Multi‐Omics Profiling Uncovers Predictive Factors for Tumor‐Infiltrating Lymphocyte Therapy in Ovarian Cancer

ABSTRACT

Background

Ovarian cancer (OC) remains a challenging malignancy with a poor prognosis, particularly in advanced stages. Tumor‐infiltrating lymphocyte (TIL) therapy shows promise but has yielded inconsistent results in OC patients. This study aimed to identify factors predicting successful TIL isolation and expansion in OC.

Methods

We performed multi‐omics profiling on tumor samples from 10 OC patients, including whole‐exome sequencing, bulk RNA sequencing, and single‐cell RNA sequencing.

Results

Genomic analysis revealed heterogeneity in tumor mutation burden, copy number variations, and homologous recombination deficiency status across patients. Single‐cell profiling uncovered distinct cellular compositions in tumors, yielding successful TIL products (TIL+) vs. those that did not (TIL‐). TIL+ tumors exhibited higher immune cell infiltration, particularly CD8⁺ effector and effector memory T cells. Metabolic profiling revealed enhanced activity in critical T cell subtypes within TIL+ tumors. Analysis of myeloid cells and fibroblasts identified subpopulations uniquely associated with TIL+ or TIL‐ samples. Cell‐to‐cell communication network analysis highlighted potential prognostic markers and interactions influencing TIL isolation outcomes.

Conclusion

Our findings provide insights into factors affecting TIL therapy efficacy in OC and suggest potential strategies for optimizing patient selection and TIL manufacturing protocols.

We profiled tumors from 10 ovarian cancer patients with whole‐exome sequencing (WES), bulk RNA‐seq, and single‐cell RNA‐seq to uncover predictors of successful tumor‐infiltrating lymphocyte (TIL) isolation and expansion. TIL+ tumors exhibited enriched CD8+ Teff/Tem populations with distinctive metabolic activity, dendritic‐cell–skewed myeloid states, and pro‐inflammatory CAFs, while TIL‐ tumors showed Th2‐skewing and immunosuppressive stromal programs; cell–cell signaling (e.g., WNT9A/EPHA1) and intracellular network analyses nominate biomarkers and therapeutic levers to refine patient selection and optimize TIL manufacturing.

1. Introduction

Due to a lack of symptoms, ovarian cancer (OC) often remains undiagnosed until it has advanced [1]. In the context of the United States, the survival rate over a span of five years for cases of ovarian cancer is documented to be 46.2% [2]. This is predominantly attributed to the diagnosis of more than 70% of ovarian cancer cases at an advanced stage, specifically stages III and IV [3]. Despite recent progress with poly‐[adenosine diphosphate (ADP)‐ribose] polymerase inhibitors (PARPi) extending progression‐free survival, the overall survival (OS) of patients with high‐grade serous ovarian carcinoma (HGSOC), the most common (~70%) histological subtype, remains poor [4]. New therapeutic options are urgently needed to improve outcomes for these patients [5].

Adoptive tumor‐infiltrating lymphocyte (TIL) therapy, which has been at the forefront of new clinical trials for many solid tumors, is a personalized form of immunotherapy [6]. In this approach, a patient's own TILs are extracted, expanded ex vivo, reactivated, and reinfused to target cancers [7]. A recent study observed that the presence of CD8⁺ TILs was significantly associated with better OS, while histological subtypes and membranous PD‐L1 expression were not, highlighting the potential of CD8⁺ TILs as an independent favorable prognostic factor [8]. From a meta‐analysis, for patients with gynecologic cancers, the overall response rate was highest for TIL therapy at 41.4% [9]. This was followed by NK cell therapy at 26.7%, peripheral autologous T‐cell transfer at 18.4%, T cell receptor T cell therapy at 15.4%, and chimeric antigen receptor T cell therapy at 9.5%.

Unfortunately, the clinical efficacy of TIL therapy in OC patients has generally been disappointing. In Pedersen et al.'s pilot study, although target lesions generally remained stable or regressed after TIL infusion, new lesions developed and progressed quickly [10]. In a clinical trial aimed at enhancing tumor‐reactive TILs, administration of an immune checkpoint inhibitor (ICI) before tumor resection and during early TIL expansion led to one partial response and disease stabilization for up to 12 months in five patients [11]. These results are consistent with other immunotherapies for OC [12]. ICIs have benefited less than ~10% of OC patients [13]. Only a tiny fraction of OC patients have benefited from immunotherapy, and most eventually suffer recurrence or progression [14]. As a result, current research focuses on improving immunotherapeutic strategies by identifying limiting factors and selecting markers with prognostic value [15].

Still, there is compelling evidence that OC is an immunogenic tumor and a suitable candidate for TIL therapy [16]. For example, the tumor microenvironment (TME) of OC is often infiltrated by immune cells whose presence correlates with increased OS [17]. Specifically, a higher ratio of intra‐epithelial CD8⁺ T cells to regulatory T cells (Tregs) is associated with better prognosis [18]. Conversely, loss of MHC‐I expression or other components of the antigen presentation machinery correlates with poor survival [19]. However, the factors limiting the clinical response to TIL therapy in OC patients remain incompletely understood. Researchers have investigated many potential factors hindering TIL therapy efficacy, yet more studies are needed [20].

In this study, we aimed to evaluate multiple factors that may affect the efficacy of TIL therapy in 10 OC patients. We employed a multi‐omics approach, including whole‐exome sequencing (WES), bulk and single‐cell RNA sequencing (scRNA‐seq), to analyze tumor cells, the TME of OC, and TIL products. Our primary objective was to identify factors that could predict successful TIL expansion in OC patients, including the quality and quantity of TILs that could be extracted from a patient's tumor and the composition of the TME. Through scRNA‐seq, we sought to uncover predictive factors that could be used to optimize TIL isolation and expansion protocols. By improving our knowledge of these factors, we may be able to enhance the efficacy of TIL treatment and extend its benefits to a broader population of OC patients. Our findings could potentially inform patient selection criteria and guide the development of optimized TIL manufacturing protocols, ultimately contributing to improved outcomes in OC treatment.

2. Results

2.1. Multi‐Omics Profiling Reveals Genetic and Microenvironmental Heterogeneity in Ovarian Cancer

In this study, we collected fresh biopsies or surgical resections from 10 OC patients, including 5 with ovarian adenocarcinomas and 5 with ovarian serous carcinomas (Table S1). We comprehensively characterized the primary tumor samples using two complementary approaches: WES of tumor tissues and matched normal blood as a discovery tool for identifying specific genetic mutations causing cancer initiation, and bulk RNA‐seq of tumor tissues as a basis for classifying the TME of OC (Figure 1a). Additional tissues, where available, and ex vivo TIL expansions derived from two patients (OC020A and OC026A) were dissociated for scRNA‐seq using the 10× Genomics platform to characterize the cellular composition of the tumor tissues and TIL products at single‐cell resolution.

