Deep spatial transcriptomic profiling of ovarian clear cell carcinoma in the real-world setting

Abstract

Ovarian clear cell carcinoma (OCCC) is a rare cancer type of significant relevance to East Asian women harboring critical unmet needs for novel therapeutic options. It is a histological subtype of ovarian cancer with distinct pathological features, molecular profiles, and biological functions. Diverse heterogeneity contributing from histopathological and multiomic molecular features has yet to be translated to guide clinical care. Here, we presented a proof-of-concept study to demonstrate the feasibility of applying deep spatial transcriptomic (ST) profiling of tumor samples from an advanced OCCC patient in the real-world setting, aiming to identify therapeutic options beyond standard-of-care. Matched primary ovarian and metastatic bladder tumor sections were profiled by using GeoMx Digital Spatial Profiling and Xenium In Situ platforms. The spatial architecture and neighborhood niches were identified from GeoMx Cancer Transcriptome Atlas (CTA) and Xenium 5K Human Pan Tissue and Pathways Panel. An immunosuppressive Wnt-activating tumor microenvironment (TME) was identified by GeoMx while a tripartite spatial relationship between SLC2A1+ hypoxic cancer cells, IFIT2+ inflammatory cancer cells, and MMP12+ dendritic cells linking towards metabolism and immune responses was identified by Xenium. Our deep ST profiling findings provided significant biological insights and demonstrated feasibility to make novel discoveries, one patient at a time.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13048-026-02048-3.

Introduction

Ovarian cancer is one of the leading causes of cancer-related death among women worldwide [1]. It remains the most lethal gynecologic malignancy, largely due to its asymptomatic progression, lack of reliable early screening, and frequent diagnosis at advanced stages [1–3]. Among the histological classifications, epithelial ovarian carcinoma (EOC) accounts for approximately 90% of ovarian cancer cases and comprises distinct subtypes such as high-grade serous carcinoma (HGSC), endometrioid carcinoma, mucinous carcinoma, and clear cell carcinoma (OCCC) [4]. The major subtypes of EOC differ in histopathology, molecular landscapes, clinical behaviour, and therapeutic responses, underscoring the need for subtype-specific diagnostic and treatment approaches [5]. Among these subtypes, OCCC represents a biologically distinct entity with unique genetic alterations, clinical features, and progression patterns, differentiating it from more common subtypes such as HGSC and endometrioid carcinoma [6]. While EOC overall remains the most lethal gynecologic malignancy-primarily due to asymptomatic onset, late-stage detection, and resistance to standard chemotherapy-OCCC poses additional challenges due to its intrinsic chemoresistance and unique tumor biology [6, 3]. Specifically, OCCC shows a striking ethnic disparity, accounting for 5–10% of EOC cases in Western populations but rising to 15–30% in East Asian populations such as Japan, Korea, and Taiwan [6, 7]. Given its highly aggressive phenotype, poor prognosis in advanced stages, and limited response to standard chemotherapy, OCCC represents a critical focus for cancer research in Asian populations [6, 8–11].

Beyond clinical features, OCCC exhibits unique histopathological and molecular signatures that contribute to its therapy resistance and biological heterogeneity [12, 13]. Recent efforts to decode OCCC heterogeneity using unsupervised gene expression clustering have identified putative molecular subtypes, including an “epithelial-like” subtype (EpiCC), marked by high expression of cell junction and differentiation genes, and a contrasting “mesenchymal-like” subtype (MesCC), enriched in EMT markers, immune infiltration, therapy-resistant behavior and often correlates with worse prognosis [14, 15]. Moreover, prior studies suggest that epithelial tumor cells within OCCC exhibit transcriptional heterogeneity, spanning secretory-like, proliferative, and stem-like states [16–18]. This cellular plasticity may be a critical driver of adaptation to therapy and metastatic potential [19]. Importantly, many of these subpopulations may be spatially segregated within the tumor microarchitecture, highlighting the necessity of spatially resolved techniques, such as ST, to disentangle lineage hierarchy, stemness, and niche-specific gene regulation [20–24]. These findings underscore the spatial intra-tumor heterogeneity (ITH), warranting further investigation via spatial and single-cell profiling approaches.

Accumulating evidence suggests that immune-related features are closely associated with OCCC prognosis. Notably, immune cell signatures differ between the EpiCC and MesCC subtypes [25], with the MesCC subtype exhibiting enriched immune infiltration and more aggressive clinical outcomes. Transcriptomic analyses have further identified “immune” and “non-immune” subtypes of OCCC, with the “immune” subtype frequently aligning with MesCC and correlating with poorer prognosis [26]. Recently, immune-hot features have also been proposed as indicators of poor prognosis in early-stage OCCC, underscoring the need to define specific patterns of intratumoral heterogeneity associated with these immune phenotypes [23]. Interestingly, an “immune-cold” microenvironment may paradoxically confer more favorable outcomes in OCCC [27]. Immune-hot tumors with poor prognosis are often characterized by high expression of immune checkpoint molecules such as CTLA-4 or PD-1 [28] and enrichment of infiltrating CD8⁺ T-cells [26]. These findings raise the possibility that immune-hot OCCC tumors might benefit from immune checkpoint blockade therapies.

However, the ITH of immune-related features presents a significant challenge, as spatial variability in immune cell distribution and tumor–immune interactions may confound therapeutic responses [29]. The advent of single-cell and ST technologies has enabled more precise dissection of lineage diversity, microenvironmental interactions, and therapy-resistant niches within tumors [24, 30]. These spatially resolved approaches are now foundational for understanding the complex tumor ecosystem of OCCC. For instance, ST analyses have revealed that both immune cell infiltration and immune mimicry by neighboring tumor cells are critical determinants of OCCC immune subtype identity [22, 23, 31]. In a recent integrative study by Mori et al. (2024), the combination of single-nucleus RNA sequencing and STs uncovered a chemoresistant subpopulation of OCCC cells characterized by high HIF activity, predominantly localized to tumor regions enriched in cancer-associated fibroblasts (CAFs) with a myofibroblastic phenotype [22]. These studies collectively highlight the power of spatial technologies in unraveling tumor–immune interactions and pave the way for the development of spatially informed therapeutic strategies in OCCC.

