The flowchart of the translational pipeline to construct the multicellular interaction gene network (MIGN) and identify drug targets of immunotherapies is presented in Fig. 1. We began with experimental results of the preclinical mouse study that investigated long-term treatment efficacy of CSF-1R inhibition in gliomas [11]. After downloading the preclinical RNA-seq data from the Gene Expression Omnibus (GEO) website (https://www.ncbi.nlm.nih.gov/geo/) with accession number GSE69104, we selected mouse samples that contained cell type-specific gene expression for both tumor-associated macrophages (TAMs) and tumor cells (TCs), leading to four drug-resistant samples, six drug-sensitive samples, and five vehicle (non-treated) samples.

For both TAMs and TCs, we focused on selecting immunotherapy-related feature genes (IFGs) that expressed differentially between drug-resistant and drug-sensitive samples during CSF-1R inhibition. First, we normalized the expression level of each gene in the treated groups (drug-resistant / drug-sensitive) to its mean level in the non-treated group as follows: zij=log2xij/log2vi¯, where xij is the expression value of the i-th gene in j-th sample of the treated groups, and vi¯ is the mean expression of the i-th gene across samples in the non-treated group. Next, we compared normalized gene expression profiles between drug-resistant and drug-sensitive samples using t-test and fold change (FC). Those genes with false discovery rate (FDR)-adjusted p − value < 0.05 in t-test and |FC|>1.5 were selected as IFGs, which were then translated to those in human genome by the online tool (NCBI, https://www.ncbi.nlm.nih.gov) for further analysis.

To construct the weighted gene correlation network and perform survival analysis, we collected clinical information and RNA-seq gene expression data for LGG patients in TCGA database (https://cancergenome.nih.gov/) and CGGA database (http://www.cgga.org.cn/). After matching the clinical information and RNA-seq data for each sample, one cohort of 524 LGG patients from the TCGA database was prepared for training, and an independent cohort of 172 LGG patients from the CGGA database was prepared for validation.

Next, the TCGA dataset was leveraged to construct TAM- and TC- associated gene networks with the translated IFGs for TAMs and TCs being their node sets, respectively. Generally, weighted gene correlation network was denoted as G=(V,W), where V was the node set, and W=wij was a |V|×|V| positively weighted adjacency matrix. Let xi and xj denote the expression levels of the i-th and j-th gene across all the samples, the entry wij was encoded as the power function of the absolute value of Pearson correlation coefficient: (1) wij=wijβ=corxi,xjβ,i≠j, where β was referred to as the soft threshold parameter. Since the correlation between each gene and itself was not considered, the diagonal elements of W were set to 0. Then we could define the connectivity degree for the i-th gene: (2) ki(β)=∑j=1|V|wij(β),i=1,2,⋯,|V| and the connectivity frequency for the network: (3) fk=∑i=1VIkiβ∈k-1,k,k=1,2,3,···.

We assume that gene network is dominated by a few highly connected nodes. Thus, we selected soft threshold parameter β to ensure the network satisfying scale-free topology criteria [15], [16], which are characterized by power law distribution of the connectivity proportion, i.e., (4) pk∼k-γ, where p(k)=f(k)/|V|, and γ is a positive constant. After taking logarithm on both sides of (4), the log-proportion of connectivity log(p(k)) is linearly correlated with log(k). Then linear regression model fitting index R2 could be used to quantify how well a network satisfies the scale-free criteria.

Gene clustering analysis was conducted within each of the TAM- and TC- associated networks. Here, we used the hierarchical clustering method in conjunction with topological overlap matrix (TOM)-based dissimilarity to detect gene clusters within the network, as performed in the R package WGCNA [17]. In specific, the topological overlap matrix S=sij is defined based on the weighted adjacency matrix W=wij as follows (5) sij=lij+wijminki,kj+1-wij, where lij=∑u=1|V|wiuwuj, and ki=∑u=1|V|wiu. And the TOM-based dissimilarity dij is defined as (6) dij=1-sij.

The dissimilarity dij summarizes the global connection information between the i-th and j-th genes within the gene network instead of only focusing on their pairwise relationship. Thus, dij leads to more distinct gene cluster than the standard measure (e.g., 1 minus the absolute value of correlation coefficient) [18], [19]. With the TOM-based dissimilarity, we obtained a clustering dendrogram using the average linkage hierarchical clustering method. The dynamic tree cut method [20] was adopted to cut branches off the dendrogram, and thus we obtained several distinct gene clusters within the network. In the R package WGCNA, the parameter deepSplit controls how finely the branches should be split. The higher value of deepSplit gives smaller clusters. Another parameter minClusterSize limits the minimum number of genes in each cluster.