FIGURE 1.

Genetic and Microenvironmental Heterogeneity in Ovarian Cancer. (a) Schematic representation of the experimental design. (b) Oncoplot depicting mutational landscape in ovarian cancer (OC) patients derived from whole‐exome sequencing (WES) data. (c) Heatmap illustrating copy number variation (CNV) events identified through WES analysis in 10 OC patients. (d) Homologous recombination deficiency (HRD) assessment results for OC patients. The dashed line indicates the clinical threshold of 42. Patients with HRD scores ≥ 42 (above threshold) exhibit better prognosis and a greater likelihood of response to PARP inhibitor therapy compared to those with scores < 42. Red: The loss of heterozygosity (HRD − LOH); Green: The telomeric allelic imbalance (Telomeric AI); Blue: The number of large‐scale transitions (LST). (e) Principal Component Analysis (PCA) of gene expression profiles from 10 OC samples, stratified by OC TNM staging. Each point represents an individual sample, color‐coded by primary tumor stage (left) or patient ID (right). Early‐stage (T1&2) and late‐stage (T3&4) samples are additionally circled for clarity. Right panel: Samples colored by patient ID. (f) Uniform Manifold Approximation and Projection (UMAP) visualization of cells, annotated with major cell type classifications. (g) Stacked violin plot illustrating expression patterns of marker genes across identified major cell types. (h) Comparative analysis of major cell type proportions between tumor samples with successful tumor‐infiltrating lymphocyte (TIL) isolation (TIL+) and those where TIL isolation failed (TIL‐).

From Table S1, the mutations in Ki67 and P53 were ubiquitous among the 10 OC patients. Analysis of the WES data showed that the patients have diverse tumor mutation burden (TMB) (Figure 1b) and distinct copy number variation (CNV) profiles (Figure 1c). As shown in Figure S1, other somatic mutations identified in WES data were also highly heterogeneous across patients. In a separate study, Li et al. found that homologous recombination deficiency (HRD) status can predict prognosis for both first‐line chemotherapy and PARPi maintenance therapy in Chinese OC patients [21]. As shown in Figure 1d, applying the HRD scoring threshold of 42, established by Li et al. all 10 patients in the current study could be divided into two equal groups, with 5 patients above and 5 below the threshold. These findings suggest that the patients in our cohort likely responded differently to chemotherapy and PARPi treatment. Patient‐specific factors, including variations in mutation profiles and HRD status, likely influenced the efficacy of cancer therapies for each individual. The heterogeneity in TMBs, CNVs, and HRD scores across patients indicates that treatment approaches should be carefully tailored to optimize outcomes for OC patients.

While genetic variants in the tumor tissues distinguished two groups of OC patients with different treatment responses, we hypothesized that the TME also plays an important role in affecting this categorization. We performed principal component analysis (PCA) on the corresponding bulk RNA‐seq data and found that OC samples from different stages could be divided into early and late stages by the first principal component (PC1) (Figure 1e). We further assessed the gene programs associated with these two OC progression groups by comparing their transcriptional profiles (Figure S2a, Table S2). Differential expression analysis (FDR < 0.05, FC > 1.5) revealed that PDCD1 (encoding PD‐1) expression could not distinguish these OC patients. The up‐regulated genes in early‐stage patients were enriched in metabolism and immune response‐related pathways (Figure S2b), suggesting metabolic alterations in TME supported tumor cells' rapid proliferation and survival, while the immune system was surveilling the TME and attempting to mount an anti‐tumor response. In contrast, late‐stage patients showed higher expression of genes involved in cell cycle regulation pathways like PI3K‐AKT and MAPK signaling (Figure S2c), indicating these OC tumors underwent significant genetic and molecular alterations, TME changes, and acquired resistance that drives uncontrolled tumor growth and progression.

Based on the WES and bulk RNA‐seq data analysis, changes in the OC TME at the tissue level were more likely associated with cancer stage than correlated with genetic mutation status. To delineate the interplay between dynamics in the TME and treatment responses caused by genetic mutations, we further performed scRNA‐seq analysis of tumor tissues from 4 patients and ex vivo TIL expansions from 2 of these 4 patients. This revealed the cellular landscape of the OC TME and provided insights into how microenvironmental changes may influence treatment responses.

2.2. Single‐Cell Profiling Uncovers Cellular Landscape of Ovarian Cancer

Utilizing scRNA‐seq, we profiled 4 tumor samples and 4 ex vivo expansions, initial and rapid expansions of TILs from 2 successful cases, referred to as TIL+ tumors (Figure 1a), with the 10× Genomics platform. In total, we obtained single‐cell transcriptome data for 80,866 cells. Among them, 306 cells expressing fewer than 200 genes and 3767 genes detected in fewer than 3 cells were removed, resulting in 78,430 high‐quality cells with a fraction of mitochondrial transcripts < 20% and total unique molecular identifiers (UMIs) < 11,000, which were used for downstream analysis. The median number of genes and UMIs detected per cell was 3516 and 10422, respectively.

Using scanpy [22], the gene expression profiles of 78,430 high‐quality cells were normalized and log‐transformed. PCA was performed on 4000 highly variable genes, identifying the 50 most significant PCs, which were adjusted using Harmony [23] and used to construct a batch‐balanced k‐nearest‐neighbor (BBKNN) graph [24]. The Leiden graph‐based clustering algorithm identified cell clusters with a resolution of 0.3 in this graph [25]. Differential expression analysis between clusters identified cluster‐specific marker genes, which were used to annotate cell types within the OC TME (Figure 1f), corresponding to T cell (CD3D, CD3E), NK cell (NKG7), B cell (MS4A1, BANK1), plasma cell (IGKC, IGHG1), myeloid cell (CD68, AIF1), cancer‐associated fibroblast (CAF) (DCN, MMP2), endothelial cell (RAMP2, PLVAP), and epithelial cell (KRT8, KRT19) (Figure 1g). Uniform Manifold Approximation and Projection (UMAP) [26] was performed on the same graph to project all high‐quality cells into a 2‐dimensional space.

To identify malignant cells, we assessed chromosomal copy number variations (CNVs) at the single‐cell level. We inferred CNV content for each cell from the scRNA‐seq data using a sliding window approach in the infercnvpy package [27]. Compared to reference cells like T cells, a subset of epithelial cells exhibited a higher CNV burden (Figure S3). Clustering cells based on their CNV profiles identified 8720 malignant cells and 3895 non‐malignant epithelial cells (Figure 1f). The malignant cells showed widespread CNVs, indicating higher genomic instability, while the non‐malignant epithelial cells had relatively stable genomes. This CNV‐based classification allowed us to confidently distinguish malignant from non‐malignant epithelial cells in the single‐cell transcriptome data.

