Multicellular gene network analysis identifies a macrophage-related gene signature predictive of therapeutic response and prognosis of gliomas

RNA-seq data (FPKM) of TCs and TAMs in mice gliomas were downloaded from the Gene Expression Omnibus (GEO) website (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​) under accession number GSE69104. The preclinical experiment [10] investigated the role of TAMs in glioma immunotherapy with BLZ945, a CSF1R inhibitor, where all TC-TAM paired samples were divided into 3 groups, i.e., Vehicle (Veh, 5 samples), Endpoint (EP, 6 samples, i.e., drug-sensitive), and Rebound (Reb, 4 samples, i.e., drug-resistant) tumors (Additional file 1: Table S1).

As described in detail in Additional file 1: Text S1, the expression level of each gene in the drug-resistant and drug-sensitive samples were normalized to their mean values in the vehicle samples. We selected significantly differentially expressed genes (DEGs) between the drug-resistant samples (Reb) and the drug-sensitive samples (Ep) for both TCs and TAMs by calculating the fold changes (FCs) and false discovery rate (FDR)-adjusted p values using t test. A gene with |FC| > 1.5 and adjusted p value less than 0.05 was considered as a DEG, resulting in a list of 1141 DEGs.

Based on the work of one of our authors [18], we developed a network perturbation analysis technique for multicellular gene networks. As described in detail in Text S1, first we used the reference samples (N = 6) in the Ep group to construct a sensitive network (Fig. 1a) for pairs of correlated DEGs in TCs and TAMs (|PCC| > 0.95 and p value < 0.05). We then added each single drug-resistant sample (Reb_i, i = 1, 2, 3, 4) to the reference samples in the Ep group to construct 4 sets of sample-specific perturbed samples, respectively (Additional file 1: Table S1). We used each set of perturbed samples (N = 7) to construct perturbation networks (Fig. 1b). The addition of each single Reb sample is the cause of differences between the sensitive and perturbation networks. If the gene expression profile in the added sample, Reb_i, was similar to that in the Ep samples, the perturbation of the PCC was negligible. However, if some gene expression levels were remarkably different between the single Reb sample and the Ep samples, significant changes in the PCCs of certain gene pairs were induced upon the addition of Reb_i to the reference samples.

Based on such network perturbation analysis (Fig. 1c), the robust significantly different PCCs (ΔPCC, see Text S1) of gene pairs were chosen to construct a differential network. Specifically, we selected significantly different edges, represented by ΔPCC, by setting a threshold of |ΔPCC| > 0.05 for each perturbation network versus the sensitive network. If a gene-pair was differentially correlated in at least 3 perturbation networks compared to the sensitive network, this edge was selected as a robust differential edge. We defined three types of edges in the differential network: correlation-gained edges (ΔPCC > 0), correlation-lost edges (ΔPCC < 0), and correlation-invariant edges (ΔPCC = 0). The differential network reflected the changes in both intracellular and intercellular edges in the multicellular gene networks of mice with distinct therapeutic responses to the CSF1R inhibition.

The functional enrichment of genes in the differential network was analyzed using the newest version of Metascape (http://​metascape.​org), a gene annotation and enrichment analysis database that integrates ontology sources, including the KEGG Pathway, GO Biological Processes, Reactome Gene Sets, Canonical Pathways and CORUM databases. The whole genome was used as the enrichment background. Terms of pathways or processes with a p value < 0.01 (accumulative hypergeometric test), a minimum of 3 genes, and an enrichment factor > 1.5 were collected and grouped into clusters based on their membership similarities.

The differential network might capture robust topological differences between the CSF1R inhibitor-sensitive network and inhibitor-resistant networks and reflect potential changes in the gene interactions across TCs and TAMs during the acquisition and development of drug resistance. It is therefore reasonable to speculate that the genes in the differential network play critical roles in promoting tumor growth, even under drug pressure. We hypothesized that the expression levels of genes in the differential network are associated with the survival outcomes of glioma patients. As such, we developed a differential network-based signature identification method to select prognostic biomarkers for glioma patients. We collected the clinical information and RNA-seq gene expression data from glioma patients in the CGGA database (http://​www.​cgga.​org.​cn/​) and TCGA database (https://​cancergenome.​nih.​gov/​). The names of genes in the differential network were mapped to those of genes in homo sapiens, resulting in a list of 29 candidate genes. By matching both patient sample IDs and gene names from the clinical information and the gene expression data, a learning set of 310 samples from CGGA dataset and an independent validation set of 690 samples from TCGA dataset were created for prognostic signature identification, validation and further analysis. The details of the sample information were listed in Additional file 2: Table S2.

We used a multivariate COX proportional hazards (PH) model [19] and LASSO regression method [20] to select prognostic genes from the differential gene association network (see details in Text S1). A tenfold cross-validation was performed to select the optimal values of the tuning parameter for minimizing the mean cross-validation error.

