Development of a novel prognostic signature based on cytotoxic T lymphocyte-evasion genes for hepatocellular carcinoma patient management

Abstract

Objectives

Cytotoxic T lymphocytes (CTLs) are major actors in innate and adaptive antitumor response. We attempted to apply cancer cell-intrinsic CTL evasion genes (CCGs) to identify and verify a risk stratification signature in hepatocellular carcinoma (HCC) patients to assess the prognosis and benefits of immunotherapy, sorafenib treatment and transcatheter arterial chemoembolization (TACE) treatment.

Methods

We developed a novel prognostic signature including six CCGs was developed by LASSO Cox regression. CIBERSORT, quanTIseq, and ssGSEA algorithms were used to investigated the correlation between the CCG signature and immune cell infiltration. We also assessed the performance of the CCG signature predicting immunotherapy, sorafenib treatment and TACE treatment with independent clinical mRNA sequencing data.

Results

The area under the curve (AUC) of the CCG signature for predicting 1-, 3-, and 5-year OS was 0.77, 0.70 and 0.70 in the learning cohort, respectively. In the external verification cohort, the AUCs of the CCG signature were 0.71, 0.74 and 0.75. The CCG signature was significantly positively related to both TMB and MSI. In addition, responders had a significantly higher risk score than nonresponders when the signature was applied in urothelial cancer patients with immunotherapy, and the AUC of the CCG signature for predicting the response was 0.65. We further found that responders had a significantly lower risk score than nonresponders in the sorafenib and TACE treatment cohorts, and the AUCs of the CCG signature for predicting the response were 0.87 and 0.76, respectively. Finally, we identified four small molecule compounds negatively related to differentially expressed genes (DEGs) between the two categories of HCC patients, including monensin, etiocholanolone, naringenin, and Prestwick-1103.

Conclusions

The CCG signature has some clinical significance that may enhance HCC patient outcomes and even help develop novel strategies for HCC patient management.

Introduction

Liver cancer is the sixth most common malignancy worldwide, with the fourth highest mortality rate globally [1]. Hepatocellular carcinoma (HCC) is the predominant form of liver cancer, accounting for approximately 80% of primary liver malignancies [2]. Widely recognized staging systems, such as the Barcelona Clinic Liver Cancer (BCLC) staging system and the American Joint Committee on Cancer (AJCC) system, have been integral in guiding treatment decisions and prognostic evaluations for HCC [3, 4]. However, the complexity and heterogeneity of HCC often result in substantial differences in outcomes among patients with similar clinical stages and treatment strategies. This highlights the limitations of current staging systems in accurately predicting prognosis and stratifying patients' risk, underscoring the urgent need for novel, robust prognostic signatures.

Emerging evidence indicates that the tumor microenvironment (TME) plays a pivotal role in the initiation and progression of cancer [5]. Many immune cells, including myelogenic inhibitory cells, macrophages, T cells, and others, constitute the tumor microenvironment [6]. CTLs play an essential role in the adaptive antitumor response and innate antitumor response among these immune cells [7]. Among these, cytotoxic T lymphocytes (CTLs) are crucial players in both adaptive and innate antitumor responses [8]. A recent study characterized a key set of 182 CCGs in 3 mouse cancer cell lines, including renal carcinoma, breast carcinoma and colorectal carcinoma, and independent perturbation of these CCGs enhanced the sensitivity or resistance of cancer cells to CTL-mediated antitumor toxicity [8]. Thus, these CCGs may help evaluate prognosis and stratify patients.

In the era of bioinformatics, numerous prognostic gene signatures based on gene expression profiles have been developed to improve risk assessment in HCC [9–12]. Despite their potential, these models often share notable limitations. Many include an excessive number of genes (≥ 8 genes), which complicates clinical application and increases healthcare costs. In addition, many previous studies have assessed the value of signatures for guiding therapeutic decisions based on some algorithms (e.g., TIDE, IPS, pRRophetic) and the mRNA expression of key immune effector genes, and these studies may not be reliable due to the lack of validation using real data or widely used biomarkers (TMB and MSI). Thus, we attempted to apply CCGs to construct and validate a risk stratification signature in HCC patients to assess the prognosis and benefits of immunotherapy, sorafenib treatment and TACE treatment based on independent clinical mRNA sequencing data. This study may be helpful in precision treatment and further improving the prognosis of HCC patients.

