Cell–cell communications shape tumor microenvironment and predict clinical outcomes in clear cell renal carcinoma

Single-cell mRNA sequence (scRNA-seq) data from 4 ccRCC patients with 4 tumor samples (1 sample is assigned to 1 patient), 1 VHL-mutated sample and 3 VHL-wild samples, were collected from Young, et al. cohort [13] and another dataset from Long cohort [14] including 4 samples from 4 ccRCC patients, all the patients were VHL mutated in Long cohort. After preliminary sample integration and quality control of scRNA-seq data sets, we generated a gene expression matrix with 76,118 cells. Furthermore, we downloaded the TCGA genomic, transcriptomic, and clinical data used in this study are from TCGA-KIRC database (https://​portal.​gdc.​cancer.​gov/​), containing 539 ccRCC samples. And we also obtained Braun, et al. cohort [15], consisting of advanced ccRCC patients receiving immunotherapy (n = 592). RNA-sequencing data (FPKM values) were transformed into transcripts per kilobase million (TPM) values with better comparability.

We created Seurat objects in general and for individual cell types depending on the purpose of our studies by the Seurat package. The top 2000 most highly variable genes were extracted to conduct principal component analysis (PCA), and the top 20 principal components (PCs) were used for the next classification analysis. To assign markers for the cell types, we performed t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction algorithm to further summarize the distinct clusters in t-SNE plots. We employed annotated information supported by the previous research to determine the biological types of all cells and the marker genes for each cell type are shown in Additional file 1: Table S1. In addition, we applied “AddModuleScore” function to calculate scores of gene set in all cancer cells.

We removed batch effect after integration of Young et al. cohort and Long cohort based on 5000 genes with the most significant variation using “IntegrateData” function. CellChat is a tool that can quantitatively infer and analyze intercellular communication networks from scRNA-seq data and contains ligand-receptor interaction databases (http://​www.​cellchat.​org/​). The interactions were identified and quantified based on the differentially over-expressed ligands and receptors for each cell group (p < 0.05). Function “netVisual_diffInteraction” was employed to depict the differences in strength of intercellular communication. Afterwards, to explain how multiple cell populations and signaling pathways coordinate and drive intercellular communication, non-negative matrix factorization (NMF) was performed to speculate the numbers of communication patterns by function “identifyCommunicationPatterns”. Finally, we extracted the major signaling inputs and outputs among all cell clusters and depicted them using scatter plots. By computing the Euclidean distance between any pair of the shared signaling pathways in the shared two-dimensional manifold scatter plot, we applied function “rankSimilarity” to identify the pathways with the most significant changes in VHL mutated and non-mutated samples. Lastly, we compared the communication probabilities of ligand-receptor pairs regulated by certain cell populations into other populations. This was done by setting the parameter compare in the function “netVisual_bubble”.

To explore the relationship between pseudotime trajectories and CCCs molecules, we utilized Monocle package for single-cell pseudotime analysis under default parameters. Our criteria for selecting highly variable genes are as follows: 1. mean expression ≥ 0.1, 2. Dispersion _empirical ≥ 1 * dispersion _fit. Then we used the ‘plot_pseudotime_heatmap’ function to depict the dynamic expression of CCCs molecules in the process of cell differentiation.

We applied SCENIC with the pySCENIC package (0.11.2) to investigate the dominant transcription factors in distinct tumor clusters in python. SCENIC reconstructs regulons (transcription factors and their target genes) accesses the activity of these regulons in individual cells and defines meaningful clusters of cells [16]. pySCENIC consists of two major processes. Establishment of co-expression network by GENIE3 and identification of targeted genes through motif-analysis by RcisTarget database. Matrix of tumor cells generated using Seurat was as input data. After running GENIE3, motif dataset (hg38_refseq-r80_10kb_up_and_down_tss.mc9nr. feather, hg38-tsscentered-10 kb 7species.mc9nr.feather) was used to construct regulons for each transcription factor.