The regression coefficients at the optimal tuning parameter were computed as risk coefficients, which were used to formularize a risk signature. The same risk signature was used to compute the risk scores (RSs) for patients in the independent validation set. The patients in each dataset were classified into a high-risk group and a low-risk group according to the optimal cutoff value using the ROC method. The Kaplan–Meier (K–M) curves for patients in the high- and low-risk groups were analyzed, and the statistical significance of the difference was assessed using the two-sided log-rank test. Time-dependent ROC analysis [21] was further conducted to evaluate the prognostic accuracy of the above RSs with respect to the 3- and 5-year survival predictions of patients in both the learning and independent validation sets.

The survival and gene expression data from glioma patients who received targeted therapies in the independent set were extracted to evaluate the predictive effectiveness of the macrophage-related gene signature for classifying patients into drug-sensitive and drug-resistant groups. We used the observed 3- or 5-year survival status to substitute for the latent drug-resistance status in these patients. Specifically, the 3- or 5-year survival status (alive or dead) was defined as the outcome (sensitive or resistant) of targeted drug treatment. The above risk signature was used to classify each patient into a sensitive group (i.e., low-risk group) or a resistant group (i.e., high-risk group) according to the optimal cutoff value of the RS using the ROC method. To further compare the powers of different gene signatures to predict responses to targeted therapy, we calculated the area under the curves (AUCs) of the time-dependent ROCs to assess their accuracies.

We compared the macrophage-related gene signature produced from the above multicellular gene network perturbation method with other commonly used methods, including LASSO Cox regression model [22] and correlation network-based biomarker identification method without network perturbation. First, a LASSO Cox signature was trained from the full list of 1141 DEGs based on the CGGA set. In addition, a weighted correlation network (without perturbation) was constructed by calculating the pairwise Person correlation coefficients (PCCs) across all pairs of DEGs within both TCs and TAMs for the full set of Veh, Reb and Ep samples (30 samples). Based on the ranking of node strength score (defined as the sum of the absolute values of correlation coefficients of the node with other nodes), we selected the top ranked 12 DEGs as a gene set for defining a prognostic signature using the CGGA dataset.

Moreover, to assess whether the macrophage-related gene signature was independently correlated with the prognosis of glioma patients, we conducted univariate and multivariate COX regression analyses of clinicopathological factors and available gene signatures. Clinicopathological information, including age, gender and grade, was available for glioma patients in both the learning and validation sets. We also included the following gene signatures in the multivariate COX regression analyses: signature 1—the macrophage-related gene signature newly proposed in this study; signature 2—an EGFR gene signature studied by many groups [15, 16]; and an immune-related gene signature for predicting the prognosis of glioma patients, i.e., signature 3—the Cheng et al. signature [17].

The above risk factors were further extracted for construction of a combined signature using the LASSO COX model [19, 20]. As a result, we defined the combined signature as follows: CS = (0.008974621 × Age) + (1.617859481 × Grade) + (0.940077644 × Signature_1) + (0.006408624 × Signature_3). Here, Grade = 1 for lower grade glioma, and Grade = 2 for high grade glioma.

The heatmaps of DEGs in TAMs (Additional file 1: Fig S1A) and TCs (Additional file 1: Fig S1B) across all samples demonstrated that the gene expression profile of the Reb samples was significantly different from that of Ep samples. The constructed sensitive network and sample-specific perturbation networks along with their topological attributes are shown in Additional file 1: Fig S2 and Fig S3–S6, respectively. A significant difference in the number of nodes and edges was observed between the sensitive and perturbation networks (Additional file 1: Fig S7). Other network topological attributes, including the network diameter, network centralization, and average number of neighbors, of the perturbation networks (Additional file 1: Fig S3–S6) were mainly higher than those of the sensitive network (Additional file 1: Fig S2). These results suggest that the perturbation networks had more complexity and that network rewiring might emerge during the acquisition and development of resistance to CSF1R inhibition.

A robust differential network (Fig. 2a) containing 42 nodes and 30 edges was thus constructed. The genes in the differential network are listed in Additional file 1: Table S3. Interestingly, all edges in the differential network were correlation-gained edges, indicating that the addition of Reb samples improved the correlation coefficients, which is in accordance with the above results (Additional file 1: Fig S2–S7).

Functional enrichment analysis of genes in the differential network revealed that some pathways, including the chemokine-mediated signaling and receptor-type tyrosine-protein phosphatase pathways as well as the cell-cell adhesion via plasma membrane adhesion molecule pathway, were significantly enriched during the acquired resistance to CSF1R inhibition treatment (Fig. 2b). This result suggests that cell–cell interactions between TCs and TAMs might contribute to the resistance of glioma to CSF1R inhibition, which was supported by the previous experimental results [10, 23‐25].

