Spatial multi-omics characterization of neuroblastoma reveals ferroptosis-associated metabolic features in high-risk tumors

Abstract

Background

Neuroblastoma is a tumor of the sympathetic nervous system and is the most common extracranial solid malignancy in children. It displays striking clinical heterogeneity, ranging from spontaneous regression to a more aggressive, treatment resistant disease. While previous studies have highlighted the importance of the tumor microenvironment (TME) in shaping disease behavior, how spatial organization and metabolic pathways contribute to high-risk neuroblastoma remains poorly understood.

Methods

Here, we performed spatial transcriptomics profiling using the GeoMx Digital Spatial Profiler (DSP) Whole Transcriptome Atlas (WTA) and a spatial proteomics profile using the Akoya PhenoCycler-Fusion on a cohort of human pediatric neuroblastoma samples to characterize tumor and TME regions.

Results

By using the GeoMx WTA panel, high-risk neuroblastoma tumor regions were found to exhibit upregulation of metabolic pathways associated with ferroptosis, including fatty acid metabolism and reactive oxygen species (ROS) signaling. However, these tumors also showed increased glutathione metabolism pathway and elevated GPX4 expression, consistent with a potential compensatory response that may limit ferroptosis-associated cell death. Further in vitro experiments showed that inhibition of GPX4 increased lipid peroxidation and reduced tumor cell viability, consistent with ferroptosis-related processes. Interestingly, spatial proteomic analysis revealed distinct spatial niches in high-risk neuroblastoma, including stroma-secluded immune cells and macrophage enriched areas, both of which were correlated with poor patient survival.

Conclusions

Our integrative spatial multi-omics analysis suggests that high-risk neuroblastoma tumors display ferroptosis-associated metabolic features, with GPX4 inhibition inducing neuroblastoma tumor cell death. We also identify macrophage-tumor interactions that may be linked to ferroptosis sensitivity. Collectively, our study highlights ferroptosis-associated pathways as potential therapeutic avenues in neuroblastoma patients.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-026-01622-0.

Background

Neuroblastoma is a tumor of the sympathetic nervous system and is the most common extracranial solid malignancy in children, representing approximately 10% of all pediatric cancer deaths [1–3]. It displays a broad spectrum of clinical behaviors, from spontaneous regression to aggressive progression and resistance to therapy. This variability is influenced by factors, including genetic mutations, MYCN amplification, and segmental chromosomal alterations [4]. Despite intensive multimodal therapy, patients with high-risk neuroblastoma have much poorer outcomes compared with patients with low- and intermediate-risk neuroblastoma. The overall survival rate for patients with high-risk neuroblastoma remains below 50% [5–7], and recurrences happen in approximately 60% of high-risk patients [6, 8, 9], with the overall survival rate for relapsed patients being only 10% [10].

The tumor microenvironment (TME) plays a pivotal role in determining disease behavior and therapeutic response [11]. Multiple studies have described the cellular and molecular complexity of the neuroblastoma TME, revealing immune and non-immune components of the TME as critical determinants of neuroblastoma progression and therapeutic response [12, 13]. For example, a recent study by Wienke et al. [12] revealed a detailed composition and functional profile of immune cells in neuroblastoma and demonstrated a significant benefit of anti-TIGIT/PD-L1 combination therapy in their models of neuroblastoma [12]. Additional work by Qiu & Matthay [13] has shown that neuroblastoma progression is shaped by a complex immune landscape involving natural killer (NK) cells, T cells, macrophages, and myeloid-derived suppressor cells (MDSCs), which can either promote anti-tumor immunity or foster immune evasion [13, 14]. In addition to immune components, cancer-associated fibroblasts (CAFs) have been shown to correlate with aggressive clinical features in neuroblastoma, including clinical stage, MYCN amplification, and bone marrow metastasis, underscoring their role in shaping a tumor-promoting microenvironment [15]. Despite these advances, our understanding of the intricate interplay between neuroblastoma cells and the immune system remains incomplete, highlighting the need for innovative research approaches. Particularly, the spatial architecture and intercellular communication within the TME in neuroblastoma remain incompletely understood. Additionally, the striking differences in clinical outcomes for patients with high-risk vs. low-risk neuroblastoma led us to investigate how the TME differs between these subtypes and patients.

Beyond immune contexture, emerging evidence indicates that metabolic reprogramming, particularly ferroptosis regulation, plays a role in neuroblastoma aggressiveness [16]. Ferroptosis is a regulated form of cell death characterized by iron-dependent lipid peroxidation [17]. GPX4 serves as the central regulator of ferroptosis, acting as the glutathione peroxidase capable of reducing lipid hydroperoxides in phospholipids and cholesterol hydroperoxides to their corresponding alcohols [18]. Through this activity, GPX4 preserves membrane integrity and redox balance, preventing the uncontrolled accumulation of lipid peroxides that trigger ferroptotic cell death. A recent study has shown that MYCN-amplified neuroblastoma displays altered iron metabolism and redox balance, rendering it sensitive to ferroptosis [16]. In particular, MYCN amplification increases cysteine demand through enhanced cystine uptake and transsulfuration, while GPX4 has been identified as critical survival mechanism that protect neuroblastoma cells from ferroptotic cell death by detoxifying lipid peroxides [16]. While these findings establish ferroptosis as a promising avenue for therapeutic targeting, they leave unresolved how ferroptosis-related pathways are altered within primary human tumors and how the TME modulates ferroptosis susceptibility. In particular, the spatial and microenvironmental regulation of ferroptosis remains largely unexplored.

Our study addresses this gap by investigating how ferroptosis may be impacted by the TME. We used integrative spatial multi-omics, combining NanoString GeoMx Digital Spatial Profiler (DSP) Whole Transcriptome Atlas (WTA) and Akoya PhenoCycler-Fusion proteomic profiling [19, 20], to comprehensively characterize both tumor and non-tumor regions within the TME of pediatric neuroblastoma samples. We investigated how spatially localized tumor-immune cell interactions, metabolic pathways and ferroptosis-associated signaling differ between high-risk and low-risk tumors. Our study identifies ferroptosis-associated metabolic reprogramming and highlights macrophage–tumor interactions via glutamate signaling as potential modulators of ferroptosis sensitivity. Together, these findings provide a spatial framework linking tumor metabolism and immune contexture to neuroblastoma aggressiveness, offering new avenues for ferroptosis-targeted and macrophage-directed therapeutic strategies.

Methods

Patient cohort and study design

Formalin-fixed paraffin-embedded (FFPE) neuroblastoma tumor samples, including primary tumors and matched pre- and post-treatment specimens, from 27 pediatric patients were assembled into three tissue microarrays (TMAs), comprising a total of 29 tumor cores. Samples were obtained from the institutional biobank of the Associação Hospitalar de Proteção à Infância Dr. Raul Carneiro (Hospital Pequeno Príncipe, Curitiba, Brazil). Patients were diagnosed between 2004 and 2014. The requirement for informed consent was waived by the institutional ethics committee (Parecer No. 5.918.107; CAAE 80073124.9.3001.0097/2025), in accordance with Brazilian National Health Council Resolution 466/12. Representative tumor regions were selected by a board-certified pathologist based on histopathological evaluation of FFPE specimens derived from diagnostic biopsies or surgical resections, as applicable. Selected regions were required to contain at least 70% viable tumor tissue prior to TMA construction. TMA 1 contained poorly differentiated, treatment-naïve primary tumors, and TMA 2 included differentiating, treatment-naïve primary tumors. TMA 3 consisted of matched pre- and post-treatment tumor samples, enabling comparison of treatment-induced changes. Clinical data collected included age at diagnosis, histology, sex, treatment status, recurrence status, staging, MYCN amplification status, and all patients had a minimum follow-up of five years (Supplementary Table S1). Risk stratification was performed according to the International Neuroblastoma Risk Group (INRG) classification system. Patients were categorized as low-risk, low/intermediate-risk, or high-risk. For comparative analyses, low-risk and low/intermediate-risk cases were combined and referred to as the low-risk group. In this study, we primarily compared high-risk and low-risk neuroblastoma to identify spatial and molecular differences in tumor organization, immune composition, and pathway activation. Serial sections from all TMAs were used for spatial transcriptomic profiling using the NanoString GeoMx DSP WTA platform and for multiplexed spatial proteomic profiling using the Akoya PhenoCycler-Fusion platform [19, 20].

