Identification of macrophage-enriched genes in ovarian cancer by single-cell RNA sequencing and establishment of a prognosis model Authors: Yingrui Gu, Jiajia Tan, Linggang Zhou, Di Yao, Yan Kuang Published: 2026 ABSTRACT AbstractOvarian cancer (OC) is the most lethal gynecologic malignancy, presenting insidious onset and lacking specific biomarkers. Tumor-associated macrophages (TAMs) modulate immunity via M1/M2 plasticity, yet their heterogeneity and pro-metastatic mechanisms remain unclear. Using scRNA-seq, we compared immune profiles of OC and normal tissues, extracted TAM modules with hdWGCNA, and intersected them with DEGs to screen candidate genes Univariate/multivariate Cox and LASSO refined a seven-gene prognosis model (EPB41L2, AAK1, PRPF38B, RB1, GPR34, CLEC12A, BRD2) and established a nomogram. We also analyzed the DEGs’ functional enrichment, immune signature, immune infiltration profile, and drug sensitivity. Unique immune cell infiltration landscape and gene expression profiles in the tumor microenvironment in OC were revealed. Based on the median risk score, OC patients were assigned to high-risk (HR) and low-risk (LR) groups, and the latter had far longer survival time than the former (P < 0.0001), which was validated in GEO (P = 0.018, P = 0.0063). The HR group had significantly increased Dysfunction Score (P < 0.01), MSI Score (P < 0.01), ESTIMATEScore (P = 0.045), and StromalScore (P = 0.0031) rather than TumorPurity Score (P = 0.045). The seven-gene prognosis model possesses a certain capability to predict survival rates and correlates with drug sensitivity and immunotherapy response. Identified immunoregulatory targets provide a theoretical basis for TAM-targeted treatments.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-56858-y.Keywords: Ovarian cancer, Macrophage, Single-cell RNA sequencing, Prognosis modelSubject terms: Biomarkers, Cancer, Computational biology and bioinformatics, Immunology, Oncology Introduction Characterized by insidious early symptoms and no specific biomarkers1, ovarian cancer (OC) has progressed to advanced stages in about 70% of patients when diagnosed2, making it the most fatal gynecological malignancy (a 5-year survival rate < 50%)3,4 and bringing a heavy burden to patients and society. Age, family history, genetic susceptibility (e.g., BRCA1 and BRCA2 mutations, and Lynch syndrome)5,6, reproductive history, and hormone replacement therapy are all risk factors for OC. The OC diagnosis and treatment remain challenging nowadays, and some limitations are present in available methods such as CA125, human epididymis protein 4, CA19-9, carcinoembryonic antigen, mesothelin, and other tumor markers and imaging tests7,8, which are not effective in ameliorating the prognosis9. Given the lack of alternative treatments for OC, searching for effective biomarkers and establishing prediction models for OC risk are crucial. The tumor microenvironment (TME) encompasses extracellular matrix (ECM), stromal cells, and exosomes, and it creates autocrine-paracrine pathways to modulate tumor cell invasion and metastasis10. Active immune cells recruited by tumor cells create an immunosuppressive TME by secreting molecules in the early stage of cancer development, interfering with tumorigenesis and progression. However, depletion or even remodeling will occur in relevant immune effector cells following sustained antigen stimulation and immune activation; as a result, they fail to exert tumor-killing effects and may even transform into a malignant phenotype, creating an immunosuppressive microenvironment11. Macrophages, a complex player in the TME, are phagocytic cells that engulf cellular debris to enhance a tolerant environment for tissue homeostasis. When macrophages gradually accumulate in solid tumors, they are known as tumor-associated macrophages (TAMs). Based on their function and phenotype, macrophages can be classified into M1 and M2 types. The M1 type usually possesses antitumor activity that suppresses tumor growth by generating cytokines and inducing apoptosis, whereas the M2 type contributes to tumor growth and metastasis by restraining immune responses and promoting angiogenesis and tumor cell migration. In OC, M2-type TAMs are usually dominant, which enhances tumor growth and metastasis, and as revealed by clinical trials, tumor cells (including OC) and TAMs display a positive correlation and an ongoing dynamic interaction12–14. One study suggests that cytokines (e.g., IL-4 and IL-10) can facilitate M2 polarization of TAMs, but the key molecular switches (e.g., transcription factors, and epigenetic modifications) during polarization remain unclear15. With the increasing use of single-cell omics, multiple cell types enriched in tumors have been revealed by single-cell studies16, enhancing our understanding of OC, and single-cell assessment has identified different cell subsets and important contributors to tumor heterogeneity, providing essential information for further clarifying the mechanism of