We selected prognostic genes from the differential network using the LASSO regression method (see “Methods” section and Text S1). Figure 3a shows the selection of optimal tuning parameter (λ) of LASSO regression based on tenfold cross-validation of the learning set. Figure 3b shows the LASSO coefficient profiles of the 29 genes in the differential network. Each curve corresponds to evolution of the coefficient of each gene with respect to the change of the tuning parameters during the LASSO regression. Figure 3c lists the 12 genes selected under the optimal tuning parameter of LASSO. Five genes were located in macrophages (MФ) (i.e., ANPEP, DPP4, PRRG1, GPNMB, TMEM26), and 7 genes were located in TCs (i.e., PXDN, CDH6, SCN3A, SEMA6B, CCDC37, FANCA, NETO2). The coefficients of each gene in the COX PH model and the corresponding hazard ratios (HRs) were also listed and used to define a macrophage-related gene signature for predicting the prognosis of glioma patients. Accordingly, we formulated the following RS for each patient based on the expression levels of the selected genes: RS = (0.001695826 × ANPEP) + (0.001351164 × DPP4) + (0.828492221 × PRRG1) + (0.002693736 × GPNMB) + (0.572250065 × TMEM26) + (0.011112329 × PXDN) + (0.000861924 × CDH6) − (0.877296902 × SCN3A) − (0.042307865 × SEMA6B) + (0.019673956 × CCDC37) + (1.184362541 × FANCA) + (0.101032334 × NETO2). Furthermore, the univariate COX regression analysis (Additional file 1: Table S4) demonstrated that each signature gene was significantly associated with the survival of glioma patients.

We then investigated the prognostic significance and accuracy of the macrophage-related gene signature in both the learning and independent validation sets. Figure 4a shows the K–M curves for high-risk (blue) and low-risk (red) glioma patients in the learning set; significant differences were assessed with the log-rank test (p value less than 0.0001). The high-risk group of glioma patients tended to have a shorter survival time than the low-risk group. Figure 4b shows the prognostic accuracy of the risk signature evaluated by the AUCs of the time-dependent ROCs with respect to the 3- and 5-year survival rates of glioma patients in the learning set (AUC at 3 year: 0.92; AUC at 5 years: 0.891).

Moreover, we tested the prognostic significance of the macrophage-related gene signature using the independent validation set (Fig. 4c). The K–M curves demonstrated a significant difference between the survival rates of high- and low-risk patients (log-rank test p value less than 0.0001). Figure 4d shows the prognostic accuracy of the macrophage-related gene signature validated by the AUCs of the time-dependent ROCs with respect to 3- and 5-year survival rates of glioma patients in the validation set (AUC at 3 years: 0.818; AUC at 5 years: 0.836). These results demonstrated good prognostic value of the macrophage-related gene signature in glioma patients.

Furthermore, we tested the significance of prognostic accuracy of the macrophage-related gene signature evaluated using a bootstrapping approach. We randomly selected 12-gene sets from the whole transcriptome for 1000 times and compared their prognostic accuracy with that of the macrophage-related gene signature. CGGA dataset was used for training and TCGA dataset for validation. Figure 5a shows ROC curve of the macrophage-related gene signature (red) against ROC curves of bootstrapped 12-gene signatures (blue, only 100 curves were randomly shown). Figure 5b shows the probability distributions of the AUC values of 1000 sets of random 12-gene signatures. The p value (0.001) from the permutation test justified the statistical significance of the prognostic accuracy of the macrophage-related gene signature.

For comparison, we also generated gene signatures using LASSO Cox regression model and correlation network-based method (see Method section). Additional file 1: Fig S8 shows the prognostic accuracies of LASSO Cox signature and correlation network-based signature with respect to 3-year and 5-year survival prediction. Although these two signatures performed well on the learning set, their predictive accuracies on the test sets were less than the macrophage-related gene signature from the network perturbation analysis. We further employed a bootstrapping approach to test and compare the robustness of these gene signatures derived from different methods. Figure 6 shows the AUC values of ROCs of these signatures with respect to overall survival, 3-year survival and 5-year survival. The macrophage-related gene signature exhibited better accuracy and robustness than the LASSO Cox signature and the correlation-network-based signature.

Considering the important roles of macrophages in glioma progression and drug resistance [10‐12, 26], a subset of TCGA patients who received targeted therapy were used to test the predictive power of the above macrophage-related gene signature for predicting drug sensitivity or resistance in glioma patients. The drug-resistance status of these patients was latent, we thus used the associated survival profiles as surrogate. The 3- or 5-year survival status (alive or dead) of each patient after the targeted therapy was used to evaluate the treatment outcome of the molecularly targeted therapeutics (sensitive or resistant). Each patient was predicted to be drug-sensitive (i.e., low-risk) or drug-resistant (i.e., high-risk) according to the optimal cutoff value of the RS using the time-dependent ROC method evaluated at 3-year or 5-year. Figure 7a, b shows the K–M survival curves of the predicted drug-sensitive patients (blue) and drug-resistant patients (red) as evaluated by 3-year survival ROC (Fig. 7a) or 5-year survival ROC (Fig. 7b), verifying distinct survival profiles between the two predicted groups of patients (log-rank test p values less than 0.0001).