After identifying the regulons for the transcription factors, we then used clusterProfler package to perform Gene Ontology (GO) enrichment analysis, cell components, and molecular function pathways. In addition, P-value < 0.05 and adjusted P-value (Q value) < 0.05 were considered statistically significant. The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway in the C2 category from the molecular signature database (MSigDB) was extracted for gene set variation analysis (GSVA) to speculate the enrichment scores for the samples.

Copy number analysis was performed with the inferCNV (v1.8.1) package under default parameters. Furthermore, mast cells were selected as normal reference. The tumor cells were clustered with the Seurat package (resolution = 0.9), and the expression matrix of tumor cells were then as input for the analysis.

Abundance of 22 immune cell types in the 539 primary KIRC samples and advanced ccRCC patients receiving immunotherapy (n = 592) were calculated using CIBERSORT and the LM22 signature matrix [17], and the algorithm was run for 100 permutations. Additionally, we calculate the relative abundance of cell types identified by scRNA-seq data in bulk RNA data using CIBERSORTx website (https://​cibersortx.​stanford.​edu).

Estimation of STromal and Immune cells in Malignant Tumors using Expression data (ESTIMATE) is a computerized algorithm to infer stromal and immune cells infiltration levels in tumor tissues based on bulk RNA-seq data. We adopted ESTIMATE algorithm to compute the immune and stromal score for TCGA-KIRC patients, all the parameters were set as default.

Statistical analyses of scRNA-seq data were performed by Seurat, CellChat, Monocle, SCENIC, inferCNV, CIBERSORTx online tool and GSVA. Difference analysis was performed by Student’s t-test and Wilcoxon rank sum test, respectively. Pearson analysis was adopted to calculate the correlation of CCCs molecules and relative abundance of 22 types of immune cell. All statistical analyses were undergone on the R (version 4.0.3) software, except for the SCENIC analysis. And p-value below 0.05 was considered statistically significant in this research.

We collected scRNA-seq dataset of Young cohort consisted of 1 VHL-mutated and 3 VHL-wild ccRCC patients to dissect the landscape of TME. We obtained 21,790 cells in total, and then performed t-SNE clustering. These cells were subsequently clustered into 10 cell subgroups, including kidney epithelium, CD4 + T cell, CD8 + T cell, myeloid cell, endothelial cell, cancer cell, NK cell, fibroblast, B cell, and mast cell (Fig. 1A) by identified the marker genes (Fig. 1B; Additional file 1: Table S1). we identified numerous CD4 + (expressing IL7R), and CD8 + T cell populations (expressing CD8A), natural killer (NK) populations (expressing GNLY, NKG7), small populations of B cells (expressing MS4A1/CD20) and population of myeloid (LYZ, CD14, S100A8, S100A9), Among non-immune clusters, we identified tumor cells (expressing CA9) and normal kidney epithelial cells (expressing ALDOB), and fibroblast (expressing ACTA2, THY1). In addition, visualization of the distribution of cells from VHL-mutated sample was conducted (Fig. 1C), and t-SNE plot displays the highly expressed marker genes in corresponding cell subgroups (Fig. 1D).

Previous studies revealed that VHL-HIF signal plays an important role in immune cell differentiation [18]. To address whether this biological phenomenon caused by different mode of CCCs due to VHL variation, we applied CellChat to investigate the characteristics of CCCs in these two states. Due to the small number of samples with VHL mutations in the M. D. Young cohort, we integrated this cohort with Z. Long cohort via the “Seurat” package (Additional file 8: Figure S1A, B) Using known lineage markers, we identified kidney epithelium, T cell, myeloid cell, endothelial cell, cancer cell, NK cell, fibroblast, B cell, and mast cell. By comparing VHL mutant samples with wild samples using composition plot (Additional file 8: Figure S1C, D), we found the shifts in the percentage of cell populations. We analyzed interaction numbers and interaction strength in different types of cell subpopulations (Additional file 8: Figure S2A, B; Additional file 2: Table S2), then compared interaction numbers and strength in VHL wild and mutant cells on the overall level (Fig. 2A; Additional file 8: Figure S1C), which suggested wide range of variations of CCCs in the global communication pattern.

