores of major cell types between AM and CM. B Differential gene expression analysis showing genes upregulated in AM and CM across major cell types. Genes upregulated in AM are shown in red, those upregulated in CM are shown in blue, and glycolysis-related genes are highlighted in green. C Scatter plots illustrating the incoming and outcoming interaction strength of the different cell types in AM and CM (melanoma cells not shown). D Ranking of significant signaling pathways based on differences in overall information flow within the inferred networks between AM and CM. Signaling pathways colored red are enriched in AM and colored blue are enriched in CM. E Representative ligand-receptor pairs with significant differences between AM and CM. Dot color reflects communication probabilities and dot size represents computed p-values. F Violin plots showing the expression differences of selected receptor and ligand molecules between AM and CM across specific cell types. G Scatter plots showing that NECTIN2 expression is significantly positively correlated with hypoxia and glycolysis ssGSEA scores in TCGA-SKCM dataset (left) and Farshidfar et al.’s dataset (right). Data were analyzed using a Spearman’s correlation test. H Effect of hypoxia on NECTIN2 expression: A375 and SK-MEL-28 cells were cultured under normoxic or 1% O2 conditions for 24 h. Protein levels of NECTIN2 and HIF1α were detected by Western blot, with GAPDH as the loading control. I mRNA expression of NECTIN2 and PGK1 (a HIF1α target gene) evaluated by qPCR. Data are presented as mean ± SD; statistical significance was determined by Student’s t-test. *for p < 0.05, ** for p < 0.01, *** for p < 0.001, **** for p < 0.0001. cDC, conventional DC; pDC, plasmacytoid DC; Mono, monocyte; macro, macrophage; VasEndo, vascular endothelial cell; LymphEndo, lymphatic endothelial cell Using CellChat [18], we then constructed intercellular communication networks and found that the incoming and outgoing interaction strength of vascular endothelial cells and pericytes/SMCs were enhanced in AM (Fig. 2C and Fig. S2C). Furthermore, ligand-receptor interactions such as NECTIN2-TIGIT (NECTIN), ANXA1–FPR1 (ANNEXIN), and TNFSF14-LTBR (LIGHT) were enriched in AM, whereas the IFNG-(IFNGR1 + IFNGR2) (IFN-II) interaction was enriched in CM (Fig. 2D, E). Further analysis revealed that IFNG-(IFNGR1 + IFNGR2) interaction enrichment in CM was mainly due to high IFNG expression in NK cells (Fig. 2E, F). IFNG secretion by NK cells has been reported to play a key role in anti-tumor immunity, including directly enhancing the killing function of NK cells and activating CD8+ T cells [26]. Importantly, TIGIT (also called T cell immunoreceptor with Ig and ITIM domains), an emerging immune checkpoint, was significantly upregulated in T cells in AM, and its ligand NECTIN2 (also called Nectin cell adhesion molecule 2) was widely expressed across melanoma cells, endothelial cells, pericytes, and myeloid cells, with significantly higher expression in AM (Fig. 2F). Combined with the finding that the hypoxia pathway was also broadly upregulated across major cell types in AM, we next examined the relationship between hypoxia and NECTIN2 expression. Stratifying cells by hypoxia AUCell score, the expression level of NECTIN2 in the high-hypoxia group was significantly higher (Fig. S2E). Two independent melanoma bulk transcriptomic datasets from TCGA [27] and Farshidfar et al.‘s dataset [28] also demonstrated that NECTIN2 expression was significantly positively correlated with the enrichment scores of hypoxia and glycolysis (Fig. 2G). To experimentally validate this correlation and clarify the potential regulation effect of hypoxia on NECTIN2, we simulated a hypoxic TME by culturing cells under 1% O2. Hypoxia induction was confirmed by upregulated HIF1α and its downstream target gene PGK1. Consistent with our bioinformatic findings, melanoma cell lines A375 and SK-MEL-28 exhibited increased levels of NECTIN2 mRNA and protein level under hypoxic conditions compared to normoxic conditions (Fig. 2H, I). Collectively, these results demonstrate that hypoxia positively regulates NECTIN2 expression. NECTIN2, a member of the nectin family, binds to TIGIT, exerting immunosuppressive