Transcriptomic and clinical data were integrated from three independent cohorts of clear cell renal cell carcinoma (ccRCC): TCGA-KIRC, EMTAB3267, and GSE22541. All the clinical annotations were carried out according to the platforms.

All datasets were filtered to include only primary tumor samples with complete survival data, tumor grades, and TNM staging. Raw data of RNA-seq were converted to transcripts per million (TPM) and transformed to log2(TPM+1). Microarray data were normalized using quantile normalization, and batch effects were removed using ComBat ( Supplementary Figure S1A ).

To identify distinct metabolic subtypes, unsupervised consensus clustering was performed on the TCGA-KIRC cohort using the ConsensusClusterPlus package (R v4.2.3) (27). Subsequently, principal component analysis (PCA) was applied to reduce the dimensionality of the original expression matrix, and the samples were projected into a two-dimensional space to intuitively display the separation of each metabolic subtype (Data sheet 1). K-means clustering with a Euclidean distance metric was employed, and the robustness of clustering was assessed across 1,000 iterations, each involving random subsampling of 80% of the samples. The optimal number of clusters was determined by integrating results from cumulative distribution function (CDF) plots, delta area analyses, and silhouette width metrics, thereby ensuring stable and well-defined subtype classification.

To better stratify patients into biologically and clinically relevant subtypes, the nearest template prediction (NTP) approach in the “MOVICS” package was implemented (28) based on predefined molecular templates. Subtype-specific templates were constructed based on the average gene expression profiles of typical samples. Gene symbols were harmonized according to HUGO Gene Nomenclature Committee (HGNC) standards, and only the common genes in all datasets were retained for subsequent analyses.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were conducted using the clusterProfiler R package (version 4.6.2). Enrichment analysis was applied separately for biological processes (BPs), cellular components (CCs), and molecular functions (MFs) in GO terms and for metabolic and signaling pathways in KEGG. Pathways with an adjusted p-value <0.05 were considered significantly enriched.

To estimate the absolute abundance of distinct stromal and immune cell populations from bulk transcriptomic profiles, we employed Microenvironment Cell Populations-counter (MCP-counter) by the “MCPcounter” R package (version 1.2.0, Bioconductor version 3.20) (29). This approach quantified specific cell populations based on cell type-specific transcriptomic markers. The immune cell populations assessed included the following: CD8+ T cells, total T cells, natural killer (NK) cells, cytotoxic lymphocytes, myeloid dendritic cells, monocytes, and neutrophils. Additionally, we estimated the abundance of stromal components, namely, endothelial cells and fibroblasts, for each sample. Furthermore, we performed six complementary algorithms—MCP-counter, CIBERSORTx, EPIC, quanTIseq, TIMER2.0, and xCell—based on the IOBR package (30) to decipher the immune infiltration pattern among high and low PTGES2 tumors.

To further delineate heterogeneity in immune infiltration between samples, single-sample gene set enrichment analysis (ssGSEA) was performed using the “GSVA” R package (Bioconductor version 3.20) (31). An enrichment score for each gene set was computed, enabling a quantitative assessment of immune-related gene set activity at single-sample resolution by ranking the genes and comparing the distributions.

Gene set variation analysis (GSVA) was applied to RNA-seq expression profiles to estimate pathway activities and immune cell infiltration in an unsupervised, non-parametric manner. GSVA and enrichment scores were performed using the “GSVA” R package. Hierarchical clustering of GSVA scores enabled the visualization and identification of subtype-specific patterns in the tumor immune microenvironment (TIME), which are displayed graphically in the heatmaps. To evaluate the features of Tertiary lymphoid structure (TLS), Germinal centers B cell (GC B), Follicular helper T cell (Tfh), Follicular dendritic cells (FDC), Plasma, and B cells, GSVA was performed based on related gene sets, which were obtained from a previous study (32). In addition, immune-related signaling between high and low PTGES2 tumors enrolled from the Hallmark database was compared based on GSVA algorithms.

mRCC patients from the CheckMate immunotherapy cohort (16) who received immunotherapies were analyzed. Clinical response was categorized as follows: clinical benefit (CB), defined as complete response (CR), partial response (PR), or stable disease (SD) lasting ≥6 months; and no clinical benefit (NCB), defined as progressive disease (PD) or SD lasting <6 months. Objective response rates (ORRs), including CR, PR, SD, PD, and confirmed mixed partial response (CMPR), were calculated for every subtype.