OC progression17. Moreover, the number of TAMs is highly correlated with the OC prognosis, so therapeutic strategies targeting macrophages have become a research hotspot. This study intends to analyze the molecular signature of TAMs in OC using single-cell RNA sequencing (scRNA-seq) and then develop a prognosis model. The innovation lies in the combination of advanced techniques to get a whole picture of the role of macrophages in OC, hoping to improve the OC diagnosis and prognosis. Materials and methods Data We acquired scRNA-seq and transcriptome data of human OC tissues from GEO (www.ncbi.nlm.nih.gov/geo/), and downloaded GSM5599220 and GSM5599227 for single-cell analyses. We accessed GSE102073 (November 8, 2024) for patients pathologically diagnosed with high-grade serous OC, and after those with no prognosis data were excluded, the clinical and scRNA-seq data were finally extracted from 84 OC patients. In addition, the transcriptome data of tissue samples and clinical data of 420 OC patients in The Cancer Genome Atlas (TCGA) (gdc.cancer.gov/) were acquired. The TCGA-OV samples served as a training set, and GSE102073 and GSE26193 as validation sets. No ethical approval was needed for data access to public databases. Single-cell data processing A Seurat object was created from scRNA-seq data by the R Seurat package. First, the data were subjected to quality control (cell count of 500–6000, and mitochondrial gene content < 25%). Then a principal component analysis (PCA) was conducted on the top 1000 most variable genes, and we selected 10 principal components for clustering and dimensionality reduction. The batch effect was eliminated by the R Harmony package (v1.2.3), followed by dimensionality reduction with t-SNE and UMAP. FindNeighbors and FindClusters were used for neighbor search and cluster analysis to help understand the data structure. Cell type annotation was completed using the celldex (v1.16.0) and SingleR (v2.8.0) packages, and human primary cells were extracted by HumanPrimaryCellAtlasData. Based on specific cell type annotation, macrophages were extracted from the overall scRNA-seq dataset and underwent a single-cell analysis in the same way. Using FindMarkers, differentially expressed genes (DEGs) in macrophages were identified (log2FC > 0.15 and Padj<0.05) and validated using macrophage marker genes (CD68, CD163, C1QA, C1QB, and C1QC). Enrichment analysis DEGs underwent Gene Ontology (GO) (P < 0.01), Kyoto Encyclopedia of Genes and Genomes (KEGG), and gene set enrichment analysis (GSEA) using clusterProfiler package (v4.14.4). The GO gene set in GSEA was downloaded from MSigDB (http://www.gsea-msigdb.org). Cell-cell communication and pseudo-time trajectory analyses Trajectory-based differential expression analysis was performed using Monocle (v2.34.0). A full model incorporating both the pseudo-time trajectory (across cell subsets) and batch effects was compared against a reduced model containing batch effects only using a likelihood ratio test. Genes with a false discovery rate (FDR)-adjusted P-value < 0.01 were considered significantly associated with the trajectory. Identification of key modules of TAMs by hdWGCNA The genes of TAMs in OC were identified by the hdWGCNA package (v0.4.00). Briefly, metacells were created using MetacellsByGroups (25 nearest neighbors, maximum number of shared cells between two metacells = 10) to reduce technical noise and preserve biological heterogeneity. A subset of Seurat objects was constructed to acquire specific cell subsets of interest. Then standard hdWGCNA was done by executing ConstructNetwork, TestSoftPowers, ModuleConnectivity, ModuleEigengenes, and RunModuleUMAP functions (default parameters). In the above way, the enriched modules for OC were selected, and their intersection with the DEGs was taken. Development and validation of prognostic signature From the intersecting genes, we initially screened the TCGA-OA cohort for prognosis-related genes (PRGs) by univariate Cox regression (P < 0.05). Then to minimize overfitting, we conducted LASSO Cox regression using the R glmnet package. Finally, we conducted stepwise multivariate Cox regression for convergence of variables and a risk signature was constructed based on the optimal genes. The risk score was calculated for each patient using the Predict function, and according to the median risk score, OC patients were assigned to high-risk (HR) and low-risk (LR) groups. We drew the Kaplan-Meier (K-M) curves by the survival (v3.8-3) and survminer packages (v0.5.0). The area under the ROC curve (AUC) was calculated by the timeROC package (v0.4). External validation (GSE102073 and GSE26193) was also conducted. Construction and assessment of nomogram With clinicopathologic features (grade, age, and clinical stage) in TCGA and risk scores incorporated, we conducted univariate and multivariate Cox regression to detect the influence of independent variables on prognosis. A nomogram was created using independent prognostic factors by the