Construction and validation of a cell-in-cell related prognostic signature for hepatocellular carcinoma

Abstract

Cell-in-cell structures (CICs) and their biological roles contribute to disease development, particularly in cancer. However, the roles of CIC-associated genes (CICGs) in hepatocellular carcinoma (HCC) are largely unknown. This study sought to establish a a CICGs-linked HCC signature and assess its predictive significance. We acquired gene expression profiling data for tumor and normal tissues of HCC patients from The Cancer Genome Atlas (TCGA) for training and from the International Cancer Genome Consortium (ICGC) for validation. Consensus clustering was employed to delineate patient cohorts with varying prognoses, categorizing HCC patients into two distinct groups. A selection of fifty CICGs was compiled from the literature, revealing their association with CIC development through functional studies. Six predictive genes were ultimately identified through univariate Cox proportional hazards regression (Cox) and least absolute shrinkage and selection operator (LASSO) analyses. This prognostic signature, resulting from a multivariate Cox, subsequently segmented TCGA and ICGC cohort individuals into high- and low-risk categories. Our verification of the signature’s precision involved survival analysis contrasts between those at high and low risk. Quantitative real-time PCR (qRT-PCR) analysis revealed markedly elevated expression levels of these six genes in HCC tumors compared to neighboring healthy tissue, underscoring their potential role in tumor development. This experimentally reinforces the accuracy of the genetic profile. We also examined variations in the composition of immune cells and immunological responses across high-risk and low-risk cohorts. This study established and verified prognostic variables connected to cell-in-cell (CIC), potentially improving personalized survival forecasts for patients with HCC.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03245-0.

Introduction

Hepatocellular carcinoma (HCC) is the sixth most common malignancy worldwide and the third leading cause of cancer-related mortality [1]. Despite advances in clinical and experimental cancer therapies, long-term outcomes for HCC patients remain poor, largely due to the high rates of tumor recurrence and metastasis following surgical intervention [2–5]. Therefore, a deeper understanding of the molecular mechanisms underlying metastasis is essential for improving prognostic accuracy and developing targeted therapeutic strategies for HCC [6, 7].

Cell-in-cell structures (CICs) refer to the process in which one cell, termed to host cell, contains one or more cells, termed to target cells, leading to biological effects. Recent research referred to different cell-in-cell (CIC) process, including entosis [8], cannibalism [8, 9], emperipolesis [10, 11], and emperitosis [8, 12], each with distinct mechanisms and implications. Based on the cellular context, CICs can be broadly classified into two types: homotypic CICs, which occur between cells of the same type—predominantly tumor cells—and heterotypic CICs, which involve interactions between different cell types, most commonly immune cells internalized by tumor cells.

The fates of internalized and host cells following CIC formation are diverse; however, the most common outcome is the death of the internalized (target) cell, which is typically eliminated by the host (winner) cell [13]. CIC is ubiquitous in many different types of tumors in vitro and in vivo [14], including pulmonary carcinoma, epidermal carcinoma, breast carcinoma, melanoma, and even hematologic malignancies such as leukemia, often involving the internalization of lymphocytes by tumor cells [15–19].

The formation of CICs proceeds through several key steps: recognition, adhesion, and internalization, which are accompanied by dynamic cytoskeletal rearrangements and interactions between the extracellular matrix (ECM) of host and target cells [9, 13]. These cytoskeletal changes include the polarized redistribution of α-catenin (CTNNA) and β-catenin (CTNNB), as well as the involvement of actin filaments, myosin II, Ras homolog family proteins (Rho), and Rho-associated coiled-coil-containing protein kinase (ROCK) [9]. Additionally, epithelial adhesion molecules such as E-cadherin (CDH1) and P-cadherin (CDH3) facilitate cell–cell junction formation, which contributes to the initiation and stabilization of CIC structures [20, 21]. Moreover, Ezrin, a cytoskeletal linker protein, has been shown to play a crucial role in mediating cell cannibalism during CIC formation [22].

Recent studies have identified six key genes involved in the formation and maintenance of cell-in-cell structures in hepatocellular carcinoma. CBX2 promotes cancer stem cell self-renewal and epithelial–mesenchymal transition through activation of Wnt/β-catenin, Notch, and PI3K/AKT signaling pathway [23–25]. ADAMTS5 facilitates extracellular matrix degradation and enhances tumor invasiveness by activating TGF-β and PI3K/AKT pathway [26–28]. LDHA supports metabolic reprogramming and CIC survival under hypoxia via the PI3K/AKT/mTOR and HIF-1α pathways [29, 30]. PLOD2 strengthens extracellular matrix remodeling and promotes metastasis by regulating TGF-β and HIF-1α signalin [31, 32]. PIGU enhances tumor metabolic activity and immune evasion through PI3K/AKT, MAPK, and NF-κB pathways, contributing to CIC maintenance and therapeutic resistance [33–35]. SPP1 promotes tumor migration, extracellular matrix remodeling, and PD-L1 upregulation through activation of PI3K/AKT and TGF-β pathways [36, 37]. Together, these genes contribute to an interconnected regulatory network that supports the stemness, metabolic adaptation, immune escape, and metastatic potential of CICs. A clearer understanding of their biological roles may offer valuable insights into the molecular mechanisms of hepatocellular carcinoma progression and inform the development of targeted therapies.

