Identification of immune-related biomarkers associated with tumorigenesis and prognosis in cutaneous melanoma patients

A total of 46 melanoma and 46 normal tissues were obtained from 92 patients at the Department of Burn and Plastic Surgery, the First Affiliated Hospital of Soochow University (FAHSU, Suzhou, China) from March 2015 to August 2019. None of the patients had received radiotherapy or chemotherapy before operation. Tissue samples, including cutaneous melanoma and normal tissue, were collected during surgery and fixed in 4% paraformaldehyde, available from FAHSU tissue bank. Clinical data was available to obtain from hospital records. This research was supported by the Independent Ethics Committee (IEC) of the FAHSU and all patients were well informed of storing and upcoming use of their resected specimens for further research purposes.

Expression profiling of SKCM patients with clinical information was obtained from the Gene Expression Omnibus (GEO) database (http://​www.​ncbi.​nlm.​nih.​gov/​geo) [9]. Among the inclusion criteria were (a) diagnosis of patients with primary melanoma (PM) and normal skin (NS), (b) detection of gene level in tissue or blood samples. Exclusion criteria included: (a) clinical data without survival time and outcome, and (b) datasets with small sample sizes (n < 15). Finally, three datasets were eligible: accession numbers GSE15605 (46 PM samples and 16 NS samples), GSE46517 (31 PM samples and 8 NS samples) and GSE114445 (16 PM samples and 6 NS samples). For validation, gene expression profiles of 472 melanoma patients were downloaded from the TCGA data portal (https://​tcga-data.​nci.​nih.​gov/​tcga/​) [10]. Clinical characteristics of patients were also obtained, including age, Clark level, Breslow depth, survival time, and outcome. All of the clinicopathological characteristics of SKCM patients were shown in Table 1.

The DEGs between primary melanoma and normal skin were identified using GEO2R (http://​www.​ncbi.​nlm.​nih.​gov/​geo/​geo2r) based on limma package with the threshold of|logFC| > 1 and P-value < 0.05. Next, the online Venn software (http://​bioinformatics.​psb.​ugent.​be/​webtools/​Venn/​) was applied to detect the overlap DEGs among three datasets.

To elucidate the biological functions of the overlapping DEGs, we performed functional annotation and pathway enrichment analysis via DAVID (The Database for Annotation, Visualization and Integrated Discovery, http://​david.​ncifcrf.​gov/​,version 6.8) online tool [11]. P-value < 0.05 was considered as the cutoff value. GO enrichment and KEGG pathway results were visualized as bubble charts by R software.

STRING (http://​string-db.​org, version 11.0) database was used to predict the PPI network of DEGs and analyze the interactions between proteins [12]. An interaction with a combined score > 0.4 was recognized as statistical significance. The molecular interaction networks were visualized using the Cytoscape (version 3.7.0) and the most significant model in the PPI network was narrowed down by MCODE, with the following criteria: degree cutoff = 2, node score cutoff = 0.2, k-core = 2, max depth = 100 [13, 14].

Metascape (https://​metascape.​org) was used to further verify the function enrichment of hub genes [15]. P < 0.05 was set as the cutoff value. Hub genes pathway analysis was performed and visualized by ClueGO (version 2.5.4) and CluePedia (version 1.5.4), the plug‐in of Cytoscape [16]. P-value < 0.01 was considered statistically significant.

GEPIA (http://​gepia.​cancer-pku.​cn/​index.​html) is an analysis tool containing RNA sequence expression data of 9736 tumors and 8587 normal tissue samples [17]. In this study, GEPIA was used to perform survival analysis based on the data from TCGA database. Student’s t test was used to analyze the correlation between the expression and pathological stages. P value < 0.05 was considered statistically significant.

Protein expression levels of hub genes were measured using IHC staining and rabbit polyclonal anti-CCL4 antibody (ab235978), anti-CCL5 antibody (ab9679), anti-CXCL9 antibody (ab9720), anti-CXCL10 antibody (ab9807), anti-CXCL13 antibody (ab272874). Positive or negative staining of a certain protein in one FFPE slide was independently assessed by two experienced pathologists and supervised by a clinician. Based on the staining intensity level (no staining, weak, moderate and strong staining), the score was ranging from 0 to 3, as previous described. The staining extent was graded from 0 to 4 for the coverage percentage of immunoreactive tumor cells (0%, 1–25%, 26–50%, 51–75%, 76–100%). The overall IHC score grading from 0 to 12 was evaluated according to the multiply of the staining intensity and extent score. Negative staining represented grade 0 to 4 and positive staining from 5 to 12 for each sample.

The cBioPortal database (www.​cbioportal.​org), a comprehensive web resource, can visualize and analyze multidimensional cancer genomics data [18]. The genetic alterations of prognostic genes were obtained from cBioPortal based on 475 SKCM samples in TCGA.

