Genotype-driven tumor ecosystems drive immune evasion and immunotherapy resistance in melanoma

Abstract

Background

Genetic and transcriptional alterations in cancer cells shape their interactions with immune and stromal compartments, thereby influencing tumor progression, immune evasion, and response to immune checkpoint inhibitors (ICIs). Yet, how these interactions are spatially organized within melanoma and how they relate to clinical outcome remains incompletely understood.

Methods

We used spatial multi-omic profiling on melanoma tissues from patients, compared those with loss-of-function mutations in the Neurofibromin 1 (NF1) tumor suppressor gene to those with intact NF1, and integrated these results with functional studies and subsequent gene expression changes observed in patient-derived melanoma models and an immunotherapy-resistant syngeneic mouse model.

Results

Spatial multi-omic analysis of melanoma tissues from patients identified 12 meta-niches composed of distinct cell populations with unique molecular features. Although all tumors contained these meta-niches, those enriched in immunosuppressive cancer-associated fibroblasts (CAFs) and macrophages were significantly more common in tumors with NF1 loss. Conversely, niches enriched in cytotoxic CD8⁺ T cells were significantly reduced. In both human and mouse melanoma, NF1 loss was associated with impaired antigen-presentation programs and reduced CD8⁺ T cell infiltration. Mechanistically, EGFR signaling emerged as a prominent feature of NF1-deficient melanoma ecosystems and correlated with reduced expression of antigen-presentation genes and T cell exclusion. EGFR inhibition restored the expression of antigen-presentation programs and enhanced antitumor immune responses in a syngeneic Nf1 knockdown model resistant to ICIs.

Conclusions

These findings define spatially organized, functionally distinct meta-niches enriched in NF1-mutant melanoma that may contribute to their aggressive disease features and poor ICI response. Our study links an understudied melanoma driver to specific immune-evasion mechanisms and identifies EGFR inhibition as a candidate strategy to improve outcomes in patients with NF1-mutant melanoma resistant to immunotherapy.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12943-026-02781-9.

Introduction

Immune checkpoint inhibitors (ICIs) have improved clinical outcomes and survival for many patients with cancer. However, more than 50% of tumors do not respond to these treatments, suggesting that they deploy immune-evasive mechanisms not targeted by ICIs [1, 2]. Although cancer-driving genetic mutations, tumor cell intrinsic transcriptional programs, and features of the tumor immune microenvironment have been linked to histopathology [3], immune evasion [4], and resistance to ICIs [5], it remains unclear how the spatial organization of tumor cells and their local interactions with surrounding stromal and immune cells shape specific mechanisms of ICI resistance.

Melanoma, the deadliest form of skin cancer due to its high metastatic potential and resistance to therapy [6, 7], is a valuable model for studying immune evasion and resistance to ICIs. Despite its high neoantigen load, transcriptional heterogeneity, and diverse oncogenic driver mutations, fewer than 50% of patients achieve durable responses to ICI, underscoring a critical gap in our understanding of tumor-intrinsic mechanisms of immunotherapy resistance. Neurofibromin 1 (NF1) loss-of-function mutations are observed in up to 27% of melanomas and in up to 45% of BRAF- and NRAS-wild-type tumors [8]. Despite being associated with worse disease-specific and overall survival [9], NF1-mutant (NF1^Mut) melanomas lack effective targeted therapies [8, 10, 11]. Moreover, co-occurring NF1 mutations with alterations in BRAF or NRAS confer resistance to BRAF and MEK inhibitors [8, 10–12]. As a result, clinical management of NF1^Mut melanoma remains particularly challenging.

NF1^Mut melanomas have the highest tumor mutational burden, and NF1 loss has been linked to increased programmed death-ligand 1 (PD-L1) cell-surface expression, suggesting that these tumors might respond to ICIs [5, 9, 10, 13, 14]. Paradoxically, more than half of NF1^Mut melanomas fail to respond to ICIs [8, 14]. These data suggest that NF1^Mut tumors may employ alternative, immune checkpoint-independent mechanisms to evade immune surveillance and resist treatment, highlighting the need to define these mechanisms in order to improve clinical responses.

Here, we used spatial multi-omic approaches to characterize cellular neighborhoods composed of melanoma cells in distinct transcriptional states together with cancer-associated fibroblasts (CAFs), endothelial cells, myeloid cells, T cells, and other immune and stromal populations in human NF1^Mut and NF1 wild-type (NF1^WT) melanoma tissues. We define recurrent cellular neighborhoods across tumors as meta-niches (MNs) and show that their abundance, composition, and immune regulatory functions are associated with NF1-loss-linked transcriptional programs. Finally, using patient tissues and a preclinical Nf1 knockdown melanoma model, we identify EGFR inhibition as a therapeutic strategy to overcome resistance to PD-1 blockade in NF1^Mut melanoma.

Materials and methods

Study cohort

Formalin-fixed paraffin-embedded (FFPE) tumor samples were prospectively collected from melanoma patients enrolled in the Interdisciplinary Melanoma Cooperative Group Database (IMCG) at NYU Langone Health (IRB #10362). Inclusion criteria for this clinicopathological database included patients who presented to the Laura and Isaac Perlmutter Cancer Center (2002–2018) with melanoma and were accrued within two months of diagnosis. All patients signed NYU Langone Health Institutional Review Board-approved consent authorizing the use of their specimens for research and follow-up. A board-certified dermatopathologist (G.J.) reviewed H&E-stained tumor sections from 32 patients and selected two representative areas per patient specimen, one at the tumor center and another at the tumor margin. These areas were used to assemble two tissue microarrays (TMAs).

Xenium spatial analysis

Processing, integration, and annotation of Xenium spatial transcriptomics data

Xenium analyses were performed on 42 cores from 32 patients contained in two tissue microarrays. Data from multiple cores were acquired on the same Xenium Slide and imported into Seurat (v5.0.0) [15] running on R (v4.3.2). Tissue coordinates of individual cores were imported as CSV files using the Xenium Explorer. For each core, a segmentation object was generated with CreateSegmentation(), and the core-specific field of view (FOV) was overlaid onto the full-region FOV using Overlay(). Core-specific counts were extracted to create Seurat objects, filtering out cells with zero or low transcript counts (< 50). Spatial features were reconstructed using centroid-based FOVs CreateCentroids() and CreateFOV() and molecule overlays CreateMolecules() for visualization within core boundaries. Individual core Seurat objects were first merged using the merge() function, with orig.ident retained to track the core of origin. The merged dataset was normalized and variance-stabilized using SCTransform() (selecting the top 5,000 variable features). Batch effects across patients were corrected using Harmony [16] (HarmonyMatrix), with core ID specified as the grouping variable. Harmony-corrected embeddings were then used for downstream dimensionality reduction, including principal component analysis (PCA) and UMAP, as well as neighbor graph construction and clustering. Cell identities were assigned by identifying cluster-specific marker genes using FindAllMarkers() and manually annotating clusters based on canonical marker expression. Differential abundance of clusters and cell types between NF1^Mut and NF1^WT samples was calculated as the fraction of cells per patient in each cluster, and differences were assessed using Wilcoxon rank-sum tests with stat_compare_means() ggpubr (v0.6.2) [17].

Spatial neighborhood mapping and meta-niche characterization

To characterize local cellular neighborhoods, tissue coordinates for each cell were retrieved, and the composition of neighboring cell types within a 40 μm radius was computed for each cell using a fixed-radius nearest neighbor search (frNN function, dbscan package) [18]. Cells with fewer than three neighbors were excluded. The resulting neighborhood compositions were represented as vectors of cell-type fractions and clustered across all cells to identify 12 MNs using k-means clustering [19]. MN labels were mapped back to the original Seurat objects, allowing for per-patient and per-cell assignment. Frequencies of MNs per patient were calculated, and differences between NF1^Mut and NF1^WT samples were assessed using Wilcoxon rank-sum tests. Niches that varied significantly between NF1^Mut and NF1^WT melanoma tissues were visualized using the ImageDimPlot function of Seurat with custom color schemes.

CellChat [20] was used to infer cell-cell communication within spatially defined MNs or on each individual tissue core. Cells were grouped by cell type and NF1 genotype, and interactions were predicted using the Human ligand–receptor database. Comparative network analyses identified shared, unique, and differential signaling between NF1^Mut and NF1^WT cells, with circle plots and heatmaps summarizing interaction strength and signaling roles across cell types.

Spatial proximity analysis of cell types

To quantify spatial relationships between cell types, tissue coordinates were extracted from each Seurat object. Spatial proximity was evaluated independently within each tissue core. For each source–target cell-type pair, the Euclidean distance from every source cell to its nearest target cell was calculated using the get.knnx() function from the FNN package (v1.1.3), and distances were summarized within each core before comparison by NF1 genotype. Cores lacking either the source or target cell population were excluded from the corresponding cell-pair analysis. To assess whether tissue-level sampling characteristics could influence the spatial-distance measurements, we quantified the total number of cells analyzed, estimated the analyzed tissue area, and overall cellular density for each core. The analyzed area was estimated from the spatial coordinate range of each core, and cellular density was calculated as the total number of cells divided by the estimated area analyzed. We also calculated the average number of melanoma-cell neighbors per tissue. These parameters were compared between NF1^Mut and NF1^WT tissues using two-sided Wilcoxon rank-sum tests.