The CIC phenomenon involves multiple signaling pathways that collectively regulate its occurrence. Toll-like receptors (TLRs) and Fc receptor (FcR) signaling are crucial in endocytosis of immune cells and phagocytosis of foreign targets [38]. The PI3K/Akt/mTOR pathway modulates cellular phagocytic capacity during autophagy [39], while the Fas receptor/Fas ligand (Fas/FasL) pathway contributes to cell internalization during apoptosis [40]. Additionally, the Wingless/integrated (Wnt)/β-catenin pathway - known for driving tumor proliferation and metastasis [41] - may facilitate cell-in-cell behavior among cancer cells. Furthermore, Rho GTPase-mediated regulation of endocytosis is essential for cell-in-cell formation [42]. Together, these pathways coordinate processes like endocytosis, proliferation, differentiation, and cell death to drive CIC phenomena across physiological and pathological contexts.

CIC plays a critical role in the initiation, progression, and prognosis of cancer. Previous studies have shown that chromosomal abnormalities are key drivers of tumorigenesis and serve as important hallmarks distinguishing cancer cells from their normal counterparts [43, 44]. Chromosomal instability (CIN), in particular, is associated with poor prognosis in several malignancies, including breast cancer [45] and endometrial cancer [45, 46]. Moreover, CIN has been implicated in resistance to therapeutic interventions, thereby limiting the efficacy of pharmacological treatments [47]. Notably, Krajcovic et al. demonstrated that CIC formation can promote tumor progression by inducing chromosomal aberrations [48].

There are also some reports of CIC in HCC. Beseler’s group discovered that the infused naïve autoreactive T lymphocytes (T cells) localized in the liver, infiltrated hepatocytes (or liver epithelial cells), and rapidly underwent cell death and degradation morphologically akin to entosis [49]. This phagocytosis of live T cells by tumor cells reflects a mechanism of immune cell elimination, thereby facilitating immune evasion by the tumor.

Collectively, these findings underscore the importance of CIC in liver cancer development. In this study, we compiled a set of 50 CICGs based on existing literature and investigated their prognostic relevance in HCC using comprehensive bioinformatics analyses.

Materials and methods

Data source and collecting of CICGs

Public transcriptome profiling data (RNA sequencing (RNA-seq), n = 465) and corresponding clinical records (n = 418) for HCC patients were retrieved from TCGA and implemented as the training cohort. The TCGA transcriptomic dataset comprised 407 tumor samples and 58 adjacent normal tissues. For independent validation, RNA-seq data (n = 445) and clinical details (n = 260) were sourced from ICGC, consisting of 243 tumor samples and 202 normal controls. Demographic features of HCC patients from both cohorts are summarized in Table 1. Fifty CICGs were chosen following a comprehensive literature analysis of CIC, as outlined in Table 1. Figure S1 depicts the procedure for data collecting and preprocessing.

Table 1.

The 50 cell-in-cell related genes

CDH1	ITGB7	SMG1	TUBA3C	MYO16
CDH3	ITGB8	MT2A	TUBA4B	MYO1D
EGFR	CTNNA2	NUPR1	EZR	MYO1B
AR	CTNNA3	KRAS	RAC1	MYO1F
ITGB1	CTNNB1	MYC	CDC42	MYO1H
ITGB2	ICAM1	ROCK1	TJP1	ACTN4
ITGB3	PARD3	ROCK2	MYO3B	ACTA1
ITGB4	FAK1	RHOA	MYO1G	ACTN3
ITGB5	CD8A	MLC1	MYO9A	ACTA2
ITGB6	IPO13	TUBA1B	MYO5B	ACTG2

Analysis of differential expression and functionality of CICGs

The transcriptomic data from TCGA were analyzed utilizing the “limma” tool in R to identify differentially expressed CICGs in HCC compared to normal tissues. CICGs were deemed differentially expressed only if they complied with key criteria: an absolute log2 fold change (FC) greater than 1 and a false discovery rate (FDR) below 0.05. Those genes that cleared both thresholds were then singled out for deeper analysis. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway studies were conducted to examine functional implications of these CICGs, with a significance threshold of P < 0.05 [50].

CICGs consensus clustering

The TCGA HCC cohort data were grouped according to the expression patterns of CICGs utilizing the “ConsensusClusterPlus” tool (version 1.58.0) in R. This clustering sought to investigate the connection between CICG expression and prognosis. To further enhance the depiction of gene expression patterns across several groups, principal component analysis (PCA) was performed. Additionally, Kaplan-Meier (KM) analysis was applied to assess overall survival (OS) and evaluate survival differences between the identified clusters.