Additionally, to further explore the basis of the pathways mediating these changes, the signaling pathways associated with CCCs networks from VHL wild and mutated samples were obtained and mapped onto a shared two dimensions manifold and clustered into groups based on functional and topological similarity (Additional file 8: Figure S2D, E). Furthermore, how multiple types of cell subgroups coordinates to function could lead to alteration of CCCs (Fig. 2B, C). And we defined three patterns for outgoing and incoming signals of non-VHL and VHL mutated groups respectively based on non-negative matrix factorization to identify the collaborative mode of CCCs (Fig. 2B, C; Additional file 8: Figure S2F, G), and the results revealed that CCCs patterns of these two different groups shared the distinct cell subpopulation constitution, incoming and outgoing signaling activations. For example, pattern 3 in VHL-wild samples is comprised of Cancer cell, Mast cell, NK cell, B cell, and T cell in the levels of types of cells, while in VHL-mutated samples, pattern 3 is constituted of Myeloid cell, NK cell, T cell, B cell. These finding potentially reveals the different mode of collaboration of various cell types in different genomic background.

Further we explored the cell types that lead to changes in CCCs patterns, according to the incoming and outgoing interaction strength, we identified the sending and receiving signal cell populations with significant changes between these two groups, especially myeloid cell, B cell, endothelial cell, and fibroblast (Fig. 2D).

In addition to the alterations in CCCs activity in non-VHL and VHL mutated groups, ligands-receptors pathways play as an important role in both mediating and altering the CCCs landscape, by calculating the Euclidean distance between any pair of the shared signaling pathways in the two groups. We found the signaling pathways like Nicotinamide Phosphoribosyltransferase-Insulin (VISFATIN), Angioprotein (ANGPT), Secreted Phosphoprotein 1 (SPP1), and Macrophage Migration Inhibitory Factor (MIF) pathways with large distance in biological functions, while VISFATIN, Galectin 9 (GALECTIN), Interferon-II (IFN-II) pathways were uncovered large distance in network structure (Fig. 2E, F). In addition, prominently changes in their information flow of pathways in non-VHL mutant group as compared with VHL mutant group with following conditions: (i) turn off (such as CCL, CD137, TNF), or (iv) decrease (such as IFN-II, COMPLEMENT). The information flow for a given signaling pathway is defined by the sum of communication probability among all pairs of cell groups in the inferred network (Additional file 8: Figure S2H) [19]. We plotted and compared the L-R pairs and the communication probabilities of each cell group pair between the two groups, to display an overall picture of pathway variability (Fig. 2G). Next, the superimposed effects of the outgoing and incoming signals were identified and compared in both groups (Fig. 2G; Additional file 8: Figure S2I, J). Lastly. We discovered the alteration of ligand-receptor pathways (e.g., VEGF, VISFATIN) pathways (Additional file 8: Figure S2K). VEGF pathway mediates intercellular communication between a wider range of cell types in VHL mutant samples, which is also present in VISFATIN.

CCCs coordinates the activities of multiple cell types required for immune response, therefore It is also important to investigate its effect on immunity to kidney cancer tumors. Mononuclear/macrophage reprogramming was identified to be strongly associated with tumor angiogenesis, immunosuppression, tissue remodeling and metastasis [20]. To further investigate the effects of tumor cell communication on the differentiation and infiltration of mononuclear/macrophage system, we extracted the myeloid cells in our data for classification and reannotation (Fig. 3A), which contains tumor associated macrophage1 (TAM1), TAM2, TAM3, inflammatory macrophage, proliferating macrophage, classic monocyte, and non-classic monocyte.

To address these questions, we used Monocle to perform differentiation trajectory analysis in mononuclear/macrophage system, based on the expression tendencies of FCGR3A (differentiation), S100A8 (stemness), S100A9 (stemness) (Additional file 8: Figure S3A). The results revealed that developmental hierarchy started with classical monocytes and non-classic monocytes and progressing towards SPP1 + TAM1 and FOLR2 + TAM2 (Fig. 3B, C). To confirm the CCCs-related molecules which probably modulate the process of mononuclear/macrophage system differentiation, we applied CellChat to visualize the communication strength, number and extract the key signaling pathways (Fig. 3D, E; Additional file 3: Table S3). We then analyzed changes in the expression of these receptor/ligand molecules during mononuclear/macrophage differentiation, and these molecules with remarkable variation were depicted on the heatmap (Fig. 3F), which indirectly suggested that these ligand/receptor molecules potentially influence the differentiation of mononuclear/macrophage system.