To quantify the dynamics of the cellular compositions between samples, we performed a t‐test on the fractions of each major cell type in TIL+ and TIL‐ tumor samples (Figure 1h). The differential abundance analysis showed a correlation between elevated B and T cell infiltrations in the TME of OC and successful TIL isolation, albeit limited by the small sample size.

2.3. Characterizing Changes in T Cell Subpopulations in OC TME and Derived TIL Expansion

The detailed characterization of T cell subpopulations provides critical insights into the immune landscape of OC tumors and the dynamics of TIL expansion. We sub‐clustered all T cells using the same strategy to identify subpopulations in T cells with known subtype markers (Figure 2a,b). As shown in Figure 2c, the higher fraction of most T cell subtypes in TIL+ tumor samples compared to TIL‐ samples suggests that a more diverse and abundant T cell infiltrate may be conducive to successful TIL isolation and expansion. Visual inspection of the proportions suggests a difference in CD4⁺ central memory T (Tcm) and Treg populations between the two rounds of ex vivo expansion (Figure 2c). However, given the very small sample size (n = 2 for each round), no formal statistical testing was performed, and these observations should be interpreted as descriptive trends.

FIGURE 2.

Characterization of OC TME and TILs through Single‐Cell Genomics. (a) UMAP visualization of the T cell population, with annotations denoting distinct T cell subtypes. (b) Stacked violin plot illustrating the expression patterns of marker genes across identified T cell subtypes. (c) Proportions of T cell subtypes in TIL+ (blue) and TIL– (orange) tumor samples, and following the first (Expansion‐r1; green) and second (Expansion‐r2; red) round of ex vivo TIL expansion. Each panel shows one T cell subtype. Boxplots indicate median and interquartile range; individual data points are overlaid. (d) Stacked bar plot depicting the proportions of predicted cell sources within each T cell subtype from tumor samples. Cells labeled as “rejected” could not be confidently assigned to a specific source by the machine learning algorithm employed. Samples OC020 and OC026 are derived from TIL+ tumors, samples OC025 and OC027 are derived from TIL‐ tumors. (e) Stacked bar plot showing the proportions of predicted cell sources within each T cell subtype in ex vivo expansion samples. As in (d), “rejected” cells indicate those that could not be confidently assigned to a specific source.

The use of the PENCIL, a supervised machine‐learning framework on the scRNA‐seq data [28], to identify phenotype‐associated subpopulations within each T cell subset is a novel approach that could enhance our understanding of T cell functionality in the context of TIL therapy. We identified high‐confidence phenotype‐associated subpopulations within each T cell subset (Figure 2d,e). Figure 2d demonstrates that predicted TIL+ cells were consistently present in TIL+ samples, while predicted TIL‐ cells were found exclusively in TIL‐ samples. Interestingly, CD4⁺ Th2 cells comprised over 50% of cells in both TIL‐ samples, while CD8⁺ effector T (Teff) and CD8⁺ effector memory T (Tem) cells exceeded 50% in TIL+ samples. Figure 2e shows that those three subtypes have high fractions of expansion round‐specific cells. The predominance of CD4⁺ Th2 cells in TIL‐ samples and CD8⁺ Teff and CD8⁺ Tem cells in TIL+ samples suggests that the balance between these T cell subsets may be crucial for successful TIL isolation and expansion.

2.4. scRNA‐Seq Reveals Enhanced Activity of Critical T Cell Subtypes in TIL+ Tumors and Rapid Expansion Culture

T cell activation triggers extensive metabolic reprogramming to support the energetic and biosynthetic demands associated with cellular activation and proliferation. Given the higher T cell infiltration rate observed in TIL+ tumors, we hypothesized that these T cells might exhibit an activated metabolic state. To investigate this, we employed Compass, which provided valuable insights into the functional state of T cells across different contexts. We applied Compass [29] to quantitatively assess the metabolic state of each phenotype‐associated subpopulation identified in our scRNA‐seq data. The algorithm evaluated 1497 reactions across 79 metabolic subsystems within T cell populations (Figure 3a–c). We first compared the metabolic profiles of CD4⁺ Th2 that were predicted to be associated with TIL+ and TIL‐ samples. As shown in Figure 3a, CD4⁺ Th2 cells linked to TIL+ samples exhibited a higher number of active reactions compared to their TIL‐ counterparts. Notably, we observed enhanced activity in the AMP Pyrophosphate pathway within purine catabolism (Figure 3d), potentially reflecting increased nucleotide metabolism necessary for rapid cell division within the TIL+ TME. Meanwhile, the majority of metabolic subsystems displayed heightened activity in CD8⁺ Tem (Figure 3b) and CD8⁺ Teff (Figure 3c) cells associated with TIL+ samples. Specifically, eicosanoid metabolism, exemplified by carbonyl reductase (NADPH) activity, was exclusively activated in TIL+ CD8⁺ Tem cells (Figure 3e). In TIL+ CD8⁺ Teff cells, we observed significant upregulation of fatty acid oxidation and tyrosine metabolism pathways (Figure 3f,g). These metabolic shifts suggest alterations in energy production and inflammatory signaling within the TIL+ TME, potentially contributing to enhanced T cell functionality.

FIGURE 3.

Differential Metabolic Activity in T Cell Subpopulations Associated with TIL Isolation Outcomes. (a–c) Metabolic state analysis of T cell subpopulations correlated with TIL isolation results, based on scRNA‐seq data. The Compass algorithm was employed to assess differential activity across 1497 reactions in 79 metabolic subsystems between predicted groups of: (a) CD4⁺ Th2 cells, (b) CD8⁺ Tem cells, and (c) CD8⁺ Teff cells. The 20 metabolic subsystems with the largest median absolute effect sizes (using Cohen's d, among reactions with adjusted P value < 0.05) are shown. Dots represent individual reactions; color indicates direction (pink: TIL+; green: TIL‐), transparency indicates significance (opaque: Adjusted P value < 0.1). (d–g) Volcano plots depicting the relationship between –log10 adjusted P values and effect sizes for specific metabolic reactions in: (d) Purine catabolism in CD4⁺ Th2 cells, (e) Eicosanoid metabolism in CD8⁺ Tem cells, (f) Fatty acid oxidation in CD8⁺ Teff cells, and (g) Tyrosine metabolism in CD8⁺ Teff cells.

Comparative analysis of metabolic activity between initial and rapid expansion phases revealed significant differential activation across all metabolic subsystems in CD4⁺ Th2, Treg (Figure S4a,b), CD8⁺ Tem, and CD8⁺ Teff (Figure S4c,d) populations. This comprehensive metabolic rewiring underscores the dynamic nature of T cell metabolism during ex vivo expansion. The observed shifts likely reflect the cells' adaptation to meet the increased energy and biosynthetic demands associated with rapid proliferation and effector function acquisition.

