Introduction

Glioma accounts for more than 80% of all primary malignant brain tumors and ~30% of all central nervous system tumors1. One of its distinguishing characteristics is its invasive behavior toward the surrounding brain tissue2. The prognosis for gliomas varies greatly depending on the tumor grade, with median overall survival (OS) ranging from 1 to 15 years3. Despite the use of various treatment methods such as surgery, radiation, and chemotherapy, patients with glioblastoma multiforme (GBM) typically have a median survival of around 14 months4. Gliomas develop within a tumor microenvironment that is both immunosuppressive and heterogeneous, which greatly influences their malignant progression5.

Various immune cells and stromal cells, including endothelial cells and TAMs, play a significant role in the intricate tumor microenvironment6. TAMs make up ~30% of the tumor mass7 and consist of two distinct populations in glioma: tissue-resident microglia and macrophages derived from the bone marrow8. TAMs are categorized into M1 and M2 macrophage phenotypes based on their cell markers and functions. M1 macrophages are associated with inflammation and contribute to anti-tumor activity, whereas M2 macrophages promote tumor growth by secreting angiogenic factors and immunosuppressive cytokines9. This promotes the creation of an immunosuppressive environment in gliomas. Studies have shown that TAMs reduce pro-inflammatory cytokines and essential molecules involved in T-cell co-stimulation, such as CD40, indicating their ability to restrict T-cell response to glioma10. Through their interactions with cancerous cells, TAMs establish a supportive environment for tumor development, allowing glioma to evade immune surveillance and promote proliferation. Consequently, understanding the relationship between glioma and TAMs is crucial to develop enhanced therapeutic approaches11. However, the precise interaction between glioma and TAMs remains uncertain. To achieve a comprehensive understanding of glioma, it is essential to study TAMs. Previous research has identified CD163 as a marker that differentiates between M1 and M2 macrophages and is associated with prognosis12. In addition to its role in regulating the stemness of glioma, CD163 is associated with a significant decrease in the infiltration of CD163+ macrophages in glioma following anti-PD-L1 antibody treatment. Therefore, targeting CD163 could potentially be used as a therapeutic approach for glioma.

The formyl peptide receptors (FPRs) are a group of chemoattractant receptors known for their association with G proteins and possession of seven transmembrane domains. In the human FPR family, there are three members: FPR1, FPR2, and FPR313,14. FPR1 can be found in macrophages, astrocytes, and immature dendritic cells, while FPR2 is present in astrocytoma cells and neuroblastoma cells. FPR3 is detected in monocytes, myeloid cells, and specific macrophage subpopulations localized in certain tissues, but it is not observed in human neutrophils13,14,15. A multitude of studies has been conducted to investigate the biological function of FPR1 and FPR2 in various tumors. However, there is limited research on the potential role of FPR3 in cancer16,17. Despite having a 56% and 83% similarity in amino acid sequence with FPR1 and FPR2, respectively, FPR3 has minimal ligand overlap with either receptor18,19. Therefore, FPR3 may have a unique functional role. FPR3 has been shown to act as a detector of pathogens when stimulated with a bacterial endotoxin20. Additionally, FPR3 is involved in the movement of CD4+T cells21. Considering its predominant expression in monocytes and its association with tumor grade and outcomes, FPR3 has the potential as a new biomarker for glioma. However, the overall role of the FPR3 receptor, particularly concerning glioma, is largely unknown.

In this study, we found that FPR3 is expressed as a marker on CD163+ macrophages and is elevated in gliomas. High expression of FPR3 predicts a poor outcome for glioma patients. Through single-cell RNA sequencing (scRNA-seq) analysis, we identified CD163+FPR+ macrophage subsets. Using machine learning, we constructed a prognostic model based on the risk score of these CD163+FPR+ macrophage subsets.