Subsequently, we confirmed these findings via bulk RNA-seq data. The mononuclear/macrophage system infiltration promotes cancer fibrosis and regulates the sensitivity of chemotherapy [21]. Applied to CIBERSORT revealed relative immune cell composition. Then, we used the “pearson” algorithm to calculate the correlation of M0, M1, M2 and monocytes with the expression of the screened receptor ligands (Fig. 3G). The analysis suggested the broad correlation between receptor/ligand gene expression and mononuclear/macrophage infiltration. Comprehensively, CCCs related molecules including C3, EGFR, HAVCR2, and so on, could influence the differentiation state and number of infiltrating myeloid in TME of ccRCC.

To confirm whether the similar biological efficacy mediated by CCCs occurs on T cell, we then clustered all the CD4 + T cells into CD4 + effector helper cell, CD4 + Treg, CD4 + cytotoxic T cell, CD4 + HSPA1A + T cell, and CD4 + NKT cell. And CD8 + T cells were clustered into CD8 + exhausted, CD8 + NK-like T cell, CD8 + exhausted immediate-early genes (IEG), and CD8 + Proliferating, repectively (Fig. 4A, B). We identified direction of differentiation of CD4 + T cells according to the expression shift of GZMK (activated), GZMA (activated), and CD27 (exhausted) (Additional file 8: Figure S3B), and TBX21 (stemness), TCF7 (stemness), and TOX (differentiation) expression levels were used to determine the differentiation direction of CD8 + T cells (Additional file 8: Figure S3C). Based on differentiation trajectory analysis, we demonstrated the divergent trajectory from HSPA1A + CD4 + T cell, cytotoxic CD4 + T cell to CD4 + Treg, and also defined the differentiation trajectory from CD8 + exhausted IEG to CD8 + exhausted (Fig. 4C). We subsequently conducted the visualization of the communication landscape between CD4 + and CD8 T + cells and tumor cells, respectively (Fig. 4D; Additional file 3: Table S3), and extracted the core receptor-ligand pathway for further analysis. We then respectively showed the expression trajectories of these ligands/receptors on CD4 + /CD8 + T cell differentiation using heatmaps (Fig. 4E, F). Correlation analysis revealed the wild associations between various types of T cell and CCCs related molecules (Fig. 4G). Among them, GZMA, FASLG, and CD27 were found to be most positively correlated with infiltration of CD8 + T cell, and negatively correlated with T cells CD4 memory resting. In general, the above results illustrated the profound significance of CCCs in T cell differentiation and infiltration.

Tumor cells are important components in TME and are inevitably influenced by small molecules, immune factors and receptor-ligand pathways in the surrounding environment, we therefore aimed to investigate the biological characteristics and potential clinical translational significance of this cluster of tumor cells. Following tumor cells classification (Fig. 5A), we determined 8 clusters by t-SNE algorithms. To investigate whether these clusters are genomic differences, inferCNV analysis was applied to reveal that the eight tumor clusters are characterized by different copy number variants (Fig. 5B). We further found that the malignant cells have more copy number expansion signatures and deletions compared to other mast cells. Furthermore, malignant cells in cluster 3 indicates that there are more copy number deletions in chromosome (chr)3, chr9, chr13, and chr14, while the copy number gains in chr6, chr10, and chr11 are higher than other malignant cell clusters. Subsequently, we aimed to confirm the tumor cluster that was vulnerable to their surrounding environment. The result revealed that the ligand scores was highest in the tumor cluster 3 (Fig. 5C). We performed GSVA analysis to investigate the biological functions of tumor cluster 3, which demonstrated significant activation of B cell mediated immunity, mesenchymal epithelial transition, cell–cell adhesion, protein exit from endoplasmic reticulum, etc. (Fig. 5D, Additional file 4: Table S4), furthermore, we calculated and compared the GSVA scores of the pathways from hallmarks gene set from Molecular Signature Database (Additional file 8: Figure S3D) and then we conducted the CellChat analysis to depict the strength of this cluster's communication with myeloid cells and T cells, the results illustrated that tumor cluster 3 has an extensive and active ligand receptor interactions with these immune cells (Fig. 5E; Additional file 5: Table S5).