effects within the TME [29]. Previous studies have elucidated that NECTIN2 is also expressed on endothelial cells, fibroblasts, macrophages, etc., in the neuroblastoma TME, and its binding with TIGIT leads to T cell dysfunction [30]. Therefore, we speculate that hypoxia-induced upregulation of NECTIN2 expression may contribute to T cell dysfunction via the NECTIN2-TIGIT axis, representing a potential mechanism of immunosuppression in AM. Immune cell heterogeneity analysis reveals functional differences in T/NK/DC cells and their link to immune suppression in AM and CM To further dissect the cellular subset basis of immune suppression in AM, we reclustered 48,034 lymphocytes and identified 13 distinct subtypes, including: B cells (MS4A1), Plasma cells (JCHAIN), KLRF1+ NK cells, XCL1+ NK cells, cytotoxic CD8+ T cells (GZMK), exhausted CD8+ T cells (HAVCR2), stressed CD8+ T cells (Heat shock proteins), IFIT1+ T cells, Tregs (FOXP3), CXCL13+ CD4+ T cells, ANXA1+ tissue-resident T cells, and TXNIP+ tissue-resident T cells (Fig. 3A). Ro/e analysis revealed that stressed CD8+ T cells, IFIT1+ T cells, and XCL1+ NK cells were significantly enriched in AM, while KLRF1+ NK cells and TXNIP+ tissue-resident T cells were more abundant in CM (Fig. 3B). Functional enrichment analysis showed that stressed CD8+ T cells exhibited upregulation of hypoxia, glycolysis, cell cycle, and unfolded protein response (UPR) pathways, indicating that they were subjected to significant external stress, which is consistent with the more hypoxic TME in AM (Fig. 3C). Despite showing high cytotoxic scores and strong GZMA/GZMK expression (Fig. 3D and Fig. S3A), stressed CD8+ T cells displayed intermediate exhaustion markers higher than cytotoxic CD8+ T cells but lower than exhausted CD8+ T cells (Fig. 3D and Fig. S4A), suggesting they may represent an early stress-response state preceding exhaustion. This subset resembles the stress-response state (STR) T cells previously defined by Chu et al., which were characterized by high expression of heat shock genes and associated with immunotherapy resistance across cancers [31]. Similarly, IFIT1+ T cells showed strong interferon-stimulated gene (ISG) signatures and IFN-α pathway enrichment, resembling an ISG-high T cell subset defined by Chu et al. linked to metastasis and poor prognosis [31]. Although similar subsets have been reported in the pan-cancer scRNA-seq analysis of T cells, our work is the first to highlight the preferential enrichment of stressed CD8+ T cells and IFIT1+ T cells in AM compared to CM. These findings underscore that poor prognosis and immune suppression in AM may involve not only classical T cell exhaustion but also additional dysfunctional states shaped by stresses—particularly hypoxia and aberrant interferon signaling.Fig. 3Characterization of immune cell differences between AM and CM. A UMAP plot showing the landscape of lymphocytes (n = 48,034 cells). B Heatmap showing the tissue preference of each lymphocyte subset as indicated by the Ro/e index. C Heatmap showing the average AUCell scores of the hallmark pathways across T and NK cell subsets. D Violin plots showing the AUCell scores for cytotoxicity, exhaustion, and stress-related gene signatures among T cell subsets. E Lollipop plot illustrating enriched pathways in KLRF1+ NK cells and XCL1+ NK cells. Bubble size indicates significance, and the color gradient represents enrichment score (NES). F Volcano plot showing differentially expressed genes between KLRF1+ NK cells and XCL1+ NK cells. Genes upregulated in KLRF1+ NK cells are shown in purple, and those upregulated in XCL1+ NK cells are shown in blue. G UMAP plot showing the landscape of myeloid cells (n = 17,806 cells). H Heatmap showing the tissue preference of each myeloid cell subset as indicated by the Ro/e index. I Heatmap showing the average AUCell scores of the hallmark pathways across myeloid cell subsets. J Violin plots showing the antigen processing and presentation, Treg and Th2 traffic AUCell scores among DC subsets. K Bubble diagram showing the genes expression for DC subsets. Bubble size reflects expression proportion, while the color gradient from blue to red signifies higher expression levels Characterization of