Results

FPR3, a marker expressed on CD163+ macrophage

The initial examination involved a scRNA-seq dataset (GSE182109) where the largest proportion of cells consisted of glioma and myeloid cells (Fig. 1A). Through scRNA-seq analysis, a total of 498 differentially expressed genes (DEGs) were identified when comparing glioma cells to myeloid cells, with a log2FC > 0.25 in absolute value. Given that macrophages can be polarized into M0, M1, and M2 phenotypes22, we utilized the WGCNA approach to identify modules associated with these three macrophage phenotypes, employing the CIBERSORT algorithm (Fig. 1B). In order to establish a scale-free network (Supplementary Fig. 1A, B), we selected a soft threshold power β = 5 (resulting in a scale-free R2 of 0.90). Interestingly, the yellow module displayed both negative and positive correlations with M0, M1, and M2 macrophage infiltration in glioma (Fig. 1B). To gain a better understanding of the infiltration populations of TAMs, we examined the single-sample gene set enrichment analysiss (ssGSEA) score of the M2-like TAMs signature in the TCGA glioma cohort. DEGs were assessed in gliomas with high and low scores (Fig. 1C). Ultimately, 232 common genes were identified (Fig. 1D). Notably, FPR3 emerged as the highest-ranked gene that was drastically upregulated (Fig. 1C). Moreover, analysis of single-cell sequencing data demonstrated a predominant expression of FPR3 in TAMs across additional datasets (Fig. 1E and Supplementary Fig. 2). Further investigation included an examination of the co-localization of FPR3 with CD163 (a marker of TAMs). The UMAP plots revealed that all CD163+ TAMs in the UMAP are not also FPR3+ (Fig. 1E). In four databases, a significant positive correlation was found between FPR3 and CD163, as well as the M2-like TAMs score (Fig. 1F, G and Supplementary Fig. 3). To explore the potential association between FPR3 and M2-like TAMs in more depth, frozen sections of human glioma were subjected to co-immunostaining with CD163. The results indicated a co-localization of CD163 with FPR3 (Fig. 1H). Similarly, the examination of IHC showed a significant positive correlation between FPR3 and CD163 in patients with glioma (Fig. 1I). Collectively, these findings strongly supported the notion that FPR3 serves as a marker expressed on CD163+ macrophages.

Fig. 1: FPR3, a marker expressed on CD163+ macrophage.
figure 1

A UMAP plot displaying identified clusters in glioma. B WGCNA identifies M0, M1, and M2-like macrophage correlated modules. C The volcano plot showing differentially expressed genes between gliomas with high- and low-M2 macrophage ssGSEA score. D Venn diagram demonstrating the overlapping of genes in brown modules and differentially expressed genes. E Expression of CD163 and FPR3 overlaid on the UMAP space. F Correlation analysis of FPR3 and M2 macrophage ssGSEA score in glioma. G Correlation analysis of FPR3 and CD163 expression in TCGA cohort. H IF staining of FPR3 and CD163 in human glioma were shown. I Representative IHC staining of FPR3 and CD163 in glioma microarray; scale bar represents 100 µm.

Full size image

FPR3 was elevated in gliomas and its high expression predict poor outcome

The expression of FPR3 mRNA in tumor tissue was significantly increased compared to normal tissue in all three datasets (P < 0.05, Fig. 2A–C). This finding was further confirmed in our own cohort using quantitative real-time polymerase chain reaction (RT-qPCR) (P < 0.05, Fig. 2D). Additionally, FPR3 levels were significantly elevated in higher-grade gliomas compared to lower-grade gliomas across various datasets (P < 0.05, Supplementary Table 1). According to the TCGA dataset, gliomas without an IDH mutation exhibited higher levels of FPR3 expression compared to gliomas with an IDH mutation, and this finding was also confirmed in the CGGA and GSE16011 cohorts (P < 0.05, Supplementary Table 1). Moreover, GBM consistently showed a high prevalence of FPR3 with elevated expression levels across all four cohorts (P < 0.05, Supplementary Table 1). In 14 paired tumor tissues, FPR3 showed a significant increase compared to the corresponding peritumoral tissues (P < 0.05, Fig. 2E). Additionally, we examined FPR3 expression in tissue microarray (TMA) using IHC staining and observed elevated levels of FPR3 in gliomas (P < 0.05, Fig. 2F). FPR3 also exhibited a substantial increase in 30 paired tumor tissues when compared to corresponding peritumoral tissues (P < 0.05, Fig. 2G). In a pan-cancer analysis, we investigated the relationship between FPR3 expression and patients’ clinical outcomes. Notably, we found a significant association between FPR3 and OS in seven distinct cancer categories, including glioma, THYM, and UVM (P < 0.05, Supplementary Fig. 4). Kaplan–Meier curves were used to assess the relationship between FPR3 expression and OS, disease-specific survival (DSS), and progression-free survival (PFI). Elevated FPR3 levels were indicative of an adverse prognosis in the TCGA cohort (P < 0.05, Fig. 2H). This prognostic importance of FPR3 expression was further confirmed in the remaining five cohorts (P < 0.05, Fig. 2H). We observed that glioma patients with elevated FPR3 levels generally had shorter OS and PFI durations compared to patients with low-FPR3 expression, as determined by qPCR (P < 0.05, Fig. 2I). Similar differences were observed in another internal cohort using IHC. Patients with glioma who had higher levels of FPR3 had relatively shorter survival times compared to those with lower FPR3 levels (P < 0.05, Fig. 2J). These findings suggest that FPR3 may have the potential to serve as a predictive marker for patients with glioma and could play a risky role in outcomes.