To investigate the differential drug sensitivity of RCC, the OncoPredict algorithms implemented in MOVICS were utilized alongside drug response data from the Genomics of Drug Sensitivity in Cancer (GDSC) database (Release 8.5, October 2023). The sensitivity of a range of therapeutic drugs was analyzed, and the half-maximal inhibitory concentration (IC₅₀) values for each drug were estimated within the context of three distinct subtypes.

Paraffin-embedded RCC tissues were deparaffinized in xylene and rehydrated through graded ethanol. Antigen retrieval was performed by boiling the sections in Tris-EDTA buffer (pH 9.0) for 15 minutes in an electric cooker, and then the sections were naturally cooled. Endogenous peroxidase activity was quenched using 3% hydrogen peroxide for 10 minutes at room temperature. Next, tissue sections were blocked with 3% bovine serum albumin (BSA) for 30 minutes and then incubated with primary antibodies (anti-PTGES2; dilution 1:200, Proteintech, Wuhan, Hubei, P.R.C, Cat No. 10881-1-AP) overnight at 4°C. The next day, sections were washed three times and incubated with Horseradish Peroxidase (HRP)-conjugated secondary antibodies (Absin Bioscience, China, No. abs996, general concentration) for 30 minutes at room temperature. Detection was performed using Diaminobenzidine (DAB) substrate, and counterstaining was performed using hematoxylin. Images were acquired using a Leica microscope and imaging system.

Human RCC cell lines 786-O and 769-P were obtained from the American Type Culture Collection (ATCC, Mansas, Virginia, USA). To confirm the identity of the cell lines, short tandem repeat (STR) analysis was used to identify the two cell lines. STR testing was conducted by Shanghai Zhong Qiao Xin Zhou Biotechnology Co., Ltd. ( Supplementary Materials 2 , 3 ). Cells were cultured in RPMI-1640 medium (Gibco, Thermo Fisher Scientific, Shanghai, China, A4192301) supplemented with 10% fetal bovine serum (FBS; Gibco, A5256701) and 1% penicillin–streptomycin (100 U/mL and 100 μg/mL, respectively). All cells were maintained in a humidified incubator at 37 °C with a 5% CO₂ atmosphere and were routinely tested to be free of mycoplasma contamination. Cell passage occurred when cells grew to a density of 70%–80%. Notably, cells within 20 passages were fit for all experiments to ensure phenotypic stability.

To generate stable knockdown RCC cell lines, two short hairpin RNA (shRNA) sequences targeting human PTGES2 were designed, and the sequences were as follows: shPTGES2-1: 5′-GAAGCCGAATCTCGCTGATTT-3′, shPTGES2-2: 5′-CGGCAATAAGTACTGGCTCAT-3′. A scrambled non-targeting RNA (5′-CAACAAGATGAAGAGCACCAA-3′) was used as a negative control (shCtrl). All oligonucleotides were cloned into the pLKO.1-puro vector to generate two knockdown plasmids, and PTGES2 [Coding sequence (CDS) region was obtained from https://www.ncbi.nlm.nih.gov/nuccore/NM_025072.7] was cloned into a pCDH vector to generate an overexpression plasmid (33).

Lentiviral particles were produced by co-transfecting HEK293T cells with the plasmids and packaging plasmids (psPAX2 and pMD2.G) using Lipofectamine 3000 (Thermo Fisher Scientific). Viral supernatants were collected at 48 and 72 hours post-transfection.

For transduction, cells were seeded in six-well plates (2 × 10⁵ cells/well) and infected with lentivirus at a suitable multiplicity of infection (MOI; 30–50) value in the presence of 8 μg/mL polybrene (Sigma-Aldrich, Beijing, China). After 24 hours, the medium was replaced with fresh complete medium containing 2 μg/mL puromycin for 7 days to select stably transduced cells.

