2.1. Data source

ScRNA-seq data were obtained from the Gene Expression Omnibus (GEO) database under accession number GSE135045, comprising 7 GBM samples. Glioma bulk RNA-seq data and corresponding clinical features were retrieved from The Cancer Genome Atlas (TCGA), the Chinese Glioma Genome Atlas (CGGA) and GEO database. Formalin-fixed paraffin-embedded samples from TCGA were excluded due to their inherent limitations in sequencing analyses [18,19]. Normal tissue bulk RNA-seq data were retrieved from Genotype-Tissue Expression (GTEx) and Human Protein Atlas (HPA). The TCGA and GTEx combined cohort was downloaded from the UCSC Xena. The summary of genomic alteration frequency was downloaded from cBioPortal. The RNA expression microarray datasets GSE74187, GSE107850, GSE147352 and GSE16011 were retrieved from the GEO database for expression validation.

2.2. Single-cell transcriptome sequencing data analysis

The R package “Seurat” (version 4.3.0) [20] was utilized to process the scRNA-seq. This involved conducting data quality control, normalization, scaling, identifying hypervariable genes, performing principal component analysis (PCA), and applying uniform manifold approximation and projection (UMAP). Initially, the CreateSeuratObject function was employed to construct a Seurat object, and low-quality cells were filtered out based on the following criteria: (1) cells with fewer than 500 unique molecular identifiers (UMI) or more than 8000 UMI; (2) cells with more than 20% mitochondrial UMI counts; and (3) cells with more than 50% ribosomal UMI. Subsequently, the ‘LogNormalize’ method of the NormalizeData function was applied to normalize the data with a scaling factor of 10,000. The FindVariableFeatures function was then utilized to identify 2000 highly variable genes for further analysis. The RunPCA function was employed for data dimensionality reduction, selecting the first 30 principal components for downstream analysis. Ultimately, the UMAP algorithm was used to visualize the dimensionality reduction of cells.

Batch effects were adjusted using the “harmony” package (version 0.1.1) [21]. Cell types were initially annotated by manual curation using canonical marker genes (see S1 Table). To validate the manual annotations, we applied automated reference-based annotation using SingleR (version 2.4.1) with default parameters and the “HumanPrimaryCellAtlasData” reference. In addition, we compared our cell type proportions with published GBM single-cell atlases [22] to ensure consistency. The CellCycleScoring function in Seurat was employed to calculate the cell cycle score of each cell. Proliferation and stemness module scores were computed using the AddModuleScore function [23]. Copy number variations (CNVs) were estimated using the “infercnv” package (version 1.10.1) [24], using T cells as a reference. HLA genes on chromosome 6 were excluded due to their potential to exhibit CNVs in immune cells [25].

2.3. Differential expression analysis

Differential expression analysis was performed using the “DESeq2” package (version 1.34.0) [26]. Genes were retained if they exhibited non-zero expression in more than 80% of samples and had an average raw count greater than 1. Genes meeting the criteria of an absolute log2 fold change of 0.585 or greater and an adjusted p-value less than 0.05 were considered differentially expressed genes.

2.4. Identification of malignant prognostic signature

Firstly, univariate Cox regression analysis was conducted to identify significant prognostic genes. Subsequently, the Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to further reduce the number of candidate genes and construct the optimal prognostic signature. The “glmnet” package (version 4.1.7) [27] was utilized to determine the optimal penalty parameter (lambda.min) through 1000 iterations of 10-fold cross-validation. Patient-specific risk scores were calculated by integrating the coefficients and expression levels of the prognostic genes, as follows: ∑(Coef_i × Expr_i), where Coef_i denotes the coefficient of gene i, and Expr_i represents the expression level of gene i.

To evaluate the discrimination ability, Kaplan–Meier analysis and the concordance index (C-index) were performed using the “survival” package (version 3.2.11). Subsequently, calibration curves were generated to assess the agreement between predicted and observed outcomes using the “rms” (version 8.1.0) package, and time-dependent receiver operating characteristic (ROC) curve analysis was performed using the “timeROC” package (version 0.4) [28]. Samples with survival times exceeding 10 years were excluded.