rms package (v6.9-0)17 to assess the 1-, 2-, and 3-year survival status. Subsequently, the concordance between the predicted and actual results was explored by calibration curves, and we made a decision curve analysis (DCA) to assess the model’s net clinical benefit. Immune infiltration and immune microenvironment (IME) signature To assess the immune infiltration status, the infiltrating immune cell score and immune-related pathway activity were calculated by the CIBERSORT package (v0.1.0)18,19. The level of infiltrating immune and stromal cells was inferred by ESTIMATE19 to quantify the proportion or abundance of immune, stromal, and tumor cells associated with the TME. In addition, Tumor Immune Dysfunction and Exclusion (TIDE) scores were given by TIDE (https://tide.dfci.harvard.edu/). Drug sensitivity analysis As described in previous studies20,21, a drug sensitivity score was calculated and converted to half maximal inhibitory concentration (IC50), a measure of drug effectiveness, by the pRRophetic package (v0.5). Then the HR and LR groups were compared for drug sensitivity, with a threshold at P < 0.01. Statistical analysis R4.4.2 was adopted for analyses. P < 0.05 suggested statistical significance. K-M survival analyses were conducted. We compared quantitative variables by the Wilcoxon rank-sum test, and determined independent risk factors by univariate and multivariate Cox regression. The association of genes with immune cells was calculated by Spearman correlation analyses. Results Single-cell transcriptome of normal and OC macrophages To investigate the heterogeneity of macrophages in human OC tissues, we collected and analyzed scRNA-seq data of an OC patient (27,984 cells expressing 8823 genes) and a normal control (27,984 cells expressing 6662 genes). Following quality control, reduction of batch effects, dimensionality reduction, and cluster analysis, the “0” cell cluster was annotated as NK cells using the singleR package, but it was found that T cell marker genes (CD3D, CD3E, TRAC, and CD8A) and NK cell marker genes (NKG7, GNLY, KLRD1, and NCAM1) were all enriched in the “0” cell subset. Given the partial overlapping features of T and NK cells, the “0” cell subset was defined as NK/T cells, while the remaining six cell subsets as “Fibroblasts”, “Macrophage”, “Endothelial_cells”, “Epithelial_cells”, “MSC”, and “B_cell” (Fig. 1A and B and S1-S2). To validate the accuracy of the above annotations, heatmaps were generated for markers in each cell subset (Figure S3), and the expression of marker genes for macrophages (CD68, CD163, C1QA, and C1QB), B cells (IGHG1, MZB1, SDC1, and CD79A), endothelial cells (PECAM1 and VWF), chondrocytes (SPIB and MS4A1), epithelial cells (EPCAM and PAX8), fibroblasts (COL1A1 and PDGFRA), MSCs (TOP2A and GTSE1), and NK/T cells (CD3D, CD3E, TRAC, CD8A, NKG7, GNLY, KLRD1, and NCAM1) was searched (Figure S4). In addition, we calculated the proportion of each cell subset (Fig. 1C), and their DEGs were visualized (Fig. 1D). Visualization of various cell subsets in OC and normal tissues and single-cell analysis of heterogeneity in TAMs in OC. (A) UMAP plot of clustering distribution of cells in the two samples. (B) UMAP plots before and after batch correction by Harmony. (C) Distribution of different cell subsets in OC and normal tissues. (D) Volcano plot for marker genes. (E) Macrophages clustered into five subgroups. (F) Visualization of “OC-enriched macrophages” and “Normal-enriched macrophages”. (G) KEGG analysis of DEGs. (H) GO analysis of DEGs (BP, CC and MF). (I) GSEA of DEGs. Then the t-SNE analysis was repeated for hierarchical clustering of macrophages. After cell type annotation, we identified five subsets of macrophages (Fig. 1E). In the “0–3” subgroup, the number of TAMs significantly rose in OC tissues, so it was annotated as “OC-enriched macrophages”, while the remaining subgroups as “Normal-enriched macrophages” (Fig. 1F), suggesting that these specific cell subsets in “0–3” may play key roles in OC development and progression. We identified 820 DEGs between OC-enriched macrophages and Normal-enriched macrophages. These DEGs underwent the KEGG analyses, and they were of high significance in the ribosome, phagosome, antigen processing, and presentation-related pathways (Fig. 1G). The enrichment analyses revealed that upregulated DEGs were of high significance in the phagosome, antigen processing, and presentation-related pathways (Supplementary Figure S5A), while downregulated DEGs were of high significance in the ribosome and leukocyte transendothelial migration-related pathways (Supplementary Figure S5B). The GO analyses revealed that the DEGs were mainly involved in “co-translational protein targeting to membrane”, “SRP-dependent co-translational protein targeting to membrane”, and “protein targeting to ER” (Fig. 1H). Upregulated DEGs were involved in biological functions such as the processing and presentation of peptide antigens via MHC class II molecules, nuclear speckle, and DNA-binding