Total mRNA of cells was extracted using TRIzol Reagent (Invitrogen, Shanghai, China), and then mRNA was reverse-transcribed into cDNA using PrimeScript RT Master Mix (Takara Bio, Beijing, China) and quantified by qRT-PCR on a QuantStudio 7 Flex System. The following primers were used:

The run method is shown below:

The relative mRNA expression is calculated using the following method:

Cells were lysed in Radio Immunoprecipitation Assay Lysis buffer (RIPA) supplemented with protease and phosphatase inhibitors (Beyotime, Shanghai, China) on ice for 30 minutes. Lysates were then centrifuged at 13,000 g for 15 minutes at 4°C, and protein concentrations were quantified using the BCA assay kit (Thermo Fisher Scientific). Equal amounts of protein (20–30 μg) were separated using the Sodium dodecyl sulfate - polyacrylamide gel electrophoresis (SDS-PAGE) gel (10%) and were transferred to Polyvinylidene Fluoride (PVDF) membranes (Millipore, Shanghai, China; 0.45 μm). Membranes were blocked in 5% non-fat milk for 1 hour at room temperature and incubated overnight at 4°C with primary antibodies (anti-PTGES2, dilution 1:1,000, Proteintech, Cat No. 10881-1-AP; loading control, beta actin, dilution 1:5,000, Proteintech, Cat No. 20536-1-AP).

The next day, membranes were washed three times and then incubated with HRP-conjugated secondary antibodies (HRP-conjugated Goat Anti-Rabbit IgG(H+L), dilution 1:5000, Proteintech, Cat No. SA00001-2) for 1 hour at room temperature. Signals were detected using enhanced chemiluminescence (ECL) reagents and visualized using an imaging system (Tanon chemiluminescence image analysis system).

Cell viability was assessed using the Cell Counting Kit-8 reagent (Dojindo, Kumamoto, Japan) according to the manufacturer’s instructions. 786-O and 769-P cells were seeded into 96-well plates at a density of 3 × 10³ cells/well. After the cell adhesion was completed, 10 μL of CCK-8 reagent was added to each well and incubated for 2 hours at 37°C. Absorbance at OD 450 nm was measured. It was noteworthy that the cells should be measured at a fixed time every day and five times in total.

The Transwell assay was used to test the cell migration capabilities; 2 × 10⁴ cells suspended in serum-free medium were seeded into the upper Transwell chamber (8-μm pore size; Corning, Shanghai, China) in 24-well plates. The lower chamber contained 800 μL of complete medium with 10% FBS. After incubating for 24 hours at 37°C, cells on the upper membrane surface were gently removed using a cotton swab. The migrated cells on the bottom surface were fixed with 4% paraformaldehyde and then stained with 0.1% crystal violet for 30 minutes. Cell counting was carried out using an inverted microscope (Olympus, Beijing, China) in three random fields per well.

To assess cell proliferative capacity, 786-O and 769-P cells were seeded into six-well plates at a low density of 1,000 cells/well. After culturing for 10–14 days, colonies were fixed with 4% paraformaldehyde for 30 minutes and stained with 0.1% crystal violet for 30 minutes. Visible colonies were counted.

The Kaplan–Meier survival curves of defined metabolic subtypes were generated using the survival and survminer R packages. Differences in overall survival (OS) among subtypes were evaluated using the log-rank test. Multivariate Cox proportional hazards models were adjusted for age, gender, grade, TNM stage, and metastasis. The HRs and 95% CIs were calculated, and the proportional hazards assumption was strictly tested using Schoenfeld residuals. Correlations between clinical and molecular features were tested using χ² or Fisher’s exact test (grade, stage, and metastasis) for categorical variables, and the Kruskal–Wallis or analysis of variance (ANOVA) (age and tumor size) was utilized for continuous variables.

The R software (v4.2.3) and GraphPad Prism (v9.0) were used for all statistical analyses. p-Values <0.05 were considered statistically significant.