2.5. Cell culture and mouse model establishment

GL261 cells were cultured in DMEM medium (containing 10% FBS and penicillin/streptomycin). The cells were cultured in a humidified atmosphere containing 5% CO₂ at 37°C.

All experimental procedures were approved by the Institutional Animal Care and Use Committee of the Animal Experimental Center at Soochow University. Animal breeding was managed by specialized personnel from the animal center. This study was conducted in strict accordance with the recommendations of the guide for the care and use of laboratory animals issued by the Ministry of Science and Technology of China.

Five female C57BL/6J mice, weighing 18 ± 2g and aged 3–4 weeks, were purchased from Changzhou Cavens Experimental Animal Co., Ltd. An intracranial GBM mouse model was established through stereotactical inoculation of GL261 cells (200,000 cells per mouse in 2 μL of serum-free media) into the brain (1.8 mm lateral and 0.6 mm anterior to the bregma; 2 mm of depth) of the mice that were anesthetized using 1% pentobarbital sodium. The tumor size was evaluated weekly through the animal imaging system. Daily monitoring included activity, weight loss (>20%), and distress signs. Veterinarian supervised endpoint determinations. Housing: All mice were housed in the animal laboratory of Soochow University (temperature 20 ~ 24 ℃, 40% ~ 45% humidity, and a 12 h light-dark cycle), and the mice were free to eat and drink during the rearing period. All mice were euthanized on day 28 after inoculation or reached Tumor Volume Endpoint: Mice were euthanized when tumor burden reached 1000 mm³ (calculated as 0.5 × length × width²), and complete excision of brains was performed, followed by isolation of the tumor from each mouse.

Mice were deeply anesthetized via intraperitoneal injection of 1% pentobarbital sodium. Adequate anesthetic depth was confirmed by the absence of the pedal withdrawal reflex. After achieving full anesthesia, mice were humanely euthanized by cervical dislocation. Subcutaneous tumors were then immediately dissected and harvested under aseptic conditions. As tumor collection was performed as a terminal procedure following euthanasia, no additional analgesics were required. All procedures were carried out rapidly to minimize discomfort, and every effort was made to alleviate animal suffering during handling and experimentation. All animal protocols were approved by the Institutional Animal Care and Use Committee (IACUC) and performed in accordance with institutional guidelines for the care and use of laboratory animals.

After intracranial injection of GL261 cells and tumor development, glioma tissues along with adjacent normal brain tissues were harvested from the mice for Western blot.

2.6. Western blotting

The harvested glioma tissues and adjacent normal mouse brain tissues were lysed and subsequently used for Western blot analysis. Separate total proteins using SDS-PAGE (Sodium Dodecyl Sulfate Polyacrylamide Gel Electrophoresis), transfer to the nitrocellulose membrane, block with 5% skimmed milk, and sequentially incubate with primary antibodies and secondary antibodies coupled with horseradish peroxidase (HRP). Visualize using chemiluminescence. The primary antibodies used for protein immunoblotting are as follows: RARRES2/Chemerin (10216–1-AP, Proteintech, 1:2000); Tubulin (66031–1-Ig, Proteintech, 1:2000); HRP-conjugated Goat Anti-Rabbit IgG(H + L) (SA00001–2, Proteintech, 1:2000).