With the understanding of the heterogeneity of OCCC, one crucial question would be how to apply these molecular features to help the real-world oncologists to make clinical decisions. Translating the contexual biology to the clinic [32] is a crucial step that will bridge the gap between the laboratory discoveries and the clinical applications. However, with the complexity of these molecular features at the spatial level, the integration of biology, profiling readouts, and clinical information could appear to be a daunting task. Foundation models might provide solutions to this task for the near future, but exploring the application at the individual patient level at the real-world setting could assist in streamlining and deducing information. Here, we report a proof-of-concept case study on applying advanced technologies in ST by using GeoMx Digital Spatial Profiling (DSP) and Xenium In Situ in an advanced OCCC patient. This study began with two tissue samples (ovarian and bladder tumors) analyzed using the GeoMx DSP CTA platform. Subsequently, serial sections from the same tissue block of this patient were profiled with the Xenium 5 K assay. For validation, we re-analyzed ST data obtained from the Visium platform, including 2 published OCCC samples from 2 patients (GSE224335) [22], 4 in-house samples from 2 patients [31] and 16 published HGSC samples from (GSE211956) [33], GSE288483 [34]. We aimed to identify actionable thereapeutic targets for possible intervention beyond standard of care.

Materials and methods

NanoString GeoMx Digital Spatial Profiling Cancer Transcriptomic Atlas analyses

This Institutional Review Board-approved study (IRB No. 202008022RINB) was conducted at the National Taiwan University Hospital. Tumor tissue Sect.  (5 μm) were obtained from archived formalin-fixed paraffin-embedded (FFPE) blocks. The cut FFPE sections were then baked at 60 °C for 1 h, followed by sequential deparaffinization and rehydration with 100% and 95% ethanol. After washing with PBS, the tissue sections were incubated in 100 °C Tris EDTA for 15 min and in 37 °C Proteinase K (1 mg/mL) PBS solution for 15 min respectively to retrieve and expose RNA targets. The tissue sections were then incubated overnight in a hybridization solution, containing 1,812-plexed GeoMx Cancer Transcriptomic Atlas panel (NanoString, Seattle, WA, USA) at 37 °C and covered with HybriSlip Hybridization Covers (Grace BioLabs, 714022). The sections were then soaked in 2X SSC with 0.1% Tween-20 to remove the HybriSlip covers, and two rounds of stringent washes at 37 °C were performed.

Tissue sections were next placed in a humidity chamber and incubated in blocking buffer for 30 min at room temperature. Incubation of visualization markers (VM), including SYTO 13 (1:10) (Thermo Fischer, Waltham, MA, USA), PanCK-Cy3 (1:40) nucleic stain and CD45-Texas Red (1:40) fluorescently labeled antibodies, was performed for 1 h. The stained sections were loaded into a GeoMx DSP system, followed by the selection of region of interests (ROIs). The OCCC samples in the study were annotated by a pathologist on the H&E slides. H&E-stained sections were used for morphology confirmation prior to sample preparation. Images of VM-stained fluorescence of DNA (blue), PanCK (green), and CD45 (red) were labeled with selected areas of illumination (AOIs). The ROI selections were based on the annotation of their respective H&E slides. The ROI sizes ranged from 200 μm to 700 μm in diameter. The ROIs were further compartmentalized into PanCK-positive (PanCK-pos, tumor cell) and PanCK-negative/CD45-positive (PanCK-neg/CD45-pos, tumor microenvironment) AOIs. UV light was projected onto each defined segment., and UV-photocleavable oligonucleotide barcodes were collected and dispensed into the corresponding wells of a microtiter plate for each AOI. Library preparation was performed with NanoString SeqCode primers (NanoString, Seattle, WA, USA), and AMPure XP beads (Beckman Coulter, Fischer Scientific, Waltham, MA, USA) were used for the pooling and purification of the polymerase chain reaction products. The constructed libraries were sequenced on a NextSeq 550 System (Illumina, San Diego, CA, USA), and the generated FASTQ files were then converted to raw counts with NanoString NGS Pipeline (Version 2.3.3.10, NanoString, Seattle, WA, USA). GeoMx data is publicly available at zonedo (10.5281/zenodo.17212542).

10x genomics xenium In Situ 5 K profiling

Xenium sample preparation

FFPE blocks of primary ovarian and bladder metastasis were acquired followed by pre-screening for RNA quality and histology according to 10x Genomics recommendations. For DV200 analysis, RNA was extracted from two to four 10-µm thick FFPE tissue sections using RNeasy FFPE kit (Qiagen, #73504). FFPE sections with DV200 value greater than 30% were used. In addition, corresponding H&E was performed for these FFPE sections to ensure intact tissue morphology. 5-µm thick FFPE sections were placed onto Xenium Slides according to the 10x Xenium FFPE tissue preparation guide (10x Genomics, CG000578, Rev E). The Xenium slides containing FFPE sections were dried overnight, stored in a desiccator at ambient temperature and used in the Xenium Prime assay (10x Genomics) within 4 weeks. Xenium ST was performed as outlined in the vendor’s protocol (10x Genomics, CG000580, Rev E), and the slides underwent a series of xylene and ethanol washes for deparaffinization and decrosslinking using the FFPE tissue enhancer. Overnight in situ probe hybridization with the probes from Xenium Prime 5 K Human pan tissue & pathways panel (10x Genomics, PN1000724) and subsequent assay steps, including cell segmentation staining with Xenium cell segmentation staining reagents kit (10x Genomics, PN1000661), were performed according to the vendor’s user guide (10x Genomics, CG000760, Rev C).

