Introduction
Hepatocellular carcinoma (HCC), originating from hepatocytes, is the dominant subtype in primary liver cancer and is characterized with high incidence, delayed diagnosis and poor prognosis [
1]. Although major risk factors, such as HBV infection, HCV infection, Aflatoxin B1 exposure, liver cirrhosis, non-alcoholic fatty liver disease (NAFLD) and alcohol abuse, have been ascertained and widely publicized for a long time, the incidence of HCC has been growing in many countries, including developed countries in Europe and North America [
2,
3]. In addition, by 2025, an estimated incidence of over 1 million cases annually will cause a large burden of disease across the world [
4]. Owing to no specific symptoms at early stages, insensitivity of traditional biomarkers and the lack of large-scale imaging screening program, only about 30% of patients are clinically diagnosed in early stages [
5], which also partly contributed to the dismal overall 5-year survival rate of less than 20% [
6]. In the past two decades, more optional treatment strategies are offered to patients with HCC, mainly including immunotherapy and molecular targeted therapy [
7]. However, only a minority of patients can benefit from each regimen. Thus, tailored therapy concept is proposed and the related mechanisms to select right regimen are fiercely discussed by researchers. With the rapid growth in genomics research, high molecular heterogeneity of HCC is detected and is considered as one of the critical factors to affect prognosis and therapeutic effects [
8]. Hence, it is of great significance to investigate novel biomarkers for the prediction of prognosis and treatment effect through high-throughput sequencing data.
Tumor immune microenvironment (TIME) consists of tumor cells, various immunocytes, stromal cells and related extracellular matrix molecules [
9]. Though immune system mainly exerts functions of eliminating cancer cells, cancer cells can escape from immune killing and form an immunosuppressive microenvironment [
10]. Therefore, a growing number of studies focus on the associations between tumor phenotypes and changes in TIME and confirm that TIME acts a significant role in cancer initiation and development [
11,
12]. Natural killer (NK) cell is one of classical cytotoxic cells and innate immune members that can identify and eliminate damaged or stressed cells [
13]. In liver tissues, NK cells account for 50% of the lymphocyte population [
14]. Unlike acquired immunity, NK cells identify target cells via a cascade of germline-encoded surface receptors, and the functions of NK cells is tightly modulated by activating and inactivating signals from these receptors [
15]. In tumor immunity, NK cells rapidly detect tumor cells, directly kill tumor cells and promote immune response mediated by T cells, thus inhibiting cancer occurrence and development [
13]. Previous studies revealed decreased infiltrating levels of NK cells in many human cancer types, including gastric cancer, esophageal cancer, breast cancer and colorectal carcinoma [
16‐
18]. In addition, high NK cell infiltration levels in tumor tissues are considered as biomarkers correlated to better prognosis [
19‐
21]. In addition, high NK cell activity in peripheral blood is related to reduced risk of malignancy [
22]. As for cancer immunotherapy, NK cell-based treatment has grown rapidly for decades, and its safety and efficacy are widely validated by clinical trials, therefore, becoming a vital domain of immunotherapy innovation [
23]. Previous publications have systematically elucidated molecular features of NK cells in bladder cancer, low-grade glioma and neuroblastoma [
24‐
26] and have attempted to build up NK-cell-related gene signature in lung adenocarcinoma, head and neck squamous cell carcinoma, and glioma [
27‐
29], while related studies concerning HCC are still rare.
Single-cell sequencing technology offers an unprecedented opportunity to deepen our understanding on the transcriptomic, genomic, proteomic, epigenomic and metabolomic information of individual cells. In recent years, using single-cell RNA sequencing (scRNA-seq) data to identify immune phenotypes and novel immune cell-related functional biomarkers in tumor microenvironment becomes feasible and popular [
30‐
32]. In the present study, we combined scRNA-seq data and bulk RNA-seq data of HCC to illuminate molecular characteristics of tumor-infiltrating NK cells and to screen out NK cell markers. Using bulk RNA-seq data and corresponding survival data from 4 independent cohorts, an NK cell marker-related prognostic signature based on 77 fundamental or combined machine learning algorithms was next developed and validated, and its associations with immune cell infiltration, immune checkpoint blockade response, chemotherapy sensitivity were further investigated. This work may provide some new ideas on prognosis evaluation and immunotherapy of patients with HCC, thereby promoting the development of individualized treatment and improving patients’ prognosis.
