2.1 Data collection and preprocessing

In this study, TCGA (https://portal.gdc.cancer.gov/) database provided a TCGA-SKCM data set as a training cohort, containing the gene expression matrix of 459 CM samples and their corresponding clinical follow-up data. The GSE22153, GSE15605, GSE22154, GSE43955, GSE46517, GSE54467, and GSE65904 data sets were downloaded from the GEO (https://www.ncbi.nlm.nih.gov/gds/) database. As for GSE54467 and GSE15605, we downloaded log₂-transformed and quantile normalization matrix data, whereas GSE22153, GSE46517, GSE22154, GSE46517, and GSE65904 data sets were converted and normalized by log₂ after download. Usually, we used the median of the expression value as the expression of genes with multiple probes. The downloaded matrix was an mRNA expression profile in fragments per kilobase of transcript per million format. The samples used in this study had to meet the following criteria¹: samples with nonzero probe expression accounted for 80% of all samples²; patients had accurate survival status and follow-up time.

The GSE65904 data set, containing information about DSS (210 patients) and DMFS (150 patients), was used as a test cohort. In addition, the GSE22153, GSE22154, GSE46517, and GSE54467 data sets, containing only the OS data (236 patients), were integrated into the validation cohort. While integrating the validation data set, we used the combat function of R package “sva” to remove the batch effect. Table S1 shows the basic data of the data sets included in this study.

CellAge⁸ (https://genomics.senescence.info/cells/), which contains 279 aging-related genes, is a database of genes associated with cell senescence. The Molecular Signatures Database (MSigDB)^{9, 10} is a collection of annotated gene sets for use with GSEA software. The three aging gene sets (GOBP_AGING, GOBP_CELL_AGING, and GOBP_CELLULAR_SENESCENCE) were downloaded from the MSigDB database (http://www.gsea-msigdb.org/gsea/msigdb/index.jsp). These three aging-related gene sets contained nearly all the genes that play an important role in aging involved in the existing aging research (n = 304). We combined two aging-related gene data sets (CellAge and MSigDB) to obtain 313 aging-related genes. We intersected these 313 genes with TCGA-SKCM dataset and all genes in the GEO data set involved in this study using an upset analysis, and we finally obtained 221 aging-related genes. The “Upset” analysis was performed by executing the “Upset” function in the “UpSetR” package, and the VENN circle was drawn using the “ggvenn” function in the “yyplot” package.

2.2 Associations between the cluster and immunity

The immune score and stromal score of each CM patient in TCGA-SKCM cohort were calculated using the ESTIMATE¹¹ and Xcell (https://xcell.ucsf.edu/)¹² website, and the differences in the immune score and stromal score among the different groups were further analyzed using the Wilcoxon test.¹³ The estimate score is the combination of the stromal score and the immune score. It is an index to evaluate the purity of a tumor.¹¹ Based on the matrix of TCGA-SKCM cohort, we calculated the abundance of immune cells by using CIBERSORT,¹⁴ MCP Counter,¹⁵ TIMER (http://timer.comp-genomics.org/),¹⁶ EPIC, and Xcell software packages and excluded the samples with p > 0.05. Finally, the Mann–Whitney U-test was used to analyze the differences in the immune cell subtypes among the different groups.

2.3 Tumor somatic mutation analysis

The tumor mutation load (TMB) has attracted much attention in immunotherapy. TMB and PD-L1 are two important biomarkers to predict the response to PD-1 antibody therapy. We used the waterfall function of the “maftools” package to show mutations in patients with high senescence and low-risk scores in TCGA-SKCM cohort. Missense, nonsense, uninterrupted, silence, and frameshift/in-frame insertions and deletions were counted, while synonymous mutations were excluded.¹⁷ The total number of somatic mutations was used to calculate the TMB score. According to the median of the TMB score, all CM samples with somatic mutations in TCGA data set were divided into a high TMB score group and a low TMB score group.

