2.1. Datasets Acquisition and Processing

We followed the methods of Dianqian Wang and Lipeng Pei et al. 2023 [16, 17]. Figure 1 illustrates the workflow of our study. Bulk RNA-seq data and genomic mutation data for ovarian cancer (n = 375) were obtained from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). Two additional GEO datasets, GSE14764 (n = 80) and GSE140082 (n = 380), were utilized as validation sets. IMvigor210, a Phase 2 study targeting PD-L1/PD-1 in metastatic urothelial carcinoma, enrolled 298 patients received one or more doses of atezolizumab (68 responders and 230 nonresponders) [18]. GSE91061 contained 58 RNA-Seq samples of advanced melanoma received ipilimumab (CA209-038 study, 16 responders and 42 nonresponders) [19]. In these studies, responders were defined as patients achieving a complete response (CR) or partial response (PR), while nonresponders were classified as those with progressive disease (PD) or stable disease (SD). IMvigor210 and GSE91061 were used to explore the predictive value of MRS in predicting immunotherapy benefits.

2.2. CIBERSORT and Weighted Gene Co-Expression Network Analysis (WGCNA)

The CIBERSORT algorithm was employed to estimate the abundance of cell types and gene expression from the RNA-seq data of the TCGA ovarian cancer dataset [20]. Following the quantification of M2 macrophage abundance, we conducted WGCNA to identify gene sets associated with M2 macrophages. WGCNA is a robust method for pinpointing gene sets of interest from large and diverse gene pools, facilitating correlation analysis with specific phenotypes [21]. In this analysis, the β value was set to 0.9 during network construction to minimize weak correlations between genes in the adjacency matrix. To identify the most representative genes, module eigengenes (MEs), we used a Pearson’s correlation coefficient threshold of 0.25. Genes within modules that showed significant positive correlations (Cor > 0.3, p < 0.001) with M2 macrophages were designated as M2 macrophage-related genes. This approach ensured the selection of gene sets strongly associated with M2 macrophages, providing a solid foundation for further analysis.

2.3. Machine Learning-Based MRS

2.4. Evaluation of the Performance of MRS

Ovarian cancer cases were divided into two groups—high-risk and low-risk—based on the median value of the risk score. Using the “survival” package, we generated overall survival (OS) curves. To assess the predictive value of the MRS for clinical outcomes in ovarian cancer, we constructed time-dependent ROC curves and clinical ROC curves using the “timeROC” package. We collected a total of 55 prognostic models, including both mRNA and lncRNA-related models, and calculated their C-indexes using the “CompareC” package. Univariate and multivariate Cox regression analyses were performed to identify significant prognostic risk factors in ovarian cancer. Additionally, we developed a nomogram based on the MRS and clinical characteristics using the R packages “rms” “nomogramEx” and “regplot” This nomogram integrates these factors to provide a comprehensive prediction tool for ovarian cancer prognosis.

2.5. TME and Genetic Mutation Landscape

The ESTIMATE algorithm was employed to calculate the immune microenvironment score for each ovarian cancer case. The infiltration levels of immune cells were quantified using the “immunedeconv” R package, which estimates immune cell fractions from bulk RNA-sequencing data through seven computational methods: CIBERSORT, MCPcounter, QUANTISEQ, XCELL, CIBERSORT-ABS, TIMER, and EPIC [27]. To investigate the biological functions associated with the high-risk and low-risk groups, we conducted Gene Set Enrichment Analysis (GSEA). Additionally, we utilized the “maftools” package to generate a waterfall plot of single nucleotide variants (SNVs).

2.6. Evaluation of the Performance of MRS in Predicting Drug Sensitivity

The immunophenoscore (IPS) and TIDE score for each ovarian cancer patient were retrieved from The Cancer Immunome Atlas (TCIA, https://tcia.at/home) and the TIDE platform (https://tide.dfci.harvard.edu), respectively. Using the gene expression data and corresponding coefficients of the MRS, we calculated the risk score for each patient in the IMvigor210 and GSE91061 cohorts. This allowed us to investigate the correlation between the risk score and immunotherapy response. To further analyze drug sensitivity, we utilized data from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/). We then calculated the half-maximal inhibitory concentration (IC50) values for common chemotherapy and targeted drugs using the “oncoPredict” package.

2.7. Statistical Analysis

Statistical analyses were conducted using R software (version 4.2.1). For comparing continuous variables, we employed the Wilcoxon rank-sum test or Student’s t-test, as appropriate. The correlations between two continuous variables were assessed using Spearman’s rank correlation analysis. To evaluate differences in Kaplan–Meier survival curves, we applied the two-sided log-rank test. Nomogram was tested by proportional hazard assumption.

3.1. WGCNA Identified Macrophage M2-Related Genes in Ovarian Cancer

We determined the optimal soft-threshold power to be 5 by setting the β value for the degree of independence at 0.9 (Figure 2(a)). The clustering tree of the 80 differentially colored modules is presented in Figure 2(b). Supporting Figure 1 showed the association between modules and immune cells. In order to find more macrophage M2-related genes, we selected the top two most significant modules for further analysis, including brown module (correlation = 0.34, p = 2e − 11) and salmon (correlation = 0.43, p = 1e − 18) module. The correlation coefficient between gene significance (GS) and module membership (MM) was 0.55 for the brown module and 0.77 for the salmon module, respectively (Figures 2(c) and 2(d)). Both the brown and salmon modules collectively encompassed 1582 macrophage-related genes (Supporting Table 1).