Xenium analyzer instrument

On the Xenium Analyzer, image acquisition was performed in cycles by automatically cycling in reagents and labeled probes for detecting RNA. These were incubated on the sections, imaged, and removed by the instrument. After binding fluorescent oligos to the amplified barcode sequence, 36 cycles of fluorescent probe hybridization, imaging, and probe removal were performed. The Z-stack images spanning 0.75-µm step size across the entire tissue thickness were taken.

Xenium cell segmentation

For cell segmentation, the Xenium cell segmentation staining reagents kit (10x Genomics, PN1000661) was applied in the Xenium Prime workflow. The kit targeted a variety of cell types and tissues and includes a cocktail of antibodies targeting the membranal proteins (anti-ATP1A1/CD45/E-cadherin), antibodies targeting the cell interior proteins (anti-alphaSMA/Vimentin), and a universal interior label against Ribosomal RNA (18 S rRNA). Cell segmentation was performed using Xenium Onboard Analysis (XOA) v3.1.0.4 algorithm, which uses custom deep learning models trained on Xenium data. Xenium data is publicly available at zonedo (10.5281/zenodo.17212542).

10x genomics cytassist visium spatial gene expression analyses

Cut tumor tissue FFPE Sect.  (5 μm) were baked at 42 °C for 3 h and stored in a desiccated condition. Deparaffinization and H&E staining were both performed according to the manufacturer’s instructions. H&E tissue images were obtained using an Imager Z2 (Zeiss) at 10x objective magnification. RNA targets were released from the tissue samples by decrosslinking, and further probe hybridization, probe ligation, probe release and extension, and library construction were performed; all were performed according to the manufacturers’ instructions. Quantification of the pooled libraries was assessed with KAPA SYBR FAST qPCR Master Mix (KAPA Biosystems). Sample index PCR was performed with proper cycles suggested by qPCR amplification plot. The constructed libraries were sequenced by Illumina NovaSeq 6000, with a dual-indexed setup for 150 base-pair paired-end. Samples were sequenced with the recommended depth of approximately 50,000 reads per spot. Visium data is publicly available at zonedo (https://zenodo.org/records/15378232).

Bioinformatics analysis pipeline

GeoMx Data normalization, analysis, and visualization with unsupervised hierarchical clustering

All analyses were conducted using R version 4.4.3 in the RStudio 2024 environment. The NanoString GeoMx DSP CTA dataset was processed using key packages, including NanoStringNCTools (v1.8.0), GeoMxTools (v3.4.0), and GeoMxWorkflows (v1.6.0). Only areas of interest (AOIs) and probes that passed quality control were retained, following the standard NanoString GeoMx QC workflow. Limits of quantification (LOQ) were estimated using the geometric mean and standard deviation of raw counts from negative control probes and applied as a threshold to exclude low-quality AOIs. Quality metrics, including gene detection rate and background signal derived from negative control probes, were examined across tissues to assess the comparability of technical performance between samples. Gene-level filtering was based on gene detection rates across AOIs to ensure robustness against low-coverage or section-specific technical artifacts. Normalization was subsequently performed using the upper quartile (Q3) method with negative control probes used to estimate background technical noise.

Residual batch effects after normalization were evaluated by examining global expression structure using unsupervised analyses, including principal component analysis (PCA) and hierarchical clustering, on quality-controlled and normalized data, with samples annotated by section identity and other relevant technical metadata. To uncover spatially distinct gene expression patterns, we selected the top 20% most variable genes across AOIs (post-normalization) and subjected them to unsupervised clustering. A correlation matrix was generated and transformed into a distance matrix by subtracting the correlation values from one. Hierarchical clustering using the average linkage method was applied via the pheatmap package (v1.0.12), with resulting clusters and expression patterns visualized as annotated heatmaps with dendrograms. Gene set variation analysis (GSVA, version 1.48.3) was used for the enrichment pathways analysis of hallmarks [35].

Xenium data normalization, analysis, and visualization

The 10X Xenium spatial transcriptomic dataset was processed using the Squidpy framework. QC was applied by removing cells with fewer than 10 total counts or fewer than 200 detected genes. Additionally, genes expressed in fewer than 5 cells were filtered out. Following QC, raw counts were stored in the counts layer for downstream reference. The dataset was normalized using total-count scaling (sc.pp.normalize_total), followed by logarithmic transformation using the natural logarithm of one plus the count (sc.pp.log1p). Highly variable genes were selected and used for PCA. Batch effects across samples were corrected using Harmony on the PCA embeddings, and the Harmony-corrected components were used for downstream analyses. Neighborhood graphs were constructed using the first 30 components (sc.pp.neighbors), followed by UMAP embedding (sc.tl.umap). Clustering was performed using the Leiden algorithm (sc.tl.leiden) with a resolution of 0.8.

Differential expression gene analysis (DEGs) between clusters was conducted using the Wilcoxon rank-sum test implemented in sc.tl.rank_genes_groups. Cluster annotation was performed using canonical marker genes and over-representation analysis [22]. Spatial colocalization analysis between cell types was conducted using the neighborhood enrichment function (sq.gr.nhood_enrichment) in Squidpy. GSVA was performed on cancer cell subsets to investigate functional enrichment [36].

Pseudotime analysis was performed using CellRank [37] implemented in Python to infer directed cell-state transitions from Xenium spatial transcriptomics data. A CytoTRACE-based kernel was applied to estimate differentiation potential at the single-cell level based on gene expression complexity, and this pseudotime information was used to construct a directed transition matrix on the k-nearest neighbor graph. The resulting transition probabilities were projected onto low-dimensional embeddings to visualize putative developmental trajectories patterns.