NanoString GeoMx WTA

NanoString GeoMx DSP WTA study design

The FFPE tissue TMA slides were processed by NanoString Technologies GeoMx DSP technology using the human WTA panel following the manufacturer’s instructions. Fluorescent morphological markers SYTO 13, CD45, CD56, and synaptophysin were used to visualize the nucleus, immune cells, and tumor cells, respectively, to select the region of interest (ROI) and area of illumination (AOI). ROI selection was guided by morphology markers to capture the AOIs: tumor compartments (CD56 + and synaptophysin+) and TME compartments (CD45+) across tissue cores. ROIs were full of tumor cells or immune cells and were selected and defined as tumor compartment (CD56+ and synaptophysin+) and TME compartment (CD45+), respectively. ROIs containing a mixture of tumor cells and immune cells were classified as either TME AOI (CD45+) or tumor AOI (CD56+ synaptophysin+) for each ROI. DNA barcoded oligo-conjugated tagged RNA detection probes on tissue cores were cleaved and collected for sequencing using Illumina next generation sequencing.

Bioinformatics analyses for NanoString GeoMx DSP WTA data

Quality control and normalization

The probeQC counts for each ROI/AOI were imported into R (v4.3.2) and the established standR (v1.2.0) workflow [21] for spatial transcriptomics was implemented to conduct quality control, normalization and downstream analysis. Briefly, a total of 69 ROIs from 24 patients and 3 slides were analyzed. Gene level filtering did not remove any low expressing genes (lowly expressed in > 90% of the ROIs, while for the ROI filtering, 2 ROIs with nuclei count < 200 were removed (Supplementary Fig. S1a). The relative log expression (RLE) plots and principal components analysis (PCA) were utilized to evaluate the overall gene distribution and to identify any confounding factors or the presence of unwanted batch effects across all 3 TMAs (Supplementary Fig. S1b). The tumor and TME ROI regions were clearly separated along PC1 and PC2 (44.23% across PC1 and 18.13% across PC2) (Supplementary Fig. S1c). When examining the risk groups, which already consider factors such as age and MYCN amplification, the differences in risk groups are mainly accounted for by PC2 (18.13% across PC2) (Supplementary Fig. S1d). The data was normalized using the Trimmed Mean of M-values (TMM) method [22] to adjust for library size variations and unwanted compositional biases.

Differential expression analysis

Differential expression (DE) analysis was performed using a linear modeling approach by utilizing the R packages edgeR (v4.2.1) [23] and limma (v3.60.4) [24]. Briefly, a linear model was fitted to the experimental design to account for biological factors of interest, clinical metadata, and patient-patient variations. The limma-voom pipeline with limma::duplicatecorrelation pipeline was used to model the variation in gene expression and to account for intra and inter patient correlations [24]. The models were then fitted using limma::lmfit to identify differential expression between comparisons of interest. The statistical significance was calculated using the Benjamini Hochberg procedure to account for multiple testing with the threshold as an adjusted P-value of ≤ 0.05 used to identify significant DE genes. The resulting statistic was an empirical Bayes moderated t-statistic, a more robust measure than the t-statistic from a classical t-test.

Gene Set Enrichment Analysis (GSEA)

Utilizing the fry method from the limma package, Gene Set Enrichment Analyses (GSEA) were conducted [24]. Pathways representing the biological themes of relevance were identified and were assessed to uncover clinical insights of relevance to the dataset and experiment.

Spatial proteomic profiling using Akoya PhenoCycler-Fusion

Image analysis

High plex spatial phenotyping TMA slides at single-cell resolution was conducted using the PhenoCycler-Fusion following the manufacturer’s instructions. This experiment was performed by Enable Medicine (US). The TMA slides were stained with 53 oligo-conjugated antibodies for immune cell, functional and structural markers (Supplementary Table S2) and underwent cycles of staining, imaging, and removing three fluorescent oligo-conjugated reporters at a time. Upon completion, the images were combined into one qptiff file. Images were visualized using Qupath (v0.4.3) [25], which in order to assess image quality control by removing TMA cores of low quality that had strong non-specific fluorescence in areas and were broken or necrotic. Cell segmentation in Qupath was the first step in image analysis. A new cell segmentation model was trained with Cellpose [26] plugin on DAPI channel based on the ‘nuc’ pretrained model. Cell segmentation was performed using the self-trained Cellpose model in Qupath. For analysis, a cell metrics table with the nuclear size, spatial coordinates, and universally unique identifier (UUID) codes for each cell, as well as the median fluorescent intensity per cell, was exported.

Cell classification

Cell matrices exported from Qupath and cell metadata were imported into Anndata [27] format for further quality control/preprocessing, cell clustering and cell phenotyping. By using a DAPI signal threshold (DAPI signal falling between 10 and 250 and followed by nucleus size exclusion < 11µm² and > 220 µm²), artifactual nuclei were removed. Markers that possessed low signal to noise and generated artifacts in unsupervised clustering were excluded. Expression matrices were transformed using arcsinh (cofactor 150), scaled within columns (markers), then scaled across rows (cells) according to recommended methods for PhenoCycler-Fusion data pre-processing [28]. Data was then integrated using the Scanpy integration of Harmony [29] and adjusted principal components used to cluster data with Leiden. Cell types were determined by manually assigning them based on protein expression patterns for individual cell clusters through hierarchical ordering on a heatmap and using a cell-typing panel (Supplementary Table S3). Cell type clusters annotated were imported back into Qupath and matched by their UUID for visual inspection and ground truth quality control.

Neighborhood analysis

Cell frequency was determined by aggregating the occurrences of each cell class and then adjusting for the total number of cells in each core to obtain cell percentages. The Cell Neighborhood (CN) analysis was conducted to identify the cell types of the 10 nearest nearby cells [30]. These cell types were subsequently assigned as the defining features of the target cell. The attributes were analyzed using an unsupervised K-nearest neighbors (KNN) algorithm and categorized into 10 groups. The relative abundance of each cell type was quantified and shown as a heatmap. The frequency of each neighborhood in each core was calculated, and t-tests were employed to identify statistically significant differences in neighborhoods regarding biological changes and clinical outcomes.

Spatial proximity analysis uses cross-type nearest neighbor distance distribution function (G-cross)

To quantify the spatial interaction strength between specific cell types within tissue sections, we applied the G-cross function [31]. The G-cross function estimates the probability that a cell of type j lies within a given radius r of a cell of type i, producing a cumulative distribution function of spatial proximity. For each pairwise cell type comparison, we computed the G-cross function up to a radius of 20 μm. The area under the resulting G-cross value (Gauc) summarizes the relative proximity of cell type j to cell type i, with higher values indicating closer and more frequent spatial association. Analyses were performed at two spatial scales: (1) across the whole tissue core per patient and (2) within spatially defined cellular neighborhoods.

TARGET database analysis and gene set scoring

RNAseq expression data and clinical characteristics for the primary tumor samples collected at the diagnosis stage of pediatric neuroblastoma patients were collected from the TARGET database (Target-NBL database, n = 112) (https://www.cancer.gov/ccg/research/genome-sequencing/target/usingtarget-data). The optimal cut-off for the single factor in the survival analysis was determined using the ctree function with the party (v1.3.17) R package [32]. The analysis was performed using the survival (v3.7.0) R package [33], and the resulting Kaplan–Meier (KM) curves were plotted using the survminer (v0.4.9) R package [34]. Gene set scoring was performed to calculate glutathione metabolism score using the singscore (v1.24.0) R package [35]. In short, genes are sorted in ascending order based on their transcript abundance. Glutathione metabolism gene signature was obtained from Human Gene Set: KEGG_GLUTATHIONE_METABOLISM [36, 37].