Fig. 2: Elevated expression of FPR3 indicates poor prognosis in glioma.
figure 2

The expression level of FPR3 in the normal tissues and glioma tissues in TCGA cohort (A), GSE16011 (B), Rembrandt cohort (C) and in-house cohort (D). E The expression of FPR3 in peritumor and glioma tissues was analyzed by RT-qPCR. F The expression level of FPR3 in the normal tissues and glioma tissues was analyzed by immunohistochemical staining. G The expression of FPR3 in glioma and peritumor tissues was analyzed by IHC. H Kaplan–Meier survival plot of FPR3 expression of glioma patients in the six cohorts. I Kaplan–Meier survival plot of FPR3 expression of glioma patients by RT-qPCR. J Kaplan–Meier survival plot of FPR3 expression of glioma patients by IHC. P < 0.05, log-rank test. (“**, ***” respectively means P < 0.01, P < 0.001).

Full size image

The correlations of FPR3 expression with immunological parameters

After analyzing the correlation between glioma malignancy and FPR3 expression, we have determined that the abnormal expression of FPR3 may potentially promote the progression of glioma. In the TCGA cohort, we identified 2902 DEGs that were significantly upregulated, as well as 1370 DEGs that exhibited significant downregulation. To further investigate the underlying pathways of FPR3, we employed GSEA. In the TCGA cohort, gliomas with elevated FPR3 levels displayed an abundance of immunomodulatory pathways (Supplementary Fig. 5A). Furthermore, gliomas with heightened FPR3 levels consistently exhibited higher immune and stromal scores compared to gliomas with low-FPR3 levels across four distinct groups (P < 0.05, Supplementary Fig. 5B). It appears that FPR3 may play a role in regulating immune and stromal cells. In order to explore the composition of immune cells in subgroups with high and low-FPR3 expression, we employed the CIBERSORT algorithm to uncover the correlation between these two subgroups and 22 subsets of immune cells (Supplementary Fig. 5C). The high-FPR3 group exhibited an increase in immune cell markers compared to the low-FPR3 group (P < 0.05, Supplementary Fig. 5D). Additionally, we discovered that the presence of FPR3 was compatible with various immune checkpoints (P < 0.05, Supplementary Fig. 5E). Furthermore, genetic abnormalities serve as conventional markers for the response to anti-PD-1/PD-L1 treatment23. The waterfall plot illustrated the mutations in the high-FPR3 category. Compared to gliomas with low-FPR3 levels (IDH1, 72% and TP53, 37%), gliomas with high-FPR3 expression (IDH1, 48% and TP53, 48%) had a lower frequency of mutations in IDH1 and TP53. On the other hand, gliomas with high-FPR3 expression displayed increased mutation frequencies in PTEN, TTN, and EGFR (PTEN, 12%, TTN, 17%, and EGFR, 10%), while those with low-level FPR3 (PTEN, 4%, TTN, 9%, and EGFR, 8%) exhibited lower mutation frequencies (Supplementary Fig. 5F). The immune-suppressive microenvironment in brain tumors is established through the cooperation of various cancer-causing pathways. Therefore, inhibiting these pathways impedes the creation of a repressive atmosphere for the immune system. We observed a significant increase in immunosuppressive oncogenic pathways in gliomas with elevated FPR3 expression (P < 0.05, Supplementary Fig. 5G). The cancer immunity cycle consists of seven stages24. In the high-FPR3 group, there was a decrease in the recruitment of various steps, including macrophage, Th1 cell, and Th17 recruitment. This decrease was observed in Step 1, Step 3, and Step 4 (P < 0.05, Supplementary Fig. 5H).

Identification of the CD163+FPR3+ macrophage subset

Myeloid cells comprise the largest immune compartment in gliomas23. To gain a comprehensive understanding of the diverse cellular and molecular characteristics of myeloid cells within gliomas, we conducted meticulous isolation of myeloid cells followed by thorough clustering analysis. This approach led to the identification of a total of 11 distinct myeloid clusters (Fig. 3A). The UMAP plots demonstrated that FPR3 was not expressed in certain CD163+ macrophages (Fig. 3B). Thus, the identification of the CD163+FPR3+ macrophage subset holds tremendous significance. We further divided macrophages with nonzero expression of FPR3 into CD163+FPR3+ and CD163+FPR3− groups. Subsequently, we obtained 142 DEGs from the CD163+FPR3+ and CD163+FPR3− signatures. In the TCGA cohort, GSEA uncovered that the genes were primarily enriched in pathways such as “IL6_JAK_STAT3_SIGNALING”, “INFLAMMATORY RESPONSE”, and “TNFA_SIGNALING_VIA_NFKB” (NES > 1, adjust P < 0.05, Fig. 3C). We examined the potential importance of FPR3 in predicting immune phenotypes and therapeutic opportunities. Based on the spatial distribution of CD8+ T cells illustrated in Fig. 3D, glioma tissues were categorized into two immune phenotypes: non-inflamed and inflamed. Tumor tissues with a non-inflamed phenotype exhibited significantly higher upregulation of FPR3 compared to those with an inflamed phenotype (P < 0.05, Fig. 3E). Within these immune phenotypes, gliomas in the FPR3 upregulated-expression group showed a greater abundance of PD-L1, PD-1, and CD40 compared to the low-expression FPR3 group (P < 0.05, Fig. 3F).