immune cell differences between AM and CM. A UMAP plot showing the landscape of lymphocytes (n = 48,034 cells). B Heatmap showing the tissue preference of each lymphocyte subset as indicated by the Ro/e index. C Heatmap showing the average AUCell scores of the hallmark pathways across T and NK cell subsets. D Violin plots showing the AUCell scores for cytotoxicity, exhaustion, and stress-related gene signatures among T cell subsets. E Lollipop plot illustrating enriched pathways in KLRF1+ NK cells and XCL1+ NK cells. Bubble size indicates significance, and the color gradient represents enrichment score (NES). F Volcano plot showing differentially expressed genes between KLRF1+ NK cells and XCL1+ NK cells. Genes upregulated in KLRF1+ NK cells are shown in purple, and those upregulated in XCL1+ NK cells are shown in blue. G UMAP plot showing the landscape of myeloid cells (n = 17,806 cells). H Heatmap showing the tissue preference of each myeloid cell subset as indicated by the Ro/e index. I Heatmap showing the average AUCell scores of the hallmark pathways across myeloid cell subsets. J Violin plots showing the antigen processing and presentation, Treg and Th2 traffic AUCell scores among DC subsets. K Bubble diagram showing the genes expression for DC subsets. Bubble size reflects expression proportion, while the color gradient from blue to red signifies higher expression levels The proportions of NK cell subsets also differed between AM and CM. KLRF1+ NK cells were more enriched in CM, while XCL1+ NK cells had a higher proportion in AM (Fig. 3B). KLRF1+ NK cells showed higher expression of cytotoxic molecules (GZMA, GZMB, PRF1) and IFNG, aligning with stronger IFN-II signaling in CM (Fig. 3E, F). They were also enriched in effector cell trafficking and lymphocyte activation pathways and expressed high levels of chemokines CCL4 and CCL5 (Fig. 3E, F). In contrast, XCL1+ NK cells were more enriched in IFN-I production, hypoxia, stress response, IL-35 signaling, and TGF-β pathways (Fig. 3E). IL-35 was recently discovered to inhibit the cytotoxicity and cytokine secretion of NK cells, triggering TGF-β autocrine and driving NK cell dysfunction [32]. Consistently, Tang et al. reported that NK cells with elevated stress responses exhibit reduced cytotoxicity and impaired anti-tumor immunity [33], suggesting the immunosuppressive NK cell component in AM. Additionally, we found that B cells were significantly enriched in CM, while AM contained more plasma cells (Fig. S4B). Similarly, myeloid cells were further subdivided into 12 subsets, with LAMP3+ DCs enriched in AM and CLEC9A+ DCs enriched in CM (Fig. 3G, H). Functional analysis revealed that CLEC9A+ DCs had the strongest antigen-presenting capacity (Fig. 3J), while LAMP3+ DCs displayed decreased antigen presentation capacity with markedly decreased expression of MHC and other antigen presentation-related molecules (Fig. 3K). Conversely, LAMP3+ DCs highly expressed CCL17 and CCL22, facilitating Treg and Th2 recruitment (Fig. 3J, K), consistent with previous reports that LAMP3+ DCs can promote immune escape by recruiting Tregs [34, 35]. LAMP3+ DCs might also inhibit NK cell function [33], aligning with the finding that AM was enriched in XCL1+ NK cells with low cytotoxicity. Correlation analysis also revealed a positive, though nonsignificant, association between LAMP3+ DCs and Treg, and a significant negative correlation with cytotoxic CD8+ T cells (Fig. S4C). Notably, LAMP3+ DCs also expressed high levels of CD274 (PD-L1) and NECTIN2 (Fig. 3K). All the above findings highlight that LAMP3+ DCs play a critical role in promoting an immunosuppressive TME in AM. Overall, comprehensive analysis of the immune microenvironment revealed that, compared to CM, AM displays a more immunosuppressive state in terms of T cells, NK cells, and DCs. Stromal cell analysis reveals enrichment of collagen-secreting pericytes and abnormal vasculature in AM After clarifying the differences in immune cell subsets between AM and CM, we next focused on stromal cells, a previously understudied component of the TME. A total of 80,423 stromal cells were subjected to further classification and annotation (Fig. 4A), identifying six fibroblast subsets, two pericyte