Visium data normalization, analysis, and visualization

The 10x Visium ST data were processed using Seurat version 4, following standard workflows for sequencing-based spatial data analysis [38, 39]. For downstream analysis, only spatial spots meeting quality control thresholds were retained-specifically, those with 1,000 to 8,000 detected genes, fewer than 50,000 total transcripts (as indicated by nCount_Spatial), and mitochondrial gene content below 30%. The SCTransform normalization method was applied to the Visium dataset to account for technical variation. GSVA (version 1.48.3) [36] was employed for projection signature from Xenium to Visium, focusing on curated gene signatures derived from SLC2A1⁺ and IFIT2⁺ cancer cell populations [36].

Results

Clinical case of advanced OCCC

A 38-year-old female presented with palpable supraclavicular and left cervical lymphadenopathy accompanied by progressive abdominal distension, prompting medical evaluation. An excisional biopsy of the cervical lymph node revealed metastatic adenocarcinoma, immunohistochemically characterized by TTF-1 negativity, PAX-8 positivity, and p53 positivity, supporting a gynecologic origin. Computed tomography of the abdomen and pelvis demonstrated a right ovarian cystic mass measuring 5.4 × 5.8 cm, ascites, extensive lymphadenopathy (including para-aortic, para-caval, porto-caval, and bilateral inguinal regions), peritoneal thickening, and multiple nodules on the peritoneum and abdominal wall-findings consistent with peritoneal carcinomatosis. Her preoperative CA-125 level was 830 U/mL. She was diagnosed with FIGO stage IVB ovarian cancer and underwent suboptimal debulking surgery, with residual tumor deposits exceeding 3 cm in the omentum, ileocecal region, and sigmoid colon. Histopathology confirmed clear cell carcinoma involving the uterus, right fallopian tube, bilateral ovaries, and bladder base. Immunohistochemical staining was positive for Napsin A and focally positive for p53, while estrogen receptor and WT1 were negative.

Adjuvant chemotherapy with weekly dose-dense paclitaxel and carboplatin resulted in a partial biochemical response, with CA-125 declining from 689 U/mL to 187 U/mL but not reaching normalization. Despite subsequent treatment with weekly paclitaxel followed by weekly docetaxel, disease progression ensued, with CA-125 levels rising above 1,100 U/mL. Concurrently, radiotherapy targeting mediastinal lymphadenopathy was administered using volumetric modulated arc therapy.

Comprehensive genomic profiling of both primary and metastatic tumor tissues using the Illumina TSO500 HRD panel revealed a pathogenic ARID1A mutation, microsatellite stability, low tumor mutation burden, and low genomic instability score. In the metastatic bladder lesion, RNA fusion analysis identified a CAPZA2-MET fusion transcript, while concurrent DNA-based profiling revealed MET amplification. Based on these findings, targeted therapy with the MET inhibitor tepotinib (Tepmetko, 225 mg daily) was initiated as a compassionate salvage treatment. Due to episodes of hepatic dysfunction, tepotinib was administered intermittently in three short courses. A biochemical response was noted, with CA-125 levels decreasing from 1,110 U/mL to 244 U/mL.

Spatial transcriptomic profiling revealed spatially resolved hypoxic tumor cells and Wnt-activating TME

To explore whether spatial profiling could help identify additional therapeutic vulnerabilities, GeoMx Digital Spatial Profiling (DSP) was performed in a translational research lab by using the Cancer Transcriptome Atlas (CTA) Panel. A total of 83 ROIs consisting of 142 AOIs were extensively profiled to cover the PanCK+ tumor cell segments and PanCK-/CD45 + immune tumor microenvironment (iTME) segments across the primary ovarian and metastatic bladder tumors (Fig. 1a, S. Fig. 1A). Significant intra-tumor heterogeneity (ITH) was found in both the PanCK+ tumor cells and CD45 + iTME. Cluster analysis revealed three subclusters (C1a, C1b, C1c) in the PanCK+ tumor cell segments and two subclusters (TME1, TME2) in the PanCK-/CD45 + iTME segments (Fig. 1B). The composite of the PanCK+ tumor cells and their surrounding paired CD45 + iTME revealed that the ovarian site was composed of C1a/TME1, C1a/TME2, C1b/TME1. Within the ovarian tumor, spatially, the C1a/TME2 composition was located at the tumor-stromal interface adjacent to a hemorrhagic area, the C1a/TME1 was located at the tumor periphery, and the C1b/TME1 was located at the tumor center (Fig. 1C). The bladder metastasis was composed purely of C1c/TME2 with homogeneous spatial distribution.

Fig. 1.

Comprehensive spatial transcriptomic profiling of an advanced OCCC case using the NanoString GeoMx DSP platform and CTA panel a Schematic of NanoString GeoMx DSP profiling for a FIGO stage IVB OCCC case, showing selected regions of interest (ROIs) from ovary and bladder tissues profiled using the 1812-plex Cancer Transcriptome Atlas (CTA) panel. PanCK⁺ epithelial ROIs and PanCK⁻/CD45⁺ stromal/immune ROIs were selected b Hierarchical clustering heatmap of all profiled ROIs identified three tumor clusters (C1a, C1b, C1c) among PanCK⁺ ROIs, and two major TME clusters (TME1, TME2) among PanCK⁻ ROIs c Spatial mapping of the clusters from (b) overlaid on matched H&E and immunofluorescence images, highlighting geospatial localization of tumor and TME subtypes d Hallmark pathway enrichment scores computed using GSVA for each cluster; heatmap shows relative enrichment scaled from − 1 to 1 e Volcano plot comparing gene signatures between TME1 and TME2; significantly enriched pathways (|log₂FC| > 1, P < 0.05) are labeled and colored by enrichment direction