Discussion
scRNA-seq technology could collect and classify the genomic data from various cell subpopulations in TME, thereby deepening our understanding of tumor heterogeneity and molecular features of tumor-infiltrating immunocytes. NK cell dysfunction had been observed in HCC tissues, which mainly contained abnormal frequency and phenotype of NK cells, and functional impairment of NK cells [
57]. Considering such phenomena, accumulating studies attempted to explore the features of NK cells in HCC and related immunoregulation of NK cells in TME. HCC patients with low intratumoral NK cells showed higher recurrence rate and shorter overall survival rate after surgical resection [
58]. In addition, NK cell markers, such as CD96, TIM-3 and TIGIT, were strongly correlated with functional exhaustion of NK cells and poor survival outcomes in HCC [
59,
60]. As for NK cell-based immunotherapy of HCC, various immunotherapeutic strategies, including antibody-dependent cell-mediated cytotoxicity (ADCC) promoter, immune checkpoint blockade, genetically modified NK cells, off target effects on NK cells, autologous NK cell transfer and allogeneic NK cell transfer, were utilized in clinical application or clinical trials, which exhibited an overall positive effect [
61]. Nevertheless, there was still a lack of studies to systematically elaborate the role of NK cell markers in prognostic prediction and immunotherapy of HCC patients.
With the rapid rise of computational medicine and bioinformatics, varied machine learning approaches were utilized to develop prediction models based on gene expression data. Nevertheless, why to select a particular algorithm and how to choose the optimal algorithm were rarely discussed in the process of model construction. As a matter of fact, a significant number of researchers chose the algorithm according to personal experience and preference, which might bias the results. To avoid this issue, we combined scRNA-Seq data, bulk RNA-Seq data and the consensus machine learning framework including a total of 77 independent or combined machine learning algorithms to develop a signature based on a total of 11 NK cell marker genes, KLRB1, CFL1, LDHA, BSG, ATP1B3, SERBP1, UBE2L3, PCBP2, ENO1, OPTN and LMO4. In GSE14520 data set, GSE76427 data set, ICGC–LIRI–JP data set and TCGA–LIHC data set, high-risk patients showed a worse OS rate than those with low risk score. In addition, the risk score was identified as the independent risk factor to affect patients’ prognosis via univariate and multivariate Cox regression analysis. ROC curves based on the survival time and clinical parameters revealed moderate to high prediction accuracy of our signature. In addition, compared to other 10 risk models published in previous studies, the leading C-index of 0.730 was also presented in our signature, displaying relatively high predictive value for HCC patients’ prognosis.
In published studies, the expression level of CFL1 was significantly increased in HCC tissue, and high CFL1 expression was strongly correlated with poorer survival outcomes. Downregulation of CFL1 suppressed proliferation, invasion and epithelial–mesenchymal transformation (EMT) in HCC cell lines [
62]. As the upstream regulators of LDHA, miR-383 and miR-142-3p exerted anti-tumor effects via negatively regulating LDHA expression [
63,
64]. BSG, also known as CD147, acted as an oncogene in HCC and was considered as prognostic indicator for HCC patients [
65]. In addition, BSG overexpression could activate ERK signaling pathway and TGF-β signaling pathway, thereby promoting HCC migration, invasion and EMT [
66,
67]. In addition, chimeric antigen receptor (CAR) therapy toward CD147 could selectively kill HCC cells and avoid severe on-target/off-tumor toxicity in mouse model [
68]. Silencing of ATP1B3 induced apoptosis and inhibited proliferation and migration in two HCC cell lines, HCCLM3 and Huh7 [
69]. SERBP1 was also detected to promote metastasis and EMT of HCC cells, while miR-218 could repressed the pro-tumor functions of SERBP1 [
70]. UBE2L3 was also identified as a pro-tumorigenic factor to stimulate HCC cell proliferation by inactivating GSK3β/p65 signaling pathway [
71]. PCBP2 expression was also increased both in HCC cell lines and tissues. PCBP2 overexpression predicted unfavorable prognosis for HCC patients and induced proliferation and sorafenib resistance in HCC cells [
72]. Ferroptosis of HCC cells was inhibited by ENO1–IRP1–Mfrn1 regulatory axis [
73]. Furthermore, elevated OPTN expression promoted the proliferation, migration and mitophagy of HCC cells, thus modulating HCC progression [
74].