3.2. Integrative Machine Learning Algorithms Constructed a Prognostic MRS

The 1582 macrophage-related genes were subjected to univariate Cox analysis, yielding 34 potential prognostic biomarkers (Figure 3(a), p < 0.05). These 34 biomarkers were then processed through a machine learning-based integrative procedure to develop an accurate and stable prognostic MRS. This resulted in the generation of 52 prognostic models (Figure 3(b)). Figure 3(b) displays the C-index of each model across the TCGA, GSE14764, and GSE140082 cohorts. The Lasso method-produced model was selected as the optimal MRS due to its highest average C-index of 0.68 (Figure 3(b)). In the Lasso regression, the optimal λ was determined when the partial likelihood deviance reached its minimum value using the LOOCV framework (Figures 3(c) and 3(d)). Ultimately, 14 macrophage-related genes were included in the MRS based on the Lasso method (Figures 3(c) and 3(d)). The coefficient of each candidate gene in MRS are shown Figure 3(e). Based on the coefficients of each gene and their expression pattern, the risk score of each ovarian cancer case was calculated as follows: risk score = (−0.0371) × PLA2G2D^exp  +  (−0.0381)  ×  SOS1-IT1^exp + 0.19 × HMGN3^exp  + (−0.0033) × RAP1A^exp + (0.0127) × ARRDC2^exp + (−0.0093) × BCCIP^exp + 0.0060 × C5AR1^exp + (−0.0074)  ×  AC018645.3^exp + 0.0006  × VASP^exp + 0.0838 × TNFAIP8L3^exp + (−0.0063) × TRBV28^exp  +  (−0.0051)  ×  WARS1^exp  +  (−0.0042) × C2^exp + (0.0020) × CD163^exp. Ovarian cancer patients were divided into two groups—high-risk and low-risk—using the median value of the risk score as the cutoff.

3.3. Evaluation of the Performance of MRS

High risk score indicated lower OS rate in ovarian cancer patients and the AUCs of 2-, 3-, and 4-year ROC curve were 0.766, 0.727 and 0.721 in TCGA cohort (Figure 4(a)), 0.745, 0.717 and 0.826 in GSE14764 cohort (Figure 4(b)), and 0.692, 0.722 and 0.919 in GSE140082 cohort (Figure 4(c)), respectively. The C-index and AUC value of the risk score regarding the MRS in predicting OS rate was higher than that of grade and stage in the TCGA, GSE14764, and GSE140082 cohort, as Figures 4(d), 4(e), 4(f) illustrates. We next gathered 55 prognostic signatures (Supporting Table 2) and determined their C-index because numerous prognostic signatures had been produced for ovarian cancer. Interestingly, our MRS’s C-index outperformed all of these prognostic indicators for ovarian cancer (Figure 5(a)), indicating that it performed well in predicting the OS of ovarian cancer patients. In the TCGA, GSE14764, and GSE140082 cohort, univariate and multivariate Cox regression analysis similarly indicated that the MRS-based risk score was an independent risk factor for ovarian cancer (Figures 5(b) and 5(c)). Next, taking into account MRS, stage, and grade, we created a survival prediction nomogram (Figure 5(d)). According to the findings, the calibration curves had a rather good predictive value for the OS rate over the next one, three, and 5 years (Figure 5(e)).

3.4. The Correlation Between MRS and TME in Ovarian Cancer

As illustrated in Figure 6(a), the IFN-γ dominant subtype (C2) was more prevalent in the low-risk group compared to the high-risk group, while the lymphocyte-depleted subtype (C4) was less common in the low-risk group (p = 0.002). Patients with low risk score exhibited higher ESTIMATE scores, immune scores, and stromal scores in ovarian cancer (Figure 6(b), all p < 0.05). Figure 6(c) highlights the significant correlation between the risk score and immune cell infiltration (p < 0.05). Further analysis revealed that patients with high risk score had elevated cell proliferation scores (Figure 6(d)) and a higher proportion of M2/M1 macrophages in the TCGA, GSE14764, and GSE140082 cohorts (Figures 6(e), 6(f), 6(g)). These findings suggest a greater likelihood of tumor progression in high-risk ovarian cancer patients.

3.5. MRS-Based Treatment Strategy for Ovarian Cancer

As shown in Figures 7(a) and 7(b), patients with low risk score exhibited higher expression levels of immune checkpoint and HLA-related genes in ovarian cancer (p < 0.05). Further analysis revealed that low-risk cases had significantly higher TMB scores, as well as elevated CTLA4 IPS, PD1 IPS, and combined CTLA4/PD1 IPS scores (Figures 7(c) and 7(d), p < 0.05). Conversely, high-risk patients were associated with higher immune escape scores, TIDE scores, and increased T cell dysfunction and exclusion (Figures 7(e) and 7(f), p < 0.05). These findings suggest that low-risk ovarian cancer patients may have a better response to immunotherapy. In the IMvigor210 cohort, nonresponders had significantly higher risk scores compared to responders (Figure 8(a), p < 0.001), with an AUC of 0.755 (Figure 8(b)). Additionally, a high risk score was linked to lower OS rates (Figure 8(a), p < 0.001). In the GSE91061 dataset, responders had significantly lower risk scores compared to nonresponders (Figure 8(b), p < 0.05), with an AUC of 0.780 (Figure 8(b)). High-risk scores were also associated with poorer OS rates (Figure 8(b), p = 0.0028). We further evaluated the differences in IC50 values of common drugs between different risk groups. As illustrated in Figures 9(a) and 9(b), low-risk ovarian cancer patients showed lower IC50 values for several chemotherapeutic agents, including 5-Fluorouracil, Cisplatin, Cyclophosphamide, Docetaxel, Paclitaxel, Crizotinib, Foretinib, KRAS Inhibitor-12, Lapatinib, and Nilotinib (all p < 0.05). This suggests that low-risk patients may be more sensitive to these treatments.