Transcriptional profiling of spatial meta-niches and association with EGFR

Following MN identification, transcriptional programs were mapped by aggregating gene expression across cells assigned to each spatial MN. Average expression profiles were computed and used to characterize niche-specific transcriptional signatures. EGFR expression was quantified across meta-niches and summarized at the patient–niche level. Genotype-dependent differences between NF1^Mut and NF1^WT samples were assessed using Wilcoxon rank-sum tests. To resolve MN-specific mechanisms, we quantified both the fraction of EGFR-positive cells and the mean EGFR expression in EGFR-positive cells, stratified by the dominant cell types within each MN. EGFR-associated transcriptional programs were assessed at the pathway level rather than by single-gene correlation. Specifically, activity scores were calculated for Gene Ontology and Hallmark gene sets, and these pathway scores were correlated with EGFR expression using Spearman’s correlation. This approach identified niche-specific pathways whose activity increased or decreased in association with EGFR expression.

Cell type–specific differential expression and pathway enrichment independent of spatial niches

Differential expression analyses were performed on melanoma cells, CD8⁺ T cells, CAFs, and myeloid cells independently of spatial niche assignment using Seurat. Cells were subset by cell type and stratified by NF1 genotype. The RNA assay was normalized using NormalizeData(), and differential expression between NF1^Mut and NF1^WT cells was assessed using Wilcoxon rank-sum tests implemented in FindMarkers(). Genes with an adjusted P-value < 0.05 were considered significant and were separated into genotype-specific up- and downregulated sets. Functional enrichment analyses were conducted using clusterProfiler (v4.10.1) with gene identifier mapping via org.Hs.eg.db (v3.18.0). Gene Ontology Biological Process enrichment was performed using enrichGO(), and pathway-level enrichment was evaluated by ranking genes by log fold change, followed by GSEA using gseGO(). In parallel, Hallmark pathway enrichment was performed using Hallmark gene sets from the Molecular Signatures Database (MSigDB; collection H, release 2025.1), accessed via msigdbr (v25.1.1). Core enrichment genes were converted to gene symbols to facilitate biological interpretation. Data were visualized using ggplot2 [21].

Bulk RNA-Seq

RNA-Seq data were analyzed as previously described [10]. Briefly, total RNA was extracted and treated with DNase using the RNeasy Plus Micro Kit (Qiagen, Cat. No.74034) following the manufacturer’s instructions. Libraries were prepared with the Illumina TruSeq Stranded Total RNA Ribo-Zero H/M/R Gold kit and sequenced on an Illumina NovaSeq 6000 at the NYU Genome Technology Center (RRID: SCR_017929). Sequencing data were demultiplexed and converted to FASTQ files using Illumina bcl2fastq. Reads were aligned to the human genome (hg19/GRCh37) or mouse genome (mm10) with the splice-aware STAR aligner, and PCR duplicates were removed using Picard (http://broadinstitute.github.io/picard/). Gene-level counts were generated with HTSeq, and DESeq2 was used to normalize counts and identify differentially expressed genes via negative binomial generalized linear models. Alignment, differential expression, and pathway analyses were performed using the Seq-N-Slide pipeline (RRID: SCR_021752, https://igordot.github.io/sns/).

CODEX spatial proteomics

45-plex immunofluorescent (IF) imaging was performed on 56 archived FFPE tissues isolated from 32 melanoma patients. This 45-plex IF panel includes markers from Akoya Biosciences Phenocycler panel to classify melanoma cells (SOX10, S100b, GP100, NGFR, CSPG4, b-catenin, Axl), immune cells (CD20, CD3e, CD8, CD4, CD57, CD68, CD14, CD11c, CLEC9a, CD1c, Mac2, CD103, Foxp3, CD45RO, TCF-1, CD21, TOX, tryptase, CD169, CD66), endothelial cells (CD31, PDPN, LYVE1, AQP1, PNAd), epithelial cells (pan-cytokeratin) and stromal cells (PDPN, SMA, Collagen IV). This panel also includes markers for proliferation (Ki67), activation (IL-33, Granzyme B, IFNɣ), antigen presentation (HLA-DR), and immune inhibition (PD-1, ICOS, PD-L1, IDO1). 7 μm FFPE tissue was stained and scanned using PhenoCycler-Fusion 2.0 (Akoya Biosciences) according to Akoya’s commercial protocol. HALO Image Analysis Platform (Indica Labs, Albuquerque, NM, USA), an AI-assisted image segmentation software, was used for quantitative image analysis. We trained the software to accurately segment cells based on their nuclear staining and identify tumor mass. This tumor classifier is trained by manually annotating tumor and non-tumor regions. The software considers both the structure and morphology of each pixel, as well as the intensities of our selected markers (DAPI, SOX10, S100b, GP100, and CSPG4), to automatically segment cell borders and annotate tumor regions. Cells within tumor nests (i.e. CD8⁺ T cells) are then annotated using markers threshold defined by pixel intensity and percent positive pixels within the cell borders. The final exported matrix contains the cell ID, tumor/non-tumor classification, average marker intensities, and the annotated cell phenotype. From this matrix, the cell densities (cells/mm2) and cell frequencies (cell count/total cells) are calculated for downstream analyses. A nonparametric Mann-Whitney test was used to calculate p-values.

Cell culture

A375 (RRID: CVCL_0132), 08-175 [10], SK-MEL-28 (RRID: CVCL_0526) and RIM3 [22] cell lines were utilized in this study [23]. A375 and Sk-MEL-28 cells were obtained from ATCC [24] and were cultured in Corning™ RPMI 1640 Medium (Mod.) 1X with L-Glutamine (Life Sciences, 10-040-CV). RIM-3 cells were obtained from the Sommer lab and cultured in DMEM (Corning, 10-017-CV). All culture media were enriched with 10% fetal bovine serum (FBS), L-glutamine from Invitrogen, sodium pyruvate solution (Sigma, S8636), non-essential amino acids (Sigma, M7145), GlutaMAX supplement (Gibco, 25050061), and normocin (Invivogen, NC9273499). All cell lines were maintained at 37 °C in 7% CO2 and were periodically tested for mycoplasma. Human cell lines NF1 and EGFR knockdowns were previously described [10]. All Lentiviral and piggyback vectors were designed and purchased from Vector Builder (Chicago, USA). shNF1 target sequence: GCCAACCTTAACCTCTCTAAT, and shEGFR target sequence: GCATAGGCATTGGTGAATTTA. NF1 ectopic expression was performed by using a piggyback vector that expresses a flag-tagged, tetracycline-inducible NF1 full-length gene (Vector builder: VB250317-1543ver), and an empty flag (VB250319-1011vbt) was used as a control. Plasmids were transfected using Lipofectamine 3000 (Invitrogen, L3000015) and expression was induced by adding 1ug/ml doxycycline (Gold-Bio, D-500-1) Western blotting and qPCR were performed as previously described [10] using antibodies against NF1 (Abcam, Cat. No. ab17963; RRID: AB_444142), EGFR (Abcam, Cat. No. ab52894; RRID: AB_869579), HSP90 (Cell Signaling Technology, Cat. No. 4877; RRID: AB_2233307), and HLA class I ABC (Thermo Fisher Scientific, Cat. No. 15240-1-AP; RRID: AB_1557426). Horseradish peroxidase-conjugated secondary antibodies were from Cell Signaling Technology (Cat. Nos. 7074 and 7076; RRIDs: AB_2099233 and AB_330924, respectively). qPCR primers were ordered from Invitrogen, USA. Primers used were: NF1: F: AACTTCTTCCTGCGACTGCG, R: CTCTGCGACAGACGTCAACA. EGFR: F: ACCTCTCCCGGTCAGAGATG, R: TGTGCCTTGGCAGACTTTCT. Live-cell imaging was performed using the Incucyte (Essen BioScience) as described previously [10]. Flow cytometry analysis was performed on cells treated with Human IFN-gamma Recombinant Protein (Thermo Fisher Scientific, Cat No. 300-02-100U). Cells were stained with PE-conjugated anti-human HLA-A/B/C antibody (BioLegend, Cat. No. 311406; RRID: AB_314875) or APC-conjugated anti-human HLA-DR/DP/DQ antibody (BioLegend, Cat. No. 361714; RRID: AB_2750316) at 4 °C for 30 min in the dark. Flow analysis was performed using BD FACS Symphony A5 analyzer and analyzed using FlowJo V10.10.0.