Single cell RNA sequencing (scRNAseq) analysis and CellChat analysis

Neuroblastoma scRNAseq dataset was downloaded from Patel et al. [38], obtained from https://humantumoratlas.org/. Using the escape (v2.2.0) R package [39], single sample Gene Set Enrichment Analysis (ssGSEA) was conducted, and hallmark gene sets were obtained from the molecular signature database (MSigDB). To estimate the cell-cell communication between different cell types, we applied the CellChat R package (v2.1.2) to scRNAseq data [40]. The significant ligand-receptor interactions were visualized by using netVisual_bubble function and netVisual_aggregate function for dot heatmap and chord diagram, respectively.

Cell culture

Neuroblastoma cell lines SK-N-AS and SK-N-BE (2) were cultured in RPMI 1640 medium supplemented with 1% penicillin-streptomycin (Lonza, #17-745E), 1% Glutamax (Gibco, #35050-061), and 10% fetal bovine serum (FBS) (Gibco, #10099141). Cells were seeded in 24-well plates and allowed to adhere overnight before treatment with either DMSO (control) or GPX4 inhibitors (10 µM ML210, FIN56, or FINO2) for overnight. Following treatment, tumor cell death and lipid peroxidation were assessed as described below, and tumor cells were stained with C11-BODIPY^581/591 for flow cytometric analysis. For lipid peroxidation measurement, cells were labeled in RPMI 1640 medium with 1 µM C11-BODIPY581/591 at 37 °C, 30 min before the experiment.

IncuCyte analysis

SK-N-AS and SK-N-BE(2) cells were seeded in a 96-well plate at a seeding density of 8 × 10³ cells/well and were allowed to attach overnight. Cells were then treated with 10 µM of ML210, FIN56 or FINO2 and propidium Iodide (PI) (BioLegend, #421301) was added to assess cell viability. The plate was placed in the Sartorius Incucyte S3 live cell analysis instrument and images were captured every 2 h for 56 h.

THP-1 cell differentiation and co-culture with neuroblastoma cell lines

THP-1 cells were differentiated into macrophage-like cells (THP-1 macrophages) by incubation in the presence of 150 ng/ml of phorbol myristate acetate (PMA) for 24 h in serum-free media, which leads to M0 macrophage. M0 macrophages were then polarized to M1 with 20 ng/ml of IFN-γ and 10 pg/ml of Lipopolysaccharide (LPS) and to M2 with 20 ng/ml of IL-4 and 20 ng/ml of IL-13 for 48 h in serum-free media Neuroblastoma cells (SK-N-AS or SK-N-BE(2)) were labeled with CellTrace Violet (CTV) for 30 min in 37 degrees. Then, THP-1 differentiated M0, M1 and M2 macrophages were co-cultured with CTV labeled neuroblastoma cells (SK-N-AS or SK-N-BE(2)) in 1:1 ratio overnight.

Preparation for bulk RNAseq and analysis

RNA was extracted from ex vivo cultured neuroblastoma cell lines using the RNeasy Plus Mini Kit (QIAGEN, #74134), according to the manufacturer’s instructions. RNA quality was checked using the Agilent Fragment Analyzer 5200 standard sensitivity assay. All had RIN scores over 7, showing high quality RNA was extracted. RNA was then normalized for prep to an input of 63ng per sample. Libraries were prepared using Illumina’s Stranded mRNA prep Kit with 10nt UDIs. Final libraries were quality checked using the Agilent Fragment Analyzer 5200 NGS assay, before normalization and sequencing on an Illumina NovaSeq 6000 to generate a minimum of 30 million reads per sample. Fastq files underwent quality checks using FastQC then trimmed using Trimmomatic (v1.4) [41] with the following parameters: removal of low-quality bases from the start and end of reads (LEADING = 20, TRAILING = 20) and a minimum read length threshold of 35 bases (MINLEN = 35). The trimmed reads were then aligned to the GRCh38 (hg38) human reference genome using HISAT2 (v2.2.1) [42]. A raw counts matrix was generated from mapped reads using featureCounts (v2.0.5) [43]. Differential expression was computed using DESeq2 (v1.38.3) [44] with the contrast set to ML210-treated tumor cells versus control tumor cells.

Statistical analysis

All other statistical analyses were performed using GraphPad Prism (v10.1). Data are presented as mean ± standard error of the mean (SEM). For comparisons between two groups, Welch’s t-test was used to assess differences between two groups. Statistical analysis in spatial GeoMx WTA and proteomics dataset, as well as scRNAseq dataset (Wilcoxon rank sum test) of marker gene expression difference and cell proportions was performed using R (v4.3.2), where Benjamini‒Hochberg was conducted.

Results

Identification of differentially expressed genes within tumor and TME regions between high-risk and low-risk neuroblastoma using spatial transcriptomics

The study cohort comprised 27 neuroblastoma patients represented by 29 tumor cores assembled into three TMAs (Fig. 1; Supplementary Table S1). We performed spatial transcriptomic profiling using NanoString GeoMx WTA platform on serial sections of the three TMAs. Regions of interest (ROIs) were segmented by staining for CD56 and Synaptophysin (tumor regions) or CD45 (non-tumor TME regions) (Fig. 2a). Gene expression analysis of PTPRC (CD45), NCAM1 (CD56) and SYP (Synaptophysin) confirmed robust segmentation of tumor and non-tumor TME regions (Fig. 2b). After quality control assessment (ROIs < 200 cells removed) and normalization, we retained 67 ROIs for analysis (Supplementary Fig. S1a-b). Principal component analysis confirmed clear segregation between tumor and TME compartments, with risk-related variance, such as age and MYCN amplification, captured primarily along PC2 (Supplementary Fig. S1c-d).

Fig. 1.

Clinical characteristic of neuroblastoma patients in our study. N = 29 patient cores, N = 27 patients. This study profiled serial sections from three tissue microarrays (TMAs) using PhenoCycler-Fusion (CODEX) and Nanostring GeoMx Whole Transcriptome Atlas (WTA). The analysis included 29 tumor tissue cores from 27 neuroblastoma patients, representing poorly differentiated (TMA 1) and differentiating (TMA 2) tumors, along with neuroblastoma samples collected before and after treatment (TMA 3). TMA 1 and TMA 2 consist of pre-treatment samples, while TMA 3 contains tissue samples collected both pre-treatment and post-treatment from the same patients

Fig. 2.

Differential gene expression analysis for comparing pre-treatment neuroblastoma patients at high-risk to low-risk. a Regions of interest (ROI) selection for the GeoMx Digital Spatial Profiler (DSP) platform. Tumor region (CD56 + synaptophysin+) in green and tumor microenvironment (TME) region (CD45+) in blue. b Representation of CD56 (left) and CD45 (right) expression in the tumor and TME regions. c and d Distribution of differentially expressed (DE) genes as a function of the average transcript expression (log2) and fold change (log2) identified in the comparison of high-risk versus low-risk in the tumor (c) and TME (d) regions. Differential expression genes were derived using the voom-limma ebayes pipeline with duplicatecorrelation to account for within sample correlations, where P < 0.05 was considered as a threshold for significant differences

We conducted differential expression (DE) analysis of high-risk versus low-risk neuroblastoma patients within tumor regions (Fig. 2c), or within the non-tumor TME regions (Fig. 2d). Within tumor regions, 130 DE genes were elevated in high-risk neuroblastoma patients, whereas 28 DE genes were downregulated (Supplementary Table S2). Within tumor regions, DE genes elevated in high-risk patients were involved in cell proliferation (MYC, RACK1, LAPTM4B, RANBP1, AKT2), metabolism (GPX1, GPX4, PRDX4, LDHB) and immune responses (IGKC, IGHG4, IGHG2, MIF) (Fig. 2c). In contrast, genes involved in cell adhesion (ITGA7, NECTIN3, CD9) and neural function such as PIRT (involved in peripheral nerve function), ACHE (a neurotransmitter) and KCNC4 (regulating neurotransmitter release) were downregulated (Fig. 2c).

In the TME regions, high-risk neuroblastoma showed downregulation of the HLA-DQB1 gene, which is implicated in antigen presentation, and increased expression of the chemokine gene CXCL9, and IFN-γ inducible gene GBP2 (Fig. 2d).