Construction of the prognostic signature

In this study, R software was used to assign scores to each prognostic-related gene based on the OS of TCGA-HCC patients. We utilized R’s ‘survival’ package (version 3.2.13) to detect differentially expressed genes (DEGs) among different clusters. Univariate Cox was utilized on DEG expression data from TCGA-HCC patients to evaluate association between CICG expression and OS, indicating significant relationships.

To identify the most critical prognostic genes from the DEGs, the LASSO function from the ‘glmnet’ package (version 4.1.3) in R was utilized. DEGs link to prognosis were chosen based on LASSO estimation (P < 0.05). Moreover, the ideal setting for the penalization coefficient lambda was established using 1,000 iterations of cross-validation likelihood (CVL).

The “survminer” package (version 0.4.9) was subsequently employed to compute the final risk assessment score. Multivariate Cox (P < 0.05) was employed to analyze the genes selected by LASSO. The risk assessment score was calculated as follows: risk score = Σ (Cox coefficient of gene x(i) × expression value of gene x(i)) [50].

Evaluation and validation of the prognostic signature

Patients from both the training and validation cohorts were stratified into high- and low-risk subgroups by implementing the median risk score as a threshold. We employed KM curves to compare OS across the study subgroups. Additionally, the model’s ability to predict outcomes was verified by receiver operating characteristic (ROC) curve analysis, which examined sensitivity and specificity via the “survival ROC” tool (version 1.0.3) in R.

To enhance the evaluation of our signature’s predictive performance, we explored its AUC values alongside traditional clinical factors. Moreover, we developed a prognostic nomogram integrating all independent clinicopathological factors identified through multivariate Cox. This nomogram offers a quantitative instrument for evaluating OS probability at 1, 3, and 5 years in individuals in HCC.

Prediction of immunocyte infiltration

Single-sample gene set enrichment analysis (ssGSEA) was performed using the ‘GSVA’ R package (v1.42.0) to assess immune cell infiltration (16 cell types) and immune pathway activity (13 pathways) in each HCC sample. The “reshape2” package, version 1.4.4, and the “ggpubr” package, version 0.4.0, in R were utilized to assess the disparities in immunocyte and immunological function between high-risk and low-risk groups.

Human tumor tissue

Fresh tumor specimens and corresponding surrounding non-cancerous liver tissues were prospectively obtained from treatment-naïve HCC patients who had curative hepatectomy at the First Affiliated Hospital of Sun Yat-Sen University between September and December 2024. The research was performed in compliance with the Declaration of Helsinki and received approval from the Institutional Ethics Committee of the First Affiliated Hospital of Sun Yat-Sen University (Approval No. [2020]339).

Clinical trial number declaration

Not applicable.

Quantitative real-time PCR

Total RNA was extracted using TRIzol reagent (Thermo Fisher Scientific, Waltham, MA, USA) according to the manufacturer’s instructions. Reverse transcription was conducted using the PrimeScript RT Reagent Kit (Takara, China), subsequently followed by quantitative real-time PCR (qRT-PCR) to assess gene expression levels. SYBR Green dye (Roche, Basel, Switzerland) was employed for qRT-PCR detection, and gene expression was measured using the 2 − ΔΔCt (cycle threshold) method. All tests were performed in triplicate (n = 3), and the primer sequences are provided in Table S1.

statistical analysis

Statistical analyses were performed utilizing R software and Perl. Cluster heatmaps and volcano plots were produced utilizing the R programs gplots and heatmap, respectively.

Hypothesis testing encompassed the examination of normally distributed quantitative data through Student’s t-test, whilst the Wilcoxon rank-sum test was employed for non-normally distributed variables. The KM test, a non-parametric approach, was employed to undertake comparisons among many groups. All tests were bilateral, with statistical significance established at P < 0.05.

Result

Differential expression of CICGs between HCC and normal sample

A total of 38 CICGs with differential expression between HCC and normal samples were identified, including 8 downregulated and 30 upregulated genes (Fig. 1A). GO enrichment analysis (Fig. 1C-D) revealed that the most significantly enriched terms were “actin filament organization” (biological process), “focal adhesion” (cellular component), and “actin binding” (molecular function). These terms are functionally associated with cytoskeletal remodeling and ECM dynamics, which play crucial roles in CIC formation.

Fig. 1.