transcription factor binding (Supplementary Figure S5C). Downregulated DEGs were closely related to cytoplasmic translation and ribosome-related functions (Supplementary Figure S5D). The GSEA also found significant enrichment of DEGs in positive regulation of B-cell activation and anti-bacterial humoral immune response, indicating that these DEGs are key players in related immune processes (Fig. 1I). Pseudo-time trajectory and cell-cell communication analyses To explore the origin of macrophages during OC development, marker genes (CD68, CD163, C1QA, C1QB, and C1QC) were used for validation in cell clusters, and they all showed strong expression (Supplementary Figure S6). A pseudo-time trajectory analysis was then conducted on the macrophages. Early-trajectory macrophages were in an immature, unpolarized state (Fig. 2A). As the pseudo-time trajectory advanced, the cells gradually acquired specific functional phenotypes, indicating a dynamic process of macrophage polarization and functional differentiation in the TME. Besides, the Normal-enriched macrophages subgroup was localized in the mid-late-trajectory, indicating that this subgroup had completed functional maturation and polarization, which may represent the terminal differentiation state of tissue-resident macrophages in microenvironmental homeostasis (Fig. 2B). Pseudo-time trajectory and cell-cell communication analyses of macrophages. (A) Early-trajectory macrophages display an immature, unpolarized state. (B) Heatmap for genes with pseudo-time values, with genes displaying similar expression trends converging into different clusters. (C) Number and strength of LR interactions between cell subset pairs. (D) Scatter plot for the changes in the strength of efferent-afferent interactions in all cell types. (E) LR cytokine pathways of macrophages and other cell subsets. (F) Interaction of MIF pathway with other cell subsets. (G) Interaction of SPP1 pathway with other cell subsets. (H) Stratified plot for cell-cell communication network of MIF pathway. (I) Stratified plot for cell-cell communication network of SPP1 pathway. We systematically analyzed the interaction network of macrophage subsets and found strong communication of macrophages with other cell subsets involved in OC pathogenesis (Fig. 2C). Macrophages and mesenchymal stem cells were more capable of secretion, which was attributed to their greater interaction than other cell subsets (Fig. 2D). In terms of the mechanism, macrophages could directly activate NK cells through cell adhesion molecule-mediated L-R axes (e.g., SPP1/CD44, MIF/CD74/CD44, and MIF/CD74/CXCR4) (Fig. 2E-I). Key modules of TAMs identified by hdWGCNA The key modules of TAMs were identified by hdWGCNA. A scale-free co-expression network was created with the optimal soft threshold of 5 (Fig. 3A), and eight gene co-expression modules were found (Fig. 3B). Notably, the green, brown, and black modules differed more greatly in normal and OC samples (Fig. 3C), and the green module was positively correlated with the other two modules (Fig. 3D). Finally, the genes in the green, brown, and black modules were intersected with DEGs, yielding 170 intersecting genes (Fig. 3E). Key modules of TAMs identified by hdWGCNA and establishment and validation of prognosis model based on TAMs. (A) A scale-free network created with the optimal soft threshold of 5. (B) Eight modules of highly variable genes categorized by hdWGCNA. (C) T-SNE plots, with colors indicating module eigengenes (MEs) in the modules identified. (D) Matrix plot visually depicting the module correlation by showing the correlation of MEs. (E) Intersecting genes of genes in key modules (green, brown, and black) with DEGs. (F) Results of LASSO analysis. (G) LASSO coefficient curves of PRGs, with log(lambda) sequence as the x-axis, displaying that the coefficients contract with increasing penalties. (H) Survival curves in the LR and HR groups. (I) Risk factors. (J) ROC curves for predictive accuracy. (K) Survival curves in the LR and HR groups in the external validation set (GSE102073). (L) Risk factors in the external validation set (GSE102073). (M) ROC curves for predictive accuracy in the external validation set (GSE102073). Prognosis model based on TAMs To explore the value of prognostic signature genes in OC, 12 PRGs were selected from the 170 intersecting genes by univariate Cox regression and LASSO (Fig. 3F-G), of which seven PRGs (EPB41L2, AAK1, PRPF38B, RB1, GPR34, CLEC12A, and BRD2) were finally determined by multivariate Cox regression. Then we calculated the risk score, based on which the HR and LR groups were set up. To further validate the prognosis model’s performance, K-M survival curves were plotted. We found a great difference in the survival rate between the HR and LR groups (P < 0.0001) (Fig. 3H), a far higher survival rate in the latter, and more deaths in the former (Fig. 3I). The prognosis model had 1-, 3-, and 5-year AUCs of 0.657, 0.730, and 0.732 (Fig. 3J). The consistent results were