Material and methods

mRNA sequencing data and CCG acquisition

To develop a prognostic CCG signature, we downloaded mRNA sequencing data and relevant clinical information from TCGA (https://gdc-portal.nci.nih.gov/) as a learning cohort. Additionally, we obtained an external verification cohort from the ICGC-LIRI-JP (https://dcc.icgc.org/) to verify the predictive prognostic power of the CCG signature. Gene expression profiles of GSE109211 (67 HCC patients with sorafenib treatment) and GSE104580 (147 HCC patients with TACE treatment) were attained from GEO (https://www.ncbi.nlm.nih.gov/geo/) to assess the predictive ability of the signature identified in those who responded to sorafenib or TACE treatment. There were 46 nonresponders and 21 sorafenib treatment responders in GSE109211. There were 66 nonresponders and 81 TACE treatment responders in GSE104580. We also collected data from a urothelial cancer cohort (348 urothelial cancer tissues) receiving immunotherapy from the R IMvigor210CoreBiologies package to test the performance of the CCG signature predicting the treatment response to immunotherapy. There were 68 nonresponders and 230 immunotherapy responders in the IMvigor cohort. Somatic mutation profiles were downloaded from TCGA to compute TMB via the R Maftools package [13]. Each gene expression was transformed into log2 and z score across patients in all cohorts. Then, 182 CCGs were acquired from previously published research [8].

Development and verification of a prognostic CCG signature

In the learning cohort, we performed differential gene analysis of transcriptome analysis data from normal liver and HCC tissues using the R stats package, setting FDR (false discovery rate) smaller than 0.05 and absolute values of log2 (fold change) greater than 1 as the cutoff values (Benjamin-Hochberg method was used for FDR calculation). Patients without information on survival status or survival time were excluded from further analysis. Univariate Cox regression analysis was conducted using the R survival package to identify CCGs significantly associated with OS. The interactive network of proteins was applied to show OS-associated CCGs via GeneMANIA [14]. Subsequently, LASSO Cox regression analysis, implemented via the R glmnet package, was performed to minimize overfitting and identify an optimal combination of genes for constructing the predictive model [15]. Based on the expression profiles of these survival-associated CCGs and their corresponding regression coefficients, a predictive prognostic CCG signature was developed [16]. The next methods were conducted simultaneously in the learning cohort and the external verification cohort. HCC patients in the learning cohort were stratified into low- and high-risk groups based on the median risk score, and this classification was applied to the validation cohort. Dimensionality reduction methods, including T-SNE and PCA (via the R Rtsne and R stats packages, respectively), were used to visualize the distribution of patients in the two risk categories. Kaplan–Meier survival analysis was performed using the R survminer package, with survival differences assessed via the log-rank test. ROC curves, generated using the R timeROC package, were employed to evaluate the predictive performance of the CCG signature for OS, with particular emphasis on its clinical relevance and utility in stratifying patients.

Assessing the independent prognostic value of the CCG signature

Patients without grade, stage information and age were also excluded from further analysis. Then, Cox regression analysis was carried out to test whether the CCG signature and the other clinicopathological characteristics, including grade, sex, TNM stage and age, were independent risk factors.

Development and verification of a predictive nomogram

We applied the R rms package to develop a predicting OS nomogram with the independent risk factors (CCG signature and TNM stage) and verified the nomogram via calibration curves [17]. ROC curves were constructed by the R timeROC package to impute the AUC of the nomogram, CCG signature and TNM stage. Decision curve analysis (DCA) was performed by the R ggDCA package to investigate the clinical benefit obtained from the predicted outcomes of the nomogram [18], CCG signature and TNM stage.

GSVA and estimation of tumor-infiltrating immune cells

In the learning cohort, GSVA [19] was applied to calculate the pathway activity score for individual HCC patients based on KEGG gene sets. In addition, the abundance of infiltrating tumors was imputed by CIBERSORT [20], quanTIseq [21], and ssGSEA [22].