Identification of differentially expressed CICGs. (A) Heatmap illustrated the expression of 50 differentially expressed CICGs between HCC tumor and non-tumor groups. (B) Protein-protein interaction network. (C, D) Bar plot and bubble plot of GO analysis showed the top 10 biological functions of the differentially expressed CICGs in the biological processes, cellular components, and molecular functions. (E, F) Bar plot and bubble plot of KEGG analysis showed the top 30 signaling pathways. ∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001

KEGG pathway analysis (Fig. 1E-F) demonstrated that these CICGs were primarily involved in the “Regulation of actin cytoskeleton,” “Focal adhesion,” “Proteoglycans in cancer,” and “Rap1 signaling pathway” - pathways known to be critically linked to CIC, immune regulation, and tumorigenesis. Furthermore, protein-protein interaction network analysis by STRING identified catenin beta 1 (CTNNB1), cell division cycle 42 (CDC42), and Ras homolog family member A (RHOA) as key hub genes among the CICGs (Fig. 1B).

CICGs consensus clustering

Consensus clustering was carried out utilizing the expression profiles of 50 CICGs sourced from the TCGA database to determine the connection between CIC and clinical outcomes in individuals with HCC. The ideal number of clusters was established to be k = 2 (Fig. 2A–B), categorizing the cohort into two distinct subgroups (Cluster 1 and Cluster 2). Markedly reduced OS was observed in Cluster 1 compared to Cluster 2. (Fig. 2C). These findings underscore the prognostic relevance of CIC in HCC, suggesting its potential role in tumorigenesis and disease progression.

Fig. 2.

CICGs consensus clustering. (A) Consensus clustering of CICGs samples from TCGA, and the optimal clustering was represented by k = 2. (B) Consensus clustering cumulative distribution function (CDF) for k = 2 to 9. (C) Survival analysis between cluster 1 and cluster 2

Construction of CICGs prognostic signature

Given the significant correlation between CIC and HCC prognosis, the differentially expressed CICGs may influence OS in HCC patients. To develop a robust prognostic evaluation system, we constructed a 6-CICGs signature. Initially, univariate Cox analysis showed 3,371 differentially expressed CICGs linked to OS (P < 0.05). Subsequent LASSO regression narrowed this down to 14 CICGs suitable for prognostic modeling (Fig. 3A-B). Finally, multivariate Cox analysis established a refined 6-CICGs prognostic signature comprising Chromobox 2 (CBX2), a disintegrin and metalloproteinase with thrombospondin motifs 5 (ADAMTS5), lactate dehydrogenase A (LDHA), procollagen-lysine, 2-oxoglutarate 5-dioxygenase 2 (PLOD2), phosphatidylinositol glycan anchor biosynthesis class U (PIGU), and secreted phosphoprotein 1 (SPP1) (Fig. 3C). The risk assessment score was derived using the following formula: Risk score = (CBX2 expression × 0.2809) + (ADAMTS5 expression × 0.5783) + (LDHA expression × 0.2872) + (PLOD2 expression × 0.2493) + (PIGU expression × 0.3696) + (SPP1 expression × 0.0777).

Fig. 3.

Construction of CICGs prognostic signature. (A, B) The Cross-Validation fit curve was calculated by lasso regression analysis. (C) The 6 CICGs construct the prognostic signature selected by Multivariate Cox regression analysis. ∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001

Assessment of the CICGs prognostic signature

To investigate the signature’s generalizability, we evaluated its predictive performance in the ICGC cohort, which served as an external validation set. Patients in both datasets were separated into two groups by applying the average risk score threshold. Risk scores and survival status was illustrated using ranked patient plots (Fig. 4A–B) and scatterplots (Fig. 4C–D) for training and test datasets, respectively. Survival analysis displayed that people with elevated risk scores had markedly inferior OS compared to those with lower risk scores, in both the training and test cohorts (Fig. 4E-F). The prognostic signature demonstrated robust predictive accuracy in time-dependent ROC analysis, with 1-year AUCs of 0.789 (training) and 0.786 (validation). At 3 years, AUCs were 0.753 and 0.758, and at 5 years, 0.758 and 0.756, respectively.

Figs. 4.

Validation of the CICGs Prognostic Signature. (A, B) Distribution of risk scores in patients in train set (A) and test set (B) with HCC. (C, D) Scatterplots of patients in train set (C) and test set (D) with HCC and different survival statuses. (E, F) Kaplan–Meier survival curves of the high-risk group and low risk in the train set (E) and test set (F). (G, H) Time-dependent ROC curve analysis of the high-risk group and low risk in the train set (G) and test set (H)

Multivariable Cox analyses, adjusted for clinical variables, demonstrated that the risk score persisted as a meaningful prognostic indicator in both the training and validation groups (Fig. 5A-D). Comparative analysis of ROC curves demonstrated that the CICGs signature outperformed traditional clinical parameters in the training cohort (Fig. 5E). However, in the test cohort, age and gender showed higher AUC values (Fig. 5F).

Fig. 5.