2.5. Dissecting Subpopulation Heterogeneity of Myeloid Cells and Fibroblasts

Both myeloid cells and fibroblasts comprise diverse subpopulations; some of these subpopulations contribute to the formation of the TME and promote tumor growth, invasion, metastasis, and resistance [30]. The identification of distinct myeloid cell and fibroblast subpopulations adds another layer of complexity to our understanding of the OC TME. Employing the same unsupervised clustering technique, we identified and characterized 10 distinct myeloid cell clusters within the OC TME (Figure 4a,b). Further analysis of the myeloid subpopulations revealed that while all dendritic cell (DC) subpopulations showed enrichment in TIL+ samples, three tumor‐associated macrophage (TAM) subpopulations were predominantly found in TIL‐ samples (Figure 4c). The enrichment of DC subpopulations in TIL+ samples is particularly interesting, given the crucial role of DCs in antigen presentation and T cell activation. This finding suggests that a more robust antigen‐presenting cell population may contribute to successful TIL isolation and expansion. The presence of these DC subpopulations could potentially create a more immunogenic microenvironment, conducive to T cell infiltration and function.

FIGURE 4.

Characterization of Myeloid Cell and Fibroblast Subpopulations in OC TME. (a) UMAP visualization of the myeloid cell population, with annotations denoting distinct myeloid cell subtypes. (b) Stacked violin plot illustrating the expression patterns of marker genes across identified myeloid cell subtypes. (c) Comparative analysis of myeloid cell subtype proportions between TIL+ and TIL‐ tumor samples. (d) UMAP visualization of the fibroblast population, with annotations indicating distinct fibroblast subtypes. (e) Stacked violin plot depicting the expression patterns of marker genes across identified fibroblast subtypes. (f) Comparative analysis of fibroblast subtype proportions between TIL+ and TIL‐ tumor samples.

The unsupervised clustering approach also revealed four distinct fibroblast clusters (Figure 4d,e). Notably, as depicted in Figure 4f, the distribution of cells within each fibroblast subpopulation varied considerably according to the sample source. Extracellular matrix‐remodeling cancer‐associated fibroblasts (CAFs), immune‐regulatory CAFs, and myofibroblasts were exclusively observed in TIL‐ samples. This exclusive presence suggests that these fibroblast subtypes may contribute to an immunosuppressive microenvironment, potentially hindering T cell infiltration and function. Conversely, TIL+ OC samples exhibited a higher abundance of pro‐inflammatory CAFs. The prevalence of pro‐inflammatory CAFs in TIL+ samples may foster a more immunologically active TME, potentially facilitating T cell infiltration and enhancing their functionality.

The differential abundances of myeloid and fibroblast subpopulations between TIL+ and TIL‐ samples underscore the complex interplay between stromal components and the immune landscape in OC. These findings suggest that the composition of myeloid cells and fibroblasts within the TME may serve as a predictor of TIL therapy outcomes and offer potential targets for therapeutic intervention to enhance TIL efficacy.

2.6. Cell‐To‐Cell Communication Network Analysis on the T Cells and Other Cell Types in the TME of OC

Given the observed metabolic distinctions in T cells and diverse cellular compositions associated with TIL+ and TIL‐ tumors, we sought to elucidate the extracellular interactions between T cells and other cell types within the TME of OC using connectomeDB2020 and NATMI [31].

We first focused on up‐regulated ligands and receptors between predicted TIL+ and TIL‐ subsets within CD4⁺ Th2, CD8⁺ Tem, and Teff populations. To ensure biological relevance, we identified ligands and receptors expressed in more than 20% of cells within other cell types, subsequently narrowing our focus to cell‐type‐specific ligands and receptors. Visualization of potential interactions between up‐regulated ligands in TIL+ CD8⁺ Tem and Teff cells and cell‐type‐specific receptors in other cell types (Figure 5a,b) revealed that these up‐regulated ligands were not cell‐type‐restricted (Figure S5), indicating non‐specific cell‐to‐cell communication patterns.

FIGURE 5.

Differential Cell‐to‐Cell Communication in T Cell Subpopulations Associated with TIL Isolation Outcomes. (a) Cell type‐specific interactions of upregulated ligands in CD8⁺ Tem cells from the TIL+ group. (b) Exclusive cell type‐specific interaction of upregulated ligands in CD8⁺ Teff cells from the TIL+ group. (c) Cell type‐specific interactions of upregulated ligands in CD4⁺ Th2 cells from the TIL‐ group. (d) Cell type‐specific interactions of upregulated receptors in CD4⁺ Th2 cells from the TIL‐ group. (e, f) Dotplots illustrating expression patterns of WNT9A and its cognate receptors in: (e) TIL+ samples, (f) TIL‐ samples. (g, h) Dotplots depicting expression patterns of EPHA1 and its cognate ligands in: (g) TIL+ samples, (h) TIL‐ samples. (i–k) Kaplan–Meier overall survival (OS) plots demonstrating the prognostic potential of: (i) WNT9A, (j) EFNA1, and (k) EFNA5 for OC patients in the TCGA cohort. The median overall survival is statistically significant (p (HR) < 0.05) for all three genes, as indicated by the text in the upper right corner of each plot.

Interestingly, TIL‐ a subset of CD4⁺ Th2 exhibited numerous up‐regulated ligands (Figure 5c) and receptors (Figure 5d) capable of interacting with cell‐type‐specific cognate receptors and ligands in other cell types. Of note, we identified the ligand WNT9A as exclusively expressed by CD4⁺ Th2 cells in TIL‐ samples (Figure 5e,f). This ligand forms an autocrine loop within Th2 cells and interacts with endothelial cells, potentially modulating vascular function or immune evasion mechanisms.

Additionally, the receptor EPHA1 was uniquely expressed in TIL‐ samples by CD4⁺ Th2 cells, forming an autocrine loop through EFNA4. Notably, all CAF subtypes can interact with Th2 cells via this receptor through other ligands such as EFNA1 and EFNA5 (Figure 5g,h). This finding suggests a complex interplay between T cells and stromal components that may influence the immune landscape.

To evaluate the clinical relevance of these extracellular cues, we utilized Gepia2 to assess their prognostic value in the Cancer Genome Atlas (TCGA) cohort [32]. Kaplan–Meier OS plots (Figure 5i) revealed that low expression of WNT9A correlates with longer survival, while its cognate receptors showed no significant association with OS. Intriguingly, ligands expressed by CAFs negatively correlated with OS, suggesting that interactions between CAFs and T cells could serve as a marker for TIL infiltration rate, thereby influencing prognosis.