Hallmark enrichment analysis for each subcluster showed preferential functional pathways for the PanCK+ tumor cells and CD45 + iTME (Fig. 1D). The tumor subcluster C1a and C1c shared many hallmarks in metabolism (REACTIVE_OXYGEN_SPECIES_PATHWAY, OXIDATIVE_PHOSPHORYLATION, PEROXISOME, ADIPOGENESIS, FATTY_ACID_METABOLISM), PI3K-Akt pathway (MTORC1_SIGNALING, PI3K_AKT_MTOR_SIGNALING), and replication, cell cycle-related pathways (MITOTIC_SPINDLE, DNA_REPAIR, G2M_CHECKPOINT, MYC_TARGETS_V1), suggesting common biology. The C1a tumor cells were characterized by additional metabolic hallmarks suggestive of hypoxic signals and metabolic switch (CHOLESTEROL_HOMEOSTASIS, GLYCOLYSIS, HEME_METABOLISM, HYPOXIA). The CD45 + TME1 showed a more inflammatory microenvironment with the significant enrichment of hallmarks in TNFα signaling and inflammatory response. TME2 was highly enriched in Wnt signaling, a known immunosuppressive pathway. Analysis of differential expression genes (DEGs) indeed showed that WNT4, AXIN2, SFRP4 were highly expressed in TME2 and HLA-DRA, HLA-DRB3, HLA-DRB4, CXCL10, CCL5 were highly expressed in TME1 (Fig. 1E).

In addition to pathway-level differences, ST analysis revealed that MET expression was predominantly enriched in PanCK⁺ tumor cell segments compared with PanCK⁻/CD45⁺ iTME segments (S. Figure 1B). MET expression further exhibited heterogeneity across tumor subclusters. Among the PanCK⁺ tumor clusters, C1c, which was specifically enriched in the bladder metastasis, exhibited significantly higher MET expression compared to C1b, which was enriched in the ovarian tumor (P < 0.05) (S. Fig. 1C), while other pairwise comparisons did not reach statistical significance. The ovarian tumor contained two clusters (C1a and C1b), indicating intra-tumoral heterogeneity, in contrast to the bladder metastasis which contained a single cluster (C1c). MET expression trended higher in C1a relative to C1b despite being non-significant, indicating the heterogeneous MET expression within the ovarian components. This finding indicated that MET amplification detected by bulk genomic profiling was in general concordant with the GeoMx ST data. Nevertheless, spatial and clonal heterogeneity at the transcriptomic level was also detected in granularity.

In short summary, mapping the enriched hallmarks onto the spatial locations within the ovarian tumor, there was a spatial pattern evolving from the less hypoxic tumor cells with inflammatory TME region at the tumor center, to a more hypoxic and glycolytic tumor cells with inflammatory TME region at the tumor periphery, towards a hypoxic and glycolytic tumor cells with Wnt-activating TME region at the tumor-stromal interface. Within the bladder metastasis, although there was no spatial pattern observed, the immune microenvironment predominantly consisted of the immunosuppressive Wnt-activating TME. Unfortunately, there was no clinically approved therapeutics targeting the Wnt pathway.

Single cell spatial transcriptomic profiling revealed spatially resolved OCCC tumor cell clones

To further explore whether there might exist unique single cell clones within the paired ovarian and bladder tumors which could yield additional information for therapeutic vulnerability, Xenium In Situ was performed in a translational research lab by using the 5 K Human Pan Tissue and Pathways Panel (Fig. 2A). A total of 1,126,954 cells were profiled from the two tumor samples with 13 single cell types including 6 cancer cell types (EPCAM+, KLRC1+, CP+, PAX8+, IFIT2+, SLC2A1+), 2 fibroblast subtypes (SCN7A+, COL5A1+), and 5 immune/stromal cell lineages being identified (Fig. 2B) with each single cell type being annotated by its most dominantly expressed marker gene (Fig. 2C). Hallmark enrichment analysis revealed the plausible functional pathways embedded in each cell type (Fig. 2D). Of note, the SLC2A2 + cells showed significant enrichment in the hallmark of hypoxia and Hedgehog signaling (Fig. 2D) which were similar to the C1a cluster identified in the GeoMx analysis. IFIT2 + cancer cells showed strong inflammatory signaling which were not clearly identified in any of the GeoMx clusters.

Fig. 2.

Single-cell spatial transcriptomics of an advanced OCCC case using the 10x Genomics Xenium platform a Overview schematic of Xenium in situ profiling with the 5 K Pan-Tissue and Pathways Panel on ovary and bladder sections, covering over 1.1 million cells from a FIGO stage IVB OCCC case b Unsupervised clustering of all cells identified 13 distinct clusters, annotated into three major cell compartments (cancer, immune, stromal) based on canonical markers and over-representation analysis c Heatmap showing representative markers for annotated cell types; expression values are scaled from − 2 to 2 to highlight distinct expression signatures d Heatmap of GSVA-based hallmark pathway enrichment scores in cancer cells across six cancer cells (e.g., IFIT2⁺, PAX8⁺), scaled from − 1 to 1, highlighting differences in metabolic, immune, and signaling pathways