Mouse models

All animal-related procedures were conducted strictly following the guidelines and approvals set forth by the Institutional Animal Care and Use Committee (IACUC# 202300101) at New York University Langone Health. RIM3 melanoma cells expressing shSCR or shNf1 were injected subcutaneously into the flank of six-week-old nude mice (Jackson Laboratory, RRID: IMSR_JAX:002019) or syngeneic C57BL/6J mice (Jackson Laboratory, RRID: IMSR_JAX:000664). Male mice were used because the sex of the mouse from which the RIM3 cell line was originally derived was not reported. For tumor-growth studies, mice were injected with1 × 10⁶ cells, whereas 2 × 10⁵ cells were used for treatment studies. Tumor dimensions were measured using digital calipers, and tumor volume was calculated as V = (π/6) x L x W²), where L represents tumor length and W represents tumor width.

For longitudinal tumor-growth studies, treatment was initiated when tumors reached approximately 100–200 mm³. Mice were then randomly assigned to vehicle control, afatinib, anti–PD-1, or combined afatinib plus anti–PD-1 treatment groups, with 12 mice per group, and treated for up to four weeks. For flow-cytometric endpoint studies, tumors were allowed to reach approximately 500 mm³ before mice were randomly assigned to the same treatment groups. 5–7 mice were treated for one week, after which tumors were collected for flow-cytometric analysis.

Afatinib (MedChemExpress, BIBW 2992, Cat. No. HY-10261) was administered by oral gavage at 20 mg/kg five days per week. Anti–PD-1 antibody (Bio X Cell, anti-mouse CD279, Cat. No. BE0146) was administered by intraperitoneal injection at 10 mg/kg twice weekly.

Mice were monitored for changes in body weight, general condition, and other signs of treatment-related toxicity daily. Animals were excluded only according to predefined criteria, including severe weight loss, and six animals were excluded. Tumor-growth curves are presented as mean ± SEM. For each mouse, the area under the tumor-growth curve (AUC) was calculated using the trapezoidal rule, and differences in AUC between treatment groups were assessed using two-sided unpaired t-tests. Mice were euthanized when tumor volume exceeded 2,000 mm³ or when a predefined experimental endpoint was reached. Overall survival was evaluated using Kaplan–Meier analysis, and survival curves were compared using the log-rank test. A two-sided P < 0.05 was considered statistically significant. To stain for immune markers, tumor masses were measured at the endpoint. Resected tumors were digested at 37 °C for 25 min with agitation by Collagenase D (1 mg/mL): (Sigma-Aldrich Cat. No. 11088866001) and DNase I (80 U/mL) (Roche Cat. No.04536282001) and passed through 70 μm cell strainers to generate single cell suspensions. Each tumor was divided into two wells and stained with a T cell panel or a Myeloid panel in addition to Ghost Dye live/dead stain (CYTEK Cat. No.130865T100).

The T-cell panel included antibodies against CD45 (BD Biosciences, Cat. No. 564279; RRID: AB_2651134), CD4 (BD Biosciences, Cat. No. 612844), CD8α (Cytek, Cat. No.65-0081-U100), CD44 (Cytek, Cat. No. 60-0441-U100), granzyme B (Thermo Fisher Scientific, Cat. No. MHGB04; RRID: AB_10372671), Ki-67 (Thermo Fisher Scientific, Cat. No. 48-5698-80; RRID: AB_11151155), FOXP3 (BioLegend, Cat. No. 126407; RRID: AB_1089116), PD-1 (BioLegend, Cat. No. 135225; RRID: AB_2563680), and TIM-3 (BioLegend, Cat. No. 75833-230).

The myeloid-cell panel included antibodies against NK1.1 (BioLegend, Cat. No. 108710; RRID: AB_313397), CD11b (BioLegend, Cat. No. 101243; RRID: AB_2561373), CD11c (BioLegend, Cat. No. 117333; RRID: AB_11204262), Ly-6 C (BioLegend, Cat. No. 128012; RRID: AB_1659241), Ly-6G (BioLegend, Cat. No. 127618; RRID: AB_1877261), I-A/I-E (BioLegend, Cat. No. 107632; RRID: AB_2650896), and F4/80 (Thermo Fisher Scientific, Cat. No. 12-4801-82; RRID: AB_465923). Tumors were stained for 1 h at 4 °C in the dark. After washing, tumors were fixed in 4% paraformaldehyde and stored at 4 °C until analysis. Immune-cell populations were quantified as frequencies of live CD45⁺ cells, and absolute cell numbers were additionally normalized to tumor mass to account for differences in tumor size. Comparisons between groups were performed using the Mann–Whitney U test.

Statistical analysis

Statistical analyses were performed with Microsoft Excel, GraphPad Prism (version 10.2.0, RRID: SCR_002798), or R statistical software (RRID: SCR_001905 version 4.3.1).

Results

Melanoma tissues are organized into 12 spatially defined meta-niches

To determine whether and how NF1 loss affects immune evasion and immunotherapy resistance, we used two tissue microarrays (TMAs) comprising samples from 32 patients with melanoma, including 17 NF1^Mut and 15 NF1^WT tumors (Fig. 1A, Table S1). Because each patient contributed two tissue cores, these TMAs contained 64 cores (34 NF1^Mut and 30 NF1^WT). We profiled 42 of these cores using the 10X Genomics 5 K Xenium platform. After image processing, cell segmentation, and rigorous quality control, we excluded cells with fewer than 50 transcripts and cores with fewer than 5000 cells. We batch-corrected and normalized the data across independent tissues, creating an integrated data set of 1,003,604 high-quality cells from 39 cores (20 NF1^Mut and 19 NF1^WT). Louvain clustering separated melanoma cells from B cells, T cells, myeloid cells, plasma cells, epithelial cells, and other stromal cell types (Fig. 1B), which were then annotated based on lineage-defining marker gene expression (Fig. 1C). Although these cell populations were detected across most melanoma samples, their spatial distributions varied markedly between tumors (Fig. 1D), suggesting that each tumor contains multiple specialized niches whose composition and organization may influence pathological features and therapy response.

Fig. 1.

Distinct spatial niches and cellular neighborhood architectures of melanoma. A Schematic representation of the study workflow, starting from isolating 64 melanoma tumors from 32 patients, generating TMAs, performing Xenium spatial transcriptomics on 42 samples and CODEX spatial proteomics on 56 samples to perform single-cell analysis, and testing of identified targets in preclinical models. B Uniform manifold approximation and projection (UMAP) plot of major cell types identified in the Xenium dataset. C Dot plot showing canonical marker genes for each cell type identified in (B). D Representative field of view (FOV) of a Xenium image, showing major cell types in multiple melanoma samples. E UMAP of lineage sub-types identified in the Xenium dataset. F Representative FOV of a Xenium image showing lineage sub-types identified in (E). G Stacked-bar plot showing the composition of each meta-niche (MN) across all cell types. Melanoma cell states are abbreviated as melanocytic (Melan), proliferative (Prol), neural-crest-like (NC), and mesenchymal (Mes). H Bar plot showing the number of patients included in each MN. I Representative FOV images of MNs identified in melanoma patients

To characterize these niches in greater detail, we re-clustered cells within each major lineage based on gene expression, identifying 18 distinct cell types, including five melanoma cell states (Fig. 1E-F, Figure S1A). We identified melanoma cells with melanocytic, proliferative, intermediate, neural crest (NC)-like, and mesenchymal (MES) features, consistent with previous reports [25]. Immune populations included cytotoxic CD8⁺ T cells and CXCL13⁺ CD8⁺ T cells, regulatory T cells (Treg), M2-like macrophages, interferon (IFN)-stimulated macrophages, activated myeloid cells, B cells, and plasma cells. Stromal populations included cancer-associated fibroblasts (CAFs) with myofibroblastic (myCAF) or inflammatory (iCAF) features, as well as pericytes, endothelial cells, and epithelial cells.

We mapped cell-type identities and transcriptional programs back to each tissue core, retrieved each cell’s centroid, and analyzed their neighborhoods within a 40-µm radius using a fast radius-based nearest-neighbor search (frNN) [18]. K-means clustering [19] grouped these neighborhoods into 12 frequently occurring meta-niches (MNs), which we defined as cellular spatial neighborhood patterns observed across multiple independent tissue samples (Fig. 1G). Each MN was annotated according to its most prevalent cell types and broadly categorized as melanoma-rich, immune-rich, stromal-rich, or vascular-rich. Notably, each cell type appeared across multiple MNs (Figure S1B), indicating that cell identity alone does not determine niche architecture.

Although individual patients exhibited significant variability in MN composition (Figure S1C), with some tissues dominated by melanoma-rich niches, the identified MNs were generally consistent across the cohort. Several MNs were found in all samples, while others appeared in more than 50% of cases (Fig. 1H-I). These results suggest that, despite differences in niche abundance among patients, the main spatial MN programs reflect recurring TME states rather than isolated, patient-specific patterns. Overall, these data demonstrate that melanoma tissues are highly compartmentalized, with distinct cell types forming recurring MNs, and that each tumor contains multiple MNs in proportions unique to each patient.