TRRUST (https://​www.​grnpedia.​org/​trrust/​) is a useful predict tool for human, and mouse transcriptional regulatory networks [19]. The TRRUST database can provide information on how these interactions are regulated, containing 8444 transcription factor (TF)-target regulatory relationships of 800 human TFs.

Prognostic gene expression data was used to measure the abundance of six types of infiltrating immune cells (B cells, CD4⁺T cells, CD8⁺T cells, neutrophils, macrophages, and dendritic cells) in SKCM patients using the Tumor Immune Estimation Resource (TIMER) algorithm [20]. Then, an integrated repository portal for tumor-immune system interactions (TISIDB, http://​cis.​hku.​hk/​TISIDB/​index.​php) was utilized to examine tumor and immune system interactions in 28 types of TILs across human cancers [21]. Spearman’s test was used to measure correlations between prognostic genes and TILs. All hypothetical tests were two-sided and P-values < 0.05 were considered statistically significant.

DGIdb (version 2.0) is an open-source project that help users mine existing resources and generate assumptions about how genes are therapeutically targeted or prioritized for drug development [22]. The parameters were set as: preset filters: FDA approved; antineoplastic; all the default.

The expression heatmap and correlation coefficient were analyzed and visualized by pheatmap and corplot packages in R software. ROC curve of prognostic genes was performed by SPSS 22.0. P-values < 0.05 were considered statistically significant in all tests.

After standardization and identification of the microarray results, DEGs (4640 in GSE15065,1096 in GSE46517 and 1895 in GSE114445) were selected. The overlap among three datasets included 308 genes as shown in the Venn diagram (Fig. 1a, Additional file 1: Table S1) between primary melanoma and normal skin.

Functional and pathway enrichment of DEGs was performed by DAVID, then displayed in bubble charts by R software (P < 0.05, Fig. 2). GO analysis indicated that changes in biologic processes were significantly enriched in the inflammatory response, immune response, regulation of transcription. As for cellular component, DEGs were particularly enriched in extracellular region, extracellular space and extracellular exosome. Changes in molecular function were mostly enriched in chemokine activity, transcriptional activator activity, protein binding and CXCR3 chemokine receptor binding. KEGG pathway analysis demonstrated that DEGs played pivotal roles in pathways in cytokine–cytokine receptor interaction, chemokine signaling pathway and pathways in cancer. These results revealed that immune response and inflammatory response play important roles in SKCM tumorigenesis.

The PPI network was constructed via Cytoscape, including 247 nodes and 792 edges (Fig. 1b). Then, the most important PPI network module was obtained using MCODE, consisted of 12 nodes and 64 edges (Fig. 1c). After selection, CCL5, PTGER3, IL6, CXCL13, CCL27, CCL4, CXCL9, NMU, CXCL2, GAL, NPY1R, CXCL10 were considered as hub genes of the network.

As expected, functional annotation obtained from Metascape suggested that hub genes were mainly enriched in regulation of T cell chemotaxis, peptide ligand-binding receptors and chemokine receptors bind chemokine signaling (Fig. 3a, b, P < 0.001). Similarly, ClueGO (Fig. 3c, P < 0.01) revealed that the most involved pathways were interleukin-10 signaling, peptide ligand-binding receptors and chemokine receptors bind chemokine signaling.

Using the Kaplan‑Meier method, the prognostic values of the hub genes in SKCM patients were determined. Seven hub genes were significantly associated with OS as shown in Fig. 4. High expression of CXCL9 (HR = 0.6; P = 2E−4), CXCL10 (HR = 0.57; P = 2.8E−5), CXCL13 (HR = 0.56; P = 2.1E−5), CCL4 (HR = 0.48; P = 8.2E−8), CCL5 (HR = 0.53; P = 3.2E−6) were significantly associated with better OS. In contrast, NMU (HR = 1.6; P = 3E−4) and GAL (HR = 1.6; P = 0.001) were found associated with poor overall survival. Furthermore, CCL4 (HR = 0.76; P = 0.026), CCL5 (HR = 0.76; P = 0.027) were also found significantly associated with better disease-free survival (Fig. 4). Except GAL (AUC = 0.536), the rest of hub genes, including CCL4 (AUC = 0.872), CCL5 (AUC = 0.775), CXCL9 (AUC = 0.834), CXCL10 (AUC = 0.786), CXCL13 (AUC = 0.807) and NMU (AUC = 0.829), were of diagnostic value (P < 0.001) (Fig. 5).

Based on data from GEPIA database, all of the prognostic genes were found differential expressed in SKCM vs normal skin tissue (P < 0.05, Fig. 6).

In addition, CXCL9 (P = 7.74E−5), CXCL10 (P = 1.05E−4), CXCL13 (P = 8.15E−4), CCL4 (P = 0.002), CCL5 (P = 0.001) were significantly correlated to SKCM pathological stages (Fig. 7). Among them, CXCL9, CXCL10, CXCL13, CCL4, CCL5 were higher expressed in stage I and decreased in the subsequent stages.