Treatment strategy based on the CCG signature for HCC

In the learning cohort, we applied the R stat package to impute the correlation between the CCG signature and TMB or MSI. Given the lack of HCC immunotherapy data, we applied the IMvigor210 cohort to explore the performance of the CCG signature in predicting immunotherapy response via the R pROC package. In addition, the GSE109211 cohort and the GSE104580 cohort were applied to explore the performance of CCG characteristics in predicting the response to sorafenib and TACE treatment.

Screening for candidate small molecule drugs associated with the CCG signature

In the learning cohort, we identified DEGs between the two types of HCC patients via the R stats package. Then, DEGs were applied to identify candidate small molecule drugs via CMap (https://www.broadinstitute.org/cmap/).

Statistical analysis

R application 4.0.5 was applied to execute all statistical analyses. The Wilcoxon test compared the differences between the two groups of continuous parameters, and the Spearman test assessed the correlation between the two continuous parameters. Survival was examined by the log-rank test and Cox regression analysis. A two-tailed P value less than 0.05 was regarded as statistically significant.

Results

Acquisition of prognostic differentially expressed CCGs

The details of these HCC patients from the learning cohort and the external verification cohort are described in Table 1. In the learning cohort, we identified 37 differentially expressed CCGs associated with OS. These genes were selected based on their overlap between the differentially expressed CCGs and those associated with OS, as illustrated in Fig. 1A–C. To visualize their interactions, these 37 OS-related CCGs were displayed in a protein–protein interaction network (Fig. 1D).

Table 1.

Clinicopathological information of the HCC patients

	Learning cohort	Verification cohort
No. of patients	365	231
Age (median, range)	61(16–90)	69(31–89)
Gender (%)
Female	119(32.6%)	61(26.4%)
Male	246(67.4%)	170(72.6%)
Grade (%)
I	55(15.1%)	NA
II	175(47.9%)	NA
III	118(32.3%)	NA
IV	12(3.3%)	NA
Unknown	5(1.4%)	NA
Stage (%)
I	170(46.6%)	36(15.6%)
II	84(23.0%)	105(45.5%)
III	83(22.7%)	71(30.7%)
IV	4(1.1%)	19(8.2%)
Unknown	24(6.6%)	0(0.0%)
Vascular Invasion
Yes	106(29.0%)	NA
No	205(56.2%)	NA
Unknown	54(14.8%)	NA
AFP
≤ 200 ng/ml	201(55.1%)	NA
> 200 ng/ml	75(20.5%)	NA
Unknown	89(24.4%)	NA

Fig. 1.

Acquisition of prognostic differentially expressed CCGs in the learning cohort. A Venn diagram showing the overlap of CCGs associated with OS and DEGs in the TCGA-LIHC cohort. B Heatmap plot illustrating the mRNA expression of 37 overlapping CCGs in HCC tissue. C Forest plots showing the results of the univariate Cox regression analysis for 37 overlapping CCGs. D Interactome map of 37 overlapping CCGs

Development of a prognostic signature

In the learning cohort, we developed a CCG signature, and the equation was as follows: risk score = 0.109*ATG10 + 0.040*HDAC1 + 0.166*PIGU + 0.069*AHSA1 + 0.138*CAD + 0.134*CEP55. HCC patients were divided by a risk score of −0.065 from the learning cohort into two categories (low-risk patients and high-risk patients) (Fig. 2A). Principal component analysis (PCA) and t-distributed stochastic neighbor embedding (T-SNE) demonstrated distinct clustering of low- and high-risk patients, indicating robust risk stratification (Fig. 2B–C). Survival analysis confirmed that low-risk patients exhibited significantly lower mortality rates compared to high-risk patients (Fig. 2D, E). The AUCs of the CCG signature for predicting 1-, 3-, and 5-year OS were 0.77, 0.70 and 0.70, respectively (Fig. 2F).

Fig. 2.

Development of a prognostic signature in the learning cohort. A Risk score distribution. B PCA plot showing the two types of HCC patients in two different areas. C. t-SNE showing that the two types of HCC patients were in two different areas. D Survival time of HCC patients in increasing order of risk score. E Kaplan–Meier plot for two types of HCC patients. F ROC curves of OS for the CCG signature

Verification of the CCG signature

In the external verification cohort, the CCG signature was applied. HCC patients were also separated into two categories based on the risk score listed above (−0.065) (Fig. 3A). Similarly, PCA and t-SNE further showed that the two types of HCC patients were in two different areas (Fig. 3B-C). Additionally, there were fewer deaths in the low-risk patients (Fig. 3D). Low-risk patients had longer OS than high-risk patients (P < 0.05) (Fig. 3E). Given that the OS of only two patients was over 5 years, the AUC of the CCG signature was 0.71, 0.74 and 0.75 (Fig. 3F). These results confirmed that it is robust for the CCG signature to assess HCC patient OS.