By integrating intersecting metadata of RCC samples from TCGA-KIRC, EMTAB3267, and GSE22541, three metabolic subtypes (C1, C2, and C3) were divided by unsupervised clustering ( Supplementary Figure S1B ). KEGG enrichment analysis indicated several significant metabolic pathways, as shown in Figure 1A . Genes in cluster C1 were mainly enriched in fatty acid degradation and metabolism of various amino acids (arginine, glycine, serine, etc.). Cluster C2 comprised galactose metabolism, cardiolipin metabolism, prostaglandin biosynthesis, and other metabolite biosynthesis processes. Cluster C3 included prostanoid biosynthesis and cyclooxygenase, arachidonic acid metabolism. These findings implicated metabolic rewiring as a key differentiator of ccRCC clinical behavior.

The Kaplan–Meier analysis showed marked survival differences among the three subtypes. Generally, cluster C1 had the better survival probability, C3 was in the middle, and C2 presented the worst prognosis (overall p=0.005, C1 vs. C2 p=0.004, C2 vs. C3 p=0.022, and C1 vs. C3 p=0.070; Figure 1B ). The detailed clinical characteristics of patients were assessed under established clusters C1–C3. The results showed that pathologic_T, pathologic_N, and pathologic_M were strongly associated with every subtype, and these three almost entirely emerged in cluster C2, but presented a low proportion in clusters C1 and C3. Meanwhile, Progression Free Interval (PFI) and OS showed a similar tendency ( Figure 1C ). Therefore, the novel three clusters based on metabolism were distinctly classified for RCC and had strongly prognostic effects on patients. Overall, these results confirmed the reproducibility, prognostic utility, and biological validity of the metabolic subtype model.

For the external datasets, NTPs were carried out to validate the classification approach and the value of prognosis. Regarding EMTAB3267, NTP analysis confirmed the reliability of metabolism C1 to C3 clusters. Overall, subtypes had high prediction confidence ( Figure 2A ). Furthermore, cluster C2 contained the worst survival probability, cluster C1 had the best, and cluster C3 was in between ( Figure 2B ). For GSE22541, almost the same consequences were demonstrated as before ( Figures 2C, D ). To sum up, these three clusters possess precise prediction; even though there was no statistical significance between clusters C2 and C3, the entire tendency was consistent and accurate.

For the C1 subtype, GO analysis revealed that BP included carboxylic acid transport, organic acid transport, and organic anion transport; CC included apical part of cell, apical plasma membrane, and brush border; MF included secondary active transmembrane transporter activity, solute:sodium symporter activity, and symporter activity ( Figure 3A ). KEGG pathway analysis found that the Peroxisome proliferators-activated receptor (PPAR) signaling pathway and mineral absorption were enriched significantly ( Figure 3B ). For the GO analysis of C2, acute inflammatory response, lipoprotein particle, and protein–lipid complex were observed ( Figure 3C ). KEGG revealed significantly enriched platelet activation, cholesterol metabolism, and arachidonic acid metabolism ( Figure 3D ). C3 mainly contained external encapsulating structure organization, immunoglobulin complex, antigen binding, etc., on GO analysis ( Figure 3E ); viral protein interaction with cytokine and cytokine receptor, and cytokine–cytokine receptor interaction were the most crucial pathways enriched by KEGG ( Figure 3F ).

Additionally, the evaluation of some classical signaling pathways suggested that cluster C2 represented a distinct difference compared to the other two subtypes. In particular, HIPPO, NRF2, PI3K, TGF-β, and RTK RAS were the typical pathways ( Figure 3G ). Among them, cluster C2 had low enrichment scores, which possibly means that genes in cluster C2 lacked response to pathway-related treatments, so the worst prognosis occurred.

All three clusters indicated the outcomes of corresponding patients, and prostaglandin biosynthesis of cluster C2 and prostanoid biosynthesis of cluster C3 were of particular attention. The prostanoids are several lipid metabolites generated from 20-carbon fatty acids, and they play a key role in the inflammatory response in vivo (34). Recent studies have shown that prostaglandins and other members could regulate the activities of T cells, B cells, and cytokines in the tumor immune microenvironment, which causes immune exhaustion, and the tumor cell apoptosis was inhibited (35). RCC patients frequently encounter resistance to immunotherapy; therefore, we emphasized the relationship between clusters and the immune microenvironment of RCC.