NF1-loss links meta-niche remodeling to immune evasion and immunotherapy resistance

To determine whether variation in MN composition was associated with cancer-driving mutations, we compared NF1^Mut and NF1^WT melanoma tissues. Although these tissues contained the same major cell populations (Figure S2A), NF1^Mut melanomas contained significantly more NC-like melanoma cells (Figure S2B-C). NF1^Mut melanomas were also significantly enriched in MNs dominated by NC-like melanoma cells, immune-suppressive M2-like macrophages, immune-suppressive iCAFs, and regulatory T cells (Fig. 2A-B, S2D-F). Conversely, NF1^Mut melanomas were depleted of MNs enriched for cytotoxic T cells, Tregs, and IFN-activated macrophages, as well as MNs enriched in pericytes and endothelial cells (Fig. 2A, C, S2D-F). These results suggest that NF1 loss reshapes the assembly of cells into specific MNs, potentially contributing to immune evasion.

Fig. 2.

NF1 mutant melanomas are enriched in oncogenic and immune suppressive meta-niches. A, C Dot plots showing the spatial niche (MN) abundance in NF1^Mut or NF1^WT melanoma. Abundance is represented as a color gradient, and Fisher’s exact P-values are represented by size. B, C Representative field of view (FOV) images of Xenium data, showing cell types and niches that are either increased in NF1^Mut (B) or NF1^WT (C) melanoma. D-G CellChat circle plots showing signaling patterns in NC-like melanoma dominated MN-2 for NF1^Mut or NF1^WT (D), signaling patterns in MN-6 (rich in CAFs and M2 macrophages) of NF1^Mut or NF1^WT (E), signaling patterns in immune dominated MN-3 of NF1^Mut or NF1^WT (F), or differential number of interactions and differential interaction strength between NF1^Mut and NF1^WT MN-3 (G). Red arrows indicate more and stronger interactions, while blue arrows indicate fewer or weaker interactions. H Heatmap showing the fraction of each MN in individual tumors, stratified by NF1 mutation status and treatment response. Color intensity indicates the fraction of neighborhoods assigned to each meta-niche (MN) within a tumor. I Effect-size plots showing differences in mean MN fractions between responders (R) and non-responders (NR) for NF1^Mut and NF1^WT tumors, including analyses restricted to patients treated with PD-1 blockade. Positive values indicate enrichment in responders, whereas negative values indicate enrichment in non-responders. Dot size represents −log10 of two-sided exact permutation test comparing tissue level MN fractions between R and NR

Cell-cell communication analyses using CellChat [20] support this idea. NC-like melanoma cells communicated more extensively with melanoma cells in other states, stromal and immune cells within MN2 (NC + M2 + CAF) of NF1^Mut than NF1^WT melanoma (Fig. 2D). Myelin protein zero (MPZ), which has been associated with tumor progression, invasion, and metastasis [26], was the predominant and most characteristic signaling pathway active within MN2 (Figure S2G). Cell-cell communication within MN6 (iCAFs+M2) was generally much stronger in NF1^Mut than in NF1^WT melanoma (Fig. 2E). Furthermore, interactions between mesenchymal melanoma cells and CAFs were consistently stronger in MN2, MN6, and MN3 of NF1^Mut than NF1^WT melanoma tissues (Fig. 2D-F).

In contrast, melanoma-immune cell interactions, especially those between activated macrophages and pericytes or endothelial cells, were stronger within the immune-active MN3 (IFN-myeloid + T cell) in NF1^WT than in NF1^Mut tumors (Fig. 2F). Differential cell-cell communication analyses within this immune-active MN3 revealed a greater number of receptor-ligand interactions among cytotoxic CD8⁺ T cells, regulatory T cells, and activated macrophages in NF1^WT than in NF1^Mut tumors (Fig. 2G). Conversely, signaling interactions between melanoma cells and CAFs in MN3 were higher in NF1^Mut than in NF1^WT tumors. CXCL signaling was significantly elevated in MN3 of NF1^WT tumors, whereas MPZ, Cell Adhesion Molecule (CADM), Protease-Activated Receptor (PAR), and Heparan Sulfate Proteoglycans (HSPG) signaling were increased in MN3 of NF1^Mut tumors (Figure S2H). These data are consistent with a more angiogenic and immunosuppressed tumor microenvironment (TME) in NF1^Mut compared to NF1^WT melanoma.

Some differences in cell-cell communication between NF1^Mut and NF1^WT melanomas were not restricted to specific meta-niches; instead, they were evident globally across tissues. NC-like melanoma cells signaled extensively to iCAFs and mesenchymal melanoma cells, while myCAFs engaged in collagen and laminin signaling with NC-like melanoma cells, in NF1^Mut but not NF1^WT melanoma (Figure S3A-B). Conversely, myCAFs showed fewer interactions with cytotoxic CD8⁺ T cells and CXCL13⁺ CD8⁺ T cells in NF1^Mut than in NF1^WT tumors.

These differences in cell-to-cell communication were accompanied by broader signaling changes across NF1^Mut compared to NF1^WT melanoma. NF1^Mut melanoma exhibited increased EGF signaling in mesenchymal melanoma cells and iCAFs; significantly higher levels of collagen, laminin, TGFβ, KIT, prostaglandin, VEGF, PDGF, and L1CAM signaling in the TME, where CXCL signaling was decreased (Figure S3C-E). It also showed increased signaling for PDGF, EGF, TGFβ, KIT, VEGF, ADGR-A, ADGR-D, tenascin, and laminin in myCAFs and iCAFs. Overall, these findings indicate that NF1^Mut melanoma is characterized by heightened stromal and growth factor signaling, a more fibrotic microenvironment, reduced vascular signals, and diminished immune-cell interactions.

The immunosuppressive TME in this NF1^Mut melanoma cohort indicated that some tumors could be resistant to ICIs. To determine whether the spatial composition of MNs correlates with anti-PD1 or other ICI response data, we stratified our NF1^Mut melanoma cohort into 7 responder (R) and 11 non-responder (NR), and our NF1^WT melanomas into 9 R and 9 NR tissues. By comparing the fraction of each MN between R and NR we uncovered distinct response-associated patterns in NF1^Mut and NF1^WT melanomas (Fig. 2H). The MN2 (NC + M2 + CAF) was highly enriched in NF1^Mut but not NF1^WT tumors that didn’t respond to ICI or anti-PD1 (Fig. 2I). myCAFs and mesenchymal melanoma cells were also enriched in the NF1^Mut NR group. In contrast, MN3 (IFN-myeloid + T cell) and MN9 (Inter+Prol+Melan) were enriched in the NF1^WT NR group. MN4 (vasc+perivascular) was more abundant in the R group. These data suggest that NC-like melanoma cells, as well as myCAFs, M2-like macrophages, and Tregs, may contribute to ICI resistance in NF1^Mut melanoma, consistent with their well-established immunosuppressive functions [27–29]. They also suggest that MNs orchestrated by melanoma cells that lost NF1 may function as spatially organized, immunosuppressive microenvironments that become enriched in NF1^Mut melanoma tissues that don’t respond to PD-1 or other ICIs.

NF1 mutant tumors accumulate dysfunctional CD8⁺ T cells

The selective enrichment of immune-suppressive MNs suggests that NF1 loss could alter the spatial structure of the tumor microenvironment, broadly impairing anti-tumor immunity. To test this idea, we examined global differences in immune cell organization and CD8⁺ T cell function in NF1^Mut melanoma by measuring distances between melanoma cells and nearby cell types. This neighborhood analysis showed that NF1^Mut melanoma cells are closer to CAFs, Tregs, M2-like macrophages, and pericytes, but are farther from cytotoxic CD8⁺ T cells, CXCL13⁺ CD8⁺ T cells, or endothelial cells compared to NF1^WT melanoma cells within tumors (Fig. 3A-B). To rule out the impact of tissue sampling differences on the spatial results, we compared the total cell counts, tissue area, cellular density, and the average number of melanoma neighbors between NF1^Mut and NF1^WT tissue samples and found no differences in these parameters (Figure S4A). These findings suggest that the microenvironment surrounding NF1^Mut melanoma cells is depleted of immune cells but rich in CAFs, while NF1^WT melanoma tissues are more infiltrated by CD8⁺ T cells and other immune cell types.

Fig. 3.