Using the scNiche framework, we delineated four spatial niches in the primary ovarian tumor and five in the metastatic bladder tumor (Fig. 3A). The ovarian niche2, bladder niche0 and niche4 had dominance of immune and endothelial cells, fibroblasts, while at the ovarian niche0, niche3 as well as the bladder niche1 and niche3, cancer cells become the dominant cell type. Neighborhood analysis further revealed distinct cellular co-localization patterns within these niches (Fig. 3B-C). The ovarian niche1 and bladder niche 3 had very few cells and did not form obvious neighborhoods and were not included in the subsequent analysis. Interestingly, EPCAM+, CP+, PAX8 + cancer cells showed strong co-localization at the ovarian niche2, forming as a major cancer cell group. This co-localization was less tight in niche0 and even disintegrated in niche3 (Fig. 3B, box 2). Conversely, the CD168 + macrophage and SCN7A+ fibroblast were almost mutually exclusive in the ovarian niche2 but started to show co-localization in niche0 and niche3 (Fig. 3B, box 1). This pattern was also observed among T cells, COL5A1 + fibroblasts, and NOTCH4 + endothelial cells that their co-localization only occurred in niceh3 in the ovary (Fig. 3B, box 3). Of note, these immune/stromal cells did not tend to co-localize with cancer cells in the niches except for niche3. Neighborhood analysis in the bladder also revealed similar cellular co-localization patterns (Fig. 3C, box 1–3). The data suggested that both cancer cells and immune/stromal cells tend to form tighter interactions among the same cell groups when they are the minority population inside the niche (e.g. macrophages and fibroblasts in ovary niche3, bladder niche1 and niche2). These tight interactions would decrease once these cell groups are the majority population inside the niches (e.g. cancer cells in ovary niche3, bladder niche1 and niche2).

Fig. 3.

Comparative niche composition and spatial colocalization across ovary and bladder tissues a Bar plots showing the proportion of total, cancer, and non-cancer cells across identified niches in ovary and bladder tissues. Subpanels visualize: all major cell types, cancer cell subtypes (e.g., SLC2A1⁺, PAX8⁺), and non-cancer populations (e.g., T cells, fibroblasts, macrophages) b Spatial mapping and colocalization analysis of ovary tissue niches (niche0, niche2, niche3), demonstrating the localization patterns of distinct cell types and their co-occurrence in tissue space c Equivalent mapping and colocalization analysis for bladder tissue niches (niche0, niche1, niche2, niche4), showing tissue-specific spatial arrangements

However, the neighborhood analysis also indicated a specific subset of cells showing consistent cellular co-localization patterns. Consistently across all niches in both tumors, T cells were always mutually exclusive from the major cancer cell group (Fig. 3B-C, hashtags) and were very likely to associate with macrophages (Fig. 3B-C, asterisks). IFIT2 + cancer clones were positioned adjacent to MMP12 + dendritic cells and frequently co-localized with SLC2A1 + cancer cells, suggesting coordinated spatial organization and potential immune-modulating interactions (Fig. 3B-C, black arrows). The co-localization between MMP12 + dendritic cells and SLC2A1 + cancer cells were stronger in the ovarian niches than those in the bladder (Fig. 3B-C, black arrows). In the ovarian niche3, bladder niche1 and niche2, T cells also started to co-localize with IFIT2 + cancer cells, MMP12 + dendritic cells, and SLC2A1 + cancer cells (Fig. 3B-C, circumflex).

Focusing on the distribution patten among these 3 cells (IFIT2 + cancer, MMP12 + dendritic, and SLC2A1 + cancer), it was apparent that the bladder tumor showed a rather homogeneous spatial distribution (Fig. 4B) while the ovarian tumor showed a heterogeneous spatial distribution of SLC2A1 + cancer cells (Fig. 4A). At the niches where the immune/stromal cells formed tighter co-localizations (ovarian niche3, bladder niche1 and niche2) (Fig. 3B-C box 3), IFIT2 + cancer cells together with MMP12 + dendritic cells and other immune/stromal cells formed stable co-localizations (Fig. 4A-D) while the SLC2A1 + cancer cells showed co-localizations only with the MMP12 + dendritic cells in the ovary (Fig. 4C) and both MMP12 + dendritic cells and IFIT2 + cancer cells in the bladder (Fig. 4D). The data revealed that the hypoxic OCCC cancer cells were situated in a rather isolated TME which could lead to refractoriness to treatments.

Fig. 4.

Spatial co-occurrence and pseudo-temporal dynamics of cancer and immune cell populations in ovary and bladder tissues a-b Spatial distribution of selected cell populations in ovary and bladder tissue. Panels from left to right showing, merged spatial map highlighting SLC2A1⁺ cancer cells, IFIT2⁺ cancer cells, and MMP12⁺ dendritic cells; spatial niche annotations (niche0, niche2, niche3); merged colocalization of the above with additional CD2⁺ T cells; spatial maps of each niche, respectively c-d Schematic illustration of inferred cell-cell spatial interaction among MMP12⁺ dendritic cells, IFIT2⁺ cancer cells, SLC2A1⁺ cancer cells, and CD2 + T cells in ovary and bladder tissue. Right panels provide zoom-in visualizations of representative niches showing these co-occurrence patterns e RNA velocity analysis of cancer cells colored by subtype with predicted directional transitions among cancer subpopulations and their pseudotime captures the continuous developmental progression from early to late cancer cell states. Spatial co-occurrence and trajectory dynamics of cancer cells

To understand how these cancer clones were derived, CellRank-based directional transition and pseudotime projection was performed (Fig. 4E). Within the cancer compartment, PAX8⁺ cancer cells were inferred to be the founder clone via pseudotime score (Fig. 4E), giving rise to two major divergent lineages, one marked by CP+, KLRC1+, and EPCAM+ expression with low hallmark enrichment for inflammatory signals, and another comprising SLC2A1 + and IFIT2 + clones exhibiting strong hallmark enrichment for inflammatory responses. The PAX8 + cancer cells also showed the shortest psuedotime values suggesting of their early clonal feature. The data suggested that the SLC2A1 + and IFIT2 + lineages might be derived after hypoxic stress. Specific niches surrounding these SLC2A1 + and IFIT2 + cancer cells might create an escape sanctuary for immune evasion.