Fig. 3: Identifcation of the CD163+FPR3+ macrophage subset.
figure 3

A A UMAP projection of de novo clustered myeloid cells. B Expression of CD163 and FPR3 overlaid on the UMAP space. C GSEA analysis of specific gene of the CD163+FPR3+ macrophage subset. D Two immune phenotypes in the glioma microarray cohort were detected using IF. E The expression level of FPR3 between glioma with two immune phenotypes. F The expression of PD-L1, CD40, PD-1, P65, and STAT3 expression in glioma with the CD163+FPR3+ and CD163+FPR3− macrophage subset. (“*”means P < 0.05).

Full size image

Development and validation of the CD163+FPR3+ macrophage-associated signature

Univariate Cox analysis identified 62 common prognostic genes from the expression profiles of 142 genes associated with CD163+FPR3+ macrophages in the six cohorts (Supplementary Table 1). Machine learning-based integrative procedures were employed to further analyze these prognostic genes. Using the LOOCV framework, we utilized 101 distinct forecasting models on the TCGA dataset and calculated the C-index for each model across all validation datasets (Fig. 4A). The optimal model comprised a combination of random survival forest (RSF) and CoxBoost techniques, achieving the highest mean C-index (0.750) (Fig. 4A). Additionally, this integrated model outperformed all other models in terms of C-index across the remaining five datasets (Fig. 4A). The CoxBoost algorithm identified the 17 genes with the highest values (IFI6, HEXB, TNFRSF11A, GPNMB, SPRED1, NRP1, RPN2, FUCA1, TGFBI, CTSC, MSR1, PRDX1, CLEC5A, SCD, TUBA1C, ARL4C, TTYH3), while the RSF algorithm selected the most reliable model (Fig. 4B, C). Subsequently, the RSF algorithm was utilized to calculate a risk score for each individual. The optimal threshold values determined by the survminer package were used to classify all patients into high-risk and low-risk groups (Fig. 4D). In the TCGA dataset, the low-risk category exhibited significantly better OS compared to the high-risk category, demonstrating a noticeable difference (P < 0.05, Fig. 4E). The same trend was observed in the remaining five validation datasets (P < 0.05, Fig. 4F–J). Univariate and multivariate Cox regression analyses demonstrated that risk scores had independent prognostic value in the TCGA cohort (P < 0.05, Supplementary Fig. 6A). Additionally, we evaluated the prognostic significance of the risk score for PFI and DSS in the TCGA cohort. Kaplan–Meier analysis revealed significant prognostic differences, as glioma patients in the high-risk category exhibited reduced survival duration (Supplementary Fig. 6B, C).

Fig. 4: Development and validation of the CD163+FPR3+ macrophage-associated signature.
figure 4

A The C-index of 101 kinds of prediction models built by machine learning algorithms for the CD163+FPR3+ macrophage-associated signatures in the six glioma cohorts. B An ensemble of 17 cell death-associated genes with Cox coefficients. C The number of trees for determining the CD163+FPR3+ macrophage-associated signature with minimal error and the importance of CD163+FPR3+ macrophage-associated signatures based on the RSF algorithm. D The identification of optimal cutoffs for survival analyses in TCGA cohort. Kaplan–Meier survival curves of OS between high-score and low-score patients identified the scoring system in TCGA (E), CGGA-693 (F), CGGA-301 (G), CGGA-325 (H), GSE16011, (I) and Rembrandt (J) datasets. P values were calculated by the log-rank test, and P < 0.05 was considered significant.