NF1 mutant melanoma is characterized by decreased T cell infiltration and activity. A NF1^Mut vs. NF1^WT differential contact heatmap showing distances between cell types. Red indicates cells are farther away, while blue indicates cells are closer in NF1^Mut tissue. B Representative field of view (FOV) images showing the spatial organization of tumor cells, CD8⁺ T cells, and stromal cells, shown by major cell type or sub-cell type in NF1^Mut or NF1^WT. C Box plots and data points of CD8⁺ T cells, CXCL13⁺ CD8⁺ T cells, and Tregs over the total number of cells per patient or the Treg/CD8⁺ T cell ratio. P-values indicate Wilcox test results. D, E Spatial heatmap (D) or FOV Images (E) of CD8⁺ T cells in NF1^Mut or NF1^WT melanoma patients. F Box plots showing differentially expressed genes comparing NF1^Mut to NF1^WT melanoma CD8⁺ T cell clusters using pseudo bulk. P-values indicate a Wilcox test. G Selected GO pathway analysis shows that response to decreased oxygen is upregulated, while T cell activation, differentiation, and NK activity are downregulated in CD8⁺ T cell clusters of NF1^Mut melanoma. H Representative images showing tumor-infiltrating CD8⁺ T cells/total number of cells per patient and the co-localization of CD8⁺ activity markers GZMB, PD-1, Ki67, CD45 RO, and TOX. I Box plots and data points of CD8⁺ T cell activity markers are shown in (H). P-values indicate Wilcox test results

Indeed, NF1^Mut melanoma cells were surrounded by fewer cytotoxic CD8⁺ T cells and CXCL13⁺ CD8⁺ T cells, and the regulatory T cell-to-CD8 + T cell ratio was higher in these tumors, consistent with a suppressed immune microenvironment (Fig. 3C-E). The few CD8⁺ T cells that infiltrated NF1^Mut tumors expressed significantly less FYN, GZMB, IRF1, TNFRSF1B, TNFSF4, and TNFRSF9 (Fig. 3F). They were also enriched in pathways associated with decreased cytokine production, reduced T cell activation and differentiation, and diminished T cell-mediated killing of cancer cells (Fig. 3G, Table S2).

To validate these data with an independent method, we stained the same cohort with a custom 45-plex antibody panel. We imaged 56 tissues on the Akoya-PhenoCycler platform and used HALO software to quantify, segment, and annotate 1,079,175 single cells. This approach captured melanoma cells, CAFs, endothelial cells, dendritic cells (DCs), macrophages, CD4⁺ T cells, CD8⁺ T cells, memory T cells, NK cells, and B cells. After quality control, we retained 1,052,420 cells from 49 samples. These immunostainings confirmed that NF1^Mut melanoma tissues contained fewer cytotoxic CD8⁺ T cells and a higher regulatory T cell-to-CD8⁺ T cell ratio. Furthermore, markers of cytotoxic (GZMB⁺), memory (CD45RO⁺), exhausted (PD-1⁺, TOX⁺), and activated/proliferating (Ki67⁺) CD8⁺ T cells were significantly decreased in NF1^Mut melanoma tissues (Fig. 3H, I). Consistent with these results, imputation of gene expression data of skin cutaneous melanoma (SKCM) that were generated by the Cancer Genome Atlas Network (TCGA) [13] with CIBERSORTx [30], using single-cell RNA sequencing data from Tirosh et al. [31] as a reference, also detected significantly fewer CD8⁺ T cells in NF1^Mut melanoma samples than in NF1^WT melanomas (Figure S4B). To test functional engagement of CD8⁺ T cells in immune responses, we measured intratumoral CD4⁺ T cell/CD8⁺ T cell/DC immune triads, which are crucial for ICI responses [32]. We quantified the presence of these triads using proximity scores based on distances between CD4⁺ T cells, CD8⁺ T cells, and DCs (CD11c⁺, HLA.DR⁺, CD14⁻, CD68⁻, SOX10⁻) and observed significantly fewer CD4⁺ T cell/CD8⁺ T cell/DC immune complexes in NF1^Mut melanoma tissues compared to NF1^WT melanoma tissues (Figure S4C-D). Moreover, these complexes were significantly higher in tumors of patients who responded to PD-1 inhibitors (Figure S4E).

NF1 loss results in decreased expression of melanoma antigen presentation mechanisms

Our analyses uncovered distinct immune evasion strategies in NF1^WT and NF1^Mut melanomas. NF1^WT tumors were enriched for immune-active MNs, in which CD8⁺ T cells followed a classical activation, memory formation, and exhaustion trajectory. By contrast, NF1^Mut melanomas were dominated by immune-silenced MNs, with broadly attenuated CD8⁺ T cell states and reduced tumor immune interactions. Consistent with these observations, inflammatory cytokine (IL1B, IL6, IL12A/B, IL17A, IL18, TNF, IFNG, IFNA1, IFNB1), chemokine (CXCL9, CXCL10, CXCL11, CXCL13), and chemokine receptor (CCR5, CCR7, CXCR3, CXCR4, CXCR5) expression scores were significantly lower in the NF1^Mut tumor immune microenvironment (Fig. 4A). These changes also correlated with suppressed immune-stimulatory (CD80, CD86, ICOS, CD28, and TNFRSF9), and immune-inhibitory checkpoint (PDCD1, CD274, CTLA4, LAG3, HAVCR2, TIGIT, and VSIR) gene expression scores. These differential expression data suggest that NF1^Mut melanomas may be less effectively recognized by immune cells, suggesting impaired antigen presentation rather than dominant checkpoint-mediated inhibition.

Fig. 4.

NF1 loss inhibits MHC class I and II antigen presentation. A Box plots and data points showing immune signature gene scores identified in the immune microenvironment of NF1^Mut and NF1^WT, normalized to the total number of cells per patient. P-values indicate Wilcox test results. B Box plots and data points of log (HLA+ tumor cells). P-values indicate a Wilcox test. C Representative images showing melanoma tumor cells stained with HLA + in NF1^Mut and NF1^WT tumor cells. D Scatter plot showing correlation and regression analysis of the number of CD8⁺ T cells and the number of HLA+ tumor cells. P-values indicate Spearman correlation. E Violin plots of the expression of MHC-related genes in melanoma cells measured by spatial transcriptomics in NF1^Mut and NF1^WT melanoma cells. P-values indicate Wilcox test results. F Western blots show changes in NF1 and HSP90 loading control or RT-PCR data showing NF1 expression changes in shNF1-1 and shNF1-2 compared to shScr expressing melanoma cells. Bar graphs indicate mean ± SD (n = 3). P-values were calculated with a two-sided t-test. G Heatmap showing differentially expressed MHC class I and II genes in A375 melanoma cells following NF1 knockdown. H Bar graphs and data points showing flow cytometry results of MHC class I or MHC class II after interferon stimulation in shNF1 or shSCR expressing SK-MEL-28 or A375 melanoma cell lines. Bar graphs indicate mean ± SD (n = 3). P-values were calculated with a two-sided t-test

Cancer cells often reduce or lose Human Leukocyte Antigen (HLA) expression as a primary immune evasion strategy [33]. Our highly multiplexed IHC analyses detected significantly lower HLA expression in NF1^Mut melanoma cells (Fig. 4B-C). This reduced HLA expression correlates directly (r_s = 0.8355) with fewer CD8⁺ T cells in tumor tissues (Fig. 4D). In addition to HLA, our Xenium data showed lower expression of TAP1, TAP2, CANX, and CIITA in NF1^Mut melanoma cells compared with NF1^WT melanoma cells (Fig. 4E). TAP1 and TAP2 are required for transporting peptides from the cytoplasm to MHC-I, and CIITA controls MHC-II expression and thus antigen presentation [34, 35].

To test whether NF1 loss directly caused reduced HLA expression in NF1^Mut melanoma, we transduced A375 and SK-MEL-28 melanoma cells with non-targeting control short-hairpin RNAs (shSCR) or NF1-inhibiting short-hairpin RNAs (shNF1) (Fig. 4F). Differential gene expression analyses showed significantly lower expression of HLA-A, HLA-B, HLA-C, HLA − F, HLA − G, HLA − DMA, HLA − DMB, HLA−DPA1, and HLA−DPB1 in shNF1-transduced A375 cells compared with shSCR controls (Fig. 4G). Furthermore, NF1-knockdown A375 and SK-MEL-28 cells expressed less MHC-I on their cell surface, and SK-MEL-28 cells also expressed less MHC-II, compared with shSCR-expressing control cells, even after IFN-ɣ stimulation (Fig. 4H). These data demonstrate that NF1 loss reduces HLA expression and that IFN-γ stimulation cannot fully restore it.