Identification of spatially resolved SLC2A1⁺ and IFIT2⁺ cancer cells in independent OCCC tissue samples

To validate whether these spatially resolved cancer cell clones could also be found in other OCCC patients, we generated gene signatures of SLC2A1⁺ and IFIT2⁺ cancer cells and applied them to both publicly available [22] and in-house ST profiling data [31]. Six OCCC Visium ST profiling data were analyzed (Fig. 5) and the projection of SLC2A1⁺ and IFIT2⁺ cancer cell signatures showed that these two cancer cell clones could be identified together in regions that were in close spatial proximity. This confirmed the generalizability of ST data derived from one single patient in independent patient samples.

Fig. 5.

Spatial enrichment score of SLC2A1⁺ and IFIT2⁺ cancer cell signatures in OCCC Visium samples. The dataset includes four OCCC tumor sections from NTUH and two publicly available samples from Yutaro Mori et al. (2024). Spatial heatmaps display the enrichment scores for each signature across tissue sections, scaled from − 1 to 1

To further assess whether this spatial relationship is conserved across other ovarian cancer histotypes, we additionally analyzed 16 published Visium ST data from high-grade serous carcinoma (HGSC) [33, 34]. Similar spatial co-occurrence patterns between SLC2A1⁺ and IFIT2⁺ cancer cell signatures were observed in HGSC samples (S. Figure 2), suggesting that these two cancer cells are not restricted to OCCC and may represent broader spatial cancer cell phenotypes within ovarian malignancies. However, this cross-cohort validation is still limited by its small scale, therefore, the true clinical impacts of how this real-world ST analysis still requires further testing in prospective studies.

Discussion

A recent integrative study in HGSC combined single-cell ST with Perturb-seq to uncover mechanisms of immune evasion [40]. By profiling over 2.5 million cells from 130 tumors, the study identified a distinct malignant cell state driven by tumor-intrinsic genetic alterations that predicted T cell/NK cell infiltration and response to immune checkpoint blockade. High-throughput screening further revealed that PTPN1 and ACTR8 knockouts induced this immunologically relevant state, while pharmacological inhibition of PTPN1/PTPN2 sensitized ovarian cancer cells to cytotoxic lymphocyte-mediated killing. This study highlights a framework linking genetic perturbation, malignant cell states, and spatial architecture to identify immune vulnerabilities in ovarian cancer. In a complementary study using Visium ST, researchers investigated the profound impact of the ITH complexity of HGSC [33]. Their ST profiling revealed multiple spatially coexisting subclones, each defined by distinct copy number alterations and divergent expression of ligands and receptors, which predicted different interactions with surrounding stromal and immune cells. Using CosMx SMI for one case, the study further resolved how individual subclones interact in situ with fibroblasts, endothelial cells, and immune populations, highlighting spatially compartmentalized cell–cell communication. Subclone-specific paracrine and autocrine signaling pathways were identified, illustrating how each subclone may sculpt its own microenvironmental niche. These findings emphasize that HGSC subclones are not only genetically distinct but also spatially organized, and that disrupting their context-specific signaling could represent a promising strategy to overcome therapy resistance.

In a multi-platform study by Siyu Xia et al. (2024), researchers integrated single-cell RNA-seq, TCR-seq, ST, and bulk RNA-seq to characterize the immunometabolic landscape of OCCC [13]. Profiling over 140,000 cells, they found that ARID1A mutations were linked to enhanced immune activation, neoantigen-reactive CXCL13⁺CTLA4⁺ CD8⁺ T cells, and FASLG-FAS signaling. In contrast, recurrent tumors showed fibrotic remodeling, angiogenesis, and fatty acid–driven metabolic reprogramming, reflecting an immunosuppressive TME. ST revealed intratumoral heterogeneity and stromal-immune interactions driving resistance. High CD36 and CD47 expression correlated with poor progression-free survival. Bevacizumab increased T cell infiltration and IFN-γ signaling, and retrospective data supported clinical benefit from combined VEGF and PD-1 blockade. This study highlights how genetic alterations, immune profiles, and metabolism shape therapy response, supporting a spatially informed immunometabolic framework for guiding combination treatment in OCCC. In an integrative study by Yutaro Mori et al. (2024), a chemoresistant OCCC subpopulation with elevated HIF activity was identified, predominantly localized in tumor regions enriched with cancer-associated fibroblasts (CAFs) exhibiting a myofibroblastic phenotype [22]. Mechanistically, co-culture experiments showed that PDGF signaling from tumor cells to CAFs supported CAF survival and induced HIF-1α, which in turn promoted chemoresistance in neighboring tumor cells. Notably, the study identified ripretinib-a receptor tyrosine kinase inhibitor-as effective in disrupting this CAF-tumor interaction. Ripretinib impaired CAF viability and, in combination with carboplatin, significantly suppressed tumor growth in vitro and in xenograft models. This work highlights the spatial and paracrine interactions between tumor cells and the stromal niche, particularly the CAF-HIF axis, in mediating chemoresistance. Targeting these interactions through spatially guided strategies offers a promising therapeutic avenue for OCCC.

Our current study of deep ST profiling in one advanced OCCC patient revealed the complexity of clonal heterogeneity and the spatial architecture among these clones. We adopted two different ST approaches in hope to find novel therapeutic vulnerability beyond standard of care. The first approach, GeoMx DSP, could be considered as a “mini-bulk” sampling method across the same tissue by using selective illumination to achieve spatial dissection [41]. The gene panel we chose was relatively selective, which only covered major cancer-related pathways. The second approach Xenium In Situ, utilizes direct hybridization of probes followed by imaging at the single cell level throughout the tissue. The gene panel we chose was a much bigger panel consisting of 5000 genes covering diverse tissue types and functional pathways. The results derived from GeoMx showed expression clusters enriched with differential functional hallmarks at selected regions within the tumor in coarse granularity. This approach is more intuitive from the histopathological perspective that the pathologists would also scan through the tissue section and select regions to zoom in for detailed inspection. In this approach, individual cell types within these ROIs need to be annotated via deconvolution. In contrast, the results from Xenium elucidated the single cell composition in exceptional details at the spatial level which has never been possible upon histopathological diagnosis. The deconvolution required for this approach was to decipher the neighborhood relationships and architecture among these single cell clones. These two approaches demonstrate different logics and philosophy in ST profiling when it comes to real-world translation. In the real-world setting when patients are always treated as “N equals to one” rather than pooled statistics, for ST to work as a tool to inform patient care, robust analytical pipelines would need to be developed.