Full size image

Evaluation of the CD163+FPR3+ macrophage-associated signature

The discriminatory ability of the CD163+FPR3+ macrophage-associated signature was evaluated using ROC analysis. The AUCs for 1-, 3-, and 5-year survival were 0.968, 0.984, and 0.984 in the TCGA cohort; 0.749, 0.784, and 0.768 in the CGGA-693 cohort; 0.749, 0.845, and 0.869 in the CGGA-301 cohort; 0.708, 0.810, and 0.769 in the CGGA-325 cohort; 0.791, 0.866, and 0.831 in the GSE16011 cohort; and 0.688, 0.760, and 0.783 in the Rembrandt cohort, respectively (Fig. 5A and Supplementary Table 2). The C-index [95% CI] values for the six cohorts were 0.936 [0.923–0.948], 0.707 [0.683–0.732], 0.719 [0.687–0.750], 0.694 [0.658–0.729], 0.726 [0.694–0.757], and 0.659 [0.627–0.692], respectively (Fig. 5B and Supplementary Table 3). Thus, the prognostic predictive ability of the risk score was compared to various clinical and molecular factors. In all six cohorts, the risk score demonstrated significantly higher accuracy compared to factors such as histology, grade, IDH mutation status, 1p19q status, MGMTp status, and TCGA subtypes (all P < 0.05, Fig. 5C–F and Supplementary Fig. 7A, B). However, there were no significant differences observed between the risk score and grade in the CGGA-325 cohort, the risk score and grade and histology in the CGGA-301 cohort, and the risk score and grade, histology, and age in the Rembrandt cohort (P > 0.05, Fig. 5F and Supplementary Fig. 7A, B). An interesting concept was to improve clinical utility by combining the risk score with commonly used clinical characteristics. The management of glioma often relies on tumor grade, and multivariate Cox regression analysis demonstrated significant results for tumor grade in multiple cohorts. Therefore, further investigations were conducted to assess the effectiveness of combining the risk score with grade. It was found that the combination of risk score and grade yielded significantly better performance compared to using the risk score or grade alone in multiple datasets (P < 0.05, Fig. 5G–J and Supplementary Fig. 7C, D). Based on these findings, it can be concluded that incorporating both the risk score and grade could enhance the predictive capacity of the model. Calibration plots were used to assess the discrimination ability of the model. The calibration plots showed strong concordance between prediction and observation for the risk score in the TCGA cohort, and this finding was further validated in the remaining five cohorts (Fig. 5K and Supplementary Fig. 7E, F).

Fig. 5: Evaluation of the CD163+FPR3+ macrophage-associated model.
figure 5

A Time-dependent ROC curves displaying the prognostic accuracy of the risk score at 1, 3, and 5 years in TCGA, CGGA-693, CGGA-301, CGGA-325, GSE16011, and Rembrandt cohorts. B C-index of the risk score in the six cohorts. The performance of risk score was compared with other clinical and molecular variables in predicting prognosis in TCGA (C), CGGA-693 (D), GSE16011 (E), and Rembrandt (F) datasets. The performance of risk score + grade was compared with risk score and grade alone in predicting prognosis in TCGA (G), CGGA-693 (H), GSE16011 (I), and Rembrandt (J) datasets. K The 1-year, 2-year, and 3-year calibration curves of the CD163 + FPR3+ macrophage-associated signature in the four datasets. The P values are labeled above each boxplot with asterisks (-, no significant, “*, **, ***” respectively means P < 0.05, P < 0.01, P < 0.001).

Full size image

Comparison of gene expression-based prognostic signatures in glioma

Significantly, the emergence of machine learning techniques, along with advancements in next-generation sequencing and big-data technologies, has greatly facilitated the generation of a substantial number of prognostic and predictive gene expression signatures. To evaluate the effectiveness of the risk score, we extensively collected published signatures for comparison, which were associated with various biological processes such as immune response, autophagy, ferroptosis, stem cell characteristics, transition from epithelial to mesenchymal cells, signaling through toll-like receptors, hypoxia, glycolysis, synthesis of lipids, aging, WNT pathway, and drug sensitivity. In all cohorts, univariate Cox regression analyses were conducted for each signature, and the results showed that only our model exhibited a statistically significant association with prognosis, indicating the enduring stability of the risk score (Fig. 6A). Moreover, we compared the C-index of the risk score with other signatures and found that the risk score performed better than almost all models in all datasets (Fig. 6B). By applying two machine learning algorithms, we were able to reduce the signature’s dimensionality, thereby improving its potential for extrapolation.

Fig. 6: Comparison between the CD163+FPR3+ macrophage-associated signature and other models in glioma.
figure 6

A Univariate Cox regression analysis of risk score and 100 published signatures in TCGA, CGGA-693, CGGA-301, CGGA-325, GSE16011, and Rembrandt cohorts. B C-index analysis risk score and 100 published signatures in TCGA, CGGA-693, CGGA-301, CGGA-325, GSE16011, and Rembrandt cohorts. The P values are labeled above each boxplot with asterisks (-, no significant, “*,**,***” respectively means P < 0.05, P < 0.01, P < 0.001).