In our cohort, we have two matched post- and pre-treatment tumor pairs, both from tumor region in TMA 3 that passed quality control, from high-risk neuroblastoma cases that had undergone conventional chemotherapy without immunotherapy. DE analysis of these paired samples revealed 68 upregulated and 22 downregulated genes following treatment (Supplementary Fig. S2; Supplementary Table S3). Post-treatment tumors exhibited increased expression of pro-survival and mesenchymal remodeling genes, including BCL2, CALM1, and CCNI (cell-cycle and survival), as well as COL6A1, SOX4, and FMNL2 (extracellular-matrix and cytoskeletal reorganization). In contrast, stress-response and differentiation-associated genes such as HSPA1A, HSPA1B, HSPB1, DDIT4, TIMP1/TIMP3, and DBH were downregulated in post-treatment high-risk neuroblastoma tumors.

Identification of pathways that might contribute to tumor progression, migration and aggressiveness in high-risk neuroblastoma

Next, we performed GSEA for tumor regions to identify pathways associated with disease aggressiveness. High-risk neuroblastoma tumors showed increased activation of TNFα signaling via NF-kB pathway, which is known to promote invasion and spread of tumor cells [45–47] (Fig. 3a). On the other hand, Gene Ontology (GO) terms related to cell-cell junctions, which is known to be involved in cell migration [48], was downregulated in the tumor regions of high-risk neuroblastoma (Fig. 3b). Moreover, within the tumor regions, several hallmark pathways regulating cell cycle and cell proliferation were upregulated in high-risk neuroblastoma, including MYC targets V1 and V2 [49], E2F targets [50, 51] and G2M checkpoint [52] (Fig. 3a; Supplementary Table S4). The upregulation of both MYC targets V1/V2 pathway (Fig. 3a) and MYC expression itself (Fig. 2c) is consistent with previous reports of MYC-driven high-risk neuroblastoma [53], validating our analytical approach.

Fig. 3.

Identification of pathways that might contribute to tumor progression, migration and aggressiveness in high-risk neuroblastoma. a Visualization of significantly upregulated Hallmark pathways in tumor regions between high-risk versus low-risk neuroblastoma. b Visualization of significantly downregulated GOBP pathways in tumor regions between high-risk versus low-risk neuroblastoma. c Schematic view of the metabolism pathways that either promote or against ferroptosis. Illustrations in c were made with ©Biorender (BioRender.com). d Ferroptosis related genes expression in tumor regions of neuroblastoma samples. P-values were calculated using the Wilcoxon rank-sum test and adjusted with the Benjamini-Hochberg. e Visualization of significantly upregulated KEGG pathways in tumor regions between high-risk versus low-risk neuroblastoma. GSEA was performed using Hallmarks, c2, c5 and c7 collections from the MsigDB, with false discovery rate (FDR) adjusted P ≤ 0.05. f Correlation between glutathione metabolism score within tumor regions and the GPX4 gene expression within tumor regions. g Kaplan–Meier overall survival analysis of paediatric neuroblastoma patients stratified by GPX4 expression levels from TARGET neuroblastoma project. A P-value < 0.05 was considered for significant differences in the Kaplan–Meier surival analysis, and the exact P value is indicated in the respective figure

Interestingly, within tumor regions, high-risk neuroblastoma showed coordinated upregulation of several metabolic pathways, including oxidative phosphorylation, hypoxia, glycolysis, fatty acid metabolism, peroxisome, and reactive oxygen species (ROS) pathway (Fig. 3a). These pathways collectively reflect an increase in oxidative and lipid metabolic activity, which are linked to key processes involved in ferroptosis regulation [17, 54] Ferroptosis is defined as a cell death that is triggered by lipid peroxidation, which is facilitated by the formation of ROS, oxidative stress, fatty acid metabolism, and iron metabolism [17] (Fig. 3c). Additionally, peroxisomes have also been demonstrated to promote ferroptosis sensitivity in ovarian and renal cancer cells [55, 56]. Here, our spatial analysis indicates enrichment of ferroptosis-related metabolic pathways within tumor regions of high-risk neuroblastoma. However, we also observed an upregulation of the glutathione metabolism pathway in the tumor regions of high-risk neuroblastoma (Fig. 3c), which is known to provide a protective antioxidant mechanism that prevents ferroptosis.

To further explore this, we assessed the expression of the genes known to be involved in either promoting or suppressing ferroptosis. High-risk tumor regions exhibited higher expression of iron metabolism regulation genes (NCOA4, TFRC, FTH1) and lipid-peroxidation enzyme, ACSL4, an enzyme that helps process long-chain fatty acids and involve in lipid peroxidation [57] (Fig. 3d), while the pro-ferroptotic gene POR was reduced (Fig. 3d). Conversely, anti-ferroptotic genes (SLC1A5, CBS, GPX4) involved in glutathione metabolism expressed higher in the tumor areas of high-risk neuroblastoma (Fig. 3d). These genes have been reported to participate in glutathione metabolism and regulation of ferroptosis in other contexts [17, 58–60].

Given the key role of GPX4 in regulating glutathione metabolism to against ferroptosis, we assessed its correlation with the glutathione metabolism score. High-risk neuroblastoma tumors displayed a higher level of glutathione metabolism score (Supplementary Fig. S3), which is consistent with the GSEA result displayed in Fig. 3e. GPX4 expression was positively correlated with the glutathione metabolism score (Fig. 3f). Comparing ferroptosis related genes expression in tumor regions of MYCN-amplified high-risk vs. MYCN non-amplified low-risk neuroblastoma samples, we found that POR expression is lower in the high-risk MYCN-amplified group. Additionally, this high-risk group has greater levels of SLC1A5 and CBS than the low-risk group (Supplementary Fig. S4). Furthermore, Kaplan-Meier survival analysis of the TARGET neuroblastoma dataset showed that pediatric neuroblastoma patients with high expression of GPX4 had a poorer overall survival (Fig. 3g).

Together, our findings suggest that high-risk neuroblastoma tumors display metabolic features associated with ferroptosis-related pathways, however elevated glutathione metabolism and GPX4 expression may reflect an enhanced antioxidant activity that could limit ferroptosis susceptibility.

Glutathione metabolism is elevated in high-risk neuroblastoma tumors

To further validate the ferroptosis-related pathways and genes identified in our GeoMx dataset, we analyzed an independent scRNAseq neuroblastoma dataset from Patel et al. 2024 (Fig. 4a) [38]. From this dataset, we extracted tumor cells only, enabling us to focus on the transcriptional changes to tumor cells rather than those contributed by the immune microenvironment. Consistent with our GeoMx WTA dataset findings, high-risk neuroblastoma tumor cells displayed significantly higher expression of the anti-ferroptotic gene GPX4 compared with low-risk tumor cells (Fig. 4b).

Fig. 4.

Validation of ferroptosis-related gene expression and pathways using scRNAseq dataset. a UMAP plot of scRNAseq data showing the maligant cells, immune cells, stroma cells and endothelial cells being sequenced from patients with neuroblastoma. b Violin plots showing expression of GPX4 gene across high-risk or low-risk maligant cells from the scRNAseq neuroblatsoma dataset. c, d, e and f Violin plots showing the relative enrichment of the indicated Hallmark gene sets, calculated using single-sample Gene Set Enrichment Analysis (ssGSEA). Enrichment scores for each gene set are plotted by risk groups, labeled as “High-risk” and “Low-risk.” Comparisons between groups were performed using Welch’s t-test with Benjamini-Hochberg correction, where P-values ≤ 0.05 was considered as significant

Next, we applied ssGSEA to compute pathway enrichment analysis for tumor cells, allowing a direct comparison of ferroptosis-associated metabolic pathways between risk groups. High-risk tumor cells exhibited increased activity in pathways linked to oxidative metabolism, including fatty acid metabolism and ROS pathways, both of which have been reported to influence ferroptosis susceptibility (Fig. 4c-d) [17, 54]. On the other hand, the high-risk tumors also exhibited increased ROS detoxification and glutathione metabolism pathways, consistent with increased antioxidant pathway activity that may influence ferroptosis sensitivity (Fig. 4e-f).