obtained in the external validation set (P = 0.018) (Fig. 3K-L), with 1-, 3-, and 5-year AUCs of 0.695, 0.603, and 0.554 (Fig. 3M). The external validation set GSE26193 also yielded consistent results (P = 0.0063) (Supplementary Figure S7A-B), with 1-, 3-, and 5-year AUCs of 0.545, 0.606, and 0.651 (Supplementary Figure S7C). Nomogram Based on the seven PRGs, a nomogram was created for predicting 1-, 2-, and 3-year OS (Fig. 4A). Good concordance between the predicted and actual results was revealed by the calibration curves (Fig. 4B). In addition, we assessed the net clinical benefit of the model by DCA (Fig. 4C) and performed GSEA on the model (Supplementary Figure S8). Construction and assessment of nomogram in the TCGA cohort. (A) A nomogram created based on the seven PRGs to predict 1-, 2-, and 3-year OS. (B) Calibration curve. (C) Decision curve. Cellular localization and immune infiltration analysis To further investigate the immune landscape of OC, we performed a quantitative analysis of the IME. First, the expression profile of prognostic signature genes across cell subsets was mapped (Fig. 5A). AAK1, BRD2, and PRPF38B were significantly expressed across cell subsets, and BRD2 and PRPF38B had strong correlations with B cells and chondrocytes (Supplementary Figure S9A). Moreover, their expression was also analyzed between normal and OC tissues. The results revealed that the expression of BRD2, PRPF38B, and AAK1 was significantly higher in OC tissues than in normal tissues, while normal tissues had stronger expression of CLEC12A (Supplementary Figure S9B). Cellular localization and immune infiltration analysis. (A) T-SNE plots for the distribution of prognostic signature genes across cells. (B) Box plot for immune cell infiltration in the two groups. *P < 0.05, **P < 0.01, ***P < 0.001: significant differences in the proportion of immune cells. (C) Proportion of immune cells in the two groups. (D) Heatmap for the association of prognostic signature genes with different immune cells. (E) Heatmap for the correlation of immune cells. We visualized the immune cell infiltration by differential abundance (Fig. 5B-C); the HR group had far higher levels of Neutrophils (P < 0.01), Macrophages M2 (P < 0.001), Mast cells activated (P < 0.05), and Monocytes (P < 0.05), and far lower levels of Macrophages M0 (P < 0.05), T cells follicular helper (P < 0.01), and T cells regulatory (Tregs) (P < 0.01). Furthermore, the association of prognostic signature genes with immune cells was investigated. We found strong positive correlations of GPR34 with Macrophages M2, Neutrophils, and Monocytes (P < 0.001), and of CLEC12A with Macrophages M2 and Neutrophils (P < 0.001) (Fig. 5D). A negative association was present between Macrophages M2 and Tregs and T cells follicular helper, and between Monocytes and Macrophages M0 (Fig. 5E). However, AAK1 and PRPF38B showed weak associations with macrophages, which may be attributed to heterogeneity in TCGA bulk data and single-cell data across platforms and batch effects. To validate the reliability of these results, more independent datasets will be incorporated for cross-validation of AAK1 and PRPF38B in future studies. Additionally, in vitro cellular experiments will be conducted to further confirm their expression profiles in macrophages. Immunotherapy response and drug sensitivity To compare the immunotherapy response between the HR and LR groups, we performed TIDE and ESTIMATE analyses. The HR group had far higher Dysfunction Score and Microsatellite Instability (MSI) Score than the LR group (P < 0.01), suggesting possible unique genomic instability (Fig. 6A). Notably, the TIDE score displayed no statistically significant difference between the HR and LR groups, suggesting a similar degree of impaired T-cell infiltration in the two groups. However, the HR group had a significantly higher Dysfunction Score, and the CIBERSORT results revealed significant enrichment of innate inhibitory cells (e.g., M2 macrophages and neutrophils) and a reduced proportion of adaptive immune cells in the HR group. This suggests that T cells may not be physically blocked outside the tumor, but rather experience functional depletion following infiltration. Immunotherapy response and drug sensitivity in the HR and LR groups. (A) From left to right: Dysfunction Score, MSI Score, TIDE Score, and Exclusion Score. *P < 0.05, **P < 0.01: statistical significance; ns: no significance. (B) From left to right: ESTIMATEScore (P = 0.045), TumorPurity Score (P = 0.045), StromalScore (P = 0.0031), and ImmuneScore (P = 0.4). (C) Box plots for IC50 values reflecting the drug sensitivity in the HR versus LR groups. The ESTIMATE analysis showed that the HR group had significantly higher ESTIMATEScore (P = 0.045) and StromalScore (P = 0.0031) but significantly lower TumorPurity (P = 0.045). Notably, no statistically significant difference was found in the ImmuneScore between the two groups (P = 0.400). This suggests that the IME changes in the HR group were