Full size image

Discussion

The field of glioma treatment has experienced diversification, thanks to advancements in molecular biology and immunology, most notably the introduction of immune checkpoint inhibitor (ICI) therapy. The presence of multiple treatment options emphasizes the need for improved individualized evaluation techniques to inform clinical decision-making. However, a persistent constraint exists due to the lack of reliable biomarkers capable of identifying patients with “cold” glioma who could potentially benefit from ICI therapy. Nevertheless, this absence of dependable biomarkers for identifying “cold” glioma patients who might benefit from ICI therapy remains a current limitation. Because of the complex and diverse tumor-immune microenvironment, although ICIs that target the immune regulators PD-L1/PD-1 result in clinical advantages, only a portion of patients achieve a sustained response to immunotherapy24,25. The utilization of scRNA-seq analysis in glioma studies has contributed to the discovery of previously unknown cell subgroups that have clinical importance, demonstrated by their distinct gene expression profiles26. In the glioma context, scRNA-seq analysis has uncovered the diverse characteristics of myeloid cells associated with glioma and has begun to clarify the dynamic interaction between different subgroups of glioma-associated myeloid cells27. This heterogeneity is considered to be the underlying cause of the ineffectiveness of targeted therapy and chemotherapy, ultimately resulting in the inevitable recurrence of GBM28.

In this study, FPR3 was mostly associated with M2-like macrophage by WGCNA module and ssGSEA algorithm. By employing scRNA-seq examination and immunofluorescence, we have identified the co-localization of CD163 with FPR3 (Fig. 1H). Moreover, the correlation assessments of FPR3 and CD163, along with the M2-like TAMs score in TCGA, CGGA, Rembrandt, and GSE16011 databases, indicated a robust and affirmative correlation (Fig. 1F, G and Supplementary Fig. 3). The data collectively showed that FPR3 was predominantly expressed in M2-like TAMs. Based on a previous study, compelling evidence suggests that FPR3 exerts a regulatory role in the inflammatory process and is detectable in monocyte-derived macrophages, mature dendritic cells, and tissue-specific macrophages13. FPR3 has been reported to be associated with macrophages in various diseases, including IgA nephropathy, diffuse large B cell lymphoma29, and breast cancer30. There is still a limited knowledge regarding the expression pattern and prognostic importance of FPR3 in glioma. In contrast, our research presents proof suggesting a significant increase in FPR3 expression in glioma and a correlation between its excessive expression and adverse prognosis in individuals with glioma (Fig. 2).

In addition, our examination uncovered that gliomas distinguished by increased levels of FPR3 consistently exhibited greater immune and stromal scores in comparison to those with decreased levels of FPR3 in four groups (P < 0.05, Supplementary Fig. 5B). The aforementioned observation implies that FPR3 might have a function in controlling immune and stromal cells in the tumor microenvironment. Furthermore, the high-FPR3 group demonstrated increased expression of markers for immune cells in comparison to the low-FPR3 group (Supplementary Fig. 5D). Significantly, our results demonstrate a distinct expression pattern between FPR3 and various immune checkpoints (Supplementary Fig. 5E). Moreover, genetic anomalies function as conventional indicators for evaluating the effectiveness of anti-PD-1/PD-L1 therapy23. Through our investigation, we discovered a notable increase in immunosuppressive cancer-causing pathways in gliomas that displayed heightened FPR3 expression (P < 0.05, Supplementary Fig. 5G). Significantly, FPR3 exhibited a robust association with the cancer immunity cycles, which could potentially result in reduced infiltration of effector TIICs in the microenvironment. Interestingly, the group with low FPR3 demonstrated a decrease in the ability of T cells to recognize cancer cells. The glioma tissues were categorized into two different immune phenotypes, non-inflamed and inflamed, based on the spatial arrangement of CD8+ T cells (Fig. 3D). Significantly, the non-inflamed tumor tissues showed a notable increase in FPR3 expression compared to the inflamed phenotype (P < 0.05, Fig. 3E). Within our internal groups, the gliomas in the high-expression FPR3 category exhibited a greater abundance of PD-L1 in comparison to the low-expression FPR3 category (P < 0.05, Fig. 3F). Therefore, the involvement of FRP3 in the control of immune cells and immune response was discovered. FPR3 is present in different types of cells, such as macrophages, which play a role in innate immunity and inflammation31. FPR3 in monocytes, macrophages, and dendritic cells and may regulate macrophage chemotaxis13,14,15,32, but its function in macrophages remains unclear. To obtain a thorough comprehension of the diversity among tissue macrophages, it is imperative to investigate the fundamental mechanisms that regulate this occurrence. Simplified models of macrophage activation fail to fully account for the observed heterogeneity in vivo33. To gain a comprehensive comprehension of the cellular and molecular heterogeneity of myeloid cells within gliomas, we undertook the isolation of myeloid cells and performed clustering analysis de novo. Consequently, we successfully discerned the presence of eleven distinct myeloid clusters (Fig. 3A). The UMAP plot revealed that FPR3 was not expressed in some of the CD163+ macrophage (Fig. 3B). We divided macrophage with nonzero expression of CD163 into CD163+FPR3+ and CD163+FPR3− macrophages.