2.7. Immunohistochemical analysis

Tumor tissues were obtained from the patients who underwent curative resection from 01/03/2022 to 01/03/2024 at Department of Neurosurgery, Nanjing medical university affiliated Suzhou Municipal Hospital. The study was conducted in accordance with the ethical principles stated in the Declaration of Helsinki and approved by the Bio-medical Research Ethics Committee of Nanjing Medical University Affiliated Suzhou Hospital (No.268(2024)). All of the patients provided written informed consent. This study utilized archived paraffin-embedded glioma tissue samples from Suzhou Municipal Hospital under approval on 28/04/2025 by Suzhou Municipal Hospital Ethics Committee. Written informed consent was obtained from all participants prior to tissue collection, including permission for future research use of anonymized data. Data accessed included [transcriptome sequencing, IHC results]. No additional clinical records beyond pathological diagnosis were used. All glioma tissue samples were processed under strict de-identification protocols: Patient IDs were replaced with anonymized codes; Sample metadata excluded name, ID, address, and exact diagnosis dates. Dewaxing of paraffin sections to water involves sequentially immersing the sections in an environmentally friendly dewaxing solution, followed by anhydrous ethanol, and then rinsing with distilled water. Sections should be heated in citrate buffer (pH = 6.0) after being submerged in 3% hydrogen peroxide (H₂O₂) for 15 min in order to retrieve the antigen. Allow the slides to cool naturally and then rinse them in PBS for three washes. For twenty-five minutes, submerge the parts in a 3% hydrogen peroxide solution at room temperature without exposure to light. Then, place the slides in PBS for 3 washes. Add 3% bovine serum albumin (BSA) evenly to cover the tissue within the histology ring. Drop the pre-diluted primary antibody onto the slide and incubated overnight at 4 °C. Rinse the slides three times in PBS. After air-drying the sections, add the secondary antibody corresponding to the species of the primary antibody, covering the tissue, and incubate at room temperature for 50 min. Rinse the slides three times in PBS. After drying the sections, add freshly prepared DAB coloration. Restain with hematoxylin for about 3 min, followed by rinsing with tap water. Finally, dehydration, gum sealing, and microscopic photography were carried out. The proportion of positive cells was determined by counting at least 1,000 cells in three random microscopic fields for each sample. The following primary antibodies were employed for immunohistochemical staining: RARRES2/Chemerin (10216–1-AP, Proteintech, 1:200).

2.8. Quantitative real‐time reverse transcription PCR (qRT‐PCR)

Total RNA was extracted using TRizol Reagent and 5 ug of RNA was used to synthesize cDNA according to the manufacturer′s instructions. The PCR conditions consisted of 5 min at 95 °C 1 cycle，30 s at 95 °C，30 s at 55 °C，30 s at 72 °C and 7 min at 72 °C 40 cycles. The sequences of primers are as follows:

RARRES2-F: TTGCTGATCTCCCTAGCCCTA;

RARRES2-R: TGGGTGTTTGTGGAACTCCTC;

IGFBP2-F: CAGACGCTACGCTGCTATCC;

IGFBP2-R: CCCTCAGAGTGGTCGTCATCA;

MDK-F: CAAGGGACCCTGAAGAAGGC;

MDK-R: CTTTGGTCTTTGACTTGCTCTTGG;

ACTIN-F: TGACTTCAACAGCGACACCCA;

ACTIN-R: CACCCTGTTGCTGTAGCCAAA.

The relative fold changes in mRNA expression were calculated using the 2^–ΔΔCT method, where the average of ΔCT values for the amplicon of interest was normalized to that of Actin.

2.9. Gene set enrichment analysis and gene set variation analysis

Gene sets of the Kyoto Encyclopedia of Genes and Genomes (KEGG), Gene Ontology Biological Process (GOBP), and HALLMARK (version 2023.2) were retrieved from the Molecular Signatures Database (MSigDB) [29]. Gene Set Enrichment Analysis (GSEA) was performed using the R package “clusterProfiler” (version 4.9.0.2) [30] to identify potential differences in biological pathway activity. Gene set variation analysis (GSVA) was performed using the “gsva” function of the GSVA package (version 1.44.5) [31].

