Cellular senescence affects energy metabolism, immune infiltration and immunotherapeutic response in hepatocellular carcinoma

Identification of different cellular senescence-related clusters

To comprehensively explore the expression patterns of cellular senescence-related genes in HCC, we downloaded the information of RNA sequencing samples and clinical information of 365 HCC patients from the TCGA database as the training set. We downloaded RNA sequencing sample information and clinical information of 461 HCC patients from ICGC database and GEO database as validation set. Based on the expression profiles of 281 cellular senescence-related genes, we stratified 365 HCC patient samples in the training set into two different clusters (233 cases in Cluster 1 (C1), 132 cases in Cluster 2 (C2) by a nonnegative matrix factorization (NMF) algorithm (Fig. 1A). In the training set, the survival of C1 patients was significantly better than that of C2 patients, and the survival curves are shown in Fig. 1B. Subsequently, we further compared the differences between clusters in terms of basic clinical characteristics, and we found that C2 patients had a greater proportion of HCC patients with Stage III and Stage IV compared with C1 patients (Log-rank test, P < 0.05), while patients with Tumor Grade G3 and G4 were significantly more represented in C2 than C1 (Log-rank test, P < 0.05). Interestingly, we found a higher proportion of old patients (age > 60) in the C2 group (Log-rank test, P < 0.05) (Fig. 1C–F). The above results suggest that there is a relationship between the expression of cellular senescence-related gene and clinical characteristics, such as Stage, Age and Tumor Grade.

Clustering of HCC patients in the training set and the relationship between the results of each cluster and clinical characteristics and survival. (A) The 365 HCC patients in the training set were clustered into Cluster1 (C1) and Cluster2 (C2) based on the expression levels of 279 cellular senescence genes. (B) Kaplan–Meier survival analysis showed a significant difference in overall survival time between C1 and C2. (C–F) The relationship between different clusters and clinical features, including age (C), tumor grade (D), M stage (E) and tumor stage (F).

Differential gene analysis and functional enrichment analysis

To further investigate the differences in gene expression and biological processes involved between different clusters, we further performed differential gene analysis between different clusters, in which 123 genes were up-regulated in C1 and 2253 genes were up-regulated in C2 (Fig. S1A–B). The GO enrichment analysis showed that the biological process (BP) of up-regulated genes in C1 was mainly enriched in small molecule catabolic process and organic acid catabolic process; the most enriched cellular component (CC) was the mitochondrial matrix and apical part of cell (Fig. S1C). The GO enrichment analysis revealed that the biological processes (BP) of upregulated genes in C2 were mainly enriched in positive regulation of cell activation and positive regulation of leukocyte activation; the most enriched cellular components (CC) were external side of plasma membrane and plasma membrane signaling receptor complex (Fig. S1D). KEGG functional enrichment analysis showed that in C1 upregulated genes were mainly enriched in Metabolism of xenobiotics by cytochrome P450 and Drug metabolism—cytochrome P450. The upregulated genes in C2 were mainly enriched in Cytokine—cytokine receptor interaction, Neuroactive ligand–receptor interaction (Fig. S1E). We then performed GSEA (KEGG) enrichment on the two clusters showing that in C1 the DRUG_METABOLISM_CYTOCHROME_P450 pathway and RETINOL_METABOLISM pathway were mainly enriched, while in C2 the OOCYTE_MEIOSIS pathway and PROGESTERONE_MEDIATED_OOCYTE_MATURATION pathway were mainly enriched (Fig. 2A). These results suggest that cellular senescence is related to cell metabolism, cytokine secretion, and immune cell activation, which encourages us to further explore the effects of cellular senescence on energy metabolism, chemokine secretion, and immune cell infiltration.

Differential expression of chemokine-related genes among different clusters. (A) GSEA enrichment analysis shows that different clusters are involved in different signaling pathways. (B) The heatmap demonstrates significant differences in the expression levels of chemokine-related genes between the different clusters. (C–F) The expression levels of chemokine-related genes were significantly higher in patients in the C2 group than in C1, including CCL26 (C), CXCL1 (D), CXCL6 (E) and CXCL5 (F).

Differences in chemokine and energy metabolism-related genes among clusters

We further compared the expression differences of chemokine and energy metabolism-related genes in C1 and C2, we found that in C2 the expression of most chemokine-related genes was significantly higher than in C1 (Fig. 2B), especially CCL26, CXCL5, CXCL6, CXCL1 (Fig. 2C–F). Similarly, we found that most of the energy metabolism-related genes were highly expressed in C2 (Fig. 3A–G). We further carried out functional enrichment analysis of the differentially expressed energy metabolism-related genes, and the results showed that these differentially expressed energy metabolism-related genes were mainly enriched in xenobiotic metabolic process, cellular response to xenobiotic stimulus, and carbohydrate biosynthetic process pathways (Fig. S2A–B). The above results further suggest that there is a link between cellular senescence and tumor microenvironment, while cellular senescence can change the energy metabolic state to some extent.