Importantly, while MET amplification identified by targeted sequencing was sufficient to guide the selection of tepotinib, spatial transcriptomic analyses revealed heterogeneity in MET expression across tumor subgroup and spatial regions. This observation provides a potential biological explanation for variable or transient responses to MET-targeted therapy and underscores the limitation of bulk genomic profiling in capturing functional oncogene dependency. By resolving the spatial and clonal heterogeneity of MET expression, which was initially identified as a singular alteration by the TSO500 panel, ST provides a biological rationale for why the patient achieved only a partial biochemical response to tepotinib before disease progression. In this context, ST does not replace established clinical diagnostics but rather offers complementary insights into spatial and clonal heterogeneity, which may be critical for understanding therapeutic resistance and informing future treatment strategies.

Although the two approaches provided us with different levels of granularity in terms of spatial resolution, the patterns of intra-tumoral heterogeneity were highly compatible and complimentary. The GeoMx analysis revealed that the bladder tumor was homogeneous in terms of tumor and TME cluster pairing. Similarly, the neighborhood niche analysis in Xenium also showed homogeneity in the spatial distribution of the mixture of niches throughout the tissue. The SLC2A1 + and IFIT2 + cancer cells identified in the Xenium analysis indicated to the hypoxic and inflammatory pathways being activated in these 2 cancer cell populations. The tripartite relationship between SLC2A1 + and IFIT2 + cancer cells and the MMP12 + dendritic cells is an interesting aspect pointing towards the interaction between antigen presenting cells (APC) inside the TME. Both the SLC2A1 + cancer cells and the MMP12 + dendritic cells expressed high HK2, the gene coding for hexokinase 2. Since HK2 is crucial in glycolysis downstream to the glucose transporter GLUT-1 (the protein coded by SLC2A1), this tripartite cellular relationship at the specific niches has revealed a TME regulated by metabolism, antigen presentation, and immune evasion [42]. HK2 might be a promising therapeutic target for advanced OCCC with small molecule inhibitors targeting HK2 are in active development [43]. These findings from the deep ST profiling of this advanced OCCC patients have provided significant biological insights which will facilitate the pursuit of novel therapeutic options for OCCC. Our study of deep ST profiling in the real-world setting has demonstrated the feasibility of making novel discoveries, one patient at a time.

Conclusion

In this study, we demonstrated the feasibility and utility of applying deep ST profiling to a real-world case for identifying clinically actionable vulnerabilities in advanced OCCC. By integrating GeoMx DSP and Xenium In Situ platforms, we identified key spatial patterns of tumor heterogeneity and tumor microenvironment interactions at both bulk and single-cell resolution. Specifically, we uncovered a Wnt-activating, immunosuppressive microenvironment and a tripartite spatial relationship among hypoxic (SLC2A1⁺), inflammatory (IFIT2⁺) cancer cells, and MMP12⁺ dendritic cells-highlighting metabolic-immune axes potentially involved in immune evasion. This proof-of-concept case study illustrates the translational potential of ST in guiding individualized therapeutic strategies and underscores the value of single-patient profiling in the era of precision oncology.

Authors’ contributions

Conceptualization, Ruby Yun-Ju Huang, Ying-Cheng Chiang, and Lin-Hung Wei; writing, Thang Truong Le, Alice Hsiang-Kuo Yang and Ruby Yun-Ju Huang; Data acquisition, Ko-Chen Chen, Yi-Chia Chiu, Sydney Rechie Necesario; spatial data analysis and interpretation, Thang Truong Le, Mqondisi Fortune Mavuso, and Tuan Zea Tan; pathology annotation and review, Wei-Chou Lin; clinical review, Lin-Hung Wei, Ying-Cheng Chiang, Alice Hsiang-Kuo Yang and Ya-Ting Tai. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Yushan Fellow Program by the Ministry of Education (MOE), Taiwan (NTU-112V0402, NTU-113V2007-1, NTU-1142007-2), the NTU Core Consortiums (NTUCC-114L890803, NTU-114L8502), and NSTC Research Project (NSTC 111-2320-B-002-090-MY3, NSTC 113-2314-B-002-007-, NSTC 114-2314-B-002-249-) to Ruby Yun-Ju Huang and the Excellent Translational Medicine Research Projects (NTU CM113C101-43) to Ying-Cheng Chiang, Lin-Hung Wei. We thank the staff of the Sequencing Core, Center for Advanced Computing and Imaging in Biomedicine, Department of Medical Research, National Taiwan University Hospital for their excellent technical support.

Data availability

The NanoString GeoMx DSP and 10X Genomics Xenium datasets are publicly available on Zenodo (10.5281/zenodo.17212542). Two published Visium OCCC samples are available from GEO (GSE224335), and four in-house Visium OCCC samples have been deposited in Zenodo (https://zenodo.org/records/15378232). In addition, 16 published Visium HGSC samples are accessible from GEO under accession numbers GSE211956 and GSE288483. Further information is available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of National Taiwan University Hospital (reference number: 20200508RIND), which waived the need for informed consent.

Competing interests

The authors declare no competing interests.