2.10. Prediction of chemosensitivity

The prediction of the half-maximal inhibitory concentration (IC50) was conducted using the “oncoPredict” package [32], with GDSC v2 data serving as the training set. A lower IC50 value indicates greater sensitivity to the drug in cancer treatment. Consistent trends across four bulk RNA-seq cohorts (TCGA-GBM, TCGA-LGG, CGGA-mRNAseq365, and CGGA-mRNAseq693) were visualized in a dot plot.

2.11. Immune cell infiltration analysis

The ESTIMATE algorithm was harnessed to derive Immune Scores (version 1.0.13) [33]. Single-sample Gene Set Enrichment Analysis (ssGSEA) and CIBERSORT were employed to quantify immune cell infiltration levels in the glioma cohort [34]. Additionally, immune cell infiltration data from multiple algorithms, including TIMER, quanTIseq, MCP-counter, xCell, and EPIC, were retrieved from the Tumor IMmune Estimation Resource (TIMER) 2.0 database [35].

2.12. Statistical analysis

All data processing, statistical analyses, and plotting were performed using R 4.2.0 software. Spearman correlation was used to assess relationships between continuous variables. For comparing two groups, the statistical significance of variables was determined using the Wilcoxon rank-sum test. Meta analysis were performed using the “meta” package (version 8.0.2). All statistical tests were two-sided, with p < 0.05 considered statistically significant.

3.1. Single-cell transcriptional profiling of glioma

A total of 27,642 cells from 7 GBM samples passed quality control filtering, and batch effects were effectively corrected using the Harmony method (Fig 1A). At a resolution of 0.3, cells were stratified into 11 clusters (C0–C10) (Fig 1B). Immune cell signature genes (CD45, CD3D, AIF1) were specifically enriched in clusters C0, C1, and C7, whereas stromal markers (PECAM1, DCN, LUM) were predominantly expressed in clusters C9 and C10 (Figs 1D, E and S1A). Compared with cluster C8, clusters C2–C6 exhibited higher copy number variation (CNV) scores within non-immune and non-stromal cell populations (Fig 1F), along with elevated proliferation and stemness scores (Fig 1G, H). These clusters also showed increased G2/M phase scores and a reduced G1 phase ratio relative to oligodendrocytes (C8), indicating active cell cycle progression (S1B–D Fig). GSVA further revealed that clusters C2-C6 were significantly enriched in pathways related to cell cycle progression and oncogenic signaling, including MYC, and E2F (S1E Fig). Consistently, known GBM tumor signature genes (SOX2, SOX4, and OLIG2) [36] were markedly upregulated in these clusters, whereas cluster C8 specifically expressed oligodendrocyte markers (MBP, PLP1, MOG) (Fig 1D, E, S2 Table). Consequently, based on marker gene expression patterns and the above features, all clusters were annotated into six major cell types: myeloid cells (C0 and C1), T cells (C7), pericytes (C9), endothelial cells (C10), oligodendrocytes (C8), and malignant cells (C2–C6) (Fig 1B, C).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Single-cell atlas of glioma tissue.

(A-C) UMAP projection of the cellular landscape colored by sample source (A), clusters (B), and cell types (C). (D) Heatmap displaying the mean expression levels of marker genes across clusters. (E) UMAP plots showing the expression of classical molecular markers. (F-H) Boxplots comparing the CNV score (F), proliferation score (G), and stemness score (H) across clusters.