subsets (COL3A1+ pericytes and RERGL+ pericytes), smooth muscle cells, lymphatic endothelial cells, and two vascular endothelial cell subsets (PODXL+ VasEndo and ACKR1+ VasEndo cells). Ro/e analysis showed pronounced differences in vascular cells between AM and CM. COL3A1+ pericytes and both vascular endothelial cell subsets were enriched in AM, whereas RERGL+ pericytes and smooth muscle cells were enriched in CM (Fig. 4B). Moreover, RERGL+ pericytes and smooth muscle cells were similar in molecular characteristics (Fig. 4Cand Fig. S4D).Fig. 4Characterization of stromal cell differences between AM and CM. A UMAP plot showing the landscape of stromal cell (n = 80,423 cells). B Heatmap showing the tissue preference of each stromal subset as indicated by the Ro/e index. C Heatmap showing the average AUCell scores of the hallmark pathways across stromal cell subsets. D Volcano plot showing differentially expressed genes between COL3A1+ pericytes and RERGL+ pericytes. Genes upregulated in COL3A1+ pericytes are shown in purple, and those upregulated in RERGL+ pericytes are shown in blue. E GSEA plots showing positively and negatively enriched pathways in COL3A1+ pericytes versus RERGL+ pericytes. F Lollipop plot illustrating enriched pathways in PODXL+ VasEndo and ACKR1+ VasEndo. Bubble size indicates significance, and the color gradient represents enrichment scores. G Violin plots displaying the expression of functional genes in PODXL + VasEndo and ACKR1+ VasEndo Characterization of stromal cell differences between AM and CM. A UMAP plot showing the landscape of stromal cell (n = 80,423 cells). B Heatmap showing the tissue preference of each stromal subset as indicated by the Ro/e index. C Heatmap showing the average AUCell scores of the hallmark pathways across stromal cell subsets. D Volcano plot showing differentially expressed genes between COL3A1+ pericytes and RERGL+ pericytes. Genes upregulated in COL3A1+ pericytes are shown in purple, and those upregulated in RERGL+ pericytes are shown in blue. E GSEA plots showing positively and negatively enriched pathways in COL3A1+ pericytes versus RERGL+ pericytes. F Lollipop plot illustrating enriched pathways in PODXL+ VasEndo and ACKR1+ VasEndo. Bubble size indicates significance, and the color gradient represents enrichment scores. G Violin plots displaying the expression of functional genes in PODXL + VasEndo and ACKR1+ VasEndo Differential expression and GSEA analyses between the two pericyte subsets revealed that COL3A1+ pericytes were significantly enriched in collagen-related pathways, angiogenesis, and the complement and coagulation cascades, and showed high expression of COL3A1, COL4A1, RGS5, and ANGPT2 (Fig. 4D, E). Among these, ANGPT2 is a marker promoting pericyte detachment from the vascular endothelium [36], while RGS5 (Regulator of G protein signaling 5, a regulator of G protein activity) and increased collagen synthesis are reported to indicate a less differentiated pericyte phenotype [37]. RERGL+ pericytes were enriched in muscle contraction pathways, showed upregulation of contractile markers CNN1 and ACTG2, and exhibited enhanced oxidative phosphorylation (Fig. 4D, E). Functionally, mature contractile pericytes promote vascular normalization and reduction of hypoxia [38], consistent with our observation that AM was more hypoxic, while CM displayed stronger oxidative phosphorylation activity. Further analysis of the two vascular endothelial cell subsets revealed that PODXL+ VasEndo cells, which were relatively more abundant in AM, showed marked upregulation of pathways associated with endothelial cell proliferation, neovascularization, NOTCH signaling (NOTCH1, NOTCH4, JAG1, PDGFB), as well as collagen degradation and basement membrane organization (MMP2, COL4A1, COL4A2) (Fig. 4F, G and Fig. S4E). In contrast, ACKR1+ VasEndo cells enhanced pathways and genes related to antigen processing and presentation (CD74, HLA-DRA, HLA-DRB1), interferon-γ response, immune cell adhesion (ICAM1, VCAM1, SELP, SELE), and cell chemotaxis (CCL2, CCL14) (Fig. 4F, G). Tumor vascular normalization is characterized by reduced angiogenesis, decreased cell proliferation, elevated adhesion molecule expression, and upregulation of antigen