2.4 Prognostic assessment using the RSS

By using the “Survival” package of R software, we used a univariate Cox analysis to calculate the prognostic value of each SRG and selected an SRG as the seed SRG for the Cox-lasso regression analysis. Then, a multivariate Cox regression analysis was performed to evaluate the prognostic characteristics of each SRG using the R packages “Survminer”, “Glment”, and “Survival”. The risk score of each patient in the training group was calculated using the Cox proportional hazards model (PH model): ${\widehat{h}}_i\left(t\right)={\widehat{h}}_0\left(t\right)\exp \left(x_i^{\prime }\widehat{\beta }\right)$ (where exp is the prognostic gene expression level; β is the multivariate Cox regression model regression coefficient; and $h_0\left(t\right)$ is the baseline hazard function. This is a function that describes the “instantaneous mortality” that specifically refers to the instantaneous mortality of the observed object at the time of survival to t). All of the CM samples were randomly divided into a high-risk score group and a low-risk score group according to the median of the risk score. The Kaplan–Meier method was used for the survival analysis, and the log-rank test was used to compare survival between the groups. The R software package, “SurvivalROC”, was used to draw the receiver operating characteristic curve (ROC) curve and calculate the corresponding area under the curve (AUC).

3.1 Cluster 1 showed great advantages in survival and nonrecurrence compared with cluster 2

Figure S1 shows the entire analytical process of the study.

First, we screened the coexisting SRGs in TCGA, GSE22153, GSE22154, GSE46517, GSE54467, and GSE65904 databases using UPSET and identified a total of 221 SRGs (Figure S2A). Then, we merged four data sets, GSE22153, GSE22154, GSE46517, and GSE54467, and eliminated the batch effect (Figure S2B). The results of the principal component analysis (PCA) before and after merging the data sets showed that the batch effect was well eliminated (Figure S2C,D). The determination of the K value is very important for the consistency of a cluster analysis. When the cumulative distribution function (CDF) reaches the approximate maximum value, the cluster analysis result is the most reliable. Typically, a K value with small decline slope of the CDF is used. Based on unsupervised clustering, TCGA-SKCM was divided into two clusters (Figure 1A,B). According to this method, it is really the best when k = 2. According to the results of prognostic analysis between the two groups, cluster 1 showed great advantages in survival and nonrecurrence (Figure 1C,D). Figure 1E shows a multivariate heat map that contains data on the expression of 221 SRGs and clinicopathological data. There was a significant difference in the multiple SRGs between the two clusters (Figure 1F). The age of patients in cluster 2 was significantly higher than that of patients in cluster 1 (Figure S3A). Therefore, we labeled the clusters 1 and 2 as the younger group (YG) and older group (OG), respectively.

3.2 YG was primarily enriched in the immune response and immunotherapy pathways

We analyzed the differential expression of 58,385 probes between the YG and OG and found that 356 protein-coding genes were upregulated in the OG, and 1048 protein-coding genes were upregulated in the YG (Figure S3B). The upregulated genes in the OG and YG were analyzed using a functional enrichment analysis of the Gene Ontology Biological Processes (GO-BP) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways, respectively. The results of the KEGG and GO-BP analyses showed that the upregulated genes in the YG were primarily concentrated in the T-cell-receptor signaling pathway, the B-cell-receptor signaling pathway, the positive regulation of lymphocyte activation, the positive regulation of biological process, PD-L1 expression, and the PD-1 checkpoint pathway in cancer, and Th17 cell differentiation (Figure 2A,B). According to the KEGG and GO-BP analysis, the upregulated genes in the OG were primarily concentrated in the skin epidermis development, the regulation of epidermis development, the regulation of epidermal cell differentiation, the IL-17 signaling pathway, the extracellular matrix (ECM)–receptor interaction, skin development, and the adaptive immune response (Figure 2C,D). The upregulated genes in the YG were primarily enriched in the immune response and immunotherapy pathways, whereas those in the OG were primarily enriched in the aging-related pathways, such as cell and epidermal development and differentiation. Therefore, we speculated that the immune landscape and response to immunotherapy between the YG and the OG may be different.