driven primarily by stromal components rather than an overall increase in immune cell infiltration (Fig. 6B). To further explore the potential therapeutic strategies for high-risk OC, the HR and LR groups were compared for their sensitivity to multiple drugs. Significant differences were present for 32 drugs, and the following drugs possessed definite therapeutic effects: LFM-A13, ispinesib mesylate, cisplatin, camptothecin, SN-38, and mitomycin-C had higher scores in the HR group, confirming that these drugs may be more effective for high-risk OC. The scores of bortezomib, BIRB-0796, AP-24,534, AMG-706, sorafenib, and midostaurin were higher in the LR group (Fig. 6C). Discussion OC is among the deadliest gynecological malignancies, and it affects over three million women annually worldwide, of which 152,000 deaths are caused, seriously threatening women’s health22. This is mainly attributed to the lack of effective means of early diagnosis. Recently, breakthroughs have been made in the single-cell sequencing technique, so a new tool is provided for interpreting the TME heterogeneity, which demonstrates unique advantages in mining diagnostic markers from complex omics data. Macrophages in the TME are implicated in OC progression and metastasis through multiple mechanisms including immunoregulation and angiogenesis, but their specific molecular signature remains to be fully elucidated. This study utilized scRNA-seq to for the first time systematically reveal the specific molecular profiles of TAMs, and establish a prognosis model based on key prognostic signature genes. From OC patients (27,984 cells expressing 8823 genes) and normal controls (27,984 cells expressing 6662 genes), we identified five differentially expressed subsets of macrophages, annotated as “OC-enriched macrophages” and “Normal-enriched macrophages”. Then the DEGs were determined from “OC-enriched macrophages” and “Normal-enriched macrophages”, and underwent functional enrichment analyses. We found that the DEGs were key players in ribosome structure composition, protein translation, and co-translational transport, emphasizing their key roles in protein synthesis and subcellular localization. The cell-cell communication analysis revealed the unique molecular signature of TAMs, especially the ability of macrophages to achieve intercellular regulation by adhering to L-R pairs. For example, a unique interaction of hepatocellular carcinoma (HCC) malignant cells with macrophages is dependent on the SPP1-CD44 axis, suggesting a role of SPP1 derived from macrophages in the HCC progression23. Meanwhile, the SPP1 expression in OC correlates with the tumor-infiltrating lymphocyte abundance, and it has a positive correlation with macrophages24. SPP1, as an ECM protein, can regulate the tumor cell adhesion to the ECM and other cells, and it, by interacting with integrins and other receptors, enhances the adhesion capability of tumor cells and contributes to tumor cell migration and invasion, thus facilitating tumor metastasis and immune escape. Moreover, SPP1 can induce M2 polarization of macrophages; the resulting M2 type with immunosuppression and pro-tumor properties can secrete various cytokines and growth factors to promote tumor cell proliferation, invasion, and angiogenesis25,26, and restrain the activity of immune cells to create favorable conditions for immune escape27. High expression of SPP1 may inhibit the activation and function, and reduce the proportion of M1 type that has strong anti-tumor activity, weakening immune defense against tumors. Such a complex intercellular interaction network offers new targets for OC treatment, and therapeutic strategies targeting TAMs are becoming a research hotspot28,29. As shown by the pseudo-time trajectory analysis, the degree of differentiation of macrophages in OC was higher at early stages, suggesting that TAMs may originate from the resident normal macrophages activated. Unlike tissue-resident macrophages that originate primarily from the yolk sac during embryonic development30, TAMs originate from circulating monocytes31. Further research is needed32 to clarify the origins of these TAMs by lineage tracing techniques and transgenic mice, contributing to the OC treatment targeting TAMs. The single-cell analysis yielded 170 intersecting genes, and a prognostic signature associated with TAMs was established by univariate and multivariate Cox regression and LASSO. We identified seven PRGs in OC (EPB41L2, AAK1, PRPF38B, RB1, GPR34, CLEC12A, and BRD2). GPR34 and CLEC12A displayed significant positive correlations with M2 macrophage infiltration. GPR34 is a metabolic immune checkpoint for type 1 innate lymphocyte-mediated antitumor immunity33, and it is noteworthy that GPR34 shows an aberrantly high expression in various malignancies (cervical cancer34, MALT lymphoma35,36, BCR/ABL+ leukemia, gastric cancer, glioma37, and colon cancer38, suggesting that it may facilitate tumor invasive growth by activating the PI3K/AKT or TGF-β/Smad