Within the TAM- and TC- associated networks, the concordance measure K-index [21] was adopted to correlate gene clusters with patient survival. Typically, gene clusters were well represented by their eigengenes (i.e., the first principal component of the expression matrix of the gene cluster). Thus, for TAM- and TC- associated networks, the eigengene-based K-index was calculated for each cluster as (7) K-index=2n(n-1)∑i≠jIRj>Ri1+expRj-Ri, where n denotes the sample size of involved patients, and Ri indicates the eigengene-based risk score for the i-th patient. Consequently, gene clusters with K-index>0.6 were considered to be highly correlated with survival outcomes in TAM- and TC- associated networks.

To build the multicellular interaction gene network (MIGN), we combined genes in those clusters that identified to be highly correlated with patient survival in TAM- and TC- associated networks. Then scale-free network was constructed with the selected genes, and clustering analysis was performed within the network. Given the intense signaling interaction between TAMs and TCs during glioma development, the selected TAM- and TC-clusters were assumed to merge in the MIGN if their underlying pathways worked together to co-influence disease progression. Fisher’s exact test was applied to examine the correspondence of gene clusters between MIGN and TAM/TC-associated networks.

Cell-cell communication was mainly mediated by secreted ligands and corresponding cell-surface receptors. We collected prior information on ligand and receptor genes from the Omni-path database [22], and screened TAM ligands and TC receptors in the TAM/TC interaction gene clusters of MIGN. Besides, hub genes were selected as those with K-index>0.6 and connectivity k(β) within the top 10% of each gene cluster in the MIGN. With the selected ligands, receptors, and hub genes, we conducted LASSO-based Cox regression on each type of candidate genes to shrink the feature size. Next, a stepwise variable selection procedure (with iterations between the ‘forward’ and ‘backward’ steps) for Cox regression was applied to identify significant features. Patient age and tumor grade were considered when selecting variables to achieve a more accurate prediction. Consequently, we constructed the MIGN signature by fitting multivariate Cox regression model using the selected signatures

Log-rank test and time-dependent ROC were used to evaluate prediction performance of the MIGN signature. For gene expression profile X in both training and validation datasets, the risk score f(X)=X′βˆ was computed, where βˆ denoted regression coefficients derived from the multivariate Cox regression model. Patients were classified into high-risk or low-risk groups by choosing the median of risk scores as cutoff in each dataset, indicating poor or good prognoses. Kaplan-Meier (K-M) curves were then generated to summarize patient survival in each risk group, and log-rank test was conducted to assess whether survival curves for high-risk and low-risk groups were significantly different. Time-dependent ROC analysis was performed to evaluate accuracy of the MIGN signature in predicting 1-year, 3-year, and 5-year survival rates. Larger value of the area under the ROC curve (AUC) indicates better prediction of patient survival.

We collected single-cell RNA sequencing (scRNA-seq) data for LGG patients from the publicly available database Gene Expression Omnibus (GEO) (www.ncbi.nlm.nih.gov/geo/) with accession number GSE182109 to verify the crosstalk between TAMs and TCs. Seurat v4.0.3 was used to perform quality control (QC), normalization, feature selection, and dimensional reduction for the scRNA-seq data. Low-quality cells containing <500 expressed genes or >20% mitochondrial transcripts or >5% ribosomal transcripts were removed. Additionally, principal components were batch-corrected for variations between tumor specimens using the Harmony R package. Potential single-cell doublets were identified using the DoubletFinder R package. Uniform Manifold Approximation and Projection (UMAP) was used for visualization of the scRNA-seq data. Moreover, multilayer signaling network analysis was performed to infer the intercellular communication between TAMs and TCs using scMLnet [23], an R package developed by our group previously.

Since the translational pipeline was originated from a mouse study that proposed IGF-1R inhibition as the treatment to drug resistance developed during CSF-1R blockade, it is of interest to investigate the potential drug sensitivity of human LGG patients to the related therapeutics. As an independently-developed bioinformatics pipeline, the oncoPredict R package [24] could predict IC50 (the half maximal inhibitory concentration) of a large number of drugs screened in cancer cell lines based on gene expression data of cancer patients. Thus, we calculated potential clinical responses (IC50) to targeted drugs for LGG patients using the oncoPredict package, with the authoritative drug response resource Genomics of Drug Sensitivity in Cancer (GDSC) [25] being the training set. We attempted to uncover treatments that exhibit significantly different sensitivity scores between the high-risk and low-risk groups of patients stratified by the MIGN signature, which may bring clinical benefit from the perspective of precision medicine.