The identification of WNT9A as a key player in TIL‐ samples aligns with the known roles of WNT signaling in immune regulation and cancer progression. The autocrine loop formed by WNT9A in CD4⁺ Th2 cells and its interaction with endothelial cells could represent a mechanism of immune evasion or altered vascular function in TIL‐ tumors. This correlative finding suggests a potential role for WNT signaling that could be explored in future functional studies as a therapeutic target. The expression of EPHA1 by CD4⁺ Th2 cells and its interaction with various CAF subtypes suggest a potential axis of communication between T cells and stromal cells that may influence the immune landscape. Ephrin signaling, previously implicated in angiogenesis and metastasis, may play a crucial role in shaping the immune microenvironment in OC.

2.7. Intracellular Signaling Networks in OC T Cells

Recognizing that extracellular interactions act as upstream perturbants of intracellular signaling networks, we sought to model the cellular mechanisms underlying these interactions. Differential analysis of phenotype‐associated T cell subpopulations for subtypes highlighted in Figure 2e revealed up‐regulated receptors in CD4⁺ Treg, CD8⁺ Tem, and Teff cells specific to rapid expansion conditions. We also identified up‐regulated transcription factors (TFs) and their target genes in these cells. Employing a decoupler [33] with the DoRothEA gene regulatory network [34], we identified the top 5 active TFs in these three subpopulations. Subsequently, a network‐optimization approach [35] was used to identify putative causal paths connecting up‐regulated receptors with these TFs.

In CD4⁺ Treg cells undergoing rapid expansion (Figure S6a), receptors IL9R, IL10RA, and IL21R all activated JAK1 and IFNGR1, suggesting a convergence on the JAK–STAT signaling pathway, which is crucial for T cell proliferation and function. This finding indicates that CD4⁺ Treg cells in rapid expansion conditions may be adapting their signaling networks to support enhanced proliferation and potentially altered suppressive functions. Conversely, in CD8⁺ Tem and Teff cells (Figure S6b,c), up‐regulated receptors suppressed MAFK, a TF involved in the oxidative stress response and potentially influencing T cell metabolism and function. This suppression of MAFK could represent a mechanism by which these cells modulate their metabolic state to support effector functions during rapid expansion.

The intracellular signaling network analysis reveals potential mechanisms underlying the differential behavior of T cell subpopulations in TIL+ and TIL‐ samples. The activation of JAK–STAT signaling in Treg cells during rapid expansion suggests a potential target for modulating Treg function in TIL products. This could be particularly relevant for optimizing the balance between effector and regulatory T cells in TIL preparations. Similarly, the suppression of MAFK in CD8⁺ Tem and Teff cells may indicate a metabolic adaptation favorable for effector function. Understanding these intracellular changes could inform strategies to enhance the anti‐tumor capabilities of expanded TILs, potentially through targeted metabolic interventions or signaling pathway modulations.

3. Discussion

Our preliminary multi‐omics analysis of the OC TME delineates cellular and molecular programs that shape TIL ex vivo expansion outcomes and provides a framework for neoantigen‐directed, precision immunotherapy. By integrating whole‐exome/RNA sequencing with single‐cell profiling, we connect tumor mutation/neoantigen landscapes with T cell states and stromal composition, thereby linking antigenicity to TIL selection and expansion.

A central insight is that distinct metabolic configurations in T cell subpopulations from TIL+ vs. TIL‐ tumors heightened purine catabolism, eicosanoid metabolism, and fatty‐acid oxidation track with successful TIL isolation and growth. Functionally, these programs may mark or enable neoantigen‐reactive CD8⁺ Teff/Tem cells, suggesting that metabolic signatures could serve as practical biomarkers to prioritize T cells for downstream co‐culture selection with autologous tumor cells or antigen‐presenting cells to enrich for neoantigen specificity as a potential future application informed by our findings, rather than a procedure used in the current study. These pathways are therefore attractive levers for metabolic modulation to boost TIL fitness during manufacturing.

Stromal heterogeneity further stratifies the OC TME into “hot” and “cold” configurations with direct implications for neoantigen presentation and T cell function. The enrichment of DC subsets in TIL+ tumors supports a more immunogenic milieu conducive to antigen cross‐presentation, while specific TAM states prevalent in TIL‐ tumors likely reinforce an immunosuppressive ecosystem. Similarly, fibroblast diversity, with pro‐inflammatory CAFs more abundant in TIL+ samples, underscores how stromal programs can amplify or dampen antitumor immunity. These observations argue that effective therapy should pair tumor‐intrinsic targeting with TME remodeling to convert “cold” tumors into “hot,” antigen‐presenting niches that support neoantigen‐driven TIL activity.

Conventional chemotherapy can profoundly reshape the OC TME. While cytotoxic therapy may transiently enhance antigen release and T cell priming, it can also promote the accumulation of immunosuppressive myeloid populations and pro‐fibrotic CAF states that contribute to immune exclusion and therapeutic resistance [13]. These dynamic changes may partially explain disease relapse following initial treatment responses and underscore the importance of timing and combination strategies when integrating TIL therapy with standard treatments.

Our cell‐to‐cell communication analysis highlights signaling axes (e.g., WNT9A and EPHA1) that are selectively engaged in TIL‐ tumors, consistent with immune evasion and perturbed vascular/structural support. Their associations with OS in TCGA reinforce clinical relevance and nominate tractable targets for combination therapy (e.g., Wnt/planar‐cell‐polarity or Eph/Ephrin pathway interference). It is critical to emphasize that the cell‐to‐cell communication interactions highlighted by our analysis, such as those involving WNT9A and EPHA1, and their association with survival in the TCGA cohort, are correlative. Their functional significance in modulating the TME and impacting TIL therapy efficacy remains to be experimentally validated using approaches such as ligand‐receptor blockade or genetic perturbation in relevant models. Complementarily, the intracellular network analysis during ex vivo expansion points to pathway‐level opportunities: JAK–STAT activity in CD4⁺ Tregs and MAFK‐linked repression in CD8⁺ T cells may influence the balance between suppressive and effector compartments. Together, these extracellular and intracellular insights suggest rational combinations‐metabolic conditioning, TME modulation, and pathway‐specific inhibitors‐layered with immune checkpoint blockade to enhance neoantigen‐specific TIL potency.

The integrated multi‐omic framework presented here provides a blueprint for personalized immunotherapy in OC. By linking tumor genomic features with T cell states, metabolic activity, and stromal interactions, this approach may enable rational patient selection, neoantigen‐directed TIL enrichment, and metabolic conditioning of TIL products. Similar strategies have guided adoptive cell therapy optimization in melanoma and gastrointestinal cancers [36], supporting the translational potential of this approach in ovarian cancer.