We illustrated the overall somatic mutation status in high- and low-risk groups and identified some highly mutated genes in HCC. In previous study, Ke et al. analyzed the somatic mutation profiles from 22 HCC patients, which revealed that TP53, MUC16 and TTN were the genes with high mutation frequency in HCC [
75]. These results were similar with ours. The correlations between immune microenvironment and risk score were also evaluated in these two groups. First, higher StromalScore and ESTIMATEScore were detected in low-risk patients. Afterward, we observed increased infiltration levels of aDCs, B cells, CD8 + T cells, mast cells, neutrophils, NK cells, T helper cells, Th2 cells and TIL cells in low-risk group, while the infiltrating level of macrophages was increased in high-risk patients. In addition, related immune function analysis revealed a decreased ssGSEA score of cytolytic activity, T cell co-stimulation and type II IFN response in high-risk patients. The anti-cancer effects of CD8 + T cells, NK cells and TIL cells in TME were widely recognized in previous study [
76]. The increase of tumor-associated macrophages in TME promoted tumor growth and resistance to sorafenib in HCC cells [
77]. Shankaran et al. reported that IFN could cooperate with lymphocytes to maintain the immunogenicity of tumor cells, thus inhibiting tumor progression [
78]. These evidences indicated anti-tumor immune response may be relatively more active in low-risk patients. Next, we also found that the expression levels of the majority of common immunotherapeutic targets were elevated in low-risk patients. In addition, IPS–CTLA4 blocker and IPS–PD1 blocker were higher in low-risk patients, while TIDE score was higher in high-risk patients. Moreover, the similar risk score distribution and survival status were also observed in the external validation cohort, PRJEB23709. Based on the distribution of cytotoxic immune cells in TME, solid tumor was classified into “hot” and “cold” tumor, and the “hot” tumor was more sensitive to ICI therapy [
79]. In the current study, the low-risk group might be identified as “hot” tumor and more likely to benefit from ICI treatment.
The curative effects of cancer targeting agents and chemotherapy drug varied greatly among individuals, which was mainly attributed to tumor heterogeneity [
80]. Thus, it had become popular in recent years to utilize tumor molecular characteristics to develop tools to identify potential beneficiary group for specific regimens. In our work, we found that low‐risk group was more sensitive to irinotecan, vorinostat, axitinib, cytarabine, oxaliplatin, leflunomide, cisplatin, topotecan, mitoxantrone, sorafenib, fludarabine and gemcitabine, which provided novel insights on clinical selection of these 12 anti-tumor drugs. In follow-up work, we are willing to explore the underlying molecular mechanisms of the differences in drug sensitivity.
Some limitations in this study should be declared. First, the data used in our work was derived from transcriptome sequencing. Therefore, some of our results may not be applicable to studies with protein level. Secondly, the underlying molecular mechanisms of the prognostic signature to influence the prognosis, drug sensitivity, ICI response, and immune cell infiltration still need further investigation. Third, the genomic data used in the current study was from retrospective studies. We deeply hope our signature can be further verified by multi-center prospective projects. Fourth, due to the lack of publicly available cohorts concerning ICI treatment of HCC patients, we enrolled the immunotherapy cohorts of melanoma patients, GSE91061 and PRJEB23709 data set, to externally validate the predictive power of our signature on ICI response. Although similar study design was detected in previous publications [
43,
44], we still expect future researchers can offer related data to supply our findings.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.