EGF-EGFR signaling is enriched in NF1-mutant melanoma meta-niches and mediates immune evasion

To identify signaling pathways hyperactivated in NF1^Mut melanoma with reduced HLA expression and amenable to therapeutic targeting, we conducted gene set enrichment analyses (GSEA) on transcripts differentially expressed in melanoma, stromal, and myeloid cells between NF1^Mut and NF1^WT tumors. These analyses revealed transcriptional programs enriched in NF1^Mut melanoma cells, including EMT, UV response, hypoxia, TGFβ, and tumor necrosis factor alpha (TNFα) signaling (Figure S5A). These pathways were predicted based on significantly increased NOTCH1, SOX2, NGFR, ERBB2, EGFR, TGFB2, TGBR2, VEGFA, VEGFC, and PDGFRB expression (Figure S5B). Conversely, NF1^WT melanoma cells were enriched for inflammatory responses, IFN-α/ɣ responses, IL2-STAT5, and IL6-STAT3 signaling, based on elevated IL10, IL21, IFNG, IFNL1, IFNA17, and TNFSF18 expression.

Furthermore, CAFs we identified in NF1^Mut melanoma upregulated genes involved in cell division, cell-cycle regulation, and EGF signaling (Figure S5C). They also downregulated genes involved in antigen processing, antigen presentation, cytokine production, inflammatory responses, T cell activation, and adaptive immune responses. Similarly, myeloid cells we identified in NF1^Mut melanoma tissues expressed transcripts associated with heightened growth factor responsiveness, increased TGFβ receptor signaling, reduced inflammatory responses, dampened immune effector processes, and diminished cytokine and immune responses (Figure S5D). Notably, EGF signaling was elevated not only in NF1^Mut melanoma cells, as we recently reported [8, 10], but also in CAFs surrounding these NF1^Mut melanoma cells, as predicted by CellChat (Figure S3C-E).

Cell-level differences in these pathways between NF1^Mut and NF1^WT melanoma were also observed at the MN level (Fig. 5A). EGFR expression was more prevalent in NF1^Mut tumors across most MNs (Fig. 5B), with the highest enrichment in NF1^Mut melanomas that didn’t respond to andti-PD-1-based therapies (Fig. 5C, S6A). To identify biological programs associated with elevated EGFR expression in NF1^Mut tumors, we computed differential pathway enrichment scores between NF1^Mut and NF1^WT melanomas. These scores correlated directly with EMT, hypoxia, and TGFβ signaling, and indirectly with inflammation, IFN responses, IL6-JAK-STAT3 signaling, IL2-STAT5 signaling, and p53 signaling (Fig. 5D).

Fig. 5.

EGF-EGFR signaling promotes an immune evasive microenvironment in NF1 mutant melanoma. A Dot plot showing differential transcriptional Hallmark gene sets enriched in each meta niche (MN) comparing NF1^Mut to NF1^WT melanoma. Color represents fold change, and size represents adjusted p-value. Melanoma cell states are abbreviated as melanocytic (Melan), proliferative (Prol), neural-crest-like (NC), and mesenchymal (Mes). B-C Dot plot showing EGFR expression in each MN, comparing NF1^Mut and NF1^WT tumors (B) or non-responders (NR) and responders (R) to anti-PD-1 based therapies (C). Dot size represents the fraction of EGFR-positive cells. Color intensity reflects mean EGFR expression. D Differential pathway activity in melanoma cells associated with EGFR, calculated as the mean difference in pathway module scores between NF1^Mut and NF1^WT. Color represents statistical significance. E Scatter plot showing correlation and regression analysis of the number of CD8⁺ T cells and EGFR immunostaining expression scores. P-values indicate Spearman correlation. F Spearman correlation analysis of EGFR expression with immune regulatory genes in melanoma clusters of NF1^Mut tissue

Elevated EGFR signaling could link oncogenic and stress-response programs to the suppression of immune-related transcriptional programs in NF1^Mut melanoma. Indeed, EGFR expression was negatively correlated (r_s = -0.622) with CD8⁺ T cell abundance in NF1^Mut melanoma tissues but not in NF1^WT melanoma tissues, indicating that tumors with the highest EGFR expression have the fewest CD8⁺ T cells (Fig. 5E, S6B). Higher EGFR expression also correlated with lower expression of antigen presentation genes (TAP1, TAP2, CANX, CIITA) and immune-stimulatory chemokines (CXCL13 and CXCL9) [36, 37], and with higher expression of TGFB1, TGFB3, IL1B, IL10, CD274, and HAVCR2 (Fig. 5F). These molecules could enhance immune evasion in NF1^Mut melanoma by inhibiting antigen presentation, decreasing T cell proliferation, or activating immune checkpoints.

To determine which immune-evasive phenotypes are attributable to NF1 loss, we compared differentially expressed genes between shNF1 knockdown and shScr-expressing A375 cells. These analyses linked gene sets associated with increased angiogenesis, hypoxia, EMT, TGFβ, and NOTCH signaling (Fig. 6A) and increased expression of VEGFA, ERBB2, EGFR, VEGFC, TGFBR2, and NOTCH1 (Fig. 6B) to NF1 loss. Conversely, ectopic expression of a doxycycline (dox)-inducible recombinant NF1 (rNF1) transgene (Fig. 6C-D) inhibited gene sets associated with these pathways (Figs. 6E-F). Consistent with these gene expression changes, NF1^Mut melanoma cells grew more slowly upon re-expression of rNF1 (Fig. 6G). These results suggest that NF1 inhibits EGFR, whereas NF1 loss activates EGFR, thereby promoting cell proliferation and immune evasion.

Fig. 6.

EGFR inhibition restores HLA antigen expression. A GSEA Hallmark pathway analysis showing EMT, TGFβ, angiogenesis, and hypoxia are upregulated after NF1 knockdown (B) Heatmap showing differentially expressed genes between shNF1 and shSCR expressing cells. C RT-PCR data showing NF1 expression changes after DOX-induced NF1 expression (rNF1) compared to control (rCon) in NF1^Mut STC-08-175 and Mewo melanoma cells. Bar graphs indicate mean ± SD (n = 3). P-values were calculated with a two-sided t-test. D Western blots showing NF1 and EGFR expression before and after DOX-induced NF1 expression (rNF1) compared to control (rControl) in NF1^Mut STC-08-175 and Mewo melanoma cells. HSP90 served as a loading control. E GSEA Hallmark pathway analysis showing IFN-α and IFN-ɣ responses are upregulated, and EMT, hypoxia and inflammatory response are downregulated as a result of NF1 ectopic expression. F Heatmap showing differentially expressed genes between cells expressing NF1 ectopically and control cells. G Growth curves of NF1^Mut STC-08-175 melanoma cells ectopically expressing rNF1 compared to control. P-values were calculated with a two-sided t-test at the experimental endpoint. Data points represent mean ± SD. (n = 12). H Western blots of IgG (control), anti-NF1, and anti-EGFR co-immunoprecipitation studies in NF1^Mut STC-08-175 or Mewo cells after DOX-induced NF1 (rpbNF1) expression compared to control (rbpControl). I Western blots showing NF1 and EGFR expression in shNf1 or shSCR expressing murine RIM-3 cells. HSP90 served as a loading control. J RT-PCR data showing Nf1 and Egfr expression changes in shNf1compared to shSCR expressing RIM3 cells. Bar graphs indicate mean ± SD (n = 3). P-values were calculated with a two-sided t-test. K Growth curves of RIM-3 melanoma cells expressing shNf1 compared to shSCR (control). P-values were calculated with a two-sided t-test at the experimental endpoint. Data points represent mean ± SD. (n = 12). L Heatmap showing differentially expressed antigen presentation genes in shNF1 expressing RIM-3 cells after EGFR knockdown. M GSEA Hallmark Gene Sets analysis showing enriched pathways in shEGFR and shNF1 vs. shNF1 expressing RIM-3 cells. N Heatmap showing differentially expressed HLA genes in shEGFR or shSCR (control) expressing NF1^Mut STC-07-127 melanoma cells. O Western blots comparing HLA and EGFR expression with and without afatinib treatment in three human NF1^Mut melanoma cell lines. HSP90 served as a loading control. P Flow cytometry analysis showing MHC class I antigen presentation difference between IFN-ɣ-stimulated or non-stimulated NF1^Mut STC-08-175 melanoma cells with and without afatinib treatment. Bar graphs indicate mean ± SD (n = 3). P-values were calculated using a two-sided t-test

BioGrid network data further indicated that NF1 may physically interact with EGFR [38]. Therefore, we co-immunoprecipitated NF1 and EGFR from cell lysates of NF1^Mut 08-175 or Mewo cells expressing rNF1, confirming that NF1 physically interacts with EGFR in melanoma cells (Fig. 6H). This interaction may inhibit EGFR activation, thereby initiating transcriptional suppression by disrupting a positive feedback loop that enhances EGFR expression [39, 40]. This autoregulatory circuit could be restored after NF1 loss.

EGFR inhibition induces an immune response that synergizes with immune checkpoint inhibitors in an NF1-depleted syngeneic melanoma mouse model

The increased expression of EGFR in NF1^Mut melanomas and its association with immune evasion mechanisms suggest that EGFR inhibition could reactivate tumor immunity and thereby improve responses to ICI therapy. To test this hypothesis, we used syngeneic RIM-3 cells [41] isolated from a melanoma that developed in Tyr::Nras^Q61K Cdkn2a^Fl/Fl mice [22]. shNf1 knockdown decreased NF1 expression and increased Egfr expression in RIM-3 cells (Fig. 6I-J) and significantly accelerated their growth rate (Fig. 6K). These results are consistent with our recent data on human melanoma short-term cultures and xenograft models [10] and now allow us to functionally test how NF1 loss and/or EGFR activation affect immune evasion and ICI responses.