While our study provides valuable insights into the complex dynamics of the OC TME and their impact on TIL therapy, it is important to acknowledge certain limitations. The sample size (n = 10) is characteristic of a pilot investigation, which, while enabling deep multi‐omics profiling, limits the statistical power and generalizability of our findings. The results presented here are, therefore, exploratory and hypothesis‐generating. Future validation in larger, independent cohorts is essential to confirm the predictive value of the identified biomarkers. Furthermore, the enzymatic dissociation process required for single‐cell analysis, although standardized across all samples, may have influenced the viability and recovery of certain cell populations, such as fragile stromal or immune cells, and could potentially alter surface epitopes. While our primary analyses relied on robust transcriptomic signatures, this technical aspect should be considered when interpreting cellular abundances. Moreover, the changes in the gene level require confirmation at the protein level by techniques such as flow cytometry, CITE‐Seq [37]. Nonetheless, the integrated design, linking genetics, transcriptomics, metabolism, and stromal architecture to intercellular signaling, yields convergent biomarkers and mechanistic targets for improving TIL manufacture and function.

4. Materials and Methods

4.1. Patients and Ethical Regulations

Ten patients with primary OC were evaluated and treated at Hainan Affiliated Hospital of Hainan Medical University. The study complies with all relevant local, national, and international regulations. In this study, all procedures involving human participants were in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants during enrollment. The Hospital IRB has approved this study under ethics approval number Med‐Eth‐RE[2021] 275.

4.2. Tumor‐Infiltrating Lymphocyte (TIL) Expansion Protocol

Fresh tumor samples were obtained from surgery or needle biopsy. TILs were cultured from tumor fragments, following a previously described approach [10]. Briefly, fresh tumor specimens were mechanically minced into about 1–2 mm³ fragments and cultured in RPMI‐1640 supplemented with 10% human AB serum and recombinant human IL‐2 (6000 IU/mL) to initiate TIL outgrowth (young TILs, yTILs) for up to 11 days. Cultures were monitored for cell expansion and viability. Successful TIL growth (TIL+) was defined by meeting all of the following pre‐specified manufacturing criteria: (i) recovery of ≥ 2 × 10⁷ viable CD3⁺ T cells after initial expansion; (ii) cell viability ≥ 70% at harvest; and (iii) predominance (> 80%) of CD3⁺ T cells with detectable CD8⁺ effector or effector‐memory subsets. Tumor samples failing to meet one or more of these criteria were classified as TIL‐. Although the absolute number of infused TILs has been positively associated with TIL therapy efficacy [38], clinical outcomes were not used to define TIL status here, as the primary objective of this study was to identify tumor‐ and microenvironment‐associated predictors of TIL manufacturability. When sufficient cell numbers were obtained, TILs underwent a single standardized rapid expansion protocol (REP) using irradiated allogeneic feeder cells at a 100:1 feeder‐to‐TIL ratio, anti‐CD3 antibody (OKT3, 30 ng/mL), and IL‐2 (3000 IU/mL) for 11 days before harvesting. All TIL products were checked for sterility, flow‐cytometric immunophenotyping (CD3, CD4, CD8, CD45RA, CCR7, PD‐1), and ELISA‐based IFN‐γ release test.

4.3. Tumor Sample Processing for DNA/RNA Sequencing

Tumor‐specific mutations and splicing variations were identified from tumor tissue samples using whole‐exome sequencing (WES) and bulk RNA‐seq. After collection, the OC tissue samples were stored in RNAlater reagent (QIAGEN) until genomic DNA or total RNA isolation. Genomic DNA (gDNA) was extracted from tumors and matched normal blood using the QIAamp Fast DNA Tissue Kit and Blood & Cell Culture DNA Kit (QIAGEN). Total RNA was isolated and purified from tumor tissues using the miRNeasy Kit (QIAGEN) after being homogenized in TRIzol.

4.4. WES Library Preparation and Sequencing

4.4.1. Evaluation of DNA Quality

The quality of isolated genomic DNA was verified by using these two methods in combination: (1) DNA degradation and contamination were monitored on 1% agarose gels; (2) DNA concentration was measured by Qubit DNA Assay Kit in Qubit 3.0 Fluorimeter (Invitrogen, USA).

4.4.2. Library Preparation

The exome sequences were efficiently enriched from 0.4 μg genomic DNA using the Agilent liquid capture system (Agilent SureSelect Human All Exon V6) according to the manufacturer's protocol. Firstly, qualified genomic DNA was randomly fragmented to an average size of 180–280 bp by the Covaris S220 Sonicator. Remaining overhangs were converted into blunt ends via exonuclease polymerase activities. Secondly, DNA fragments were end‐repaired and phosphorylated, followed by A‐tailing and ligation at the 3′‐ends with paired‐end adaptors (Illumina). DNA fragments with ligated adapter molecules on both ends were selectively enriched in a PCR reaction. After the PCR reaction, libraries were hybridized with a liquid phase with biotin‐labeled probe, and then used magnetic beads with streptavidin to capture the exons of genes. Captured libraries were enriched in a PCR reaction to add index tags to prepare for sequencing. At last, the DNA library was sequenced on Illumina for paired‐end 150 bp reads.

4.4.3. Clustering & Sequencing

The clustering of the index‐coded samples was performed on a cBot Cluster Generation System using Illumina PE Cluster Kit (Illumina, USA) according to the manufacturer's instructions. After cluster generation, the DNA libraries were sequenced on the Illumina platform, and 150 bp paired‐end reads were generated.

4.5. Bioinformatics Analysis of WES Data

4.5.1. Data Quality Control

Quality control is an essential step and is applied to guarantee the meaningful downstream analysis, which is performed by fastp (v0.23.2) [39]. The steps of data processing were as follows:

(1) Discard paired reads if either one read contains adapter contamination (> 10 nucleotides aligned to the adapter, allowing ≤ 10% mismatches).

(2) Discard paired reads if more than 10% of bases are uncertain in either one read.

(3) Discard paired reads if the proportion of low‐quality (Phred quality < 5) bases is over 50% in either read.

4.5.2. Reads Mapping to Reference Sequence

Valid sequencing data were mapped to the reference human genome (GRCh38) by Bowtie2 (v2.4.5) [40], which was selected for its efficiency and high accuracy in aligning whole‐exome sequencing data, particularly for germline variant detection. Then, SAMtools (version 1.15.1) [41] and Picard (http://broadinstitute.github.io/picard/) [42] were used to sort BAM files and do duplicate marking, local realignment, and base quality recalibration to generate a final BAM file for computation of the sequence coverage and depth. The mapping step was very difficult due to mismatches, including true mutations and sequencing errors, and duplicates resulted from PCR amplification. These duplicate reads were uninformative and shouldn't be considered as evidence for variants. We used Picard to mark these duplicates for follow‐up analysis.

4.5.3. Variant Detection

SAMtools mpileup and bcftools (v1.15.1) [41] were used to do variant calling and identify SNPs, InDels. CNVkit (v0.9.9) was utilized to do CNV detection.