https://doi.org/10.1371/journal.pone.0349749.g001

3.2. Identification of malignant prognostic signatures

Considering the higher CNV, stemness, and proliferation scores observed in clusters 2, 3, and 4, these clusters were designated as representative malignant cells, and differential analysis was performed to identify cluster-specific marker genes. Among these, 689 genes were also upregulated in GBM samples compared with normal controls in the TCGA-GBM cohort (Fig 2A, B). To construct malignant prognostic signatures for glioma patients, univariate Cox regression analysis was performed on these malignant DEGs in the discovery cohort (TCGA-GBM), and a total of 55 genes met the cutoff of p < 0.05, with genes exhibiting p < 0.01 displayed in a forest plot (Fig 2C). These genes were subsequently subjected to LASSO-Cox proportional hazards regression with 10-fold cross-validation. The LASSO coefficient profile plotted against the log(λ) sequence identified three optimal coefficients using the minimum λ criterion (Fig 2D). Ultimately, a three-gene signature comprising IGFBP2, MDK, and RARRES2 demonstrated optimal predictive performance (Fig 2D), and malignant prognostic scores (MPS) for each patient were calculated based on the LASSO model. Survival analysis indicated that patients with lower MPS had significantly prolonged overall survival (Fig 2E, log-rank test, p = 0.002). Consistently, ROC curve analysis demonstrated the prognostic performance of the MPS, with 1-, 3-, and 5-year AUC values of 0.67, 0.71, and 0.85 in the TCGA-GBM cohort (Fig 2F).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Construction and validation of the prognostic model.

(A) Volcano plot of differential analysis between GBM and normal tissues in the TCGA-GBM cohort. (B) Venn diagram of upregulated genes in GBM and malignant cluster marker genes. (C) Forest plot showing the prognostic effect on OS based on univariate Cox regression in the TCGA-GBM cohort. (D) Cross-validation of the prognostic effect using LASSO regression. (E) Kaplan-Meier survival curves comparing OS between low- and high-MPS groups in the TCGA-GBM cohort. (F) Time-dependent ROC analysis assessing the malignant score for predicting OS at 1, 3, and 5 years in the TCGA-GBM cohort. (G, I, K, M) Kaplan-Meier survival curves comparing OS between low- and high-MPS groups in the TCGA-LGG (G), mRNAarray 301 (I), mRNAseq 325 (K), and mRNAseq 639 (M) cohorts. (H, J, L, N) Time-dependent ROC analysis assessing the malignant score for predicting OS at 1, 3, and 5 years in the TCGA-LGG (H), mRNAarray 301 (J), mRNAseq 325 (L), and mRNAseq 639 (N) cohorts.

https://doi.org/10.1371/journal.pone.0349749.g002

3.3. External validation of MPS prediction efficiency

To confirm the robustness and efficacy of the MPS across diverse populations, the same calculation was applied to four external validation cohorts from the TCGA and CGGA databases (TCGA-LGG, mRNAarray 301, mRNAseq 325, and mRNAseq 693). In all cohorts, the median overall survival of patients in the high-MPS group was significantly shorter (Fig 2G, I, K, and M). ROC curve analysis showed that the AUC values for predicting 3-year survival were all above 0.75, while the AUC values for 1- and 5-year survival were all above 0.65 (Fig 2H, J, L, and N). In addition, calibration plots showed strong agreement between predicted probabilities and observed outcomes (S2 Fig), with an average C-index greater than 0.7, indicating good predictive accuracy of the model. These findings collectively demonstrate the robust prognostic performance of the MPS.

To further evaluate whether the malignant prognostic signature (MPS) provides prognostic information independent of established clinical and molecular factors, multivariable Cox regression analyses were performed adjusting for age, WHO grade, IDH mutation status, 1p/19q codeletion, and MGMT promoter methylation. After z-score normalization, MPS remained a statistically significant independent predictor of overall survival in the TCGA cohorts (TCGA-GBM: HR = 1.327, p = 0.045; TCGA-LGG: HR = 1.663, p = 0.002; S3 Table). In the CGGA cohorts, MPS showed a consistent direction of effect (HR > 1), although statistical significance was not uniformly retained. An additional multivariable model including only molecular covariates (IDH mutation, 1p/19q codeletion, and MGMT promoter methylation) was further constructed, in which MPS remained statistically significant across all five cohorts (S4 Table).