Fig. 3.

Verification of the CCG signature in the external verification cohort A. Risk score distribution. B PCA plot showing the two types of HCC patients in two different areas. C t-SNE showing that the two types of HCC patients were in two different areas. D Survival time of HCC patients in increasing order of risk score. E Kaplan–Meier plot for two types of HCC patients. F ROC curves of OS for the CCG signature

Development and verification of a nomogram with independent risk factors

In the learning cohort and the external verification cohort, Cox regression analyses were performed to assess the independent prognostic value of the CCG signature. Univariate Cox regression analyses indicated that the risk score was significantly related to OS (all P < 0.05) (Fig. 4A, B). We further removed the clinicopathological factors that were not related to OS (P > 0.05), and we found that the risk score was an independent risk factor (all P < 0.05) (Fig. 4C, D). In the learning cohort, we developed a predicting OS nomogram with the CCG signature and TNM stage (Fig. 4E). The calibration curves indicated strong conformance between the actual outcomes and the outcomes of the nomogram prediction (Fig. 4F). Both ROC curves and DCA showed that the predicting OS performance of the nomogram was stronger than the CCG signature and AJCC tumor stage, and predicting OS performance of the CCG signature was stronger than the AJCC tumor stage (Fig. 4G, H).

Fig. 4.

Development and verification of a nomogram with independent risk factors. A–B. Forest plots showing the results of the Cox regression analyses in the learning cohort. C–D. Forest plots showing the results of the Cox regression analyses in the external verification cohort. E A nomogram with the CCG signature and the TNM stage for quantifying HCC patient OS in the learning cohort. F The calibration plots of the nomogram. G. The ROC curves of the nomogram, the CCG signature and the TNM stage. H DCA plots of the nomogram, the CCG signature and the TNM stage

GSVA and immune infiltrating characteristics

In the learning cohort, we investigated the biological mechanisms underlying the prognostic signature by performing gene set variation analysis (GSVA) with KEGG pathway annotations. The NF-kappa B pathway, a key mediator of inflammatory and immune responses, promotes immune evasion in HCC by inhibiting CTL-mediated apoptosis through the upregulation of anti-apoptotic proteins and immunosuppressive cytokines such as TGF-β. GSVA revealed that high-risk patients exhibited significantly elevated activity scores for the NF-kappa B and autophagy pathways compared to low-risk patients (P < 0.05) (Fig. 5A, B). Immune cell quantification using Cibersort, quantiseq, and ssGSEA demonstrated significantly increased infiltration of regulatory T cells (Tregs) in high-risk patients compared to low-risk patients (all P < 0.05) (Fig. 5C–E). Tregs further exacerbate immune suppression by inhibiting CTL activity and promoting a tolerogenic tumor microenvironment through cell–cell interactions and secretion of inhibitory cytokines such as IL-10 and TGF-β. These findings provide a mechanistic explanation for the association between high-risk scores and immune escape in HCC, offering insights into potential therapeutic targets for high-risk patients.

Fig. 5.

GSVA and immune infiltrating characteristics in the learning cohort. A–B. Comparison of the GSVA scores of the autophagy and NF-kappa B pathways between the two types of HCC patients. C–E. Comparison of the abundance of tumor-infiltrating immune cells between two types of HCC patients based on three algorithms (C. CIBERSORT; D: quanTIseq; E. ssGSEA). *, **, ***, **** represented P < 0.05, P < 0.01, P < 0.001 and P < 0.0001, respectively