To further assess the accuracy of different gene signatures for predicting the sensitivity or resistance to the targeted therapies, we compared 3 risk signatures that were designed for glioma patients: signature 1—the macrophage-related gene signature newly proposed in this study; signature 2—a conventional EGFR gene signature studied by many groups [15, 16]; signature 3—an immune-related gene signature, i.e., the Cheng et al. signature (FOXO3, IL6, IL10, ZBTB16, CCL18, AIMP1, FCGR2B, and MMP9) [17]; Signature 4—a gene signature based on IGF1/IGF1R-mediated pathways between macrophages and glioma cells (Quail et al. [10]), including several gene families in these pathways (PIK3R1, PIK3R2, PIK3R3, PIK3R4, PIK3R5, PIK3R6, PIK3AP1, AKT1, AKT2, AKT3, IGF1, IGF1R, IL4, IL4R, NFATC1, NFATC2, NFATC3, NFATC4, NFAT5, STAT6, CSF1 and CSF1R). We calculated the AUCs of the ROCs of these signatures to quantitatively assess and compare the accuracies of different gene signatures. Figure 7c shows the AUCs of the ROCs of these signatures for predicting drug sensitivity as evaluated by the 3-year survival outcomes (AUC of signature 1: 0.8295; AUC of signature 2: 0.6475; AUC of signature 3: 0.7155; AUC of signature 4: 0.5802). Figure 7d shows the AUCs of the ROCs of these signatures for predicting drug sensitivity as evaluated by the 5-year survival outcomes (AUC of signature 1: 0.7973; AUC of signature 2: 0.6462; AUC of signature 3: 0.6718; AUC of signature 4: 0.5877). Small p values computed using a bootstrap method [27] further demonstrated a superior predictive power of the macrophage-related gene signature compared with that of other signatures.

The univariate and multivariate COX regression analyses for both the learning and validation datasets revealed that the macrophage-related gene signature retained prognostic significance for glioma patients when adjusted for both clinicopathologic risk factors (age, gender, grade) and other existing gene signatures (EGFR gene signature, Cheng et al. signature and IGF1/IGF1R pathways signature) (Table 1). These results indicated that the macrophage-related gene signature was independently correlated with the overall and 5-year survival of glioma patients.

To further explore the prognostic value of the macrophage-related gene signature in stratified cohorts, patients were first classified by 2 important clinicopathological factors, age and grade, that significantly correlated with the prognosis of glioma patients (Table 1). Figure 8a–d shows the prognostic significance of the macrophage-related gene signature in different glioma cohorts stratified by age (Fig. 8a, b) or grade (Fig. 8c, d). In all of these cohorts, patients were classified as high- versus low-risk groups using cut point from ROCs, and the high-risk patients had a significantly shorter overall survival than low-risk patients. Subsequently, patients who received pharmaceutical therapy and radiotherapy were utilized to validate the prognostic significance of the macrophage-related gene signature. Figure 8e–h demonstrates that the macrophage-related gene signature retained prognostic significance for glioma patients treated with or without pharmaceutical therapy (Fig. 8e, f) and radiotherapy (Fig. 8g, h). These results indicated that the macrophage-related gene signature could accurately identify patients with an unfavorable prognosis regardless of their clinicopathological and treatment characteristics.

We further examined whether combining the macrophage-related gene signature with clinicopathological risk factors and other existing gene signatures could significantly improve the prognostic accuracy of the risk signature. The time-dependent ROC curves (Fig. 9) compared the prognostic accuracy by age, grade, signature 1 (i.e., macrophage-related gene signature), signature 3 (i.e., Cheng et al. signature) and the combined signature (see the definition in the Methods section). The AUCs of ROC curves in both the learning dataset (Fig. 9a, b) and the validation dataset (Fig. 9c, d) showed that the combined signature outperformed other risk signatures except the macrophage-related gene signature for predicting the 3- and 5-year survival rates of glioma patients. Noticeably, the AUC of the ROC of the macrophage-related gene signature was rather close to that of the combined signature in the learning dataset (Fig. 9a, b), or even higher in the validation dataset (Fig. 9c, d). These results indicated that the macrophage-related gene signature is an independent prognostic signature and possessed convincingly strong and robust prognostic power in comparison to the clinicopathological factors and other existing risk signatures.