The expression levels of genes related to energy metabolism differed significantly between clusters. (A) Heatmap results showed that most energy metabolism-related genes were highly expressed in C2. (B–D) Energy metabolism-related genes such as CYP1A2 (B), CYP3A4 (C), CYP2A7 (D) were overexpressed in C1 compared to C2. (E–G) Energy metabolism related genes such as B3GALT5 (E), CYP24A1 (F), HS3ST6 (G) were highly expressed in C2 compared to C1.

Characterization of immune landscape in distinct cellular senescence clusters

Previous studies have demonstrated a relationship between cellular senescence and tumor immune infiltration in a variety of tumor types^23,24. We calculated the proportion of 21 immune cell species in each HCC sample using the R software CIBERSORT, while comparing the differences between immune cell components across clusters (Fig. 4A). Particularly, we found that regulatory T cells (Tregs) were markedly elevated in C2. Subsequently, we applied the ssGSEA algorithm to determine the relative ratios of 28 immune cells and immune-related pathways in each HCC sample, comparing the differences in immune cell composition between clusters (Figs. 4B, S3). We found that most immune cells were significantly enriched in C2, and interestingly, Myeloid-derived suppressor cells (MDSCs) cells (MDSCs) were significantly higher in C2 than in C1, which was concordant with previous observations linking C2 to an immunosuppressive phenotype. We further compared the effect of infiltrating immune cells on the survival of HCC patients, and we found that patients with highly infiltrated CD8T cells had a better prognosis; while highly infiltrated M0 Macrophages and M2 Macrophages cells had the opposite result (Fig. 4C–E). Interestingly, we found that the proportion of infiltrated Treg cells in the tumor microenvironment had no significant effect on the prognosis of HCC patients (p > 0.05) (Fig. 4F). In addition, we calculated the StromalScore, ImmuneScore, ESTIMATEScore, and TumorPurity for each HCC sample using the ESTIMATE function of R software, and the ImmuneScore of C2 was significantly higher than that of C1, in agreement with the previous results (Fig. S4A–D). Overall, these results confirm that cellular senescence is associated with the tumor microenvironment, and to some extent, contributes to the formation of an immunosuppressive microenvironment.

Analysis of immune cell infiltration in the tumor microenvironment. (A) Differences in the proportion of immune cell infiltration by tumor microenvironment species between clusters. (B) Differences in the proportion of 28 immune cell infiltrates in the tumor microenvironment between clusters. (C) Kaplan–Meier survival analysis showed that HCC patients in the highly infiltrated T cell CD8 group had longer overall survival compared to the low infiltrated T cell CD8 group. (D) Kaplan–Meier survival analysis showed that HCC patients in the highly infiltrated M0 Macrophages group had a shorter overall survival time compared to the low infiltrated group. (E) Kaplan–Meier survival analysis showed that HCC patients in the highly infiltrated M2 Macrophages group had a shorter overall survival time compared to the low infiltrated group. (F) Kaplan–Meier survival analysis showed no significant difference in overall survival time in the highly infiltrated Treg cell group compared to the low infiltrated group of HCC patients.

Landscape of tumor mutation and prediction of immunotherapy response in different clusters

We compared the differences in somatic mutations and tumor mutational burden (TMB) between clusters. Interestingly, C1 had the CTNNB1 gene as the major mutated gene, while C2 had the TP53 gene as the major mutated gene (Fig. 5A–B). We further compared the differences in TMB scores between different clusters, and the results showed that there was no significant difference in TMB scores between the two groups (p > 0.05)(Fig. 5C). MSI can be used as a marker for predictive immunotherapy in a variety of solid tumors. We further compared the MSI differences between different clusters. Interestingly, the results showed that there was no significant difference in MSI between the two clusters (p > 0.05), but the median MSI value of the C2 was higher than that of the C1 cluster (Fig. S5). To some extent, the C2 may be more responsive to immunotherapy than the C1. To further assess the differences in immune response between clusters, we compared the differences in TCR richness, BCR richness, and CTA scores between the two clusters. The results showed that C2 had a higher TCR richness, BCR richness, and CTA score (Fig. 5D–F), further suggesting that patients in the C2 group were more likely to respond to immunotherapy. The mRNA expression level of immune checkpoint-related genes is the basis of immunotherapy. Therefore, to further explore the complex communication between immunomodulators, immune infiltration, and cellular senescence, we explored the expression of immune checkpoint-related genes among different clusters. The results showed that the mRNA expression levels of immune checkpoint-related genes were significantly upregulated in C2 group when compared with C1 group (Fig. S6A–B). We found that immune checkpoint-related genes commonly used in liver cancer immunotherapy, including PD-L1, PD1, PD-L2, CTLA-4, TIGIT, and TIM-3, were significantly upregulated in C2 (Fig. 6A–F), further suggesting the existence of an immunosuppressive microenvironment in C2 and that immunotherapy could reverse this immunosuppressive state. Finally, we used the TIDE score to assess the clinical effectiveness of immunotherapy across clusters. In our results, C2 had the lower TIDE score and Dysfunction score (Fig. 6G–H), implying that patients in C2 could benefit more from immunotherapy than C1.