To further reveal the important role of Cluster 3 for CCCs within TME of ccRCC and potential mechanisms of differentiation. We performed SCENIC (single-cell regulatory network inference and clustering) analysis to conjecture regulon activities, which utilizes both co-expression modules between transcription factors and candidate target genes, according to the databases of DNA binding motifs of transcription factors to infer significant gene regulation (Additional file 6: Table S6) by these selected transcription factors and [22], and a transcription factors together with its target genes comprise a regulon. Based on the activities of the regulons we determined, all the tumor cells in cluster 3 were assigned to the common regions in our heatmap (Fig. 5F), which indicated the robustness and stability of our clustering. We screened out the regulons in the boxed area, which are highly expressed in cluster 3. Then, we performed the GO analysis to infer the biological functions of the gene regulatory networks (Fig. 5G; Additional file 7: Table S7). GO analysis revealed that the regulon networks enriched in the pathways associated with membrane protein synthesis and localization, this further reveals why this tumor subpopulation is highly ligand-expressing and thus may play an important role in the regulation of the microenvironment.

Restricted by lack of clinical information on our single cell data and inclusion of patients, we have difficulty in determining the clinical translational value the high-ligand-expression tumor cluster. Therefore, we applied CIBERSORTx algorithm to calculate infiltration scores of cell subpopulations we identified in single-cell data to explore the clinical significance. Furthermore, we conducted unsupervised clustering to analyze ccRCC samples and classified patients into qualitatively different subgroups. Two distinct subgroups were ultimately identified, including 306 cases in Cluster A, and 224 cases in Cluster B (Fig. 6A). The heatmap demonstrated that the infiltration of HL tumor (High ligand tumor) is higher in Cluster B compared with Cluster A, and there is notable difference between these two cluster in Macrophage Proliferating, Classical Monocyte, CD4 + HSPA1A + T cell, TAM3 MT1G + , TAM1 SPP1 + , Non-classical Monocyte, HL_Tumor and dendritic cells. Furthermore, these clusters are predictive of disease outcome across TCGA (P < 0.001) cohorts (Fig. 6B).

CCCs as an important factor influencing the TME has profound implications for immunotherapy. The two clusters we obtained have different immune cell infiltration landscapes, also suggesting two different patterns of CCCs. Therefore, we inferred that there is distinct immunotherapy response between the Cluster A and Cluster B. To prove this inference, we then performed the following research. Firstly, we discovered that there was significantly different immune cell abundance in two clusters (Fig. 6C). Secondly, we applied the ESTIMATE algorithm to, to calculate the immune score and stromal score of each tumor tissue sample based on the gene expression matrix, and all parameters are set as default.

As expected, the immune parts and stromal parts were significantly upregulated in Cluster B (Fig. 6D, E). Thirdly, Genome analysis indicated that TMB is higher in Cluster B compared with Cluster A (p = 0.02, Fig. 6F), and waterfall plot demonstrated that the mutation rates of PBRM1 (45%) in Cluster B are significantly higher than these in Cluster A (33%), and BAP1 mutation rates were ultimately decreased in Cluster B (5%) compared with Cluster A (17%) (Additional file 8: Figure S4B, C). Lastly, we utilized David A. Braun, et al. cohort to further confirm our inference, this cohort contains information on patient survival and all patients received anti-PD-1 (Nivolumab) or anti-mTOR (Everolimus) therapy. Our classification result showed the heterogeneity in immune infiltration in both clusters as well (Fig. 6G), and the overall survival between these two clusters are also notably different (p = 0.026, Fig. 6H). In summary, different patterns of CCCs in ccRCC predicts different clinical outcomes and different immunotherapy response.