3.3 Difference in the immunotherapy response and immune landscape between the OG and the YG

Based on the results of the TIDE analysis, we further found that the response of patients in the YG to the immune checkpoint blockade was significantly higher than that in the OG (Figure 3A,B). In addition, the Kaplan–Meier (KM) analysis in patients who received immunotherapy and chemotherapy showed that the response of patients in the YG to immunotherapy and chemotherapy was significantly better than that of patients in the OG (Figure 3C,D). In addition, we compared the expression of multiple immune checkpoints in the B7 family and the CD28 family between the YG and the OG. The results showed that the expression levels of CD274, CTLA4, PDCD1, CD80, CD86, CD28, ICOS, and ICOSLG in the YG were significantly higher than that in the OG (Figure 3E). To explore the mechanism between the immunotherapy response and the tumor immune microenvironment, we further evaluated the differences in the tumor immune microenvironment between the two groups. The TME contains immune, interstitial, and endothelial cells, as well as inflammatory mediators and ECM molecules.¹³ The results showed that the immune, stromal, and estimated scores of patients in the YG were significantly higher than those of patients in the OG (Figure S4A). The results of Xcell validated that the tumor immune microenvironment score of the YG was higher than that of the OG (Figure S5C). Through a KM analysis, we found that the prognoses of CM patients with higher stromal scores, immune scores, and ESTIMATE scores were better than that of CM patients with lower scores (Figure S4B–D).

Based on the fact that the prognosis of patients with different immune scores was quite different, we further explored the landscape of immune infiltration between the two groups of patients. Through a comparative analysis of the results of CIBERSORT and MCPcounter, we found that the infiltration abundances of CD8⁺ T cells, NK cells, monocyte lines, macrophage cell line, and B-cell lines in patients in the YG were significantly higher than that in patients in the OG, while the abundances of neutrophils in the YG was significantly lower than that in patients in the OG (Figure S6A,B,D). Then, by regarding immune cells as prognostic factors and combining them with prognostic data, a univariate Cox regression analysis was conducted for each type of immune cells. We found that the CD8⁺ T cells, NK cell lines, B-cell lines, monocyte lines, and macrophage cell line were associated with a better prognosis in patients with CM (Figure S6C,E). In addition, we compared the results of TIMER, EPIC, and Xcell and found that the CD8⁺ T cells, NK cells, monocyte lines, and B-cell lines in patients in the YG were still significantly more abundant than those in patients in the OG (Figure S5A–C). Moreover, the abundance of melanocyte infiltration in patients in the OG was higher than that in patients in the YG (Figure S5C). The univariate Cox analysis also revealed that the abundances of B cells, NK cells, CD8⁺ T cells, B-cell lines, and dendritic cells were associated with a better prognosis in patients with CM (Figure S7A–C). The abundance of melanocytes was associated with a poor prognosis in patients with CM (Figure S7A).

3.4 The TMB of the YG was significantly higher than that of the OG

Figure S8 shows a summary of mutations in the CM patients in TCGA database provided by “maftool”. The mutation waterfall map (Figure S9) of all the CM patients indicates that the mutation rate of BRAF in CM patients was 47%. The mutation rate of recorder NRAS was 25%. Through the somatic mutation analyses of the YG and the OG, we showed that the total mutation rate and the mutation rate of common mutation genes in the OG were lower than those in the YG (Figure S10A,B). By calculating the tumor mutation load scores of the two groups, we found that the TMB of the YG was significantly higher than that of the OG (Figure S10D). The KM analysis showed that the prognosis of the CM patients with high TMB scores was significantly better than that of the CM patients with low TMB scores (Figure S10C). In addition, the NRAS mutation rate (21%) in the YG was significantly lower than that in the OG (38%) (Figure S10E).

3.5 Construction of the RSS for CM

In total, 94 SRGs (p < 0.05) selected using the univariate Cox regression analysis were included in the LASSO regression analysis (Figure S11A). According to the standard of lambda.1se (lambda value = 4) (Figure S11B,C), four molecules, namely, IL15, B2M, FOXM1, and KIR2DL4, were selected for the multivariate Cox regression analysis (Figure S11D). We calculated the risk score of the RSS according to the formula and according to the median risk score, and the CM patients in the training cohort and validation cohort were divided into a high-risk score group and a low-risk score group (Figure S12A,B). We found a significant difference in the expression of the four SRGs between the high-risk score group and the low-risk score group in both the training cohort and the validation cohort (Figure S12C,D). It was obvious that the expression of KIR2DL4 in the low-risk score group was significantly higher than that in the high-risk score group. In contrast, the expression of FOXM1 in the low-risk score group was significantly lower than that in the high-risk score group.