Moreover, this study devised a scoring mechanism that employs distinctive genes from different subgroups to forecast the efficacy of immunotherapy in glioma patients. A combination of the RSF and CoxBoost algorithms was found to be the most optimal model, demonstrating the highest average C-index of 0.750. Significantly, this particular amalgamation model consistently surpassed alternative models in every validation dataset (Fig. 4A). The CoxBoost algorithm specifically identified 17 genes that were considered highly valuable (Fig. 4B, C). Additionally, the RSF algorithm efficiently filtered and selected the most reliable model. By utilizing integrative approaches, one can take advantage of employing various machine learning algorithms and their combinations to create a model that achieves a collective performance in forecasting the prognosis of glioma. Additionally, the combination of algorithms can effectively reduce the number of variables, resulting in a more simplified and easily applicable model. In TCGA database, patients categorized as high-risk demonstrated notably worse OS in comparison to those categorized as low-risk (P < 0.05, Fig. 4E). The same pattern was consistently seen in the remaining five validation datasets (P < 0.05, Fig. 4F–J). Significantly, our signature demonstrated independent functionality from these factors and exhibited notably superior prognostic performance as assessed by the C-index (Fig. 5A, B and Supplementary Tables 2, 3). In addition, we chose 100 published signatures at random that cover a wide range of functional gene combinations (Fig. 6A, B). Out of all these signatures, only a few have been incorporated into clinical practice, and very few have undergone thorough validation. For example, the analysis of univariate Cox regression showed that the risk score maintained prognostic importance consistently in all cohorts (Fig. 6A). Upon comparing the predictive superiority of these signatures, it was noted that the risk score demonstrated superior performance across all datasets in comparison to almost all models (Fig. 6B). It was noted that while the majority of models demonstrated satisfactory performance in their respective training datasets as well as certain external datasets, their performance was notably subpar in other datasets. The difference could be ascribed to the restricted applicability of models caused by overfitting. In contrast, our signature underwent dimension reduction through two machine learning algorithms, enhancing its potential for extrapolation. Hence, our distinctive characteristic holds potential as a viable substitute for assessing the prognosis of glioma in clinical environments.

Despite the promising clinical significance of the risk score in glioma, there exist certain limitations that need to be addressed. To validate our findings, conducting larger preclinical studies and retrospective clinical trial analyses is crucial. Second, the unavoidable intratumor or interpatient heterogeneity resulting from the utilization of glioma cohorts from various public datasets must be acknowledged.

Taken together, we have successfully identified distinct groups of macrophages characterized by the concurrent presence of CD163 and FPR3. By harnessing machine learning algorithms, we have constructed a prognostic model, denoted as the CD163+FPR3+ macrophage-related risk score, which displays the capacity to forecast both prognosis and treatment response in individuals suffering from gliomas. As a result, this model demonstrates potential as a valuable instrument in the realm of glioma research.

Methods

Collecting and analyzing data

Transcriptome data and clinical information were obtained from six publicly available datasets, namely TCGA, GSE16011, CCGA-693, CCGA-301, CCGA-325, and Rembrandt datasets. The processing and collection of these data are consistent with prior literature34.

The human specimens

The internal cohort included 144 glioma tissues and 14 tumor tissues, along with their corresponding para-tumor tissues (referred to as in-house cohort 1). The in-house cohort 2 consisted of 122 glioma tissues, 30 tumor tissues, and their corresponding para-tumor tissues. Normal brain tissues were obtained from individuals with traumatic brain injury. Those fresh samples were briefly preserved in liquid nitrogen, fixed with paraformaldehyde, and then paraffin embedded to make tissue chips and more details can be found in Supplementary Table 4.

Single-cell RNA sequencing analysis

ScRNA-seq data were obtained from Abdelfattah et al.27 (GSE182109, GEO—), and the analysis was performed using the Seurat R package (version 4.0.4). Quality control and clustering were conducted based on previous literature27, and the cells were annotated. After data preprocessing, normalization, and batch correction, myeloid cells were individually segregated based on CD163+FPR3+ and CD163+FPR3− expression. Signature genes were selected according to strict criteria, including a log2FC with an absolute value exceeding 0.25 and a P value below 0.05.