pathway. Similarly, an obvious positive correlation is present between CLEC12A and various immune checkpoint proteins that are key factors for regulating the TME and influencing the immunotherapy response. CLEC12A alterations are observed in 23 types of cancers, with the highest mutation rate in OC39. At the level of epigenetic regulation, circRNA EPB41L2 up-regulates TFRC by adsorbing miR-1270, activating the epithelial-mesenchymal transition (EMT) pathway and facilitating tumor invasion and metastasis40. In addition, highly expressed EPB41L2 in cancer-associated fibroblasts is closely linked to the exhaustion of PD-1 + Tim-3 + T cells, which enhances immunosuppression by modulating the TME remodeling and reduces the immunotherapy response31. These findings demonstrate that EPB41L2 may serve as a multifunctional hub molecule driving tumor progression by the circRNA-miRNA axis, metabolic reprogramming, and cell-cell communication. The regulatory network of multiple potential therapeutic targets has been revealed recently: AAK1 regulates endocytosis by phosphorylating EGFR, affecting tumor cell proliferation; it also mediates cytoskeletal remodeling and EMT to facilitate tumor invasion and metastasis41. In mouse models, AAK1 inhibitors can enhance anti-PD-1 efficacy (a 58% reduction in tumor volume), suggesting their potential as an immunotherapy sensitizer42. As one of the key splicing factors in OC, PRPF38B may regulate base excision repair and alternative splicing of pyrimidine metabolism-related genes43 RB1 deficiency worsens DNA damage, while platinum drugs achieve an antitumor effect by causing DNA damage. Meanwhile, RB1 plays a role in regulating immune function, and its deficiency may affect the IME in OC, which offers new ideas for immunotherapy44. BRD2 is a chromatin remodeling regulator, and it influences the cell cycle and proliferation by regulating gene expression and chromatin remodeling in breast cancer, leukemia, and prostate cancer, thus promoting tumor development. Besides, BRD2 strengthens drug resistance by RasGRP1/Ras/ERK in adult T-cell lymphoblastic lymphoma, and BRD2-targeted strategy may ameliorate the therapeutic efficacy and extend the survival period45 All the above molecules form a multidimensional interaction network by regulating immune checkpoints, metabolic reprogramming, and epigenetic modifications, jointly shaping the pro-tumor microenvironment. To sum up, immunometabolic interventions against GPR34 and CLEC12A, EPB41L2-associated circRNA-targeted therapies, and a combination of PRPF38B, RB1, AAK1, and BRD2 inhibitors with immune checkpoint blockers may be new strategies breaking through the bottleneck of current diagnosis and treatment. Notably, although the TIDE scores had no significant difference between the HR and LR groups46, both the Dysfunction Score and MSI Score were significantly elevated in the HR group, and the ESTIMATE analysis showed that the HR group had significantly higher StromalScore but significantly lower TumorPurity. These results were highly consistent with CIBERSORT results that the HR group had significantly enriched innate inhibitory cells (e.g., M2 macrophages and neutrophils) and a reduced proportion of adaptive immune cells (e.g., Tfh and Tregs)47,48. This suggests that immune escape in the HR group did not result from T cells being physically blocked outside the tumor (as indicated by no difference in the TIDE scores) but rather from T cell functional depletion mediated by local myeloid cells through metabolic competition and immunosuppressive factors (as indicated by elevated Dysfunction Scores)49. Furthermore, the elevated MSI Scores in the HR group suggested potential immunogenicity, which coexisted with a myeloid inhibitory barrier composed of M2 macrophages, creating a paradoxical state of “high immunogenicity-high immunosuppression”50–52. The prognostic signature genes GPR34 and CLEC12A both displayed significant positive correlations with M2 macrophages and neutrophils, further indicating that this prognosis model captured the transcriptomic imprinting of myeloid polarization and innate immunosuppression, which was independent of the physical barrier mechanism underlying the TIDE score53. It is worth noting that the significant elevation of the ESTIMATEScore was mainly driven by the StromalScore (P = 0.0031), whereas the ImmuneScore showed no statistically significant difference between the two groups (P = 0.400). This result was highly consistent with the CIBERSORT deconvolution analysis: Although innate inhibitory cells (e.g., M2 macrophages and neutrophils) were enriched in the HR group, the proportions of Tfh, Tregs, and M0 macrophages decreased. Due to this selective subset imbalance, the overall ImmuneScore did not reach statistical significance. Therefore, the IME signature of the HR group should be interpreted as “selective innate immunosuppression in the context of stromal remodeling”, rather than an overall increase in immune infiltration. The global ESTIMATE score