Validation of the Predictive Values of the CICGs Prognostic Signature. (A, B) The Forest plot represented the results of univariate Cox regression analysis in the training set, and test set, respectively. (C, D) The Forest plot represented the results of multivariate Cox regression analysis in the train set, and test set, respectively. (E, F) Time-dependent ROC curve analysis of the high-risk and low-risk group in the train set (E) and test set (F)

Development of predictive nomogram integrating risk score and clinical parameters for HCC patients

We created a nomogram that integrates the risk score with clinical factors, including age, tumor grade, and Tumor-Node-Metastasis (TNM) stage, to improve the precision of prognosis prediction for HCC patients in the TCGA cohort. This nomogram facilitates the computation of survival probability at 1, 3, and 5 years by aggregating the points allocated to each variable on the scoring scale (Fig. 6A).

Fig. 6.

Nomogram and immune response. (A) A nomogram consists of risk score and stage predicted the OS of HCC patients at 1-, 3-, and 5-years. The ssGSEA scores of 16 immune cells (B) and 13 immune-related functions (C) in the TCGA cohort. ∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001

Immune cells infiltration and Immune-Related pathways

To examine disparities in immunological status, we assessed the enrichment of immune cells and pathways linked to immunity between the groups with higher and lower risk. In the TCGA cohort, the high-risk group exhibited a substantial elevation in immune cells, including dendritic cells, macrophages, and regulatory T cells, while B cells, mast cells, and natural killer cells were markedly downregulated (all adjusted P < 0.05; Fig. 6B).

Regarding immune-related pathways, several processes—including antigen-presenting cell (APC) co-inhibition, C-C chemokine receptor signaling, APC co-stimulation, immune checkpoint signaling, human leukocyte antigen expression, major histocompatibility complex class I activity, and parainflammation—were markedly enhanced in the high-risk group. Cytolytic activity, type I interferon response, and type II interferon response were significantly downregulated (all adjusted P < 0.05; Fig. 6C).

Validation of genes expression levels in high-risk and low-risk group and tissues

To confirm signature’s robustness, we examined the differential expression patterns of the six CICGs between the two subgroups (Fig. 7A). Moreover, qRT-PCR examination of HCC patient samples demonstrated a substantial elevation in the mRNA levels of the six hallmark genes (CBX2, ADAMTS5, LDHA, PLOD2, PIGU, and SPP1) in tumor tissues relative to adjacent non-tumor tissues (Fig. 7B). These results align with our earlier findings, confirming the consistent upregulation of these genes in HCC tumors.

Fig. 7.

Expression differences of 6 signature genes. (A)Heatmap illustrated the expression of 6 prognostic-related genes in the high-risk group and low-risk group in the TCGA cohort. (B) Relative expression levels of CBX2, ADAMTS5, LDHA, PLOD2, PIGU, and SPP1 in tumor and para-tumor tissues. ∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001

Discussion

The management of HCC is hindered by high rates of recurrence and metastasis, emphasizing the need for reliable prognostic markers. In this study, we identified a six-gene prognostic signature associated with CICs, consisting of CBX2, ADAMTS5, LDHA, PLOD2, PIGU, and SPP1. This signature was established through a three-step analytical process involving univariate Cox regression, LASSO regression, and multivariate Cox regression, applied to both the TCGA and ICGC cohorts. To validate these findings, we further assessed the differential expression of the six genes between tumor and adjacent non-tumor tissues using qRT-PCR, thereby confirming the robustness of the signature.

CBX2, a core constituent of the Polycomb Repressive Complex 1 (PRC1), is critically involved in numerous biological processes, including cell differentiation, proliferation, and oncogenesis. Its overexpression has been shown to preserve the undifferentiated state of cancer stem cells (CSCs), thereby contributing to tumor aggressiveness and diminished patient survival. In contrast, CBX2 deficiency induces chromosomal instability, which disrupts proper cell division and hampers proliferative capacity [51]. In HCC, elevated CBX2 expression is significantly associated with poor clinical outcomes, in agreement with our findings [52].

CBX2 is also implicated in key signaling pathways related to CIC formation and tumor progression. Notably, it activates the Wnt/β-catenin pathway by promoting nuclear translocation of β-catenin, thereby sustaining Wnt-mediated transcriptional activity. This activation supports CSC self-renewal and enhances tumor cell proliferation, which collectively reinforce malignant phenotypes [25]. Furthermore, CBX2 governs the self-renewal and differentiation of stem cells through the Notch signaling pathway, and its overexpression facilitates tumor cell proliferation by augmenting Notch activation [53].

Beyond Wnt and Notch, CBX2 further facilitates oncogenic processes through activation of the PI3K/AKT axis and induction of epithelial–mesenchymal transition (EMT), a hallmark of enhanced tumor cell mobility and immune escape [24]. EMT is critical for tumor cell migration and invasiveness, enabling immune evasion through the CSC phenotype and enhancing metastatic potential. By modulating these key signaling pathways, CBX2 not only reinforces the self-renewal capacity of CSCs but also plays a key role in tumor metastasis.