Immunohistochemistry (IHC) and immunofluorescence (IF)

A TMA comprised 78 glioma tissues and 6 normal brain tissues. IHC was performed using a primary antibody targeting FPR3 (Proteintech, China). The evaluation of IHC analysis followed a previously described approach34. Additionally, based on the localization of CD8+ T cells25, tumors were classified into three phenotypes: “the immune-desert phenotype, the immune–excluded phenotype and the inflamed phenotype, among those, the immune-desert phenotype, the immune–excluded phenotype”. The IF protocol was used for CD8 staining, with CD8 antibodies similar to those used in IHC. The corresponding secondary antibody, CY5-TSA, was employed for CD8 detection. Nuclei were stained with blue DAPI. The scoring was conducted in a double-blind manner by two pathologists.

Quantitative real-time polymerase chain reaction (RT-qPCR)

RNA extraction was performed using TRIzol reagent (Life Technologies), followed by cDNA synthesis using a Reverse Transcription Kit (Thermo Fisher Scientific). The relative expression of FPR3 was determined using the 2−ΔΔCt method, with β-actin as the reference gene. The primer sequences are provided below:

FPR3 (F) 5′- TGACACTGCTGTAGAGAGGTTGAAC-3′,

FPR3 (R) 5′- TGATGGACATAGGCACGCTGAAG-3′,

β-actin (F) 5′-ACAGAGCCTCGCCTTTGCCGAT-3′,

β-actin (R) 5′- CTTGCACATGCCGGAGCCGTT-3′.

Weighted gene co-expression network analysis (WGCNA)

The construction of mRNA co-expression networks was performed using the R WGCNA package35 on the TCGA dataset. Initially, the Pearson correlation coefficient was computed between each pair of genes to derive a similarity matrix. Subsequently, WGCNA transformed the similarity matrix into an adjacency matrix using a power function. From all the soft thresholds (β) with an R2 value exceeding 0.9, we chose the automatic value β (β = 5) generated by the WGCNA pickSoftThreshold function. Following the WGCNA guidelines, we selected a network merge height of 0.25. The remaining WGCNA parameters were set to their default values.

Survival analysis

The association between FPR3 expression and the clinical outcomes of patients across 33 cancers was illustrated using forest plots (OS: overall survival; DSS: disease-specific survival; and PFI: progression-free interval). The patients were divided into two groups: the high-FPR3 group and the low-FPR3 group. Next, the Kaplan–Meier method was used to investigate the impact of FPR3 on OS in glioma patients from the six cohorts. Univariate survival analysis was conducted to calculate the hazard ratio and determine the 95% confidence intervals.

Gene set enrichment analysis (GSEA)

GSEA was performed using GSEA software (version 3.0) obtained from the Broad Institute (http://www.broadinstitute.org/gsea) to explore the underlying biological processes of the two groups based on DEGs (|log2FC| ≥ 1, q value (FDR) < 0.05). Significant findings were determined by considering a P value threshold of 0.05 and a false discovery rate (FDR, q val) below 0.25.

Immune and stromal cell infiltration analysis

We used the ESTIMATE R package to assess the immune and stromal scores, which indicate the prevalence of immune cells and stromal cells36. In each glioma sample, the CIBERSORT algorithm was employed to decipher bulk RNA-seq profiles into 22 scores representing different cell types37. To construct the glioma cell landscape, ssGSEA was used to assess the presence of different subsets of immune and stromal cells38.

Artificial intelligence-derived scoring system

In the six public datasets, a total of 142 genes underwent univariate Cox regression analysis. Out of these, 62 genes showed a significant association with prognosis, meeting the criteria of P < 0.05. To build a stable prognostic scoring system, 10 machine learning algorithms39, including RSF, CoxBoost, and SVM, were integrated into 101 algorithm combinations. We used the public cohorts to construct prognostic signatures. The risk score was calculated for the six cohorts using the signature obtained from the TCGA cohort. Additionally, we evaluated the prognostic performance of an artificial intelligence-based prognostic signature by comparing it with 100 previously published signatures.

Ethics approval

The glioma samples were collected at Xiangya Hospital, Central South University between 2015 and 2022, and approved by the Ethics Committee.

Statistical analysis

Paired or unpaired Student’s t-test was used to analyze the two-group comparisons. Correlation analysis was performed by Spearman correlation coefficient test. R software (www.r-project.org) was used for all statistical analyses. Statistical significance was determined for P values of less than 0.05, 0.01, and 0.001.