To do this, we first knocked down Egfr in shNf1-expressing RIM-3 cells and used RNA-seq to identify gene sets regulated by EGFR. We found a significant decrease in Egfr expression, along with increased expression of genes required for antigen presentation and NK and T cell activation (Fig. 6L). Pathway analyses also revealed increased IFN-α and IFN-ɣ responses, while TNF-α signaling via NFκB, EMT, and TGFβ signaling were downregulated as a result of EGFR inhibition (Fig. 6M). We observed similar changes in HLA expression when we compared shEGFR -expressing human NF1^Mut melanoma cell lines with shSCR-expressing cell lines (Fig. 6N) or when we treated short-term cultures (STCs) from human NF1^Mut melanoma patients with the EGFR inhibitor afatinib (Fig. 6O) [42]. Furthermore, MHC-I expression increased even more when we treated NF1^Mut melanoma cells with afatinib and IFN-ɣ than with IFN-ɣ alone (Fig. 6P). These data demonstrate that increased EGFR signaling compromises antigen presentation in NF1^Mut melanoma cells, and these changes could contribute to immune evasion, defective T cell infiltration, impaired cytotoxic immune tetrad formation, and thus resistance to ICI therapy.

To assess whether NF1 loss affects immune evasion, we subcutaneously transplanted equal numbers of RIM-3 cells expressing shSCR or shNf1 into immunocompromised Nude mice or syngeneic C57BL/6J mice and compared tumor initiation rates. Although both shSCR- and shNf1-expressing RIM-3 cells initiated tumors at similar rates in Nude mice (Fig. 7A), only shNf1-expressing RIM-3 cells formed tumors in C57BL/6J mice (Fig. 7B). This result supports the notion that NF1 loss enhances immune evasion, consistent with our spatial transcriptomic results from human melanoma tissues.

Fig. 7.

EGFR inhibition induces immune responses in preclinical NF1-depleted melanoma models. A, B Kaplan Meier curves showing tumor-free survival of shNf1 or shSCR expressing RIM-3 cells in either immune-compromised Nude (A) or immune-competent C57Bl/6J (B) mice. P values show log-rank test results (n = 6 per arm). C Workflow of preclinical syngeneic model testing starting with transplanting shNf1 expression RIM-3 cells into BL/6J mice, monitoring tumor growth until either ~ 100–200 mm³ for treatment or ~ 500 mm³ for flow analysis, using mice randomized into 4 treatment groups (control, EGFR inhibition, PD-1 inhibition or EGFR + PD-1 inhibition), monitoring tumor growth and collecting the tumors at endpoint to generate single cell suspensions, staining with T cell or myeloid cell markers, and FACS analyses. D Growth curves of shNf1 RIM-3 cells treated with a vehicle control, afatinib, mouse-specific anti-PD-1 antibody, or a combination of both. P-values were calculated using unpaired t-tests of AUC (n = 12 per arm). Data points represent mean ± SEM. E Kaplan-Meier curves showing overall survival (D). P values were calculated using log-rank tests. F-I Flow cytometry shown as contour plots (F, H) and quantified in bar plots (G, I) showing T cells (F, G) or myeloid cells (H, I). Panels compared mice after one week of treatment with control, afatinib, anti-PD-1 antibody, or a combination of both. Bar graphs indicate mean ± SEM. P-values were calculated using a Mann-Whitney test

To test whether these shNf1-expressing melanoma cells are resistant to immunotherapy and whether EGFR inhibition can improve anti-PD-1 response, we randomized the mice into four groups that received afatinib, anti-PD-1 antibodies, afatinib plus anti-PD-1 (combo), or vehicle control (Fig. 7C). The anti-PD-1 antibody did not significantly inhibit tumor growth on its own, suggesting that shNf1-expressing RIM-3 cells are ICI-resistant (Fig. 7D). However, afatinib significantly inhibited tumor growth compared with vehicle control, consistent with results from recently reported PDX models [10]. Combining afatinib with an anti-PD-1 antibody further enhanced the tumor-inhibitory effect of afatinib in the shNf1-expressing RIM-3 model. These responses led to prolonged survival in tumor-bearing mice treated with afatinib monotherapy or afatinib plus anti-PD-1 until the experimental endpoint, whereas mice treated with vehicle control or anti-PD-1 monotherapy reached their humane endpoints sooner due to their rapid growth rates (Fig. 7E).

To determine how these therapies affect immune responses in these tumors, we treated an independent cohort of NF1-depleted RIM-3 melanoma-bearing mice for one week and analyzed changes in myeloid and lymphoid cells by flow cytometry (FACS). We stained single-cell suspensions with antibody panels to detect and quantify T cell populations (CD45, CD4, FOXP3, CD8, CD44, GZMB, Ki67, TIM-3, and PD-1) and myeloid cell populations (CD45, NK-1.1, I-A/I-E, F4/80, CD11b, CD11c, Ly-6 C, and Ly-6G). Our FACS data revealed that afatinib treatment alone did not affect tumor-infiltrating CD4⁺ T cells or CD8⁺ T cells. In contrast, PD-1 inhibition increased CD4⁺ T cell (p = 0.0582) and CD8⁺ T cell (p = 0.0742) infiltration, although the response was variable and not statistically significant. By contrast, combined inhibition of EGFR and PD-1 led to a statistically significant increase in tumor-infiltrating CD4⁺ T cells (p = 0.0123) and CD8⁺ T cells (p = 0.0035), as well as a change in the CD8⁺ T cell/Treg ratio (p = 0.0030). Among T cell activity markers, GZMB increased (p = 0.053) in tumors treated with afatinib and anti-PD-1, but not in tumors treated with afatinib or anti-PD-1 alone (Fig. 7F-G). These results indicate that combination therapy improves the T cell immune response, whereas anti-PD-1 or afatinib alone has no significant effect. Furthermore, afatinib treatment significantly increased neutrophil (p = 0.0203) and NK cell (p = 0.0258) infiltration in Nf1 knockdown tumors (Fig. 7H-I), perhaps due to upregulation of Ulbp1 [43] (Fig. 6K). Anti-PD-1 antibodies had no significant effect on these cells, but effects on NK cells (p = 0.0363), neutrophils (p = 0.0305), and monocytes (p = 0.0166) were augmented by combination therapy (Fig. 7H, I).

In conclusion, our data indicate that increased EGFR expression resulting from NF1 loss leads to immune evasion and, consequently, ICI resistance. Therefore, combining EGFR inhibition with ICI therapy would be a rational approach to improve clinical outcomes in patients with NF1^Mut melanoma.

Discussion

We deployed spatial transcriptomic and proteomic profiling of human melanomas, complemented by functional preclinical models, to define the multicellular architecture of the TME and to reveal its link to immune evasion and immunotherapy response. We identified 12 recurrent MNs as spatially correlated assemblies of transcriptionally distinct melanoma cells that consistently interact with specific stromal and immune populations across samples and patients. These MNs comprise melanoma cells in diverse states, along with subsets of CAFs, endothelial cells, T cells, myeloid populations, and other stromal cells, revealing modular spatial ecosystems that may underlie clinical heterogeneity.

Given the unmet need for more effective treatment of NF1^Mut melanoma [8–11], we focused on how NF1 loss shapes TME organization. We report that NF1^Mut tumors were enriched for MNs marked by heightened EGF–EGFR signaling, NC-like dedifferentiation, more abundant CAFs, a collagen-rich extracellular matrix, and spatial patterns consistent with T cell exclusion. In contrast, NF1^WT tumors exhibited greater infiltration by active, exhausted, and memory CD8⁺ T cells. Our data also reveal that NF1 loss promotes EGFR expression and activation, thereby reducing MHC class I and MHC class II antigen presentation on melanoma cells and, in turn, promoting immune evasion. Consistent with these findings, we observed decreased HLA presentation in NF1^Mut tumors and an inverse correlation between EGFR expression and intratumoral CD8⁺ T cell density.

Our spatial analyses identified fewer MNs containing cytotoxic CD8⁺ T cells, activated macrophages, and Tregs, as well as fewer intratumoral CD4⁺ T cell/CD8⁺ T cell/DC immune triads [32]. Such an immune landscape is more consistent with impaired antigen presentation and T cell exclusion than with activation–exhaustion dynamics, in which exhausted and memory T cells would be expected to accumulate as more readily seen in NF1^WT melanoma tissues. These results refine prior reports that emphasized PD-L1 upregulation as the dominant immune evasion mechanism in NF1^Mut melanoma [44]. Although increased PD-L1 surface expression has been observed in NF1-depleted cultures [44], and desmoplastic melanoma (a rare melanoma subtype typically associated with NF1 mutations, a high tumor mutational burden, and favorable response to immunotherapy [5]), our spatial analyses of intact tissues also identify a more complex and immunologically “silent” phenotype in a significant subset of NF1^Mut melanomas that do not respond to ICIs.