We proposed a translational pipeline to construct the multicellular interaction gene network (MIGN), to model communication between macrophages and glioma cells for LGG patients, and to facilitate genome-wide screening of new drug targets for immunotherapies. As demonstrated in the flowchart (Fig. 1), we first selected immunotherapy-related feature genes (IFGs) for TAMs and TCs using cell-type specific RNA-seq data of existing randomized preclinical mouse trials [11]. After translating them to human genome, we obtained 255 IFGs for TAMs and 795 IFGs for TCs, respectively. Prior to building the MIGN, we constructed and analyzed TAM- and TC- associated networks (see section Methods for details).

For the TAM-associated network G1=V1,W1, the vertex set V1 contained 255 nodes that being the TAM-specific IFGs, and weighted adjacency matrix W1 was built by setting the soft threshold parameter β to 5. Thus, the network approximately satisfied scale-free topology criteria with the index R2 reaching 0.97. By setting the clustering parameters deepSplit as 3 and minClusterSize as 25, four gene clusters were identified in the TAM-associated network (Fig. 2(a)). Two clusters (TAM-brown and TAM-turquoise) were considered being highly associated with overall survival, with K-index for their eigengenes greater than 0.6 and their K-index distributions were significantly higher than those of others (Fig. 2(c)).

For the TC-associated network G2=V2,W2, the vertex set V2 contained 795 nodes that being the TC-specific IFGs, and weighted adjacency matrix W2 was built by setting the soft threshold parameter β to 5. Thus, the network approximately satisfied scale-free criteria with the index R2 reaching 0.84. By setting the clustering parameters deepSplit as 3 and minClusterSize as 30, six gene clusters were identified in the TC-associated network (Fig. 2(b)). Three clusters (TC-blue, TC-green and TC-turquoise) were considered being highly associated with overall survival, with K-index for their eigengenes greater than 0.6 and their K-index distributions were significantly higher than those of others (Fig. 2(d)).

To construct the MIGN G3=V3,W3, we combined genes in those five clusters (TAM-brown, TAM-turquoise, TC-blue, TC-turquoise, and TC-green) that identified to be highly correlated with patient survival in TAM- and TC- associated networks. Here, the vertex set V3 contained a total of 609 nodes, and weighted adjacency matrix W3 was built by setting the soft threshold parameter β to 5, making the network approximately satisfied scale-free topology criteria with the index R2 greater than 0.8 (Fig. 3(a)). With the deepSplit set to 3 and minClusterSize set to 30, four gene clusters (MIGN-blue, MIGN-brown, MIGN-turquoise, and MIGN-grey) were identified in MIGN (Fig. 3(b)). All the clusters other than MIGN-grey were highly correlated with patient survival (K-index > 0.6), and their K-index distribution was significantly higher than that of MIGN-grey (Fig. 3(c)). Additionally, there was a significantly positive correlation between K-index and connectivity in these three clusters (Fig. 3(d)), allowing us to select hub genes in subsequent analysis.

Next, contingency tables were generated to examine the correspondence of gene clusters between MIGN and TAM/TC- associated networks (Fig. 4(a) and (b)). Consequently, the clusters MIGN-blue and MIGN-brown that comprised of both TAM- and TC-clusters were referred to as the interaction gene clusters (IGCs), and the cluster MIGN-turquoise that contained single type of cell-related genes was a non-interaction gene cluster (non-IGC). Then KEGG enrichment analysis was performed on each of the gene clusters to explore their biological implications (Fig. 4(c)). Both the IGCs (MIGN-blue and MIGN-brown) were enriched in cell adhesion molecules. Despite involved in cell adhesion pathways, MIGN brown was also enriched in several cancer-related pathways (e.g., TNF signaling pathway). We screened TAM ligands and TC receptors in the IGCs since these clusters represent intercellular signaling communication between TAMs and TCs. Besides, hub genes were selected as those with K-index>0.6 and connectivity k(β) within the top 10% of both the IGCs and non-IGC.