To validate the prognostic ability of RSS, we performed ROC and KM analyses. In the training and validation cohorts, the OS and recurrence free survival (RFS) of the low-risk score group were significantly better than those of the high-risk score group (log-rank test p-value < 0.0001, Figure S13A,C,D). The results of the ROC analysis showed that the AUC of 5-year OS in the training and validation cohorts was 0.711 and 0.830, respectively (Figure S13B,E). To further understand whether RSS has the same predictive effect on DSS and DMFS in patients with CM, we performed a similar analysis in the test cohort (n = 210). Similarly, the DSS and DMFS of CM patients with high-risk scores were significantly lower than those of CM patients with low-risk scores (log-rank test p-value < 0.001; Figure S13F,H). In addition, the ROC analysis showed that the RSS predicted that the AUC of 3-year DSS and DMFS in patients with CM was 0.661 and 0.698, respectively (Figure S13G,I). These results suggest that RSS can accurately predict the prognosis of patients with CM.

Based on the results of the TIDE analysis, we further found that the response to the immune checkpoint blockade among patients in the high-risk score group was significantly lower than that in the low-risk score group (Figure 4B,C). In addition, the KM analysis of the two groups of patients who received immunotherapy and chemotherapy showed that the response of the low-risk score group to immunotherapy and chemotherapy was significantly better than that of the high-risk score group (Figure 4A,D). In addition, we compared the expression differences of multiple immune checkpoints in the B7 family and CD28 family between the low-risk score and high-risk score groups. We found that the expression levels of CD274, CTLA4, PDCD1, CD80, CD86, CD28, ICOS, and ICOSLG in the low-risk score group were significantly higher than those in the high-risk score group (Figure 4E).

3.6 Difference in the expression of four molecules in the RSS between CM and normal skin tissues

Based on the results of GEPIA2, the expression levels of B2M, KIR2DL4, and FOXM1 in CM patients were significantly higher than those in normal skin tissues, but there was no significant difference in the expression of IL15 between the two groups (Figure 5A). Next, we verified the expression of these four genes in two CM cohorts containing normal skin tissue samples. The results showed that the expression levels of B2M, FOXM1, and KIR2DL4 were still high in CM patients. Surprisingly, in the GSE46517 cohort, the expression of IL15 in the tumor group was significantly higher than that in the normal skin tissue (Figure 5B,C).

3.7 Validation of the RSS in various clinical subgroups

Clinically, the pathological stage¹⁸ is an important factor that affects the survival of patients with CM, whereas other key factors include sex and age.¹⁹ Therefore, we validated the ability of RSS to predict the prognosis of CM patients in different stages, age, and sex, both in the training and the validation cohorts. RSS accurately stratified the risk in all four stages, both sexes, and both age categories (age ≥ 65 years and age < 65 years) (Figure S14).

3.8 KIR2DL4 activated NK cells by activating the MEK/ERK pathway

We obtained 13 frozen samples of skin melanoma patients who received nivolumab (NIVO) (PD-1 inhibitors) and ipilimumab (IPI) (CTLA-4 inhibitors) treatment from March 2018 to March 2022 from the Department of Pathology of the third people's Hospital of Chengdu. The methodology of immunohistochemistry and immunofluorescence had been provided in the supplementary document of the manuscript. Immunohistochemical and immunofluorescence staining were performed on it (Figure S15A,B). Obviously, the results of immunofluorescence suggest that KIR2DL4 is indeed co-located with NK cells. The results from immunohistochemistry showed that, among the 13 specimens, 7 had high expression of KIR2DL4 and 6 had low expression. The OS time of CM patients with high expression of KIR2DL4 was significantly longer than that of patients with low expression of KIR2DL4 in CM patients (Figure S15C). In addition, the results of Kaplan–Meier analysis also suggested that OS in patients with high expression of KIR2DL4 was better than that of patients with low expression of KIR2DL4 (Figure S15D).