4.5.4. Somatic Mutation Calling

The somatic SNV was detected by Varscan (v2.4.4) [43], the somatic InDel by Strelka (Saunders, Wong et al.), and CNVkit (v0.9.9) [44] was used to detect somatic CNV (Boeva V et al.).

4.5.5. Annotation

Ensembl‐vep (v99.0) [45] is used to perform annotation for the VCF (Variant Call Format) file obtained in the previous step. The variant position, variant type, conservative prediction, and other information are obtained at this step through a variety of databases.

4.5.6. TMB Measurement Approach

The Tumor Mutational Burden (TMB) is usually defined as the total number of non‐synonymous mutations per coding area of a tumor genome, which is calculated by pyTMB [46] (GitHub‐bioinfo‐pf‐curie/TMB: Tumor Mutational Burden) with the somatic mutation calling results.

4.5.7. HRD Testing via WES

Homologous recombination deficiency (HRD) was assessed by calculating three component scores: Telomeric allelic imbalance (Telomeric AI), loss of heterozygosity (HRD‐LOH), and number of large‐scale transitions (LST). The HRD‐LOH score quantifies the number of distinct genomic regions exhibiting loss of heterozygosity exceeding 15 megabases in length that do not encompass an entire chromosome. The LST score enumerates large‐scale chromosomal transitions, defined as breaks positioned between two adjacent genomic segments each measuring at least 10 megabases, with inter‐segment distances not exceeding 3 megabases. The Telomeric AI score represents the number of allelic imbalance events extending to the telomeric terminus of a chromosome. The composite HRD score was derived from the summation of the HRD‐LOH, LST, and Telomeric AI scores. All three component scores were computed using the R package scarHRD (https://github.com/sztup/scarHRD) with default parameters.

4.6. RNA‐Seq Library Preparation and Sequencing

4.6.1. Total RNA Quantification and Qualification

Total amounts and integrity of total RNA were assessed using the RNA Nano 6000 Assay Kit of the Bioanalyzer 2100 system (Agilent Technologies, CA, USA).

4.6.2. Library Preparation for Transcriptome Sequencing

Total RNA was used as input material for the RNA‐seq library preparations. Briefly, mRNA was purified from total RNA by using poly‐T oligo‐attached magnetic beads. Fragmentation was carried out using divalent cations under elevated temperature in First Strand Synthesis Reaction Buffer (5X). First‐strand cDNA was synthesized using random hexamer primer and M‐MuLV Reverse Transcriptase, followed by treatment with RNase H to degrade the RNA. Second‐strand cDNA synthesis was subsequently performed using DNA Polymerase I and dNTP. Remaining overhangs were converted into blunt ends via exonuclease/polymerase activities. After adenylation of the 3′ ends of DNA fragments, Adaptors with hairpin loop structure were ligated to prepare for hybridization. To select cDNA fragments of preferentially 370 ~ 420 bp in length, the library fragments were purified with the AMPure XP system (Beckman Coulter, Beverly, USA). After PCR amplification, the product was purified by AMPure XP beads, and the library was ready for sequencing.

4.6.3. Clustering and Sequencing

After the library is qualified, the different libraries are pooled according to the effective concentration and the target amount of data off the machine, and then are sequenced on the Illumina NovaSeq 6000. The end reading of 150 bp pairing is generated.

4.7. Bioinformatics Analysis of Bulk RNA‐Seq Data

4.7.1. Quality Control

The image data measured by the high‐throughput sequencer is converted into sequence data (reads) by CASAVA base recognition. Raw data (raw reads) of fastq format were first processed by fastp (v0.23.2). In this step, clean data (clean reads) were obtained by removing reads containing adapters, reads containing N base and low‐quality reads from the raw data. At the same time, Q20, Q30, and GC content of the clean data were calculated. All the downstream analyses were based on the clean data with high quality.

4.7.2. Reads Mapping to the Reference Genome

Reference genome and gene model annotation files of the GRCh38 reference genome were downloaded directly from the genome website. An index of the reference genome was built using Hisat2 (v2.2.1) [47], and paired‐end clean reads were aligned to the reference genome using Hisat2 as well.

4.7.3. Quantification of Gene Expression Level

Transcript‐level abundance was estimated using Kallisto (v0.48.0) [48] to obtain Transcripts Per Million (TPM) values, which were used for gene expression visualization and normalization.

4.7.4. Tumor Digest Preparations for Single‐Cell RNA‐Seq

After resection, the fresh tumor tissues were immediately stored in MACS Tissue Storage Solution (Miltenyi Biotec) and kept at 2°C –8°C during transportation. Within 24 h, samples were transferred into RPMI‐1640 medium (Gibco) and enzymatically digested with a Gentle MACS Tumor Dissociation Kit (Miltenyi Biotec) following the manufacturer's protocol. The tissues were incubated at 37°C for 60 min using standardized concentrations of Enzyme H, R, and A across all samples. The dissociated cells were then passed through a 70‐mm cell‐strainer (BD). Red blood cells were removed before single‐cell RNA‐Seq analysis.

4.8. Single Cell cDNA Library Preparation and Sequencing

Single‐cell RNA‐seq libraries were prepared using the Chromium Next GEM Single Cell 5 Kit v2 from 10× Genomics, following the manufacturer's instructions. In brief, single cells were resuspended in PBS containing 0.04% BSA to a final concentration of ~500–1,200 cells/ml. About 10,000 cells were captured in droplets to generate nanoliter‐scale Gel bead in EMulsion (GEMs). GEMs were then reverse transcribed. After reverse transcription and cell barcoding, emulsions were broken, and cDNA was isolated, followed by PCR amplification. Amplified cDNA was fragmented and end‐repaired, PCR‐amplified with sample indexing primers. Constructed libraries were sequenced on an Illumina HiSeq X‐Ten sequencer with 150 bp paired‐end reads.

4.9. Bioinformatics Analysis of Single‐Cell Sequencing Data

The Cell Ranger toolkit (version 7.0.0) provided by 10× Genomics was applied to aggregate raw data, filter low‐quality reads, align reads to the 10× Genomics‐provided human reference genome (GRCh38), assign cell barcodes, and generate the unique molecular identifier (UMI) matrix.

A Python‐based toolkit, Scanpy (version 1.9.1) [22], was used for analyzing the single‐cell RNA‐Seq (scRNA‐seq) data. Specifically, the raw UMI matrix was processed to filter out genes detected in fewer than 3 cells and cells with fewer than 200 genes. We further filtered the data to retain high‐quality cells with a mitochondrial gene fraction below 20% to exclude cells that are damaged or undergoing apoptosis, and a total UMI count below 11,000 to exclude potential doublets or multiplets, which can artifactually exhibit very high UMI counts. Scrublet [49] was then applied to each sequencing library to remove potential doublets with the expected doublet rate of 6% (not removed, just marked). The normalized expression matrix was calculated based on the raw UMI counts after normalizing total counts per cell (library size) and then scaled by 1e6 and logarithmically transformed.