Treatment strategy based on the CCG signature for HCC

In the learning cohort, the CCG signature showed positive and highly significant correlations with TMB and MSI (all P < 0.05) (Fig. 6A, B), indicating an indirect association with the efficacy of immunotherapy. Given the lack of HCC immunotherapy datasets, we further assessed whether the CCG signature had robust performance in predicting immunotherapy response in the IMvigor210 cohort. Responders had a significantly higher risk score than nonresponders when the signature was applied to urothelial cancer patients treated with immunotherapy (Fig. 6C), and the AUC of the CCG signature was 0.65 (Fig. 6D). Some studies have demonstrated that sorafenib or TACE treatment may affect the activation of CTLs [23, 24]. Herein, we also assessed whether the CCG signature had robust performance in predicting sorafenib and TACE treatment response. Responders had markedly poorer risk scores than nonresponders (Fig. 6E) (P < 0.05), and the signature could effectively distinguish responders from nonresponders to sorafenib treatment (AUC = 0.87) (Fig. 6F). Similarly, responders had significantly lower risk scores than nonresponders (Fig. 6G) (P < 0.05), and the signature could effectively distinguish responders from nonresponders with TACE treatment (AUC = 0.76) (Fig. 6H). Taken together, sorafenib and TACE treatment might be appropriate for patients with lower risk, and immunotherapy might be recommended for patients with higher risk.

Fig. 6.

Treatment strategy based on the CCG signature for HCC. A–B. Relationship of TMB, MSI, and risk score. C. ROC curves showing the performance of the CCGs in predicting immunotherapy response in the IMvigor cohort. D. Comparison of the risk score in the nonresponse and response in the IMvigor cohort. E. ROC curves showing the performance of the CCGs predicting sorafenib treatment response in the GSE109211 cohort. F Comparison of the risk score in the nonresponse and response in the GSE109211 cohort. G ROC curves showing the performance of the CCGs predicting TACE treatment response in the GSE104580 cohort. H Comparison of the risk score in the nonresponse and response in the GSE104580 cohort

Identification of related small molecule drugs

To further improve HCC patient outcomes with higher risk scores, we used CMAP to identify potential small molecular compounds. As shown in Table 2, four small molecule compounds were negatively related to DEGs containing monensin, etiocholanolone, naringenin, and Prestwick-1103.

Table 2.

Results of CMap analysis

Cmap name	Mean	n	Enrichment	p	Specificity	Percent nonnull
Monensin	−0.333	6	−0.77	0.00026	0	50
Etiocholanolone	−0.341	6	−0.714	0.00115	0.0455	66
Naringenin	−0.55	4	−0.823	0.00189	0.0081	75
Prestwick-1103	−0.288	4	−0.822	0.00189	0	50

Discussions

Considering the substantial heterogeneity in prognostic outcomes among HCC patients, it is imperative to develop an accurate and robust classifier to optimize clinical decision-making, enhance individualized treatment strategies, and improve timely risk stratification. Despite increasing efforts to unravel the complex mechanisms of HCC, gaps remain in understanding key prognostic factors, the tumor microenvironment (TME), and therapeutic targets. To address this critical gap, we developed a CCG-based signature aimed at providing more precise prognostic predictions and improving clinical outcomes in HCC patients.

With the rapid development of high-throughput sequencing technologies, we are better able to identify key genes associated with tumorigenesis [25]. To our knowledge, this is the first study to assess the relationship between CCG and HCC. The mRNA expression and prognosis of the 182 CCGs were comprehensively explored. Then, we developed a CCG signature and performed external validation for its performance. Moreover, the CCG signature had independent prognostic value, and the AUC was greater than or equal to 0.7. ROC curves also showed that the AUC of the CCG signature was superior to that of TNM stage. Thus, the signature for assessing HCC patients OS has stability and applicability.

Cytotoxic T lymphocytes (CTLs) are known to induce tumor cell death through two primary mechanisms: Fas ligand-mediated apoptosis and granule exocytosis [26]. However, the NF-kappa B pathway has been demonstrated to inhibit cancer cell apoptosis [27, 28] and reduce the cytotoxicity of CTLs by transforming growth factor beta [29]. Furthermore, regulatory T cells (Tregs) play a crucial role in inducing CTL exhaustion within the tumor microenvironment [30]. In our study, the NF-kappa B and autophagy pathway scores based on GSVA differed between the two categories of HCC patients. The NF-kappa B and autophagy pathway activity score and the immune infiltration abundance of regulatory T cells (Tregs) in high-risk patients was significantly increased compared with that in low-risk patients. These results revealed that the antitumor function of CTLs might be suppressed in HCC patients with higher risk scores and partially account for the value of this prognostic signature.