processing and presentation [12, 39, 40]. Moreover, endothelial cells play an important role in regulating immune cell trafficking. The extravasation of immune cells from the circulation into tumor tissues is mainly achieved through paracellular diapedesis or transendothelial migration, and downregulation of key adhesion genes, including ICAM1, VCAM1, SELP, SELE, etc., impairs immune cell extravasation and recruitment to the tumor mass [12]. Collectively, these findings indicate active angiogenesis in AM, but the lack of mature pericytes coverage likely leads to vascular abnormality and defective endothelial-mediated immune cells trafficking, ultimately impairing anti-tumor immunity in AM. Spatial transcriptomics validates colocalization of abnormal vasculature and low immune infiltration Based on the above results, we further explored the association between vascular cells and immune cells. Correlation analysis revealed that COL3A1+ pericytes were positively correlated with PODXL+ VasEndo cells but negatively correlated with cytotoxic CD8+ T cells (Fig. 5A, B). RERGL+ pericytes correlated positively with CCL19+ fibroblasts, Langerhans cells, and CLEC9A+ DCs, and ACKR1+ VasEndo cells also showed a positive association with Langerhans cells (Fig. 5A, B and Fig. S5A). Given that Langerhans cells serve as one of the main antigen-presenting cells in melanoma [41] and CCL19+ fibroblasts promote antitumor immunity [42], these results suggest that RERGL+ pericytes and ACKR1+ VasEndo cells, both indicative of vascular normalization, may facilitate the recruitment of anti-tumor immune cells, whereas COL3A1+ pericytes, associated with abnormal vasculature, may impair the infiltration of cytotoxic CD8+ T cells.Fig. 5Spatial transcriptomic validation of the role of pericyte subtypes in the immune microenvironment. A Bar plots showing correlation coefficients between the proportion of COL3A1+ pericytes (left) and RERGL+ pericytes (right) and other tumor microenvironment cell subsets in melanoma. Significant correlations are colored red for positive correlation and blue for negative correlation. Data were analyzed using a Spearman’s correlation test. B Scatter plots showing correlations between the proportions of two pericyte subsets and selected cell types. C UMAP plot showing the two pericyte subsets of cells (n = 2,462 cells). D Violin plots displaying the expression of marker genes in COL3A1+ pericytes and MYH11+ pericytes. E UMAP plot showing the two vascular endothelial cell subsets of cells (n = 4,925 cells). F Violin plots displaying the expression of marker genes in ACKR1+ VasEndo and KDR+ VasEndo. G violin plots showing the differences in neighboring cell proportions between the two pericyte subsets based on spatial analysis. H Representative FOV (field of view) showing the colocalization of COL3A1+ pericytes and ACKR1+ VasEndo. I Representative FOV (field of view) showing increased cytotoxic CD8+ T-cell infiltration around MYH11+ pericytes relative to COL3A1+ pericytes Spatial transcriptomic validation of the role of pericyte subtypes in the immune microenvironment. A Bar plots showing correlation coefficients between the proportion of COL3A1+ pericytes (left) and RERGL+ pericytes (right) and other tumor microenvironment cell subsets in melanoma. Significant correlations are colored red for positive correlation and blue for negative correlation. Data were analyzed using a Spearman’s correlation test. B Scatter plots showing correlations between the proportions of two pericyte subsets and selected cell types. C UMAP plot showing the two pericyte subsets of cells (n = 2,462 cells). D Violin plots displaying the expression of marker genes in COL3A1+ pericytes and MYH11+ pericytes. E UMAP plot showing the two vascular endothelial cell subsets of cells (n = 4,925 cells). F Violin plots displaying the expression of marker genes in ACKR1+ VasEndo and KDR+ VasEndo. G violin plots showing the differences in neighboring cell proportions between the two pericyte subsets based on spatial analysis. H Representative FOV (field of view) showing the colocalization of COL3A1+ pericytes and ACKR1+ VasEndo. I Representative FOV (field of view) showing increased cytotoxic CD8+ T-cell infiltration around