Tumor somatic gene mutation and immunotherapy response prediction. (A) Mutation landscape of tumor somatic cells in patients with cluster C1. (B) Landscape of tumor somatic cell mutations in patients with C2 clusters. (C) Differences in the distribution of tumor mutation burden among clusters. (D–F) Differences between different clusters in TCR Richness (D), BCR Richness (E) and CTA scores (F).

Relationship between mRNA expression levels of immune checkpoint-associated genes and different clusters. (A–F) Differences in mRNA expression levels of six immune check-related genes commonly used in HCC immunotherapy between C1 and C2, including PD-L1 (A), PD1 (B), PD-L2 (C), TIM-3 (D), CTLA4 (E) and TIGIT (F). (G-H) TIDE scores (G) and Dysfunction scores (H) were used to assess immunotherapy response.

Taken together, our comprehensive analysis showed that cellular senescence clusters are significantly associated with energy metabolism, chemokines, tumor microenvironment, patient prognosis, and immunotherapy response, which may provide new insights into HCC classification system.

Construction of the cellular senescence score for overall survival in HCC patients

To better reflect the characteristics of C1 and C2, we constructed a risk score signature to differentiate HCC patients. First, we performed differential gene analysis for C1 and C2 with the screening criteria of p-value < 0.05 and logFC > 1 or logFC < -1. A total of 8341 differential genes were obtained. Subsequently, we selected 68 genes that overlapped with the set of genes associated with cellular senescence (Fig. 7A). We identified 36 prognosis-related cellular senescence genes using Cox univariate regression analysis (Fig. 7B); we then used LASSO Cox regression analysis and multifactorial Cox regression analysis to build CSS (Fig. 7C–D). Four prognostic genes (CENPA, CXCL8, EZH2, and G6PD) were identified in the training set used to construct the CSS (Fig. 7E). The risk score was calculated using the following formula: risk score = (0.22281 × CENPA gene expression) + (0.10830 × CXCL8 gene expression) + (0.19533 × EZH2 gene expression) + (0.18866 × G6PD gene expression). Patients in the training set were divided into high-risk group and low-risk group based on the median risk score. Patients in the low-risk group had significantly higher survival rates than those in the high-risk group (Fig. 7F). As shown in the heatmap, the mRNA expression levels of 4 cellular senescence-related genes were significantly higher in the high-risk group than in the low-risk group; the mortality rate of patients gradually increased with increasing risk scores (Fig. 7G). We evaluated the predictive efficacy of risk scores on the prognosis of HCC patients using ROC curves (Fig. 7H). To further verify the superiority of the CSS, we compared the CSS with other clinical features in terms of predicting the prognosis of HCC patients, and we found that the CSS ad the largest area under the curve in predicting the prognosis of HCC patients, suggesting that the CSS was significantly superior to other clinical features (Fig. S7A). We used the validation set to determine the robustness and prognostic value of CSS. To further evaluate the validity and robustness of the CSS, two independent external data sets were used to validate the CSS, and the clinical characteristics of the 2 independent datasets are presented in Supplement Table 1. We used the same calculation formula to calculate the risk scores of each HCC patient in the validation set, and divided the HCC patients in the validation set into high-risk group and low-risk group according to the median value of the risk scores. The Kaplan–Meier analysis revealed that the overall survival time of high-risk patients was considerably shorter than that of low-risk patients in two external cohorts: GSE14250 (Fig. S7B, HR = 1.687, 95% CI 1.092–2.606, p = 0.0185) and ICGC (Fig. S7C, HR = 2.802, 95% CI 1.458–5.387, p = 0.002). In two independent external validation sets, ROC curves confirmed that risk scores can effectively predict 1-, 3-, and 5-year survival in HCC patients (Fig. S7D–E).

Construction and validation of a prognostic signature based on cellular senescence-related genes. (A) Venn diagram demonstrating the acquisition of 68 differentially expressed cellular senescence-related genes. (B) Results of univariate Cox regression analysis of differentially expressed cellular senescence-associated genes. (C) The coefficients of genes calculated by multivariate Cox regression using LASSO. (D) The partial likelihood deviance of genes. (E) Results of multivariate Cox regression analysis of four differentially expressed cellular senescence-related genes. (F) Kaplan–Meier curves were used to compare the overall survival of HCC patients between the high-risk and low-risk groups. (G) The association of risk scores with survival status and gene expression in HCC patients.(H) ROC curves of the prognostic signature for predicting the risk of death at 1, 3, and 5 years.

