2.1 Acquisition of liver cancer sample data

Gene expression and clinical data of patients were obtained from the UCSC Xena (https://xenabrowser.net/datapages/), the Cancer Genome Atlas (TCGA) Liver Hepatocellular Carcinoma (https://portal.gdc.cancer.gov/), and International Cancer Genome Consortium (ICGC) Liver Cancer-RIKEN, Japan databases (https://dcc.icgc.org/). We use the Strawberry Perl software and tidyverse R software package to integrate gene expression data in the “TCGA-GTEX-Gene-Counts” file downloaded from the UCSC Xena database and the “LIRI-JP” file downloaded from the ICGC database. Differentially expressed genes were identified among normal and tumor samples (p < .05, FDR < 0.05, and $\left|\log 2\mathrm{FC}\right|>1.0$) based on transcriptomic data analysis results obtained by the DESeq2 package. According to the differentially expressed genes, the “TCGA-LIHC” data in TCGA database and the “LIRI-JP” data in the ICGC database were integrated to generate an FPKM gene matrix for further analysis.

2.2 Identification of cellular senescence-associated lncRNAs

We logged onto the GeneCards website of the human gene database (https://www.genecards.org/), and used “cellular senescence” as the keyword withgift score > 10.0 and relevance score > 4.0 as screening criteria to determine the cellular senescence genome. Subsequently, we took the intersection of cellular senescence genome and FPKM gene matrix, and analyze the intersection genes with the limma R package (p < .05, FDR < 0.05, and $\left|\log 2\mathrm{FC}\right|>1.0$) to identify the differentially expressed cellular senescence genes and draw a volcano map. Finally, we use the Strawberry Perl software and limma R package to perform a Pearson correlation analysis on the differentially expressed cellular senescence genes and lncRNAs to determine cellular senescence-related lncRNAs (PCCs > 0.4 and p < .001) and draw a network diagram.

2.3 Establishing the risk signature

2.4 Validating the risk signature

To compare the prognosis of low- and high-risk subgroup samples in each cohort, risk score and survival status evaluation and survival analysis curve construction were performed in three independent cohorts (i.e., training, validation, and entire cohort) by implementing the survival, survivor, and phenotmap R packages, respectively. Temporal receiver operating characteristic (ROC) curves of the model at 1, 2, and 3 years were subsequently constructed in three independent cohorts by executing the survival, timeROC, and survminer R packages, respectively. To further verify the effectiveness of the model score, R packages such as survival, caret, and glmnet were used to perform Cox regression analysis to determine whether the risk score and other clinical features were independent variable factors; ROC curves were constructed to compare the accuracy of outcomes predicted by different clinical characteristics. The nomogram including risk scores, gender, and stage was subsequently used to predict patient OS at 3 years. The calibration curves and the concordance index (C-index) were used to determine matching between the predicted and the actual results. In addition, to observe whether missing samples (60 of 562 samples are missing) have an impact on the cohort, we perform cross-validation on subgroup clinical characteristics to test cohort independence. Finally, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were performed on differentially expressed genes in the low- and high-risk groups using the clusterProfiler, enrichplot, and GOplot R packages (p < .05).

2.5 Tumor mutation burden analysis

Patients with poorer prognosis are often accompanied by more gene mutations. Therefore, nucleotide mutation data of the TCGA Liver Hepatocellular Carcinoma and ICGC Liver Cancer-RIKEN, Japan, cohorts were selected for tumor mutation burden (TMB) analysis to observe the differences in TMB values between low- and high-risk subgroups in an entire cohort (62 of 562 samples are missing). Using limma and ggpubr R packages and Strawberry Perl software, samples were divided into low (L)-TMB and high (H)-TMB groups (TMB = 3.00), and differences in TMB were compared between samples from the low- and high-risk subgroups; survival curves were then constructed for these groups using the survival and survminer R packages. Furthermore, limma and ggplot2 R packages were used to construct a scatter plot to identify correlation between risk scores and TMB. Finally, survival curves were constructed for the risk subgroups in combination with TMB to further compare the OS of subgroup patients.

2.6 Analysis of the TME and tumor-infiltrating immune cells

To explore differences in the immune microenvironment between low- and high-risk subgroup samples, we analyzed the TME scores and proportion of tumor-infiltrating immune cells in the subgroups. TME scores were calculated for differentially expressed cellular senescence genes in the low- and high-risk groups using the estimate R package. The preprocessCore and limma R packages were then used to evaluate the proportional difference of tumor-infiltrating immune cells between the low- and high-risk groups and draw a violin chart (p < .05); the CIBERSORT algorithm was used, and Kaplan–Meier survival curves were constructed for the immune cell scores (p < .05).