MYH11+ pericytes relative to COL3A1+ pericytes To validate our findings, we analyzed a publicly available spatial transcriptomics dataset from CM [43]. Through cell-type annotation, we identified in melanoma spatial transcriptomic data the same stromal and immune subsets as in single-cell analysis, including collagen-secreting pericytes (RGS5+/COL3A1+), mature pericytes (MYH11+ pericytes, analogous to the RERGL+ subset identified in scRNA-seq), angiogenesis-related endothelial cells (KDR+ VasEndo cells, similar to PODXL+ subset identified in scRNA-seq), immune cell trafficking-related endothelial cell (ACKR1+ VasEndo cells), as well as CCL19+ fibroblasts and LAMP3+ DCs (Fig. 5C–F, and Fig. S5B, C). These findings confirmed the accuracy and reproducibility of our prior single-cell annotations. By assessing the neighboring cell types of the two pericyte subsets, we found results consistent with those observed in the scRNA-seq data: RGS5+/COL3A1+ pericytes mainly colocalized with KDR+ VasEndo cells, whereas MYH11+ pericytes were mainly adjacent to ACKR1+ VasEndo cells (Fig. 5G–I and Fig. S5D, E). Moreover, the number of infiltrating cytotoxic CD8+ T cells and CCL19+ fibroblasts around MYH11+ pericytes was significantly greater than around RGS5+/COL3A1+ pericytes (Fig. 5G–I and Fig. S5D, E). These findings suggest that the collagen-secreting RGS5+/COL3A1+ pericytes and KDR+ VasEndo cells, which are more enriched in AM, can impair immune cell infiltration and anti-tumor immunity, and this may be one potential reason for the poorer efficacy of immunotherapy in AM compared to CM. To verify this hypothesis, we conducted subsequent analyses. Bulk transcriptomic deconvolution validates vasculature abnormality in AM and identifies RGS5+/COL3A1+ pericytes as potential targets for combined immunotherapy We performed deconvolution using BayesPrism [20] on a bulk RNA-seq dataset containing both AM and CM samples. Consistent with our findings in single-cell transcriptomics, RGS5+ /COL3A1+ pericytes and KDR+/PODXL+ VasEndo cells were significantly enriched in AM compared with CM, accompanied by higher expression of the markers RGS5, KDR, and PODXL (Fig. 6A and Fig. S6B; marker identification workflow in Fig. S6A). RERGL+/MYH11+ pericytes and ACKR1+ VasEndo cells showed no significant subtype differences (Fig. S6C). Bulk-level analysis further confirmed that the proportion of RGS5+/COL3A1+ pericytes was significantly positively correlated with KDR+/PODXL+ VasEndo cells, and the expression of RGS5, PODXL, and KDR correlated positively with HIF1A (Fig. 6B, C).Fig. 6External bulk RNA-seq datasets validating the findings from single-cell transcriptomic analysis. A Violin and box plots showing differences in COL3A1+ pericytes/PODXL+ VasEndo proportion and expression levels of their marker RGS5/KDR between AM and CM in the Farshidfar et al.’s dataset. B Scatter plots showing correlations between the proportions of COL3A1+ pericytes and PODXL+ VasEndo. C Scatter plots showing correlations between the expression of HIF1A and RGS5, PODXL, and KDR. D Kaplan-meier plots showing overall survival in CM patients receiving immunotherapy stratified based on the expression level of RGS5 (left) and KDR (right) genes. Data were analyzed using cox regression. E Violin plots showing the differences in RGS5 expression between clinical benefit and non-benefit groups in AM patients receiving immunotherapy. F Scatter plots showing correlations between the expression of RGS5 and CD8+ tumor-infiltrating lymphocytes (TIL) determined by immunohistochemistry (IHC). G Graphical abstract illustrating the distinct characteristics of the tumor microenvironment of AM and CM External bulk RNA-seq datasets validating the findings from single-cell transcriptomic analysis. A Violin and box plots showing differences in COL3A1+ pericytes/PODXL+ VasEndo proportion and expression levels of their marker RGS5/KDR between AM and CM in the Farshidfar et al.’s dataset. B Scatter plots showing correlations between the proportions of COL3A1+ pericytes and PODXL+ VasEndo. C Scatter plots showing correlations between the expression of HIF1A and RGS5, PODXL, and KDR. D Kaplan-meier plots showing overall survival in CM patients receiving