Meanwhile, we further explored the relationship between CSS and different clusters, and we found that most patients in C2 cluster were in the high-risk group; similarly, most patients in the low-risk group belonged to C1 cluster (Fig. S7F). These results suggest that CSS can effectively respond to cellular senescence clusters. We used ROC curves to verify the mapping relationship between CSS and cellular senescence clusters, and the results showed that the CSS had better predictive efficacy for cellular senescence clusters (AUC = 0.893)(Fig. S8A). Therefore, we believe that the risk score of this signature can well reflect the characteristics of cellular senescence in HCC.

Identification and validation of prognostic signature as an independent prognostic factor

To further investigate the prognostic value of the prognostic signature for patients with HCC, we performed univariate Cox regression analysis and multivariate Cox regression analysis of risk scores with other clinical characteristics. The results suggested that Stage and risk score were independent risk factors affecting HCC patients (Fig. 8A–B). To better predict the probability of survival in HCC patients, we created a predictive nomogram based on the integration of risk scores and Stage (Fig. 8C). ROC curves confirm that nomogram can effectively predict the survival of HCC patients at 1, 3, and 5 years (Fig. S8B). Calibration curves further confirm that nomogram can effectively predict the overall survival time of HCC patients (Fig. S8C). This indicates high accuracy of our nomograms. In addition, decision curves confirmed that nomograms constructed based on risk scores and Stage significantly outperformed nomograms constructed on other clinical characteristics (Fig. S8D).

Construction of nomograms and validation of immunotherapy responses in multiple independent external datasets. (A) Results of univariate Cox regression analysis of risk scores and clinical characteristics. (B) Results of multivariate Cox regression analysis of risk scores and clinical characteristics. (C) Nomogram constructed from Stage and risk score. (D) In the IMvigor210 dataset, patients who responded to immunotherapy (CR/PR) had significantly higher risk scores than patients in the low-risk group. (E) In the GSE91061 dataset, patients who responded to immunotherapy (CR/PR) had a higher median risk score than patients in the low-risk group. (F) In the IMvigor210 dataset, the percentage of patients who responded to immunotherapy (CR/PR) was significantly higher in the high-risk group than in the low-risk group. (G) In the GSE91061 dataset, the percentage of patients who responded to immunotherapy (CR/PR) was higher in the high-risk group than in the low-risk group.

Independent external immunotherapy data validate the prediction of immunotherapy response

To further validate the predictive value of the CSS for immunotherapy response, we downloaded two independent external immunotherapy data for a comprehensive analysis (IMvigor210, GSE9106). We calculated risk scores for patients receiving immunotherapy in both immunotherapy datasets using the same formula as in the training set, and divided patients receiving immunotherapy into high-risk and low-risk groups based on the median value of the risk scores. We found a significant correlation between risk score and immunotherapy response (Fig. 8D–E), with a significantly higher proportion of CR/PR patients in the high-risk patients than in the low-risk group in the two independent external immunotherapy data (Fig. 8F–G). These validation results all confirm that patients in the high-risk group may be more responsive to immunotherapy and more suitable for immunotherapy. Thus, our novel CSS can effectively predict the response to immunotherapy in HCC patients.

qRT-PCR confirms overexpression of four cellular senescence-related genes and two chemokine-related genes in HCC tissues

To validate the robustness of the CSS, we collected tumor tissue samples and paraneoplastic tissue samples from 12 patients with HCC confirmed by postoperative pathological pathology from the clinic and verified the mRNA expression levels of four cellular senescence-related genes and two chemokine-related genes using qRT-PCR. The qRT-PCR results demonstrated that four cellular senescence-related genes (CENPA, CXCL8, EZH2, and G6PD) and two chemokine-related genes (CCL26 and CXCL5) were overexpressed in tumor tissues from HCC patients compared with paraneoplastic tissues (Fig. 9). These results suggest that our novel prognostic signature constructed based on cellular senescence-related genes is highly robust, and this robustness was validated in clinical samples.

The mRNA expression of four cellular senescence-related genes and two chemokine-related genes in HCC tissues was confirmed by qRT-PCR. (A–B) The qRT-PCR results confirmed for 2 chemokine-related genes overexpressed in HCC tissues, including CXCL5 (A), CXCL26 (B). (C–F) The qRT-PCR results confirmed that four cellular senescence-associated genes used to construct the prognostic model were overexpressed in HCC tissues, including CENPA (C), CXCL8 (D), EZH2 (E), and G6PD (F).