3.4. Expression dysregulation and prognostic impact of MPS genes

Next, to further evaluate the role of each MPS gene, differential expression and survival analyses were conducted for the three MPS genes in TCGA and CGGA glioma cohorts. In all three CGGA cohorts, expression levels of all three genes increased significantly with higher tumor grade (Fig 3A–C). Compared to IDH wild-type samples, IDH-mutant samples exhibited significantly lower expression of IGFBP2, MDK, and RARRES2 (Fig 3D–F). Additionally, IGFBP2 and RARRES2 expression were reduced in samples with 1p/19q codeletion (Fig 3G–I). Because IDH mutation and 1p/19q codeletion are well-known markers of prolonged survival and delayed malignant progression [37,38], these expression patterns further highlight the relevance of the three MPS genes to glioma progression.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Expression distribution of MPS genes across clinicopathological characteristics.

(A-C) IGFBP2 (A), MDK (B), and RARRES2 (C) expression levels are significantly increased in higher-grade gliomas in the CGGA database. (D-F) IGFBP2 (D), MDK (E), and RARRES2 (F) expression levels are significantly increased in gliomas without IDH mutation in the CGGA database. (G-I) Expression distribution of IGFBP2 (G), MDK (H), and RARRES2 (I) across different 1p/19q status groups in three CGGA cohorts.

https://doi.org/10.1371/journal.pone.0349749.g003

Moreover, in both the TCGA and CGGA glioma cohorts, IGFBP2, MDK, and RARRES2 were identified as significant prognostic risk factors (Fig 4A). The AUC values for RARRES2 in predicting 1-, 3-, and 5-year survival all exceeded 0.6, and the average AUC values for IGFBP2 in predicting 1-, 3-, and 5-year survival were 0.73, 0.78, and 0.75, respectively (Fig 4B–F). MDK also demonstrated robust predictive performance across at least four cohorts. Decision curve analysis (DCA) indicated that combining tumor grade with IGFBP2, MDK, or RARRES2 improved prognostic accuracy (S3 Fig). Together, these findings suggest that IGFBP2, MDK, and RARRES2 can serve as effective complements to traditional clinical indicators in glioma prognosis.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Prognostic impact of MPS genes.

(A) Univariate Cox regression analysis of IGFBP2, MDK, and RARRES2 in the TCGA-GBM, TCGA-LGG, mRNAarray 301, mRNAseq 325, and mRNAseq 639 cohorts. (B-F) Time-dependent ROC analysis assessing IGFBP2, MDK, and RARRES2 in the TCGA-GBM (B), TCGA-LGG (C), mRNAarray 301 (D), mRNAseq 325 (E), and mRNAseq 639 (F) cohorts.

https://doi.org/10.1371/journal.pone.0349749.g004

3.5. Efficacy of MPS genes in predicting drug sensitivity

To explore the relationship between the three prognostic signatures and drug sensitivity, the half-maximal inhibitory concentration (IC50) values of each drug for glioma samples in four sequencing cohorts were calculated. The landscape of correlation and significance between drug sensitivities and malignant signature expression is depicted in S4 Fig.

Positive correlations were observed between the IC50 values of entinostat and the expression of IGFBP2 and MDK. Additionally, the IC50 values of doramapimod, gefitinib, erlotinib, LY2109761, AZD1208, SB505124, vorinostat, and afatinib were positively correlated, while the sensitivities of MG-132, SCH772984, PLX-4720, PD0325901, AZD5582, trametinib, and AZD1332 exhibited negative correlations with IGFBP2 expression. The IC50 values of camptothecin, luminespib, MG-132, irinotecan, 5-fluorouracil, SCH772984, talazoparib, topotecan, rapamycin, and epirubicin were found to be negatively correlated with MDK expression. Furthermore, three consistent pairs of correlations between drug sensitivity and RARRES2 expression, including PCI − 34051, PRIMA−1MET, and venetoclax, were observed across the four cohorts. These findings suggest the potential existence of additional FDA-approved drugs that may hold promise for the treatment of glioma.