immunotherapy stratified based on the expression level of RGS5 (left) and KDR (right) genes. Data were analyzed using cox regression. E Violin plots showing the differences in RGS5 expression between clinical benefit and non-benefit groups in AM patients receiving immunotherapy. F Scatter plots showing correlations between the expression of RGS5 and CD8+ tumor-infiltrating lymphocytes (TIL) determined by immunohistochemistry (IHC). G Graphical abstract illustrating the distinct characteristics of the tumor microenvironment of AM and CM We next assessed the association between the identified vascular cell subsets and ICB response in immunotherapy cohorts. In a cohort of CM patients treated with anti-PD-1 [44], high RGS5 and KDR expression was associated with poor survival, whereas high expression of RERGL and ACKR1 was linked to survival benefit (Fig. 6D and Fig. S6D). Similarly, in an AM cohort receiving immunotherapy [45], RGS5 expression was higher in patients without clinical benefit and negatively correlated with CD8+ tumor-infiltrating lymphocytes (Fig. 6E, F). These results suggest that RGS5+/COL3A1+ pericytes and KDR+/PODXL+ VasEndo cells, which are indicative of abnormal vasculature and are more enriched in AM, may impair the efficacy of immunotherapy (Fig. 6G). Moreover, markers of collagen-secreting pericytes and abnormal vasculature such as RGS5 may serve as potential biomarkers of immunotherapy resistance and targets for combination immunotherapy. However, the current analysis relies on limited validation cohorts. These findings require prospective validation in broader, independent cohorts to determine their potential clinical utility. Overall, our analysis provides important insights into the potential mechanisms underlying the poor efficacy of immunotherapy in AM. Additional experimental validations are needed to clarify the functions of key cell types and genes in tumor immunity. Discussion In this study, we integrated multiple published datasets to construct a comprehensive scRNA-seq atlas of AM and CM. This atlas allowed us to systematically delineate subtype-specific differences in melanoma TME cellular composition and molecular features, and identify novel potential stromal contributors of AM’s poor response to immunotherapy—addressing a critical gap in prior research focused primarily on immune cell heterogeneity or tumor-intrinsic traits. A key finding of our work is the pronounced hypoxic state in AM, in contrast to more active oxidative phosphorylation (OXPHOS) metabolism observed in CM. This aligns with established literature linking tumor hypoxic to immunosuppressive TME remodeling [46], and extends prior proteomic work by Michal et al., who demonstrated that OXPHOS enrichment correlates with better immunotherapy responses in advanced melanoma [47]. Our data thus reinforce the clinical relevance of metabolic differences between AM and CM. In the context of the hypoxic TME in AM, we observed coordinated upregulation of the NECTIN2-TIGIT axis. Hypoxia was associated with increased NECTIN2 expression and may act as an upstream factor contributing to activation of the NECTIN2-TIGIT pathway. The NECTIN2-TIGIT axis is a well-validated immunosuppressive pathway in cancers such as neuroblastoma, ovarian cancer, and leukemia, and combination therapy with anti-PD-1 and anti-TIGIT antibodies has shown enhanced efficacy [30, 48, 49]. Our study provides supportive evidence that hypoxia-associated NECTIN2 upregulation may contribute to immune suppression in AM. However, it should be noted that functional validation of the hypoxia-NECTIN2 axis was performed only in cutaneous melanoma cell lines. To confirm the robustness and generalizability of this mechanism, further validation in additional models (including acral melanoma cell lines and tumor organoids) and functional co-culture assays with T cells would be required. Tumor hypoxia is tightly coupled to vascular dysfunction [12, 50], and our analysis revealed that pericytes—key regulators of vascular stability and hemodynamics [51]—exhibited the most prominent transcriptomic differences between AM and CM. Specifically, AM was enriched in RGS5+/COL3A1+ pericytes, which spatially colocalized with KDR+/PODXL+ VasEndo