GSVA delineated clear immunophenotypic segregation among the three defined subtypes. Notably, C2 was characterized by a marked enrichment of immunosuppressive cellular populations, including interleukins, cytokines, B-cell functions, T-cell functions, NK cell functions, antigen processing, and macrophage functions; C3 displayed a relatively balanced TIME ( Figure 4A ). MCP-counter validated quantitative enrichment of suppressive cell populations in C2, which contained a higher portion of endothelial cells, neutrophils, NK cells, T cells, CD8 T cells, and so on ( Figure 4A ). In contrast, C1 was characterized by nearly absent immune infiltrates. Concerning samples in clusters, ssGSEA was conducted to explore the relative abundance of T cells. Consistent with GSVA findings, ssGSEA-derived enrichment scores reinforced the TIME dichotomy among the three subtypes: C2 and C3 involved more quantities in Th17, Tcm, Tem, Th1, Th2, and Treg cells ( Figure 4A ). A few immune-suppression cells, like Th1, Th2, and Treg cells, were highly enriched in clusters C2 and C3, while immune-activation cells, such as neutrophils and Th17 cells, were poorly enriched ( Figure 4B ), suggesting the poor outcomes of C2 and C3, which were consistent with the immune status.

Immune checkpoints are tightly linked to the clinical efficacy of RCC patients, and RCC is regarded as a kind of “hot” immunological tumor, so we explored the association between three subtypes and the expression of eight immune checkpoints. The results showed that cluster C2 exhibited a higher expression of CD274 than clusters C1 and C3; however, PDCD1LG2, TNFRSF9, and TNFRSF4 were at extremely low levels compared to the others ( Figure 4C ). We analyzed the downregulation of co-stimulatory molecules such as TNFRSF and TNFRSF4 in the C2 subtype, which may weaken T-cell activation and immune effects in the tumor microenvironment, thereby partially explaining why the immune efficacy remains poor despite high PD-L1 expression.

We also retrospectively reviewed all patients’ outcomes after accepting immunotherapy in the CheckMate immunotherapy cohort. RCC patients in cluster C1 achieved significantly prolonged survival compared to patients in C2 and C3 (p=0.033); the trend favored C1, suggesting durable immune engagement ( Figure 4D ). In particular, 47% patients in cluster C2 showed NCB, far more than those in C1 (30%) and C3 (35%). The frequency of CB in C2 was 24%, which was lower than that in C1 (29%) and C3 (31%). Meanwhile, the ORRs were similar to the previous benefits. Of patients in cluster C2, 48% were PD, but only 32% in C1 and 38% in C3 ( Figure 4E ). Hence, this classification is meaningful to patients’ effects of immunotherapy.

To strengthen the relation between subtypes and TLS, we calculated and explored TLS score, GC B, Tfh, FDC, Plasma, and B cell_score features among RCC patients, and the results showed that C3 presented higher expression of these TLS-related features, including CR2, FCER2, PAX5, CD19, MZB1, JCHAIN, DERL3, CCL19, CCL21, CXCL13, and BLC6. Furthermore, the C3 cluster also presented high enrichment scores of TLS score, GC B, Tfh, FDC, Plasma, and B cell_score features, indicating a high TLS formation tendency ( Supplementary Figure S2 ).

Dozens of IC₅₀ estimates were derived for each of the drugs from independent replicates. The distribution of IC₅₀ values was visualized using violin–box hybrid plots to assess both the central tendency and variability across cell subtypes for each drug. Notable variability in drug sensitivity was observed across the three subtypes, with specific subtypes demonstrating distinct patterns of drug resistance or susceptibility.

Many kinds of drugs showed antitumor properties; herein, certain representative drugs were chosen. For the most frequently targeted medicine, TKI was observed to have a significant sensitivity gradient among the three subtypes. For the receptor TKI sunitinib, cluster C2 exhibited the highest estimated IC₅₀ value, indicating resistance. One-way ANOVA confirmed that the observed differences in sunitinib sensitivity were statistically significant (p =1.2e−12) ( Figure 5A ). Saracatinib was an effective Src inhibitor, NSC-87877 was a Shp inhibitor, and both were non-receptor TKIs. Cluster C2 appeared to have the highest IC₅₀ value as well (saracatinib p=0.029, NSC-87877 p=0.00082) ( Figures 5B, C ).