3.6. Function analysis and expression verification of RARRES2

To further explore the biological basis of the prognostic marker, GSEA was performed between the high- and low-expression groups. This analysis revealed that multiple immune-related pathways were enriched in the high-expression groups, including IL6-JAK-STAT3 signaling, IL2-STAT5 signaling, interferon-α response, inflammatory response, and the complement system (Fig 5A). Among these, RARRES2 expression showed a consistent positive correlation with the immune score across four cohorts (Figs 5B and S5). Furthermore, using seven different deconvolution algorithms from TIMER 2.0, we analyzed correlations between specific immune cell abundances and gene expression levels. Significant correlations were identified between RARRES2 expression and the abundance of CD8 ⁺ T cells, NK cells, and especially M2 macrophages (Figs 5B and S6). Cross-algorithm analysis indicated that estimates for CD8 ⁺ T cells, monocytes, and macrophages were strongly correlated across methods, supporting the robustness of their association with RARRES2 (S7 Fig). GSEA based on immune cell signatures showed that microglia/macrophage (both M1 and M2 subtypes) were consistently enriched in the high-RARRES2 groups (S8A–B Fig). Notably, M2 macrophages constituted the largest proportion among macrophages and were the primary contributors to changes in macrophage infiltration (S8C–F Fig). Using the square root of fold change × proportion as a relevance metric [39], we found that M2 macrophages, rather than M1 macrophages, were more closely associated with RARRES2 expression (S8G–J Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Functional analysis of RARRES2.

(A) Hallmark gene set enrichment analysis between high- and low-expression groups of IGFBP2, MDK, and RARRES2 in the TCGA-GBM, TCGA-LGG, mRNAseq 325, and mRNAseq 639 cohorts. Circle size represents the FDR value of the enriched term in each cancer, and color indicates the normalized enrichment score (NES). (B) Correlation analysis between RARRES2 expression levels and immune cell infiltration levels. (C-E) UMAP plots showing the expression distribution of RARRES2 receptors: CMKLR1, GPR1, and CCRL2. (F) qPCR validation of RARRES2 in paired glioma and normal tissues. (G) Expression distribution of RARRES2 in normal brain and glioma tissues in three independent GEO cohorts. (H) Forest plot of meta-analysis for RARRES2 based on univariate Cox regression of PFS. (I) RARRES2 expression levels in paired glioma and adjacent normal tissues detected by Western blot. (J) Representative IHC images and the percentage of RARRES2-positive cells in glioma samples. Scale bars: 50 μm (left) and 20 μm (right). LGG samples, n = 6; GBM samples, n = 6.

https://doi.org/10.1371/journal.pone.0349749.g005

Previous studies have identified CMKLR1, GPR1, and CCRL2 as primary receptors of RARRES2 [40]. CMKLR1 and CCRL2 were predominantly expressed in myeloid cells, while GPR1 was mainly expressed in malignant cells (Fig 5C–E). In the GTEx-TCGA, GSE16011, and GSE147352 cohorts, RARRES2 expression levels were significantly higher in GBM than in normal brain tissue (Fig 5F). Similarly, in our in-house cohort, both transcriptional (Fig 5G) and protein-level (Fig 5I) analyses showed elevated RARRES2 expression in GBM. Immunohistochemistry confirmed a significantly higher proportion of RARRES2 ⁺ cells in the malignant subtype (Fig 5J), consistent with single-cell analysis (S1H Fig). Like RARRES2, the expression of CMKLR1, GPR1, and CCRL2 was higher in GBM than in normal brain or LGG tissues (S9A–C Fig). A meta-analysis of progression-free survival (PFS) using transcriptomic data from TCGA and GEO (datasets with >50 cases) identified RARRES2 as a significant prognostic risk factor (Fig 5H).

Additionally, GSVA and correlation analyses revealed that the activity of glycolipid metabolism–related pathways (such as the PPAR signaling pathway, glycosphingolipid biosynthesis, and glycolysis/gluconeogenesis) was positively correlated with RARRES2 expression (S9D Fig). However, it should be noted that in most glioma cohorts of TCGA and CGGA, RARRES2 expression was significantly negatively correlated with key genes of the fatty acid synthesis pathway (including SREBF2, FASN, ACACA, and ACSL1) but positively correlated with genes involved in long-chain fatty acid synthesis (S9E Fig).