Subsequently, all of the samples from TCGA were ordered according to the expression level of CXCL9, and statistical analysis was conducted to compare the head 80 with the tail 80 samples. The heatmap (Fig. 8b) and statistical results suggested that higher expression levels (tail 80 samples) of CXCL9, CXCL10, CXCL13, CCL4, CCL5 were related to more patients of lower Breslow depth (0–1.5 mm, 51.2%), BRAF mutation (60%) and lower Clark levels (I–III 45% and IV–V, 55%), reduced T4 stage (T4, 25%). With the decrease levels of these five chemokines (head 80 samples), there were more patients with higher Breslow depth (0–1.5 mm, 15% and > 1.5 mm, 85%), BRAF WT (wild type, 56.3%), higher Clark levels (I–III 17.5% and IV–V, 82.5%) and T4 stage (T4, 47.5%).

All of the data above suggested that five chemokine family members play critical roles in cutaneous melanoma tumorigenesis and progression.

As a result, there were nearly 6% (CXCL9), 5% (CXCL10), 3% (CXCL13), 4% (CCL4), 5% (CCL5), 3% (NMU), 4% (GAL) of SKCM samples had genetic alteration. The most common genetic change among five chemokines (CXCL9, CXCL10, CXCL13, CCL4, CCL5) was enhanced mRNA expression (Fig. 8a). While genetic change in GAL and NMU were mainly related to amplification.

Then, we assessed the correlation coefficients in prognostic genes. As expect, significant positive correlations were observed between five chemokines (Fig. 8c). Among them, CXCL9 and CXCL10 had the highest positive correlation with a Spearman’s correlation coefficient of 0.92. CXCL9 exhibited the tightest association with all other chemokines, with a median Spearman’s correlation coefficient of 0.85.

Using TRRUST database, we explored potential transcription factor targets of the five chemokines. As a result, IRF7, IRF3 (FDR = 3.09E−05); RELA, NFKB1 and IRF1 (FDR = 1.5E−04) were found to be the key transcription factors for CCL4, CCL5, CXCL10 (Table 2).

Therefore, IHC was performed to validate the expression of five chemokine family members. Consistent with the mRNA expression, we found significantly elevated CXCL9, CXCL10, CXCL13, CCL4 and CCL5 proteins expression in the SKCM than in the normal tissues. The results and the plots of IHC score were illustrated in Fig. 9.

We performed correlation analysis between prognostic genes expression and immune infiltration level for SKCM. There was a positive correlation between CXCL9 expression and the infiltration of B cells (Cor = 0.238, p = 3.58e−07), CD8+ T cells (Cor = 0.643, p = 1.94e−52), CD4+ T cells (Cor = 0.299, p = 1.19e−10), macrophages (Cor = 0.247, p = 9.59e−8), neutrophils (Cor = 0.612, p = 7.03e−48), and dendritic cells (Cor = 0.648, p = 1.56e−54; Fig. 10a). Similar results were obtained for CXCL10, CXCL13, CCL4 and CCL5 (Fig. 10b–e). While, the correlation between NMU/GAL and immune infiltration is not obvious (Fig. 10f, g). Additionally, we also found significant correlations of these seven genes with 28 types of TILs across human heterogeneous cancers (Additional file 2: Fig. S1). CXCL9 significantly positive correlated with abundance of 28 types TILs such as activated CD8 T cells (Act_CD8 T cells; rho = 0.801, P < 2.2e−16), regulatory T cells (Treg; rho = 0.693, P < 2.2e−16), macrophage (rho = 0.655, P < 2.2e−16), natural killer T cells (NK T cells; rho = 0.74, P < 2.2e−16), myeloid derived suppressor cells (MDSC; rho = 0.744, P < 2.2e−16), immature B cell (Imm_B; rho = 0.78, P < 2.2e−16). Similar results were found in CXCL10, CXCL13, CCL4 and CCL5 (Additional file 2: Fig. S1).

Additionally, in the survival model of TIMER, results showed that high levels of B cells (Log-rank P = 0), CD8+ T cells (Log-rank P = 0), neutrophils (Log-rank P = 0), dendritic cells (Log-rank P = 0), CXCL9 (Log-rank P = 0), CXCL10 (Log-rank P = 0), CXCL13 (Log-rank P = 0), CCL4 (Log-rank P = 0), CCL5 (Log-rank P = 0), were significantly related to longer survival time in SKCM patients (Additional file 3: Fig. S2).

We obtained 42 drug-gene interaction pairs in DGIdb, including genes (IL6, CXCL2, CXCL10, CCL4, CCL5, PTGER3, GAL, NPY1R) and 69 drugs, as shown in Fig. 11. This result may help develop new targets for SKCM therapy.