Collectively, our data suggest that while high-risk neuroblastoma tumors display transcriptional features consistent with metabolic processes associated with ferroptosis, they also upregulate protective antioxidant mechanisms, highlighting a complex ferroptosis-related equilibrium.

GPX4 inhibition suppressed cell growth and induced lipid peroxidation in neuroblastoma cells

Given that both GeoMx WTA and scRNAseq datasets showed higher GPX4 expression in high-risk tumors compared to low-risk tumors, we next investigated how a GPX4 inhibitor would impact neuroblastoma cells. MYCN-amplified (SK-N-BE(2)) and MYCN-nonamplified (SK-N-AS) neuroblastoma cells were treated with the GPX4 inhibitor ML210, followed by bulk RNAseq analysis to assess transcriptional changes. To determine the ML210’s effect on the neuroblastoma tumor cells, DE analysis was performed comparing ML210-treated versus control in both cell lines. The DE genes due to ML210 treatment of two cell lines and the overlaps between them are shown in the Venn diagram (Fig. 5a).

Fig. 5.

The impact of GPX4 inhibitors on human neuroblastoma cell lines. a Venn diagram illustrating the overlapping differential expression genes between ML210 treated SK-N-BE(2) cells versus control and ML210 treated SK-N-AS cells versus control. b and c Distribution of differentially expressed (DE) genes as a function of the average transcript expression (log2) and fold change (log2) identified in the comparison of ML210 treated SK-N-BE(2) cells versus control (b) and ML210 treated SK-N-AS cells versus control (c). d-f Neuroblastoma cell lines, SK-N-AS and SK-N-BE(2), were treated with GPX4 inhibitors ML210 (d), FIN56 (e) and FINO2 (f) for 56 h. Cell death was assessed by counting propidium iodide-positive (PI+) cells using live-cell imaging with the IncuCyte S3 system. g–i The area under the curve (AUC) for PI+ cells was calculated for each sample and compared between groups. Statistical significance of AUC differences was determined using a one-way ANOVA, followed by Tukey’s multiple comparison test. j-I Representative histogram showing the generation of lipid peroxidation determined by the C11-BODIPY (581/591) probe. m-o Bar plots showing the Lipid oxidates C11-BODIPY levels. Each dot represents one technical replicate. Statistical significance was determined by Welch’s test, where * P < 0.05, and ** P < 0.01, *** P < 0.001, **** P < 0.0001

In SK-N-BE(2) cells, ML210 treatment upregulated genes involved in iron metabolism (TFRC, FTH1, FTL, HMOX1) and lipid peroxidation gene (CPT1A) (Fig. 5b), consistent with increased expression of genes involved in iron and oxidative metabolism. Compared to control samples, ML210-treated SK-N-BE(2) samples showed downregulation of genes related to cell cycle regulation (CCNA1, SKP2, CDK1, GTSE1) and mitosis and cell division (KIF20A, KIF11, KIF4A, MELK, BUB1, NEK2, INCENP), as well as genes associated with cell proliferation (HMGB3, SPIN1) and cell survival (MCL1, RET) (Fig. 5b).

In SK-N-AS cells, ML210 treatment also reduced expression of cell survival gene (RET) and gene associated with cell migration (RHOC) (Fig. 5c). Conversely, genes are known to inhibit cell growth and migration, TRIM16 and TRIM16L [61], were upregulated following ML210 treatment (Fig. 5c). Furthermore, ML210-treated SK-N-AS exhibited increased expression of oxidative stress response gene (OSGIN1) and the detoxifying enzyme (NQO1), genes previously reported to be associated with oxidative stress responses [62] (Fig. 5c). Our findings collectively showed that the GPX4 inhibition downregulates genes linked to cell growth and proliferation in both neuroblastoma cells, with iron metabolism genes preferentially upregulated in MYCN-amplified neuroblastoma cells.

To further assess the impact of GPX4 inhibition, we evaluated several ferroptosis inducers in neuroblastoma cells. A panel of ferroptosis inducers by targeting or inhibiting GPX4, including ML210 [63], FIN56 [64], and FINO2 [65] were tested for their effects on cell viability using a Incucyte™ live-cell imaging system and quantification of PI-positive cells (dead cells). ML210 induced significant cell death in SK-N-BE(2) cells but had minimal effects on SK-N-AS cells (Fig. 5d, g; Supplementary Fig. S5a). Other GPX4 inhibitors, FIN56 and FINO2, caused significant cell death in both SK-N-BE(2) and SK-N-AS cells, with a stronger effect in SK-N-BE(2) cells (Fig. 5e-f, h-i; Supplementary Fig. S5b-c). We next evaluated lipid peroxidation, a hallmark of ferroptosis, by flow cytometry and found that ML210 and FINO2 increased lipid peroxidation in both SK-N-BE(2) and SK-N-AS cells (Fig. 5j-k, m-n), while FIN56 also enhanced lipid peroxidation in SK-N-AS cells (Fig. 5l, o). These findings demonstrate that GPX4 inhibition increases lipid peroxidation and reduces cell viability in neuroblastoma cells, consistent with ferroptosis-associated processes. However, since rescue experiments using ferroptosis inhibitors (e.g., Ferrostatin-1) or iron chelators (e.g., deferoxamine) were not performed, further validation will be important to confirm whether the observed effects are iron-dependent and lipid ROS–mediated.

Spatial proteomic profiling of neuroblastoma TME indicates neighborhoods of macrophages and stroma-secluded immune cells correspond to poorer survival

To investigate how spatial organization differs between high-risk and low-risk neuroblastoma, we performed multiplexed spatial proteomics profiling of the neuroblastoma TME using a 53 marker PhenoCycler-Fusion panel (Supplementary Table S6). This panel identified 12 major cell types in the neuroblastoma TME, including tumor cells, neutrophils, stromal cells, blood vessels, CD4 T cells, CD4 memory T cells, CD4 T follicular helper (TFH) cells, CD8 T cells, CD8 memory T cells, B cells, and CD68 + macrophages and CD163 + M2-like macrophages (Supplementary Fig. S6a-d; Supplementary Table S6).

Cell proportion analysis revealed that patients with high-risk neuroblastoma tend to have a higher proportion of CD163 + M2-like macrophages and CD68 + M1-like macrophage compared to low-risk patients (Supplementary Fig. S7a-b), both of which are associated with worse 5-year survival rates (Supplementary Fig. S8a-b). No such associations were observed for other immune or stromal cell populations (Supplementary Fig. S7c-k and S8c-k). When comparing MYCN-amplified high-risk and MYCN non-amplified low-risk group, we detected no significant differences in cell type proportions (Supplementary Fig. S9).

To investigate the spatial architecture, we conducted cellular neighborhood (CN) analysis, identifying 10 distinct cellular neighborhoods (CN0 – CN9) (Fig. 6a-b; Supplementary Fig. S10). Cellular neighborhood CN0 (tumor) was more prevalent in the low-risk patients. No other cellular neighborhoods (CN1–CN9) differed significantly between groups (Supplementary Fig. S11). When comparing cellular neighborhood proportion between MYCN-amplified high-risk and MYCN non-amplified low-risk, no significant differences were observed between groups (Supplementary Fig. S12). Importantly, cellular neighborhood CN4, a macrophage-rich niche, and CN6, consisting of stroma-secluded immune cells, were both associated with worse 5-year survival rates and were more prevalent in high-risk patients (Fig. 6c-e). Moreover, patients who experienced tumor recurrence displayed higher proportions of these macrophage-rich and stroma-secluded immune cell neighborhoods compared with those who did not recur (Fig. 6f–g), suggesting that these spatial features may serve as predictive markers of poor prognosis and recurrence risk.

Fig. 6.