We first applied LASSO-based Cox regression on each type of the candidate genes (ligands, receptors, and hub genes) to shrink the feature size. Then we applied stepwise Cox regression to identify the significant features, in which significance levels for entry (SLE) and stay (SLS) were set at 0.15 to be conservative. Patient age and tumor grade were considered when selecting variables to achieve a more accurate prediction. Additionally, the significant TAM ligands (Fig. 5) were always included in the model, though they may lose significance due to the masking effects of intermediate variables when combined with receptors and hub genes. As a result, we obtained the MIGN signature that comprised of four ligands (EFNB2, C3, IGF1, TNF), four receptors (LRRC4, FXYD6, GFRA1, TNFRSF21), seven hub genes (ASPM, SOX6, ACVR2B, ATCAY, MSR1, PAK7, DUSP26), and two clinical factors (Age and Grade). The forest plot together with the coefficient, p-value and hazard ratio of the MIGN signature in the multivariate Cox regression model were presented in Fig. 6.

We performed K-M and time-dependent ROC analyses to evaluate the performance of MIGN signature with respect to predicting patient survival. For both the training and validation datasets, the risk score was calculated for LGG patients to rank the sample in descending order. We assigned the top 50% of patients as the high-risk group and the bottom 50% as the low-risk group according to the median risk scores. K-M curves for those two groups of patients were well separated and the difference was significant according to the log-rank test in both the training and validation datasets (Fig. 7(a) and (b)). Time-dependent ROC analysis was then performed to assess prognostic accuracy. For the training dataset, the time-dependent AUCs with respect to 1-year, 3-year, and 5-year survival rates were 0.923, 0.902, and 0.83, respectively (Fig. 7(c)). For the validation dataset, the time-dependent AUCs with respect to 1-year, 3-year, and 5-year survival rates were 0.881, 0.867, and 0.839, respectively (Fig. 7(d)). We tested prognostic significance of the MIGN signature using a bootstrapping approach. We randomly selected 15-gene sets from the whole transcriptome 1000 times and compared their predictive accuracy with that of the MIGN signature. The ROC curves of the MIGN signature against bootstrapped 15-gene signatures for the validation set of CGGA cohort justified the prognostic significance of the MIGN signature Fig. 7(e). These results suggest that the MIGN signature has good prognostic significance and accuracy in both the training and validation datasets.

Moreover, we compared the predictive accuracy of the MIGN signature with three other risk signatures designed for glioma patients. For clarity, the MIGN signature proposed in this study is denoted as signature 1; signature 2 denotes the macrophage-related gene signature proposed by Sun et al. [26] based on 12 genes (ANPEP, DPP4, PRRG1, GPNMB, TMEM26, PXDN, CDH6, SCN3A, SEMA6B, CCDC37, FANCA, and NETO2); signature 3 represents the immune-related gene signature proposed by Zhang et al. [27] with 6 genes (CANX, HSPA1B, KLRC2, PSMC6, RFXAP, TAP1); signature 4 denotes the clinical factor signature (Age and Grade). To achieve a fair comparison, we adjusted signature 2 and signature 3 with the clinical factors (Age and Grade) and re-trained their coefficients using the TCGA dataset, which was exactly the training set for the MIGN signature. Consequently, the time-dependent ROCs for the validation set of CGGA cohort indicated that the MIGN signature outperformed all the other risk signatures (Fig. 8(a)–(c)).

To investigate biological functions of the MIGN signature, we conducted a literature survey on the associated genes. As demonstrated in Fig. 8(d), all the signature genes were engaged in biological processes of tumor growth and metastasis or have been reportedly correlated with neuronal function. For example, IGF1 is known to regulate proliferation and migration of glioma cells, and IGF1 together with its receptor IGF1-R appears to trigger LGG progression to GBM [28]. Thus, the IGF1 signaling may provide therapeutic opportunities for LGG patients. In addition, some of our signature genes have been validated as druggable targets in glioma studies (Fig. 8(d)), further highlighting potential utilities of the MIGN signature to develop novel treatments to extend survival of LGG patients.

To further explore the prognostic value of the MIGN signature in stratified cohorts, we classified LGG patients in the TCGA dataset according to their treatment status and assessed the survival difference between high-risk and low-risk patients within each treatment group. Consequently, the MIGN signature retained prognostic significance for LGG patients treated with or without pharmaceutical therapy (Fig. 9(a) and (b)) and radiotherapy (Fig. 9(c) and (d)), indicating that the MIGN signature could accurately identify patients with an unfavorable prognosis regardless of their treatment characteristics. Additionally, we looked into IDH1, IDH2 and TP53 mutation status with the information available from CGGA dataset. It appears that there is a significant difference between risk scores evaluated by MIGN signature of mutant and wild-type groups of patients (Fig. 9(e)), indicating the association between MIGN signature and genomic alterations in low-grade gliomas.