Our spatial analyses suggest that CAFs could function as central orchestrators of the NF1^Mut TME. Melanoma cells are in closer spatial proximity to CAFs in NF1^Mut than in NF1^WT tumors, and their extracellular matrix is more collagen-rich. Ligand–receptor analyses indicated significantly stronger interactions of CAF populations with melanoma cells and other neighboring stromal and immune cells in NF1^Mut tumors. Together, these data suggest that NF1 loss remodels the extracellular matrix and amplifies CAF-driven signaling [45], creating an immune-suppressive, pro-angiogenic niche that physically and functionally restricts lymphocyte access to tumor cells.

Our results also extend the broader literature implicating EGFR in tumor immune evasion across other solid tumors [46–48], highlighting this pathway as a potential target for combination therapy in NF1^Mut melanoma. EGFR was enriched in MNs associated with heightened oncogenic signaling and immune escape. Increased EGFR expression also correlated with immune-evasive transcriptional programs in NF1^Mut melanoma. Functionally, we found that EGFR inhibition can restore the expression of genes required for antigen presentation in NF1^Mut melanoma cells and enhance anti-tumor immunity. In addition to upregulating the MHC machinery [49], EGFR inhibition increased ULBP1 expression, potentially facilitating NK cell recruitment and innate immune responses [50]. These observations, together with the more permissive immune context in NF1^WT tumors, support a therapeutic strategy that combines EGFR inhibition with PD-1 inhibitors to overcome antigen presentation defects and immune cell exclusion, thereby improving therapeutic responses in patients with NF1^Mut melanoma [51]. Results from our syngeneic mouse model support the feasibility and enhanced efficacy of this combinatorial treatment approach.

Our work has immediate implications for clinical translation. First, it delineates genotype-specific spatial ecosystems in melanoma, linking NF1 loss to an EGFR-dependent immune-evasive program marked by diminished antigen presentation, CAF-driven matrix remodeling, and T cell exclusion. Second, it identifies EGFR inhibition as a rational partner to PD-1 blockade in NF1^Mut disease, providing a mechanistic basis for combination therapy. Third, it highlights spatially resolved biomarkers, including EGFR activity, HLA antigen presentation signatures, CAF density and proximity, collagen enrichment, and T cell exclusion metrics, that could guide patient stratification and pharmacodynamic monitoring.

Although NF1^Mut melanoma has been associated with immunosuppressive features, NF1 mutations do not define a uniformly immune checkpoint inhibitor–resistant subtype, as clinical response rates remain broadly comparable to those reported for other genetic melanoma subtypes [6, 9, 10]. NF1 loss also results in increased RAS/MAPK pathway activity [11], providing a rationale for MAPK-directed therapy. However, the clinical activity of MEK and related MAPK pathway inhibitors in NF1^Mut has been limited [52], potentially due to incomplete pathway suppression, adaptive reactivation of the pathway, and compensatory receptor tyrosine kinase signaling. Our results implicate EGFR/ERBB signaling as a potential adaptive dependency in NF1^Mut melanoma. Thus, EGFR inhibition may complement rather than replace MAPK-directed therapy, and combined EGFR and RAS/MAPK pathway inhibition may provide more effective suppression of tumor-intrinsic signaling.

Whether and how immunotherapy should be incorporated into such combinations remains to be determined. Our preclinical data provide a rationale for combining EGFR inhibition with immune checkpoint blockade. However, additional studies will be necessary to directly compare EGFR inhibition with combinatorial EGFR-RAS or EGFR-MEK inhibition and their combinations with ICIs to determine the most effective therapeutic strategy.

We acknowledge some limitations. The Xenium panel tests only ~ 5,000 genes and some signaling pathways may have been missed. Although our dataset presents the largest spatial multi-omic comparison of NF1^Mut and NF1^WT tissues to date, larger patient cohorts will need to be analyzed to better assess the effects of variant-specific NF1 mutations and co-occurring driver mutations. Finally, the TME programs we describe might reflect convergent signaling pathways downstream of NF1 loss or EGFR activation, and similar outcomes could result from alternative mechanisms. Although these mechanisms might be identified in our spatial transcriptomic data, their testing may require additional preclinical model development.

Conclusion

In conclusion, NF1 loss is associated with spatially distinct meta-niches in melanoma characterized by reduced MHC antigen presentation, CAF enrichment, matrix remodeling, and T cell exclusion, features collectively associated with immune evasion. By linking these features to EGFR-dependent programs and demonstrating that EGFR inhibition can restore tumor immunogenicity, our findings provide a rationale for biomarker-driven clinical trials combining EGFR inhibitors with PD-1 blockade in NF1^Mut melanoma, or for adding EGFR inhibitors if NF1^Mut melanomas do not respond to PD-1 monotherapy.

Abbreviations

ADGR

Adhesion G protein-coupled receptor

AUC

Area under the curve

CAF

Cancer-associated fibroblast

DC

Dendritic cell

ECM

Extracellular matrix

EGF

Epidermal growth factor

EGFR

Epidermal growth factor receptor

EMT

Epithelial-to-mesenchymal transition

FFPE

Formalin-fixed, paraffin-embedded

FOV

Field of view

GO

Gene Ontology

GSEA

Gene set enrichment analysis

HLA

Human leukocyte antigen

ICI

Immune checkpoint inhibitor

IF

Immunofluorescence

IFN

Interferon

iCAF

Inflammatory cancer-associated fibroblast

IHC

Immunohistochemistry

MAPK

Mitogen-activated protein kinase

MEK

Mitogen-activated protein kinase kinase

MES

Mesenchymal

MHC

Major histocompatibility complex

MN

Meta-niche

myCAF

Myofibroblastic cancer-associated fibroblast

NC

Neural crest

NF1

Neurofibromin 1

NF1^Mut

NF1-mutant

NF1^WT

NF1 wild-type

NF-κB

Nuclear factor kappa B

NK

Natural killer

NR

Non-responder

PCA

Principal component analysis

PD-1

Programmed cell death protein 1

PD-L1

Programmed death-ligand 1

PDGF

Platelet-derived growth factor

qPCR

Quantitative polymerase chain reaction

R

Responder

STC

Short-term culture

TGFβ

Transforming growth factor beta

TMA

Tissue microarray

TMB

Tumor mutational burden

TME

Tumor microenvironment

Treg

Regulatory T cell

UMAP

Uniform manifold approximation and projection

Authors'' contributions

Project concept and design: M. Ibrahim, A. W. Lund, M. Schober, and I. Osman.Clinical evaluation, pathology evaluation, and selection of the melanoma patients for the experiments: G. Jour and I. Osman.Experiments execution and acquisition of data: M. Ibrahim, I. Illa-Bochaca, T. Muijlwijk, I. Delclaux, K. S. Ventre, P. Angulo-Salgado S. Qiu and A. DuttData analysis and interpretation, figure preparation: M. Ibrahim, T. Muijlwijk, I. Delclaux, K. S Ventre, P. A. Salgado, A. W. Lund, M. Schober, and I. Osman.Writing, reviewing, and/or revising the manuscript: M. Ibrahim, M. Schober, and I. Osman.All authors reviewed the manuscript and gave final approval for publication.

Funding

This project was supported by an NIH Melanoma SPORE grant NCI P50 CA225450 to (I.O.), U54 CA2630001 (to M.S., A.W.L., and I.O.), Melanoma Research Foundation - Fellow Research Grants 1287389 (to M.I.), Dr. Keith Landesman Memorial Fellow Cancer Research Institute (to T.M.), and F30 CA288142-01 A (to K.S.V.). The Genomics Technology Center and the experimental pathology cores are supported by the Cancer Center Support Grant “NIH/NCI 5 P30CA16087”.

Data availability

Xenium spatial dataset is available at GEO dataset (GSE316387): and RNA-Seq datasets: (GSE316551). Custom R code used to reproduce the Xenium spatial transcriptomic analyses and associated manuscript figures is publicly available through GitHub (miladebrahim-afk/xenium-spatial-analysis-ffpe-melanoma) and has been permanently archived in Zenodo (version v1.0.1; DOI: 10.5281/zenodo.22259135).

Declarations

Ethics approval and consent to participate

All protocols were approved by NYU Institutional Review Board at NYU Langone Health (IRB #10362) and were conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from participants to allow using their specimens for research and to conduct follow-up.

Consent for publication

Not applicable as this manuscript does not contain any individual-level data requiring consent for publication.

Competing interests

The authors declare no competing interests.