Spatial organization of tumor microenvironment (TME) of neuroblastoma. a Cellular neighborhood (CN) analysis identified 10 distinct neighborhoods in neuroblastoma tumor microenvironment. b Representative spatial image showing the identified CNs in neuroblastoma TME. c Representative images showing 5 cell-typing markers on the high-risk tumor core. e and d Kaplan–Meier 5-year survival analysis of paediatric neuroblastoma patients stratified by mean of CN4: macrophage-rich neighborhood precentage (d) and CN6: stroma-secluded immune cells neighborhood precentage (e). f and g Percentage of identified cellular neighborhood CN4 (f) and CN6 (g) compared between non-recurrence and recurrence group. A Wilcoxon test was performed, and P value was adjusted with the Benjamini-Hochberg. A P-value of < 0.05 was considered for significant differences, and the exact adjusted P value is indicated in the respective figures

We next quantified spatial interaction strength using the G-cross function, which computes the likelihood of different cell types being located near tumor cells within 20 μm. Across whole sample cores, tumor cells showed strongest proximity to stromal cells (mean Gauc = 5.304) and blood vessels (mean Gauc = 2.578), indicating preferential tumor-stroma and tumor-vascular co-localization (Supplementary Fig. S13a). Within the tumor-rich neighborhood (CN0), tumor-stroma proximity was high (mean Gauc = 0.8266), whereas interactions between tumor cells and other cell types were weaker (mean Gauc < 1), suggesting that tumor cells preferentially associate with stromal components rather than immune cells (Supplementary Fig. S13b).

In the macrophage-rich neighborhood (CN4), M1 macrophages most closely interacted with M2 macrophages (mean Gauc = 7.635), followed by stromal cells (mean Gauc = 2.733), tumor cells (mean Gauc = 2.451), and CD4 T cells (mean Gauc = 2.077). Similarly, M2 macrophages showed the strongest proximity to M1 macrophages (mean Gauc = 3.912), followed by stromal cells (mean Gauc = 3.278), CD4 T cells (mean Gauc = 2.713), and tumor cells (mean Gauc = 2.261) (Supplementary Fig. S14a-b). These findings indicate distinct TME architectures, with particularly enriched M1–M2 co-localization in macrophage-rich neighborhood, CN4, which was associated with poor prognosis (Fig. 6c-e).

Neuroblastoma tumors can contain two main subpopulations: mesenchymal (MES) and adrenergic (ADRN) subtypes [66]. Therefore, we examined whether distinct MES and ADRN tumor cell regions exist within our patient samples. Using published MES and ADRN gene signatures [67], we calculated MES and ADRN scores for each ROI in our GeoMx WTA dataset using the singscore R package [35]. Based on percentile-based thresholds, ROIs were classified as ADRN-enriched, MES-enriched, or ADRN/MES-mixed. MES and ADRN scores clearly distinguished tumor from non-tumor TME ROIs, yet most tumor ROIs exhibited mixed ADRN/MES profile rather than forming distinct spatially segregated MES- or ADRN-enriched regions (Supplementary Fig. S15). Only three tumor ROIs were ADRN-enriched, and two non-tumor TME ROIs were MES-enriched (Supplementary Fig. S15). Consistent with a previous study, we also observed MES tumor cells had an association with immune cells in the TME while ADRN tumors were less likely to be adjacent to immune cells [38]. These findings suggest that MES and ADRN populations may coexist within the TME, without clear spatial compartmentalization in our GeoMx WTA dataset.

Macrophages increase neuroblastoma tumor cell sensitivity to GPX4 inhibitor–induced cell death

To investigate whether macrophages influence tumor cell sensitivity to ferroptosis, we conducted ligand-receptor interaction analysis using the neuroblastoma scRNAseq dataset from Patel et al. (2024) (Fig. 4b) [38]. This analysis demonstrated cell-cell communication networks and provided evidence for significant ligand-receptor interactions between macrophages and neuroblastoma tumor cells (Fig. 7).

Fig. 7.

CellChat analysis of neuroblastoma scRNAseq dataset reveals macrophages interact with neuroblastoma tumor cells via glutamate signaling. a Chord plot showing the glutamate signaling pathway network among different immune, stroma cells and tumor cells. b Significant glutamate signaling pathways targeting neuroblastoma tumor subtypes. c Violin plots showing the expression levels of glutamate signaling genes in the cell clusters identified in the scRNAseq dataset. d-f Co-culture of neuroblastoma cell lines SK-N-BE(2) and SK-N-AS with M0, M1, M2 in the presence of GPX4 inhibitor FINO2. Tumor cell death was measure by flow cytometry. Statistical significance was determined by Welch’s test, where * P < 0.05, and ** P < 0.01, *** P < 0.001, **** P < 0.0001

CellChat analysis revealed glutamate signaling as a key pathway mediating communication between macrophages and multiple tumor subtypes, including adrenergic, sympathoblast and mesenchymal populations (Fig. 7a). When assessing interactions using various immune and stroma cells identified as sources and tumor subtypes as targets, we found that macrophages uniquely interact with tumors through the Glu-(SLC1A3 + GLS)-GRIA2 pathway (Fig. 7b). Moreover, macrophages express SLC1A3, a glutamate transporter responsible for glutamate, whereas tumor cells do not express SLC1A3, but express GLS (glutaminase), which converts glutamine to glutamate, potentially leading to glutamate accumulation (Fig. 7c). Accumulation of glutamate is known to inhibit system Xc⁻ on tumors, reducing cystine uptake and glutathione synthesis, which can increase ROS levels and contribute to lipid peroxidation, a hallmark of ferroptosis [68]. Our results suggest a potential link between macrophage-tumor interaction via glutamate signaling, whereby macrophages may modulate glutamate metabolism, potentially influencing ferroptosis sensitivity in neuroblastoma.

Next, we investigated whether direct contact of macrophages with neuroblastoma tumor cells affects tumor susceptibility to ferroptosis. To test this, we co-cultured neuroblastoma tumor cells (SK-N-BE(2) and SK-N-AS) with THP-1-derived M0, M1 or M2 macrophages in the presence of GPX4 inhibitor FINO2. Tumor cells were pre-labeled with CTV, and tumor cell death was quantified by flow cytometry. Tumor cell death was gated on single cells/CTV+ cells/dead tumor cells as shown in Supplementary Fig.S16. Results demonstrated that co-culturing tumor cells with macrophages significantly increased FINO2-induced tumor cell death in both SK-N-BE(2) and SK-N-AS cells compared with tumor cells cultured alone (Fig. 7d-f). In the absence of FINO2, co-culture with macrophages resulted in minimal tumor cell death, with less than 10% cell death for SK-N-AS cells and less than 20% cell death for SK-N-BE(2) cells (Fig. 7d-f). Taken together, these findings suggest that direct macrophage–tumor interactions may enhance tumor cell sensitivity to GPX4 inhibitor-induced cell death in vitro. While our results show a potential role for macrophages in modulating ferroptosis susceptibility in neuroblastoma, additional studies will be required to define the precise molecular pathways involved.

Discussion

Patients with high-risk neuroblastoma continue to experience poor survival outcomes despite intensive multimodal treatment regimens [69], underscoring the urgent need to better understand the biological mechanisms driving disease progression and treatment resistance. In this study, we utilized an integrated spatial multi-omics approach combining spatial transcriptomics and spatial proteomics to map the TME and molecular landscape of neuroblastoma across different clinical risk groups. Our findings showed that high-risk neuroblastoma tumors exhibit enrichment of metabolic pathways associated with ferroptosis regulation, such as the ROS and fatty acid metabolism pathways, alongside an upregulation of the glutathione metabolism pathway and increased GPX4 expression, which are associated with antioxidant activity and have been reported to limit ferroptosis in other contexts. Furthermore, GPX4 inhibitors stimulated lipid peroxidation and cell death in human neuroblastoma cells. We also identified spatial niches (macrophage-rich and stroma-secluded immune cell neighborhoods) within the TME of neuroblastoma, which are correlated with poorer survival outcomes.

To capture the broader biological context beyond ferroptosis, we analyzed the gene profiles in both tumor and non-tumor TME regions of low-risk and high-risk neuroblastoma. In tumor regions, we observed upregulation of cell proliferation genes (MYC, RACK1, LAPTM4B, RANBP1, AKT2), which may explain why tumor cells in high-risk neuroblastoma are more aggressive than in low-risk neuroblastoma. Importantly, the correspondence between MYCN amplification and MYC upregulation serves as a positive internal control, confirming that our analysis recapitulates expected oncogenic patterns in neuroblastoma.