Other kinds of kinase inhibitors were committed to lethal effects on cancer cells. For the serine/threonine kinase inhibitors, such as VX-680, which were a part of evolutionary conserved families (Aurora-A, Aurora-B, and Aurora-C), controlling events of cell cycles, cluster C2 possessed a higher IC₅₀ value than others (p=4e−14) ( Figure 5D ). Furthermore, multi-kinase inhibitors like sorafenib (p=5.9e−07) and AP24534 (p=1.7e−05) were also resistant to cluster C2 ( Figures 5E, F ).

Except for kinase targets, inhibitors specific to diverse pathways were considered beneficial in the extinction of cancer cells. Therefore, some drugs were estimated in three subtypes. Rapamycin and phenformin blocked the mTOR signaling pathway, inducing cancer cell autophagy and apoptosis, respectively. In cluster C2, both showed high IC₅₀ values (rapamycin p=0.00092, phenformin p =1.1e−06) ( Figures 5G, H ). As a Nuclear Factor-κB (NF-κB) inhibitor, parthenolide was resistant in cluster C2 (p=0.00019) ( Figure 5I ), as well as the FH535, referred to as a Wnt/β-catenin pathway inhibitor (p=4.7e−09) ( Figure 5J ).

Otherwise, chemotherapeutics for cancers that target DNA replication are always available. Etoposide and doxorubicin were the common interventions inhibiting topoisomerase-II, thereby terminating DNA replication in cancer cells. As expected, drug resistance was the same in all the above drugs (etoposide p =3.2e−05, doxorubicin p=0.00028) ( Figures 5K, L ).

The overall drug response patterns corroborate our hypothesis that the molecular and phenotypic profiles of C1, C2, and C3 contribute to distinct therapeutic vulnerabilities. In particular, the cluster C2 phenotype appeared to be associated with enhanced drug resistance, which was in line with the poor prognosis of cluster C2 as mentioned before.

For further validation of cohorts, patients were stratified into high and low prostanoid/prostaglandin expression groups based on the median expression of the metabolism gene signature (PGE2). In the TCGA-KIRC cohort, patients with high prostanoid/prostaglandin expression exhibited significantly worse survival (prostanoid p=0.013, HR=1.46, 95% CI: 1.09–1.97; prostaglandin p=0.012, HR=1.47, 95% CI: 1.09–1.97). Similarly, in the FUSCC cohort, patients with high prostanoid/prostaglandin expression had worse survival (prostanoid p=0.061, HR=1.55, 95% CI: 0.96–2.5; prostaglandin p=0.05, HR=1.58, 95% CI: 0.98–2.55) ( Figures 6A–D ).

Patients in the high prostanoid group with PD were more than those in the low prostanoid group (26% vs. 23%), the same as in the high/low prostaglandin group (28% vs. 21%). In terms of progression, the high prostanoid and low prostanoid groups showed no difference (87% vs. 87%), but patients in the high prostaglandin group were more likely to progress than others (91% vs. 83%) ( Figures 6E, F ). These results suggested that high prostanoid/prostaglandin RCC harbors a more suppressive and therapy-resistant microenvironment, with lower immunogenicity and metabolic reprogramming.

To evaluate the clinical relevance of prostaglandin biosynthesis in the context of therapeutic response, immunohistochemistry (IHC) analysis was performed on RCC tissue microarrays (TMAs) obtained from patients with differential responses to targeted therapy. A total of 12 tumor specimens were analyzed and stratified into four groups according to clinical response criteria: CR (n = 3), PR (n = 3), SD (n = 3), and PD (n = 3). For analytical purposes, CR and PR were classified as the non-progressive disease (non-PD) group, while SD and PD were considered PD based on radiological and clinical progression metrics.