cells displaying pro-angiogenic features but reduced immune adhesion molecules, suggesting an abnormal vascular niche associated with hypoxia in AM. This abnormal vascular niche may impair anti-tumor immunity: our spatial transcriptomics data confirmed reduced immune infiltration around such vessels, and high expression of RGS5+/COL3A1+ pericytes and KDR+/PODXL+ VasEndo cells markers (e.g., RGS5) correlated with poor immunotherapy response in external cohorts, suggesting their potential utility as biomarkers for patient stratification. Importantly, despite higher expression in AM, the prognostic value of RGS5 for immunotherapy response was validated in both AM and CM, implying its potential is not confined to the acral subtype. Clinically, these observations align with a recent trial (CAP 03-NEO) of AM patients treated with camrelizumab (anti-PD-1), apatinib (VEGFR-2 inhibitor), and temozolomide, which achieved a 66.7% ORR, substantially higher than the <20% ORR of anti-PD-1 monotherapy in AM [52]. This trial validates that vascular normalization therapy can enhance immunotherapy efficacy in AM. Notably, most current vascular normalization strategies primarily target endothelial cells, whereas emerging evidence suggests that modulating pericyte phenotypic states may achieve more durable vascular normalization [37, 38]. In this context, our study provides a rationale for further exploring RGS5+/COL3A1+ pericytes as a potential therapeutic target. Even so, whether RGS5+/COL3A1+ pericytes causally impair anti-tumor immunity and influence immunotherapy response remains to be further experimentally validated. Given these insights and limitations, further analyses and experiments are warranted. First, our analysis focused on TME heterogeneity (stromal and immune cells) but did not dissect crosstalk between these components and tumor cells—for example, how melanoma cell-derived factors (e.g., cytokines, growth factors) regulate pericyte differentiation or NECTIN2 expression should be investigated. Second, our conclusions are primarily based on bioinformatic integration of public datasets. Additional functional validation is needed to support the mechanistic links proposed in AM, such as in vitro co-culture assays to confirm hypoxia-related NECTIN2-TIGIT immunosuppressive axis activation and assess T-cell migration or activation in the presence of RGS5+/COL3A1+ pericytes, and in vivo models to test pericyte-targeting strategies. Third, while we leveraged spatial transcriptomics to validate cellular colocalization, future studies with multi-omics spatial profiling (e.g., spatial proteomics) could provide deeper insights into the molecular interactions between abnormal vasculature and immune cells, especially the mechanism by RGS5+/COL3A1+ pericytes or abnormal vasculature promoting immune exclusion. Finally, given that our validation analyses were conducted in limited external cohorts with modest sample sizes, subsequent investigations in larger, comprehensively annotated AM immunotherapy cohorts are warranted to further explore the predictive relevance of RGS5+/COL3A1+ pericytes and related markers (e.g., RGS5), while systematically considering additional influencing factors. Conclusions In conclusion, our work highlights a potential hypoxia-vascular-immune regulatory axis in AM: RGS5+/COL3A1+ pericyte-associated vascular abnormality may exacerbate hypoxia, which in turn activates the NECTIN2-TIGIT immunosuppressive axis, collectively contributing to AM’s poor immunotherapy response. These findings not only advance our understanding of AM-specific TME dysfunction but also identify RGS5+/COL3A1+ pericytes and the NECTIN2-TIGIT axis as promising targets for combination immunotherapy, with direct implications for improving clinical outcomes in AM patients. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary material 1 Supplementary material 2 Supplementary material 3 Supplementary material 4 Supplementary material 5 Supplementary material 6 Supplementary material 7 Supplementary material 8 Supplementary material 1 Supplementary material 2 Supplementary material 3 Supplementary material 4 Supplementary material 5 Supplementary material 6 Supplementary material 7 Supplementary material 8