ADAMTS5 is a metalloproteinase that significantly contributes to tumor cell migration and invasion by degrading components of the extracellular matrix (ECM). Its overexpression is closely associated with activation of the TGF-β signaling pathway, which promotes EMT and endows tumor cells with enhanced migratory and invasive properties, thereby facilitating the formation and maintenance of CICs [54]. By breaking down collagen and other ECM components, ADAMTS5 enables tumor cells to detach from the primary site and infiltrate surrounding tissues, promoting metastasis. Moreover, ADAMTS5 has been shown to engage the PI3K/AKT pathway, with its activation enhancing CIC survival and proliferation [55].

In addition to ECM remodeling, ADAMTS5 plays a role in modulating immune responses [27]. TThrough regulation of cell adhesion and ECM degradation, it influences tumor cell motility and intercellular interactions essential for CIC formation, underscoring its relevance in cancer progression. In glioblastoma, ADAMTS5 has been reported to enhance cellular invasiveness [56]. Furthermore, the related protein ADAMTSL5 is implicated in hepatocellular carcinoma progression and drug resistance; its depletion has been shown to increase sensitivity to therapeutic agents [28].

Therefore, by activating signaling pathways such as TGF-β and PI3K/AKT, ADAMTS5 not only contributes to tumor cell migration and metastasis but also supports the survival of CICs in the tumor microenvironment, further promoting tumor expansion and progression.

LDHA is crucial in the energy metabolism of tumor cells, especially during metabolic reprogramming. Tumor cells often acquire energy by enhancing anaerobic glycolysis, a phenomenon known as the Warburg effect. By catalyzing lactate production, LDHA sustains energy supply for tumor cells in hypoxic environments, which is crucial for tumor growth, metastasis, and immune evasion [57]. Moreover, LDHA facilitates tumor cell proliferation and survival by activating the PI3K/AKT/mTOR pathway [58, 59]. It also participates in the adaptive response of tumor cells via the hypoxia-Inducible factor 1 alpha (HIF-1α) pathway, enabling tumor cells to survive under conditions of hypoxia and nutrient deprivation, which is essential for the maintenance of CICs [58]. Overexpression of LDHA helps CICs survive in unfavorable microenvironments and enhances their resistance to chemotherapy. Furthermore, LDHA is closely associated with plasticity of breast CSC and the recruitment of tumor-associated macrophages [29]. Studies have shown that LDHA phosphorylation facilitates cancer cell invasion, antioxidant defense, and tumor metastasis. LDHA expression is markedly increased in pancreatic cancer, prostate cancer, and gliomas compared to normal tissues, with especially high levels in HCC tissues [60].

Therefore, LDHA is not only a key regulator of tumor metabolism but also plays a critical role in the formation and maintenance of CICs, providing survival support for CICs within the tumor microenvironment.

PLOD2, an essential enzyme in collagen synthesis, significantly contributes to tumor growth by facilitating collagen cross-linking to reinforce the extracellular matrix (ECM). The activation of PLOD2 promotes the EMT process by modulating the TGF-β signaling pathway, endowing tumor cells with enhanced migratory and invasive capabilities [61]. EMT is a process in which tumor cells transition from an epithelial state to a more migratory mesenchymal state, which is crucial for the formation and metastasis of CICs.

Beyond the TGF-β pathway, PLOD2 also regulates tumor cell migration and invasion through the HIF-1α pathway and microRNA-26a/b signaling. Under hypoxic conditions, HIF-1α upregulates PLOD2, promoting collagen cross-linking and increasing the stiffness of the tumor microenvironment, thereby further enhancing tumor cell migration and invasion [62]. Additionally, microRNA-26a/b suppresses PLOD2 expression, reducing collagen cross-linking and weakening tumor cell invasiveness.

High expression of PLOD2 is significantly associated with increased tumor cell aggressiveness and metastatic potential. By altering the structural organization of the tumor microenvironment, PLOD2 provides a supportive niche for the survival and expansion of CICs [32]. Through enhanced collagen cross-linking stability, PLOD2 improves the tumor growth environment, facilitating CIC survival during metastasis. Studies have demonstrated that PLOD2 overexpression correlates with poor patient prognosis, particularly in liver, lung, and breast cancers [63, 64]. PLOD2 not only facilitates tumor cell migration and invasion but also is essential for the production and sustenance of CICs.

PIGU is essential for intracellular lipid and glucose metabolism. It augments the proliferative and metabolic capabilities of tumor cells via modulating the PI3K/AKT/mTOR and MAPK signaling pathways [65–67]. Metabolic reprogramming in tumor cells enables CICs to survive under hypoxic and nutrient-deficient conditions. PIGU promotes adaptive metabolic responses in cells, strengthens the self-renewal ability of CICs, and aids tumor cells in surviving therapeutic stress. Furthermore, PIGU is closely associated with tumor cell resistance to chemotherapy and radiotherapy. By boosting metabolic activity in tumor cells, it enhances the survival of CICs, thereby serving as a key gene regulating their persistence and expansion.