2.7 Reclustering model-related genes

To observe the expression of cellular senescence genes associated with model lncRNAs (AGAP11 and FAM182B) in the samples, the limma and reshape2 R packages were used to perform correlation analysis on model lncRNAs and these genes. Cellular senescence genes associated with the AGAP11 and FAM182B genes were listed in the form of a correlation heatmap (p < .05), and the samples were clustered according to these genes using the ConsensusClusterPlus R package. A survival analysis and an analysis of immune composition were then performed on the recluster groups. Finally, a Sankey diagram was plotted for AGAP11 and FAM182B genes versus target genes and presented in boxplots using the ggpubr and limma R packages ($\left|\text{PCCs}>0.4\right|$, p < .001).

2.8 Potential drug target predictions

To clarify the level of immune escape ability of the samples, immune escape scores were compared between the low- and high-risk groups, which were obtained from the tumor immune dysfunction and exclusion (TIDE) website (http://tide.dfci.harvard.edu/). The pRRophetic R package was used to assess treatment response in the low- and high-risk groups based on half-maximal inhibitory concentrations of the samples (p < .05). Information pertaining to the genes was obtained from the CellMiner website (https://discover.nci.nih.gov/cellminer/home.do), and the potential of model lncRNAs as therapeutic targets was predicted by evaluating the correlation between genes and drug targets (p < .05).

3.1 Acquisition and identification of cellular senescence-associated lncRNAs

The study flowchart is shown in Figure 1. Transcriptome data for 614 tumor samples and 252 normal tissue samples were obtained and integrated from the UCSC Xena, TCGA, and ICGC databases. The analysis results of Count gene matrix by the DESeq2 R package showed that 3959 genes (including 59 lncRNAs) with $\left|\log 2\mathrm{FC}\right|>1.00$, FDR < 0.05, and p < .05 were considered differentially expressed genes (Data S1). Using the limma R package to analyze the intersection FPKM gene matrix of differentially expressed genes and cellular senescence genome in normal and tumor samples, 109 differentially expressed cellular senescence genes with $\left|\log 2\mathrm{FC}\right|>1.00$, FDR <0.05, and p < .05 were identified; 35 of these genes were downregulated, while the others were upregulated (Figure 2A; Data S2). On subsequent correlation analysis of 109 differentially expressed cellular senescence genes, 18 lncRNAs with PCCs > 0.4 and p < .001 were selected as cellular senescence-related lncRNAs for this study (Figure 2B; Data S3).

3.2 Construction of the prognostic model

The Uni-Cox regression analysis used the bootstrap resampling method based on the optimal risk coefficient and p-value to screen a group of survival-related lncRNAs from cell senescence-related lncRNAs (Figure 3A); eight lncRNAs were selected to be associated with OS (p < .05). Overfitting of the prognostic signature was avoided by performing a LASSO regression analysis (Figure 3B,C) to further screen for OS-related lncRNAs when the first-order value of log (λ) demonstrated the least likelihood of bias (p < .05). The Multi-Cox regression analysis of the LASSO regression results using a bootstrap resampling method identified the best survival-associated model lncRNAs, including AGAP11 and FAM182B (Figure 3D). Finally, the median risk score calculated from the risk score formula was used to establish low-risk and high-risk subgroups (Data S4).

3.3 Verification of the prognostic model

Based on subgroup analysis, the entire cohort was divided into training and validation cohorts. In three independent cohorts, the low-risk group performed better in terms of risk score distribution (Figure 4A), survival status (Figure 4B), and survival time (Figure 4C). The heatmaps show the model gene expression for the low-risk and high-risk groups (Figure 4D). Cross-validation showed that there were no significant differences in clinical characteristics between the cohorts (p > .05), proving that the cohorts were independent of each other (Table 1). Kaplan–Meier survival curves based on clinical characteristics of the entire cohort showed statistically significant differences between the low- and high-risk groups (Figure 5A–L).

3.4 Assessment of clinical factors of the prognostic model

Cox regressions were used to evaluate the clinical characteristics of the entire cohort; in the Uni-Cox (Figure 6A) and Multi-Cox regressions (Figure 6B), the hazard ratios for the risk scores were 1.456 and 1.534, respectively, and the p value were less than .05, which proved that risk scores could be recognized as an independent risk factor. Notably, except for cases with meaningful risk scores, the p value for the tumor stage was <.05. The C-index (Figure 6C) was >0.6 for both the risk score and tumor stage, which proves that the cellular senescence-related prognostic model indicated good accuracy and validity in predicting the OS of samples. Finally, a nomogram for the 1–3-year OS was generated based on the model (Figure 6D).