Conversely, cell adhesion genes (ITGA7, NECTIN3, CD9) were significantly downregulated in high-risk tumors. Loss of CD9 and NECTIN3, both known as suppressors of metastasis and invasion, could facilitate tumor dissemination and invasion [70, 71]. In addition, downregulation of neural function genes PIRT (affecting peripheral nerve function) [72], ACHE (a neurotransmitter) [73] and KCNC4 (regulating neurotransmitter release)[74, 75], indicates elevated oxidative stress [76] and reduced neuronal features, aligning with the more undifferentiated state of high-risk tumors.

Within TME regions, CXCL9 and GBP2 were upregulated, indicating an inflammatory TME in high-risk neuroblastoma [77, 78]. However, reduced HLA-DQB1 expression suggests impaired antigen presentation. Collectively, these gene-level findings show that high-risk neuroblastoma is characterized by proliferative activation, immune dysregulation, and loss of adhesion, all of which may support tumor aggressiveness.

High-risk neuroblastoma tumors displayed enrichment of iron metabolism and ROS-associated pathways, including upregulation of TFRC, FTH1, NCOA4, and ACSL4, which are critical regulators of lipid peroxidation and ferroptosis. Interestingly, both pro-ferroptotic genes (POR) and anti-ferroptotic genes (e.g., SLC1A5, CBS, GPX4) were upregulated in high-risk neuroblastoma tumors. Examining genes involved in iron metabolism (NCOA4, TFRC, FTH1), these genes are known to involve in ferritinophagy, the process of ferritin degradation that controls cellular iron homeostasis. TFRC promotes iron uptake into the cell, increasing the intracellular iron pool available for ferroptosis. While NCOA4 binds to FTH1, a ferritin subunit, and transports it to lysosomes for degradation, thereby releasing free iron that elevates intracellular ROS levels and enhances ferroptosis sensitivity [79–81]. Rather than being contradictory, this dual expression likely reflects a complex transcriptional pattern involving both pro- and anti-ferroptotic pathways in high-risk neuroblastoma.

Notably, ACSL4 expression was also higher in high-risk neuroblastoma tumor. ACSL4 plays a crucial role in determining ferroptosis sensitivity by activating polyunsaturated fatty acids (PUFAs) and catalyzes PUFAs to PUFA-CoA, which serve as key substrates for lipid peroxidation and ferroptotic damage [57]. Nevertheless, the pro- ferroptotic gene POR was expressed at lower levels in high-risk neuroblastoma tumors than low-risk tumors. This POR gene is essential for the last stage of ferroptosis induction because it facilitates the production of hydrogen peroxide by transferring electrons from NAD(P)H to oxygen, which contributes to phospholipid peroxidation in ferroptosis [82]. Reduced POR expression may indicate an adaptive mechanism by which high-risk tumor cells limit excessive ROS accumulation and lipid damage, maintaining redox equilibrium despite elevated iron metabolism.

Furthermore, the observed strong activation of glutathione metabolism and antioxidant enzymes (GPX4, GPX1, PRDX4, CBS) in high-risk neuroblastoma tumors, suggesting that increased GPX4 expression may be associated with enhanced antioxidant capacity in high-risk neuroblastoma [83]. This is consistent with reports that MYCN-amplified neuroblastoma cells are intrinsically sensitive to ferroptosis due to perturbations in cysteine metabolism and redox homeostasis [16]. By using glutathione as a reducing agent to convert lipid hydroperoxides into non-toxic alcohols, GPX4 suppresses lipid peroxidation and thereby prevents ferroptotic cell death. Supporting this, GPX4 knockdown has been shown to induce ferroptosis in MYCN-amplified neuroblastoma cells, while inhibition of GPX4 or cystine uptake reduces tumor growth in vivo [16, 84]. In our study, high GPX4 expression correlated with poor prognosis, supporting its role in patient survival. These findings suggest that high-risk neuroblastoma may rely on GPX4-mediated detoxification of peroxidized lipids for survival under oxidative stress conditions.

Additionally, we found that anti-ferroptotic genes, SLC1A5 and CBS were expressed higher in high-risk compared to low-risk neuroblastoma tumors. SLC1A5 is a glutamine transporter that promotes glutamine metabolism and can directly contribute to the production of glutathione (GSH), a crucial antioxidant that aids in preventing lipid peroxidation and ferroptosis [59]. Thus, SLC1A5 can increase glutamine uptake and inhibit ferroptosis via the GPX4-related pathway [58]. Consistent with this, MYCN has been shown to induce the transsulfuration enzyme CBS as part of an adaptive mechanism to maintain cysteine availability and suppress ferroptosis under oxidative stress [16]. The elevated CBS expression we observed in high-risk neuroblastomas likely reflects this adaptation, enabling sustained GSH synthesis to support GPX4 activity. From the scRNAseq analysis, we observed the consistent upregulation of GPX4 gene expression and glutathione metabolism pathway in high-risk neuroblastoma tumor cells. Together, these results indicate that high-risk neuroblastomas display transcriptional features consistent with ferroptosis-associated metabolic stress, alongside increased GPX4 expression that may contribute to antioxidant defense. These features may represent a metabolic vulnerability that warrants further functional and preclinical investigation.

Our bulk RNAseq analysis revealed that GPX4 inhibitor ML210 treatment induces iron metabolism genes (TFRC, FTH1, FTL, HMOX1) specifically in MYCN-amplified neuroblastoma tumors, while lipid peroxidation genes (CPT1A) are upregulated in both MYCN-amplified and non-amplified neuroblastoma tumors. This suggests iron metabolism in response to GPX4 inhibitor in MYCN-amplified neuroblastoma tumors. Previous study has linked MYCN amplification to altered iron accumulation and ferroptosis sensitivity in neuroblastoma [85]. Our findings provide further evidence that MYCN-amplified neuroblastoma tumors may display greater involvement of iron metabolism pathways, potentially making them more vulnerable to ferroptosis-inducing therapies.

Next, we demonstrated that GPX4 inhibitors (ML210, FINO2, FIN56) reduced neuroblastoma cell viability and increased lipid peroxidation. These findings align with previous reports where GPX4 inhibitors have been demonstrated to diminish tumor growth and reprogram the tumor microenvironment in a breast cancer mouse model [54]. While our findings are consistent with GPX4 inhibition increasing lipid peroxidation and reducing tumor cell viability in vitro, additional preclinical validation will be necessary to determine whether targeting GPX4 represents a viable therapeutic strategy in neuroblastoma. Rescue experiments using ferroptosis inhibitors (e.g., Ferrostatin-1, Deferoxamine) would also be essential to confirm ferroptotic cell death is induced by GPX4 inhibition, and dose–response studies are needed to precisely quantify ferroptosis sensitivity to these inhibitors. Additionally, lipidomics mass spectrometry quantification of lipid peroxidation is required to validate the lipid ROS detection with C11-BODIPY staining. It would be worth testing other classes of ferroptosis inducers (e.g., erastin and cysteine depletion) in neuroblastoma, to evaluate whether alternative mechanisms of ferroptosis induction could serve as potential therapeutic strategies in this tumor. Future work incorporating these assays will be critical to validate the translational potential of targeting GPX4 in neuroblastoma. Moreover, a recent review has discussed that ferroptosis sensitivity can vary between in vitro and in vivo contexts [86], thus in vivo testing in xenograft models would further strengthen our conclusions. However, establishing humanized neuroblastoma models remains technically challenging and of limited translational relevance. These models often have incomplete reconstitution of the human immune system, replacement of human stromal elements with murine counterparts, inadequate modelling of metastatic progression, and the inability to fully replicate the complexity of the pediatric TME. Consequently, these limitations make it difficult to draw direct correlations between in vivo experimental results and our spatial omics findings.