The overexpression of PIGU is notably prevalent in HCC and is associated with reduced patient survival. Research has shown that PIGU expression is markedly elevated in HCC tissues, and its overexpression promotes the proliferation, motility, and invasive potential of HCC cells while inhibiting apoptosis [68]. PIGU also promotes the NF-κB pathway, enhancing immune evasion and facilitating tumor cell survival, migration, and invasion, thus further advancing HCC progression [68].

Consequently, PIGU is essential for tumor metabolic regulation and is crucial for the creation and maintenance of CICs, positioning it as a potential target for therapeutic intervention [69].

SPP1 is an essential protein in the extracellular matrix that facilitates cell adhesion, migration, and invasion. SPP1 modulates tumor cell adherence to the matrix and augments their migratory ability by interacting with signaling molecules, including Integrin and focal adhesion kinase (FAK). The overexpression of SPP1 may indirectly enhance the self-renewal and preservation of CIC characteristics by activating the PI3K/AKT and TGF-β pathways [36, 70].

Furthermore, SPP1 alters the structure and rigidity of the ECM, increasing tumor cell invasiveness and facilitating immune evasion. Increased SPP1 expression has been noted in multiple malignancies and correlates with unfavorable prognosis [71]. In HCC, SPP1 upregulates PD-L1 expression, aiding tumor cells in evading immune surveillance [72]. Additionally, SPP1 enhances tumor metastasis by remodeling the ECM and promoting tumor cell migration. Through these mechanisms, SPP1 not only supports the survival of CICs within tumors but also drives tumor progression and metastasis, establishing itself as a key regulator of tumor advancement [36, 73, 74].

The genes CBX2, ADAMTS5, LDHA, PLOD2, PIGU, and SPP1 cooperatively promote the self-renewal, metabolic reprogramming, invasiveness, and metastatic potential of CICs by regulating multiple key signaling pathways. Through enhancing the Wnt/β-catenin, Notch, TGF-β, and PI3K/AKT/mTOR pathways, they improve tumor cell survival and drug resistance, thereby driving tumor advancement and metastasis. Comprehending the functions of these genes in CICs and their regulatory mechanisms not only clarifies the basic attributes of tumors but also offers prospective tactics for future therapeutics aimed at targeting cancer stem cells.

Furthermore, we analyze the differential expression between the groups with higher and lower risk levels. Figure 8 illustrates that the KEGG analysis indicated these high-risk groups were predominantly linked to the tumorigenesis and metastasis of HCC. The pathways enriched in the “carboxylic acid catabolism,” “fatty acid metabolism,” and “collagen-containing ECM.” It suggested that CIC would affect the generation, development and prognosis of HCC through carboxylic acid catabolic, fatty acid metabolic and ECM composition. Fatty acid metabolism is abnormally activated in most tumors [75], and fatty acid production is crucial for the advancement and progression of breast cancer [76]. And as we know, collagen and ECM are necessary for cell growth, migration and communication, and ECM deposition is considered a hallmark of cancer [77].

Fig. 8.

(A, B) KEGG analysis of the different expressions between the high-risk group and low-risk group in the HCC patients from TCGA

In light with no particularly effective way for HCC treatment, our signature may provide some clues for further clinical research. Previous research reported that prognostic signatures with related genes from various biological processes such as immune microenvironment and autophagy [78–80].The signature exhibited excellent predictive performance, with elevated AUCs for 1-, 3-, and 5-year OS. It largely due to CICs closely correlated with poor prognosis [81]. Therefore, exploring CICs in clinical implication may provide new potential target during cancer diagnosis and therapy.

The six-gene signature identified in this study, including CBX2, ADAMTS5, LDHA, PLOD2, PIGU, and SPP1, demonstrates significant prognostic value in hepatocellular carcinoma (HCC) and shows potential for future clinical application. First, this gene signature may serve as a useful tool for risk stratification, helping clinicians to identify patients at higher risk of tumor recurrence or poorer survival outcomes. Such stratification could support more personalized treatment planning, including decisions regarding adjuvant therapies and postoperative monitoring schedules, ultimately contributing to improved patient care. Second, several genes within the signature are implicated in key oncogenic pathways, suggesting possible avenues for targeted therapies. For example, LDHA, a regulator of glycolytic metabolism, could potentially act as a biomarker for metabolism-focused treatments [82]. SPP1, which influences the tumor immune microenvironment, might help predict immunotherapy responsiveness [83]. In addition, CBX2 and PLOD2, associated respectively with tumor proliferation and extracellular matrix remodeling, may represent promising molecular targets warranting further investigation [84, 85]. In summary, this six-gene signature not only provides valuable prognostic insight but also establishes a biologically meaningful framework that may support the advancement of precision medicine approaches in HCC. Nevertheless, extensive experimental validation and prospective clinical studies are required to confirm its robustness, generalizability, and practical utility in clinical settings.