and the subset resolution of CIBERSORT complemented each other and jointly depicted the complex microenvironmental heterogeneity of the HR group. The following limitations are worth noting. First, cell annotation methods all had deficiencies. When the SingleR package was utilized for cell annotation, NK and T cells had overlapping features at the transcriptomic level (such as marker genes CD3D, NKG7, and GNLY). Macrophages and monocytes were similar and belonged to the myeloid cells, and they were annotated by the SingleR package into functional cell clusters. More tissue samples will be harvested for experimental analyses in the future. Second, the prognosis model created in this study demonstrated limited performance in external validation mainly due to three factors: (1) Variations in sample processing, sequencing platforms, and follow-up protocols across multicenter validation cohorts may introduce technical biases and thus affect the model’s generalizability. (2) This model was developed based solely on gene expression profiles, without integrating clinicopathological parameters (e.g., FIGO stage, histological grade, and BRCA mutation status), resulting in insufficient interpretability. (3) Due to the high heterogeneity of OC, it is difficult for a single signature gene to capture the differences in the IME across histological subtypes. Therefore, although the model’s robustness was enhanced by cross-validation across multiple datasets, its clinical translational value remains limited as can be seen from the medium AUC. In the future, further optimization is required by large-scale prospective cohorts and integrated clinical-genomic models. Third, M1 and M2 polarization states of macrophages were not systematically distinguished in the IME analysis. M1 and M2 types were defined primarily based on differences in cell proportions between normal and OC tissues. As a result, macrophage functional polarization was insufficiently characterized. Moreover, due to the absence of classic markers, CD4⁺ and CD8⁺ T cell subsets could not be reliably distinguished, restricting our ability to analyze the interactions between macrophages and specific T cell subsets. Furthermore, although AAK1 and PRPF38B were identified as macrophage-enriched genes in the single-cell analysis, their association with macrophages was limited as suggested by TCGA bulk data. This discrepancy may be attributed to heterogeneity in TCGA bulk data and single-cell data across platforms and batch effects. Therefore, more external datasets and in vitro cellular experiments are required for further validation. Finally, the number of macrophages was imbalanced between normal and OC tissues due to the limited sample size, compromising the statistical power of intergroup comparisons. In the future, we will actively expand independent retrospective cohorts, collect prospective data, combine deeper sequencing or TCR sequencing to elucidate the mechanism of interaction between macrophages and T-cell subsets, and incorporate more self-test samples to validate and optimize the conclusions of this study. Conclusion In conclusion, we created a prognosis model based on prognostic signature genes using scRNA-seq data, and confirmed the unique value of scRNA-seq in the analysis of the TME. Systematically identifying specific markers for macrophage subsets not only offers a novel biomarker combination for early diagnosis of OC but also reveals the underlying molecular mechanism of the regulation of TAM plasticity. These findings lay a theoretical basis for formulating TME-based precise diagnostic and therapeutic strategies, and validate the importance of multidisciplinary methods in translational medicine research. This study paradigm can be extended to the microenvironment of other malignancies, presenting new ideas for the discovery of cancer diagnosis and therapeutic targets. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Supplementary Material 8 Supplementary Material 9 Author contributions YR G: Conceptualization, Data curation, Methodology, Project administration, Writing – original draft, Writing – review & editing. JJ T: Data curation, Formal analysis, Investigation. LG Z: Writing – review & editing. DY: Writing – review & editing.YK: Conceptualization, Supervision, Writing – review & editing. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Funding This research was supported by National Natural Science Foundation (82260566). Data availability The datasets analysed during the current study are available in the GEO (www.ncbi.nlm.nih.gov/geo/), and downloaded GSM5599220, GSM5599227, GSE26193 and GSE102073 (November 8, 2024). Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Supplementary Material 6 Supplementary Material 7 Supplementary Material 8 Supplementary Material 9 Data Availability Statement The datasets analysed during the current study are available in the GEO (www.ncbi.nlm.nih.gov/geo/), and downloaded GSM5599220, GSM5599227, GSE26193 and GSE102073 (November 8, 2024).