Spatial proteomics revealed distinct immune architectures between low- and high-risk neuroblastoma. High-risk tumors tend to have more macrophage-enriched spatial niches and stroma-secluded immune niches that were associated with poor survival. High-risk patients also tend to have more M2-like macrophages, and this cell type was associated with poorer survival. M2-like macrophages can induce tumor invasion and metastasis [87], the presence of M2-like macrophages in high-risk tumors may contribute to a more aggressive phenotype and is associated with poorer patient outcomes, as they enhance various aspects of tumor progression and metastasis. Additionally, our data showed that high-risk tumor has upregulation of hypoxia pathway, which are prone to macrophage infiltration [88]. The increased hypoxia signature in high-risk tumors may further promote macrophage recruitment and polarization, reinforcing immune suppression. The stroma-secluded immune niches identified in high-risk tumors may act as physical barriers preventing immune infiltration into tumor cores, which may contribute to the aggressive tumor behavior in high-risk neuroblastoma. More importantly, the association of both macrophage-rich and stroma-secluded immune cell neighborhoods with poorer patient survival further supports that the organization and compartmentalization of immune cells, rather than overall immune cell abundance, influence tumor progression. Collectively, these observations highlight the prognostic significance of spatial organization of immune and stromal components in neuroblastoma. Although the current cohort size constrains statistical power, the enrichment of macrophage-rich and stroma-secluded immune niches in recurrent patients supports their prognostic significance. Future studies with larger patient cohorts will be critical to confirm these patterns and further elucidate how spatial organization impacts disease progression.

Single-cell ligand-receptor analysis revealed that macrophages interact with neuroblastoma tumor cells through the Glu-(SLC1A3 + GLS)-GRIA2 signaling axis. Additionally, macrophages express SLC1A3, a glutamate transporter. However, neuroblastoma tumor cells do not express SLC1A3, but instead express GLS, which converts glutamine to glutamate. Also, in other cancers, such as lung adenocarcinoma, the accumulation of endogenous glutamate following the inhibition of the cystine-glutamate antiporter system X_C⁻ determines ferroptosis sensitivity, making the tumor cells more prone to ferroptosis [89]. Thus, we hypothesize that this interaction may regulate extracellular glutamate accumulation, which could inhibit the cystine/glutamate antiporter (system Xc⁻), deplete intracellular glutathione, and sensitize tumor cells to ferroptosis.

Our co-culture experiments further showed that direct contact between macrophages and neuroblastoma tumor cells enhanced the sensitivity of tumor cells to GPX4 inhibitor-induced cell death, suggesting a possible association between macrophage–tumor interactions and GPX4 inhibitor sensitivity. However, we acknowledge that this proposed mechanism requires additional functional evidence. Future studies assessing glutamate and glutathione levels, gain- and loss-of-function manipulations of SLC1A3 or GLS, and rescue assays using ferroptosis inhibitors in the co-culture system will be essential to validate this pathway.

Conclusions

In conclusion, our integrative spatial multi-omics analysis identifies metabolic and immune features that distinguish high-risk from low-risk neuroblastoma. High-risk neuroblastoma tumors display transcriptional enrichment of pathways linked to ferroptosis-associated metabolic stress, alongside enhanced glutathione-associated antioxidant defenses, including elevated GPX4 expression. In vitro experiments demonstrate that GPX4 inhibition increases lipid peroxidation and reduces neuroblastoma cell viability, consistent with ferroptosis-associated mechanisms, although validation of iron-dependent ferroptotic cell death will require additional functional studies. Furthermore, spatial analysis reveals distinct macrophage-rich and stromal immune architectures associated with poor clinical outcomes and enhanced sensitivity to GPX4 inhibitor–induced cell death in co-culture models. Collectively, these findings highlight ferroptosis-associated pathways and spatial immune organization as biologically relevant features of high-risk neuroblastoma that require further mechanistic and translational investigation.

Abbreviations

ADRN

Adrenergic

AOI

Area of illumination

CAFs

Cancer-associated fibroblasts

CTV

Cell Trace Violet

CN

Cellular neighborhood

DE

Differential expression

DSP

Digital Spatial Profiler

G-cross

Cross-type nearest neighbor distance distribution function

FBS

Fetal bovine serum

FFPE

Formalin-fixed paraffin-embedded

GEO

Gene Expression Omnibus

GO

Gene Ontology

GSEA

Gene Set Enrichment Analysis

GSH

Glutathione

INRG

International Neuroblastoma Risk Group

KNN

K-nearest neighbors

KM

Kaplan–Meier

LPS

Lipopolysaccharide

MES

Mesenchymal

MSigDB

Molecular signature database

MDSCs

Myeloid-derived suppressor cells

NK

Natural killer

PMA

Phorbol myristate acetate

PUFAs

Polyunsaturated fatty acid

PCA

Principal components analysis

PI

Propidium iodide

ROS

Reactive oxygen species

ROIs

Regions of interest

RLE

Relative log expression

ssGSEA

Single sample Gene Set Enrichment Analysis

scRNAseq

Single-cell RNA sequencing

SEM

Standard error of the mean

TMAs

Tissue microarrays

TMM

Trimmed Mean of M-values

TME

Tumor microenvironment

WTA

Whole Transcriptome Atlas

Authors’ contributions

C.T.: Conceptualization, data curation, formal analysis, investigation, visualization, writing–original draft, writing–review and editing. C.W.T., A.C.A., G.A-C., N.O., G.H., J.M., K.C., A.M., A.S., O.V., A.M., C.M.S., F.A.B.P., S.E-E., W.N., L.N., N.J., H.N.H. and Y.C.T.: Investigation and Resources. F.S.F.G., and A.K.: Conceptualization, funding, resources, supervision, writing–original draft, writing–review and editing. All authors read and approved the final manuscript.

Funding

The Guimaraes Laboratory is funded by a grant (#2019485) awarded through the Medical Research Future Fund (MRFF, with the support of the Queensland Children’s Hospital Foundation, Microba Life Sciences, Richie’s Rainbow Foundation, Translational Research Institute (TRI) and UQ), and generous philanthropic funding from the Cooper Rice-Brading Foundation, The Tie Dye Project, The Kids Cancer Project, Tour de Cure, Bricks & Smiles and the PA Research Foundation. Cui Tu, Gabrielle Antonio-Carreon and Kimberly Chung were funded by the Australian Government Research Training Program Scholarships. Cui Tu is funded by a 2025 Australia and New Zealand Sarcoma Association (ANZSA) Project Grant. Natacha Omer is funded by a NHMRC Postgraduate Scholarship (#2021932) and a Children’s Hospital Foundation Mary McConnel Career Boost Program for Women in Pediatric Research Grant (#WIS0092024).

Data availability

The NanoString GeoMx DSP spatial transcriptomic raw sequencing data generated in this study are available in the Gene Expression Omnibus (GEO) repository under accession number GSE270234 [90]. The multiplex spatial proteomic dataset generated using the Akoya PhenoCycler-Fusion platform is available in the Zenodo repository at 10.5281/zenodo.11444750 [91]. The bulk RNA sequencing dataset generated in this study is available in the GEO repository under accession number GSE296939 [92].

Declarations

Ethics approval and consent to participate

The study was approved by The University of Queensland Human Research Ethics Committee (2023/HE000027) and the institutional ethics committee of the Associação Hospitalar de Proteção à Infância Dr. Raul Carneiro (Hospital Pequeno Príncipe, Curitiba, Brazil) (Parecer No. 5.918.107; CAAE 80073124.9.3001.0097/2025). The requirement for informed consent was waived by the institutional ethics committee, as this study utilized de-identified archival formalin-fixed paraffin-embedded tissue samples from the institutional biobank with no direct patient contact, in accordance with Brazilian National Health Council Resolution 466/12. This study was conducted in accordance with the principles of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

FSFG is a Board Member of Cure Cancer Australia Foundation and member of the Scientific Advisory Committee of ANZSA. WN is a member of the board of directors of ANZSA. Authors AM is the CSO and Co-Founder of Enable Medicine. AK is on the Scientific Advisory Board for Omapix Solutions, Predxbio, Molecular Instruments, and Visiopharm. Sanofi TSH, Microba Life Sciences, and Cartherics sponsor research in the laboratory of F.S.F.G. None of these companies have a commercial, proprietary, or financial interest in this study. The remaining authors declare that they have no competing interests.