While this study provides potentially meaningful insights, several limitations should be acknowledged to offer a more balanced perspective. First, the analysis relies on retrospective public datasets, which may introduce selection bias due to inconsistencies in sample collection, clinical annotation, and data processing. Additionally, heterogeneity in ethnicity, underlying etiology, and geographic origin among patient cohorts may affect the generalizability of the findings. Although the prognostic signature was validated in independent datasets and supported by qRT-PCR analysis, the mechanistic roles of the identified genes in CIC formation and tumor progression remain to be elucidated. Further experimental studies are required to clarify their functional relevance. Moreover, future research involving large, multi-center cohorts with diverse clinical characteristics is essential to confirm the reproducibility and clinical applicability of this CIC-related gene signature.

Conclusion

We identified and validated a CICGs prognostic signature across multiple datasets, which demonstrated superior performance in predicting OS in HCC compared to six existing gene signatures. Furthermore, our data demonstrated substantial disparities in the infiltration levels of six immune cell types across the groups categorized by risk levels. Additionally, seven immune-related activities were markedly increased in the high-risk group, while two displayed a contrary tendency.

Abbreviations

ADAMTS5

A Disintegrin And Metalloproteinase with Thrombospondin Motifs 5

APC

Antigen-Presenting Cell

AUC

Area Under the Curve

B cell

B lymphocyte

CBX2

Chromobox 2

CCR

C-C Chemokine Receptor

CDC42

Cell Division Cycle 42

CDH1

E-Cadherin

CDH3

P-Cadherin

CIC

Cell-in-cell

CICGs

Cell-in-cell associated genes

CICs

Cell-in-cell structures

CIN

Chromosomal instability

Cox

Cox proportional hazards regression

CSC

Cancer Stem Cell

Ct

Cycle threshold

CTNNA

α-Catenin

CTNNB

β-Catenin

CTNNB1

Catenin beta 1

Cvl

Cross-validation likelihood

DEGs

Differentially expressed genes

ECM

Extracellular matrix

EMT

Epithelial-mesenchymal transition

FAK

Focal Adhesion Kinase

Fas/FasL

Fas receptor/Fas ligand

FC

Fold change

FcR

Fc receptor

FLCV

Folliculin

FDR

False discovery rate

GO

Gene Ontology

HCC

Hepatocellular carcinoma

HIF-1

Hypoxia-Inducible Factor 1 Alpha

HLA

Human Leukocyte Antigen

ICGC

International Cancer Genome Consortium

KEGG

Kyoto Encyclopedia of Genes and Genomes

KM

Kaplan-Meier

LASSO

Least Absolute Shrinkage and Selection Operator

LDHA

Lactate Dehydrogenase A

MHC

Major Histocompatibility Complex

OS

Overall Survival

PI3K/Akt/mTOR

The phosphatidylinositol 3-kinase/protein kinase B/mammalian target of rapamycin

PD-L1

Programmed Death-Ligand 1

PCA

Principal Component Analysis

PIGU

Phosphatidylinositol Glycan Anchor Biosynthesis Class U

PRC1

Polycomb Repressive Complex 1

PLOD2

Procollagen-lysine,2-oxoglutarate 5-dioxygenase 2

qRT-PCR

Quantitative Real-Time PCR

Rho

Ras Homolog

RhoA

Ras Homolog Family Member A

RNA-seq

RNA sequencing

ROC

Receiver Operating Characteristic

ROCK

Rho-associated Coiled-coil Containing Protein Kinase

SPP1

Secreted Phosphoprotein 1

T cell

T lymphocyte

TCGA

The Cancer Genome Atlas

TLRs

Toll-like receptors

TNM

Tumor-Node-Metastasis

Author contributions

M-FH, R-ZW, WL and QS conceptualized the research endeavor and authored the manuscript. HZ, DW, and Y-SC conducted the majority of the experiments and data analysis. HZ, D-HZ, C-XW and R-QL assisted with experiments. All authors evaluated and approved the final version of the paper.

Funding

This research received funding from the National Natural Science Foundation of China (No. 82373069), GuangDong Basic and Applied Basic Research Foundation (No. 2022A1515220130), Science and Technology Program of Guangzhou (2024A04J6489), GuangDong Medical Products Administration Science and Technology Project (No. 2023ZDZ02, 2023YDZ02 and 2023YDZ04), Medical Science and Technology Research Foundation of GuangDong Province (No. B2022285, B2022321 and B2023363), the CAMS Innovation Fund for Medical Sciences (2021-I2M-5-008).

Data availability

The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Human Research Ethics Committees of the First Affiliated Hospital of Sun Yat-sen University (approval No. [2020]339), in accordance with the Helsinki Declaration.Informed consent, including both Consent to Participate and Consent to Publish, was obtained from all individual participants involved in the study. All participants were 18 years of age or older at the time of consent.

Consent for publication

All authors have read the manuscript and approve of its submission.

Competing interests

The authors declare no competing interests.