Recently, immunotherapy has attracted wide attention in the clinical treatment of cancer and represents a novel direction to enhance the prognosis of patients with tumors [31]. However, only a small percentage of HCC patients respond to immunotherapy [32]. Consequently, identifying patients who are most likely to benefit from immunotherapy is a critical clinical priority. The CCG signature showed positive and highly significant correlations with TMB and MSI, and responders had a significantly higher risk score than nonresponders when the signature was applied in the IMvigor210 cohort. Consequently, immunotherapy might be recommended for patients with higher risk. At present, TACE and sorafenib treatment are the main therapies for patients with advanced HCC [33], but sorafenib or TACE treatment still lacks effective biomarkers. Thus, we assessed the predictive value of the signature for quantifying the benefits of sorafenib and TACE. The ROC curves demonstrated that the signature has a high clinical value in guiding the selection of TACE and sorafenib treatment for HCC patients. Responders had a significantly lower risk score than nonresponders in the sorafenib and TACE treatment cohorts. Therefore, the signature has the potential to help differentiate HCC paitents who might benefit from immunotherapy, sorafenib treatment and TACE treatment.

Previous studies on HCC prognostic signatures have predominantly focused on algorithms such as TIDE, IPS, and pRRophetic, or relied on the mRNA expression of key immune effector genes, often without using independent clinical mRNA sequencing data or validated biomarkers like TMB and MSI. For example, the TIDE algorithm is only suitable for assessing immunotherapy response in melanoma or non-small-cell lung cancer [34], but many studies have applied it in HCC patients [35–37]. In addition, pRRophetic algorithmwas based on the mRNA expression data of cancer cell lines with drug treatment [38], but it has not been validated in real HCC patients. Thus, many studies might not be reliable due to the application of these algorithms in HCC9 [35]. In contrast, our study utilized independent clinical mRNA sequencing data and validated biomarkers such as TMB and MSI to comprehensively evaluate the CCG signature, demonstrating its robust potential to guide therapeutic decisions effectively.

Finally, we used CMAP to identify potential small molecular compounds that could further improve HCC patients with higher risk scores. Among the top 4 drugs, monensin was considered to be the most promising. Monensin is an ion carrier antibiotic obtained from Streptomyces cinnamomi [39]. Monensin has been proven to inhibit the NF-kB and autophagy pathways, which synergistically contribute to CTL evasion [8, 40, 41]. Therefore, monensin might enhance the antitumor function of CTLs in HCC.

While this study presents a powerful and clinically relevant prognostic signature, several limitations warrant further investigation. First, although independent external validation was conducted, the study cohorts did not encompass HCC patients from diverse ethnic and geographical backgrounds, potentially limiting the generalizability of the findings. Consequently, future research should focus on validating this CCG signature in larger, multicenter cohorts with broad demographic representation to ensure its universal applicability and clinical relevance. Second, the underlying molecular mechanisms of the six CCGs and monensin in HCC progression and treatment efficacy require more in-depth investigation through experimental and preclinical research. Despite these limitations, the CCG signature provides a strong foundation for improving prognostic accuracy and offers a promising avenue for developing novel therapeutic strategies for personalized HCC management.

Conclusions

Collectively, we constructed and validated the CCG signature, which had independent prognostic value. The prediction performance of the CCG signature was stronger than that of the AJCC tumor stage, and the CCG signature can offer promising ability for risk stratification. Moreover, the signature has substantive clinical value that may enhance OS and even help develop novel strategies for HCC patient management.

Author contributions

Suya Liu, Chunqian Yang, and Jingsong Tang designed the study. Ting Wei and Shiping Liao performed data collection and data analysis. Qinmei Zhu performed visualization. Ting Wei and Shiping Liao wrote and proofread the paper.

Funding

This study was conducted without fund.

Data availability

The transcriptomic data utilized in this investigation are publicly accessible through the following repositories: 1. The Cancer Genome Atlas (TCGA): http://cancergenome.nih.gov/; 2. International Cancer Genome Consortium (ICGC): https://dcc.icgc.org/; 3. Gene Expression Omnibus (GEO): https://www.ncbi.nlm.nih.gov/geo/.

Declarations

Competing interests

The authors declare no competing interests.