Dimension reduction and unsupervised clustering were performed according to the standard workflow in Scanpy. In brief, unwanted sources of variation, such as total counts, were regressed out from the normalized expression matrix. After scaling each gene to unit variance and clipping values exceeding a standard deviation of 10, principal component analysis (PCA) was performed on the variable gene matrix to reduce noise, and the top 50 components were used for downstream analyses. A k‐nearest neighbor (KNN) graph was then constructed based on the Euclidean distance in PCA space, and the edge weights between any two cells were refined based on the shared overlap in their local neighborhoods. The Leiden algorithm was applied to such KNN graphs to detect communities and find cell clusters. Note, the same principal components were also used for non‐linear dimension reduction to generate the Uniform Manifold Approximation and Projection (UMAP) for visualization.

After the first round of unsupervised clustering, marker genes and differentially expressed genes were detected using the rank_genes_groups function with parameters of method = ‘lwilcoxon’. We annotated each cell cluster according to these detected cell markers. Next, we performed a second round of unsupervised clustering on some major cell types to obtain the subpopulations of these cell types in OC. The second‐round clustering procedure was the same as the first‐round clustering, both of which started from the normalized expression matrix, and then calculated the PCA matrix, detected cell clusters by the Leiden algorithm, performed dimensionality reduction for visualization, and identified subpopulations based on the detected marker genes.

To identify malignant cells, we assessed CNV profiles in the dataset based on the reasoning that epithelial cells with a high degree of CNVs are likely malignant. We used a sliding window approach executed in the ‘infercnvpy’ package to obtain CNV profiles. Normal reference cells were T cells.

PENCIL v0.7 [28] was utilized to predict T cells associated with a specific tumor source or expansion round. The UMAP coordinates, log‐transformed expression matrix, and sample source for each T cell were integrated into PENCIL for analysis. Cells with a confidence score above 0 were labeled with the predicted sample source, while others were marked as ‘Rejected’, indicating a lack of group‐specific characteristics. For each subtype, the rank_genes_groups function was applied once more with a Wilcoxon test to discern significantly up‐ and down‐regulated genes (false discovery rate (FDR) < 0.05) in the phenotype‐associated subpopulation.

We ran Compass [29] on the normalized count data between the cells with confident PENCIL labels in the critical T cell subtypes. Reaction scores for every single cell were processed and visualized using the Python script provided by the authors.

NATMI was run on the normalized expression data from two patient groups, respectively, with default settings, ligands, and receptors in connectomeDB2020, with a detection rate lower than 20% in a cell type, which are considered as not expressed in the corresponding cell type. The expression level of a detected ligand/receptor within a cluster is the mean expression level of the ligand/receptor across all cells of the same cell type. Using NATMI, we also compared and visualized the cell‐connectivity‐summary networks of top‐ranked cell‐type pairs for each dataset.

The relationships between the expression of selected genes and overall survival in the TCGA OV cohort were analyzed by the “survival analysis” module of GEPIA2.

We ran LIANA+ [50], an all‐in‐one cell–cell communication analysis framework, to infer signaling networks from prior knowledge, linking up‐regulated receptors from connectomeDB2020 to downstream intracellular signaling pathways revealed by the decoupler Python package and transcription factors involved in DoRothEA.

4.9.1. Statistical Analysis

Statistical analyses were performed using R (v 4.2.1) and Python (v 3.9). Comparisons of cell‐type proportions between TIL+ and TIL‐ tumor samples were conducted using two‐sided Student's t‐tests, given the exploratory nature of the study and limited sample size. Differential gene expression analyses were performed using Wilcoxon rank‐sum tests with false discovery rate (FDR) correction (Benjamini–Hochberg). For single‐cell metabolic and cell–cell communication analyses, effect sizes were calculated using Cohen's d. P‐value < 0.05 was considered statistically significant unless otherwise stated.

Funding

This study is supported by Joint Program on Health Science & Technology Innovation of Hainan Province and Clinical Translational Innovation Cultivating Fund 550 Project of Hainan General Hospital WSJK2024MS125 and 2021CXZH03.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Summary of somatic mutation in OC patients based on WES data. (a) The missense mutation is the most common mutation in the variant classification of OV patients. (b) SNP is the most common genetic variant type. (c) C > T is the most common SNV transition. (d) The mutational load per OV patient sample. (e) The variant classification summary represents a box plot of the numbers across all OV patients. (f) The top 10 frequently mutated genes across the cohort. Figure S2: Differential Gene Expression and Pathway Enrichment Analysis of RNA‐seq data of OC samples. (a) Volcano plot illustrating differentially expressed genes between two OC stages (early vs. late). Genes with absolute log2 fold change (|logFC|) > 1 and false discovery rate (FDR)‐adjusted p < 0.05 are highlighted. Red and blue points represent significantly up‐ and down‐regulated genes, respectively. (b, c) Bar plots depicting the top significantly enriched biological process KEGG (Kyoto Encyclopedia of Genes and Genomes) terms for genes up‐regulated in early‐stage OC (b) and in late‐stage OC (c). The x‐axis represents the negative log10 of the adjusted p‐value, and the y‐axis lists the enriched pathways. Figure S3: Inferred CNV profiles for all cells obtained by scRNA‐seq data. Figure S4: Metabolic State Analysis of T Cell Subpopulations Correlated with Ex Vivo Expansion Rounds Based on scRNA‐seq Data. The Compass algorithm was employed to assess differential metabolic activity across 1497 reactions (represented by dots) in 79 metabolic subsystems between predicted groups of: (a) CD4⁺ Th2 cells, (b) CD4⁺ Treg cells, (c) CD8⁺ Tem cells, and (d) CD8⁺ Teff cells. Effect sizes were determined using Cohen's d. Figure S5: Differentially Activated Cell‐to‐cell communication in CD8⁺ Tem and Teff a–b. Dotplots depicting expression patterns of identified ligands and receptors highlighted in Figure 5a,b: (a) In TIL+ samples, (b) In TIL‐ samples. Figure S6: Causal Intracellular Signaling Networks of Upregulated Receptors and Downstream Transcription Factors in Rapid Expansion‐Associated T Cell Subpopulations a–c. Intracellular signalling networks in: (a) CD4⁺ Treg, (b) CD8⁺ Tem, and (c) CD8⁺ Teff.

Table S1: Summary of treatment histories at time of biopsy, clinicopathological features, and genomic features across profiled samples.

Table S2: Differentially expressed genes (DEGs) in two OC subtypes divided by PC1.

Data Availability Statement

The data that support the findings of this study are openly available in Zenodo at https://zenodo.org/, reference number 18382163.