3.5 ROC curves and calibration curve validation of the prognostic model

Temporal ROC curves were utilized to assess the sensitivity and specificity of the model in predicting prognosis; the area under the curve (AUC) was also calculated. The AUCs for the training cohort (Figure 7A), validation cohort (Figure 7B), and the entire cohort (Figure 7C) were all >0.65, which indicated that the model was reliable in predicting the OS of patients. Clinical ROC curves were also used to evaluate the clinical characteristics of the model; the AUCs for the risk score and total nomogram score were 0.674 and 0.729, respectively, further illustrating the sensitivity of the model (Figure 7D). Finally, verification using calibration curves showed the nomogram to have considerably reliable predictive performance (Figure 7E).

3.6 GO and KEGG enrichment analysis of risk subgroups

To investigate differences in genes between risk subgroups, GO and KEGG analyses were performed on differentially expressed genes (p < .05) (Figure 8A), with multiple pathways of GO enrichment possibly related to tumor invasion and immune infiltration; the included pathways positively regulated cell adhesion and T cell activation. A gene set enrichment analysis was then performed for the risk subgroups based on KEGG pathways. In the low-risk group, 7 of the top 12 pathways (p < .05, FDR < 0.25, and $\left|\text{normalized enrichment score}\right|>1.5$) with enrichment showed high correlation with immune infiltration and tumor invasion; these included the vascular endothelial growth factor pathway and the T cell receptor signaling pathways (Figure 8B).

3.7 Analysis of TMB in the prognostic model

Calculation of TMB values for the risk subgroups revealed a significantly higher TMB value for the high-risk group (Figure 9A). The survival curve showed that grouping into L-TMB and H-TMB groups impacted OS (Figure 9B); in addition, correlation analysis showed a positive correlation between risk scores and TMB values (Figure 9C). Combining TMB groups with risk subgroups markedly affected survival, with the H-TMB + high risk group demonstrating significantly poorer prognosis than the L-TMB + low risk group (Figure 9D). High-risk subgroups had higher TMB values ​​and shorter OS values. The combined analysis of risk subgroups and TMB values ​​further verified the reliability of the model to a certain extent.

3.8 Differences in the TME between risk subgroups

An analysis of differences in the TME between risk subgroups divided based on the model established with FAM182B and AGAP11 genes revealed that the low-risk group had higher scores (Figure 10A, Data S5). An analysis of tumor-infiltrating immune cells showed statistically significant differences between the low- and high-risk groups in nine groups of immune cells, including CD8 and CD4 memory T cells (Figure 10B, Data S6). Subsequent survival analysis revealed that CD8 T cell (Figure 10C), memory B cell (Figure 10D), γδ T cell (Figure 10E), and CD4 memory T cell (Figure 10F) scores had a significant impact on survival. Furthermore, survival analysis showed that the tumor-infiltrating immune cell types with poorer prognosis were more prevalent in the high-risk group.

3.9 Reclustering analysis based on the correlation of model lncRNAs

The correlation heatmap shows the differentially expressed genes associated with model lncRNAs (Figure 11A; Data S7), and the network diagram shows the association between the model lncRNAs and the differentially expressed genes (Figure 11B). Samples consisting of these differentially expressed genes were reclustered and divided into two groups (as the optimal choice) (Figure 12A–C); the survival curves indicated that this grouping had an effect on survival time (Figure 12B). An analysis of immune composition between the recluster-groups showed that four groups of immune cells, including macrophages and tumor-infiltrating lymphocytes, exhibited statistically significant differences (Figure 12C). Finally, a Sankey diagram (Figure 12D) and boxplot (Figure 12E) were constructed to visualize the differentially expressed genes closely related to the model lncRNAs ($\left|\text{PCCs}>0.4\right|$, p < .001).

3.10 Immune escape and drug correlation analysis

There were significant differences in the immune escape scores between the low- and the high-risk groups, and a higher immune escape score in the high-risk group indicated a higher possibility of immune escape and a worse prognosis, which was consistent with the prediction results of the model (Figure 13A). Then the difference between the groups in terms of half-maximal inhibitory concentrations of the drug was used to screen for potentially beneficial therapeutic drugs (Figure 13B); potential drugs predicted from the expression of model lncRNAs had a p value of less than .05, which demonstrated the potential of model lncRNAs as therapeutic targets to some extent (Figure 13C).