To provide further mechanistic evidence on the interplay between TAMs and TCs in LGG patients, we leveraged the scRNA-seq data of human LGG tissues [29] to infer cell-cell interactions in the tumor microenvironment. After cell filtration based on QC metrics, we obtained 25986 cells from 4 tissues for subsequent analysis. PCA was applied to reduce the dimensionality of data, and the top 2000 most variable genes were used for cell clustering (Fig. 10(a)). The computed PCs were batch-corrected for variations between tumor specimens using the Harmony R package (Fig. 10(b)), and doublet cells were removed using the DoubletFinder R package (Fig. 10(c)). After re-clustering the remaining single cells, we obtained nine cell clusters with distinct gene expression patterns (Fig. 10(d)). Using the marker genes specifically expressed in each cluster, we classified cell clusters as either macrophage (CD14, AIF1, FCER1G, TYROBP, CSF1R), glioma cells (PDGFRA, SOX2, NES, GFAP, EGFR), oligodendrocytes (MBP, TF, MAG, MOG, CLDN11), stromal cell (PDGFRB, ACTA2, PROM1, PECAM1, TEK) or T cells (CD2, CD3D, CD3E, CD3G) (Figs. 10(e) and 11(a)).

Cell-cell communication networks globally depicted the intercellular ligand-receptor interactions between the above cell types, which indicated that TAM-TC interactions were among the most abundant cell-cell communication in the tumor microenvironment (Fig. 11(b)). Moreover, the MIGN signature-derived ligands and receptors (IGF1, EFNB2, TNF, TNFRSF21, LRRC4) were verified to mediate TAM-TC interactions as shown in the multilayer signaling network (Fig. 11(c)), where the downstream TFs (e.g., EGR1, MYC, and TP53) activated target genes related to PI3K signaling pathway in glioma cells [30]. Collectively, these results indicated that the MIGN signature plays a vital role in regulating intercellular communication between TAMs and TCs, and it thus implicates opportunities to block LGG progression by biological interventions targeting TAMs.

Potential clinical responses of LGG patients to target therapies were calculated by the oncoPredict R package [24]. Genomics of drug sensitivity in cancer (GDSC), an authoritative resource for drug response data, was used as the training set. Using gene expression data of LGG patients in the TCGA dataset, we found three IGF-1R inhibitors (Linsitinib, NVP-ADW742, and GSK1904529A) that exhibited significantly different responses between high-risk and low-risk groups classified by the MIGN signature (Fig. 12(a)). Besides, those IGF-1R inhibitors displayed markedly discriminative sensitivity scores for high-risk and low-risk groups in the CGGA cohort (Fig. 12(b)). For both the TCGA and CGGA cohorts, the low-risk group consistently presented lower IC50 compared to the high-risk group, indicating that patients in the low-risk group were more sensitive to the IGF-1R inhibitors and may thus benefit a better therapeutic effect from this treatment.

On the other hand, gliomas are characterized by substantial heterogeneity among tumors, underscored by the identification of several molecular subtypes (classical, mesenchymal, neural, and proneural) [31]. Since proneural gliomas are always associated with better survival than other subtypes [32], [33], we compared sensitivity scores among molecular subtypes of LGG patients to examine whether the proneural patients possess potentially more backup resources for the IGF-1R inhibition. As LGG patients are all missing information on molecular subtypes in the TCGA database, the drug sensitivity exploration was hence focused on LGG patients from the CGGA database. For the IGF-1R inhibitor NVP-ADW742, with a relatively lower IC50 than two other inhibitors (Linsitinib and GSK1904529A) in both the TCGA and CGGA cohorts (Fig. 12(a) and (b)), the predicted value of IC50 for proneural patients was significantly lower than those of other subtypes (Fig. 12(c)). Furthermore, correspondence analysis was conducted to reveal that the proneural patients were highly enriched in the low-risk group classified by the MIGN-based signature (Fig. 12(d)). Taken together, these results support our hypothesis that IGF-1 signaling, the critical pathway that connecting TAM and TC, might provide useful targets for developing novel therapies in human gliomas.