3.7. Pan‑cancer analysis of RARRES2 in TCGA cohorts

Finally, given that GTEx and HPA data show relatively low RARRES2 expression in normal brain tissue compared to other tissues (S10 Fig), we assessed its dysregulation and prognostic impact across 33 cancer types. Compared to normal tissues, RARRES2 expression was lower in 12 tumor types, including bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical and endocervical cancer (CESC), cholangiocarcinoma (CHOL), colon adenocarcinoma (COAD), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), rectum adenocarcinoma (READ), thyroid carcinoma (THCA), and uterine corpus endometrial carcinoma (UCEC), but higher in five tumor tissue types: glioblastoma multiforme (GBM), head and neck squamous cell carcinoma (HNSC), kidney renal clear cell carcinoma (KIRC), kidney papillary cell carcinoma (KIRP), and liver hepatocellular carcinoma (LIHC) (Fig 6A). Univariate Cox regression indicated that RARRES2 could serve as a prognostic risk factor in six cancers, including uveal melanoma (UVM), low-grade glioma (LGG), pancreatic adenocarcinoma (PAAD), glioblastoma multiforme (GBM), stomach adenocarcinoma (STAD), and kidney renal clear cell carcinoma (KIRC) (Fig 6B). Across the TCGA pan-cancer cohort, RARRES2 expression was significantly correlated with the immune score in various cancer types (Fig 6C). Notably, RARRES2 showed both immunostimulatory and immunoinhibitory associations: in cancers like BLCA, BRCA, HNSC, LGG, LUAD, LUSC, PCPG, PRAD, and THCA, RARRES2 was positively correlated with most immunostimulators (e.g., CD40LG, CCL12, TNFSF14), while also positively correlated with certain immunoinhibitors (e.g., CSF1R, IL10, TGFBR1) in BLCA, BRCA, COAD, HNSC, LUAD, LUSC, PRAD, and THCA (S11A–C Fig). Furthermore, RARRES2 showed concurrent positive correlations with both anti-tumor (e.g., activated T cells, NK cells, type 1 helper T cells) and pro-tumor/immunosuppressive (e.g., Tregs, Th2 cells, MDSCs) immune cell infiltrates in most cancer types (S11D Fig). These coordinated patterns suggest a RARRES2-mediated immunoregulatory feedback loop, potentially explaining the paradoxical coexistence of immune activation and suppression in the tumor microenvironment.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Pan-cancer analysis of RARRES2.

(A) RARRES2 expression distribution between tumor and normal tissues across 33 cancer types. (B) Summary of the relationship between RARRES2 expression and OS, DSS, DFI, and PFI based on univariate Cox regression. Red indicates RARRES2 as a risk factor, and blue indicates a protective factor. Only P values < 0.05 are shown. (C) Correlation analysis between RARRES2 expression and immune score. (D) Dot plot showing differential infiltration levels between high- and low-RARRES2 expression groups. Circle size represents the P value of the correlation. Red indicates upregulation in the high-expression group, and blue indicates downregulation. (E) Correlation analysis between RARRES2 expression and tumor mutation burden. (F) Analysis of RARRES2 alteration frequency across cancers. (P < 0.05, *P < 0.01, **P < 0.001, ***P < 0.0001).

https://doi.org/10.1371/journal.pone.0349749.g006

Additionally, RARRES2 was significantly correlated with tumor mutation burden (TMB) in ACC, LGG, LIHC, PAAD, PCPG, PRAD, STAD, THCA, and THYM (Fig 6E). Genomic alteration analysis revealed that RARRES2 alterations are relatively uncommon across cancers, with ovarian serous cystadenocarcinoma showing the highest alteration frequency (mostly amplifications) (Fig 6F).