We isolated NK cells from melanoma tissue using flow cytometry (Figure S15A) (sorting–culture–re-sorting–re-nourishment). We first explored whether KIR2DL4 directly regulated the cytolytic activity of NK cells. We knocked down KIR2DL4 from the isolated NK cells with siRNA, and then we washed and co-cultured with melanoma cells to evaluate the cell lytic activity using a lactate dehydrogenase (LDH) assay. As shown in Figure S15B, the cytolytic activity of the NK cells decreased significantly compared with the control group, and the maximum killing rate was 1/4 of that of the control group. Next, we used enzyme-linked immunosorbent assay (ELISA) to evaluate the effect of KIR2DL4 on the cytokine production in the NK cells. As shown in Figure S15C,D, the KIR2DL4 knockdown significantly reduced the production of IFN-γ and TNF-α by the NK cells. To explore how KIR2DL4 activates NK cells, we constructed an overexpression plasmid of KIR2DL4 and transfected it into the NK cells. The results showed that the classical NK cell activation pathway, namely, the MEK/ERK signal pathway, was significantly activated, and the phosphorylated MEK/ERK protein was significantly upregulated after overexpression of KIR2DL4 (Figure S15E).

3.9 Establishment of a nomogram based on the RSS and clinical features

To further stratify the risk in CM patients and validate the prognostic predictive ability of RSS, we constructed a prognostic nomogram. We included four variables, including age, sex, stage, and the risk score, into univariate and multivariate Cox regression analyses (Figure 6A,B). The risk score was independently associated with poor prognosis of patients with CM (HR = 2.613, p < 0.05). Next, we used these four features to construct a prognostic nomogram, and each subtype of the clinical features in this model corresponded to a specific score (Figure 6C). The vertical red line in the figure indicates the score of a CM patient (TCGA-DA-A960-01) included in the model. She was female, 73 years old, the AJCC grade was grade II, and the risk score was 1.711751, belonging to the high-risk score group. After calculation, her total risk score was 180, and the probability that her survival time was less than 3 years, 5 years, and 7 years was 0.517, 0.694, and 0.832 respectively. In fact, her survival time was 804 days, less than 3 years. The above results verified that our prognosis prediction model was accurate and effective. Finally, the total score of all clinical factors corresponded to a specific 3-, 5-, and 7-year survival probability. The C index of the nomograms of the training cohort and the validation cohort were 0.697 (95% CI, 0.695–0.699) and 0.762 (95% CI, 0.746–0.778), respectively. The calibration curve of the training cohort showed that the predicted values were consistent with the observed values of the 3-, 5-, and 7-year OS (Figure 6E). We calculated the total risk score of each patient according to each predictor in the nomogram model and took the median total risk score as the cut-off value to divide all CM patients into high-risk patients and low-risk patients. The KM analysis showed that the survival of high-risk patients was significantly poorer than that of low-risk patients (Figure 6D). The ROC analysis showed that the AUC values of 3-year, 5-year, and 7-year prognosis of CM patients by the total risk score reached 0.767, 0.758, and 0.781, respectively (Figure 6G). In addition, the AUC value of the total risk score was always higher than that of the AJCC stage. Finally, the clinical decision curve analysis showed that the patient benefit rate of the model excluding the risk score was significantly lower than that of the model excluding the AJCC stage (Figure 6F). Unsurprisingly, the model that included both the risk score and the AJCC stage showed the highest clinical benefit rate. In addition, The validation cohort can accurately validated the model (Figure S17).

3.10 GSEA

The results of GSEA showed that the cell cycle pathway was enriched in the high-risk score group (Figure S18). However, the pathways that were enriched in the low-risk score group were mainly the immune response-related pathways, such as the T-cell-receptor signaling pathway, B-cell-receptor signaling pathway, NK cell-mediated cytotoxicity, primary immunodeficiency, cell adhesion molecules (CAM), cytokine–cytokine receptor interaction, NOD-like receptor signaling pathway, PD-L1 expression, and the PD-1 checkpoint pathway in cancer, JAK/STAT signaling pathway, and Toll-like receptor signaling pathway. This suggested that the patients in our low-risk score group were in a state of immune activation, and the effect of immunotherapy was better.