Here, we found prostaglandin E synthase 2 (PTGES2), one key enzyme catalyzing the synthesis of prostaglandin E2, which was associated with the progression of some tumors, such as colorectal cancer (CRC), hepatocellular carcinoma (HCC), and RCC (36–38). IHC staining revealed a distinct expression pattern of PTGES2 across the four groups. In the PD subgroup (PD and SD), PTGES2 exhibited markedly higher expression, characterized by strong staining intensity in RCC cells ( Figure 7A ). In contrast, the non-PD subgroup (CR and PR) demonstrated weak-to-moderate PTGES2 staining, with a substantial reduction in both staining intensity and percentage of positive cells. These findings suggested that PTGES2 expression correlated positively with therapy resistance and tumor progression status in RCC, which was in accordance with cluster C2. In addition, PTGES2 was inversely associated with TLS programs: TLS (r = −0.17, p =6e−5), GC B (r = −0.31, p = 3.6e−13), and Plasma (r = −0.26, p = 3.4e−9); FDC showed no association (r = −0.054, p=0.22); Tfh showed a weak positive correlation (r=0.15, p =6e−4) ( Supplementary Figure S3A ). Consistently, high PTGES2 tumors had lower TLS features (all p ≤ 0.05), with no change in FDC (p=0.59) and only a marginal increase in Tfh (p=0.075) ( Supplementary Figure S3B ), indicating that high PTGES2 is linked to reduced TLS enrichment. Across six deconvolution methods, high PTGES2 tumors consistently exhibited increased Tregs and M2-like macrophages with relative depletion of CD8+ T cells and NK cells ( Supplementary Figure S4A ). Interferon-α/γ responses, cytokine–cytokine receptor interaction, and immune system signaling were enriched in high PTGES2 ( Supplementary Figure S4B ), consistent with an inflamed yet functionally suppressive microenvironment. We observed modest negative associations with PD-L1 and CTLA-4, a weak positive association with TIGIT, and no material association with PD-1 or LAG-3 ( Supplementary Figures S4C, D ).

To further investigate the role of PTGES2 in RCC cell proliferation and metastasis, we performed a series of in vitro assays following the stable knockdown and overexpression of PTGES2 in the 786-O and 769-P cell lines. The efficiencies of the knockdown and overexpression of PTGES2 were verified at both the mRNA and protein levels by quantitative polymerase chain reaction (qPCR) and Western blotting (WB) analysis, respectively. The results showed that PTGES2 was successfully knocked down and overexpressed in both 786-O and 769-P cells ( Figures 7B–D ).

The CCK-8 assay was used to explore cell proliferation. Knockdown of PTGES2 resulted in a significant reduction in cell reproductive capacity in both 786-O and 769-P cell lines; in contrast, PTGES2 overexpression contributed to the obvious proliferation of RCC cells, and all differences were statistically significant (p < 0.05) ( Figure 7E ). The colony formation capability of PTGES2 knockdown cells was significantly decreased compared to that of shCtrl cells ( Figure 7F ), the colony counts in the shPTGES2 groups were decreased under the microscope, and the difference was statistically significant (p < 0.05). The results were reversed while PTGES2 was overexpressed in RCC cell lines ( Figures 7F, G ). These data indicate that PTGES2 can promote the proliferation of RCC cells.

To evaluate the metastasis of PTGES2 in RCC cells, a Transwell assay was conducted to determine the cell migration ability. The results suggested that shPTGES2 cells exhibited significantly reduced migration compared to the control cells ( Figure 7H ). The number of migrated cells was decreased visibly in shPTGES2-treated cells after cell counting, and the difference was statistically significant (p < 0.05) ( Figure 7I ). Certainly, PTGES2 overexpression was observed to motivate the migration of RCC cells, and the difference between the vector and overexpression groups was statistically significant as well (p < 0.05) ( Figures 7H, I ). These results revealed that PTGES2 drove RCC cell migration and may be positive for the metastasis of RCC.

In general, these in vitro experiments strengthened the association between PTGES2 expression and clinical response. Our observations implicated the translational value of PTGES2, which could serve as a potential biomarker and a therapeutic target for RCC.