Identification of anoikis-related subtypes in hepatocellular carcinoma and construction of prognostic model: Construction of prognostic model

Abstract

This study aimed to investigate the distinct subtypes of anoikis in hepatocellular carcinoma (HCC) and their underlying molecular mechanisms, and to construct a prognostic risk model. The gene expression profiles of HCC were downloaded from the cancer genome atlas and gene expression omnibus database, while anoikis-related genes were obtained from the GeneCards database. Unsupervised clustering algorithms were applied based on the expression of differentially expressed anoikis genes to identify related subtypes. Survival curves were used to analyze the differences in survival between subtypes, and the enrichment pathways and immune microenvironment differences were explored. Limma analysis, Cox proportional hazards regression analysis, and Least absolute shrinkage and selection operator regression algorithms were utilized to identify genes affecting the prognosis of subtypes and to construct a prognostic model, while also classifying high and low-risk groups for immune correlation analysis. The study identified 2 subtypes, C1 and C2, which showed significant differences in survival probability, enriched pathways, and immune microenvironment. SFN, BUB1, BSG, and HMOX1 were identified as prognostic genes, and a prognostic model was successfully constructed. There were significant differences in prognosis, expression of prognostic genes, and immune microenvironment between the high-risk and low-risk groups. The identification of different subtypes of anoikis in HCC and the construction of a prognostic model provide new insights and directions for the treatment and prognosis research of HCC.

1. Introduction

Hepatocellular carcinoma (HCC), the main type of primary malignant tumor of the liver, accounts for up to 90% and ranks fifth among global tumors, being the second leading cause of cancer- related deaths.^[1,2] HCC patients generally have a poor prognosis and a short overall survival period, and most die due to tumor metastasis and recurrence, which is closely related to its significant heterogeneity and invasiveness.^[3]

Anoikis, a form of programmed cell death associated with the loss of extracellular matrix, is a self-defense mechanism of organisms to prevent the growth of abnormal cells.^[4] Studies have shown that tumor cells can escape anoikis through paracrine mechanisms, thereby promoting tumor invasion and metastasis, and inhibiting the resistance of tumor cells to anoikis may effectively prevent tumor formation and metastasis.^[5] Although previous studies have pointed out the connection between anoikis and HCC, there is still a lack of in-depth analysis of the subtype differences of anoikis in HCC and their impact on prognosis.

Leveraging publicly available transcriptomic datasets from the cancer genome atlas (TCGA)-LIHC and gene expression omnibus, in conjunction with curated anoikis-related geness sets from the GeneCards database, this study aimed to deeply explore the subtypes of anoikis in HCC, its molecular mechanism, and construct a prognostic model, in order to provide a clearer understanding of the intrinsic heterogeneity of HCC and provide a scientific basis for treatment decisions and prognostic evaluation.

2. Material and methods

2.1. Data acquisition

Gene expression profiles of HCC were downloaded from TCGA database (https://portal.gdc.cancer.gov/, accessed on May 20, 2024) in FPKM (Fragments Per Kilobase of transcript per Million mapped reads) format. The FPKM values had been pre-normalized for library size by standardized RNA-Seq pipelines. The TCGA-LIHC dataset comprised 50 normal liver tissues and 374 tumor specimens (from 371 patients). To ensure comparability across samples and mitigate heteroscedasticity effects associated with highly expressed genes, log₂transformation [log₂(FPKM + 1)] was additionally applied. Differential expression analysis between tumor and normal tissues was conducted using the R package “limma” (version 3.40.6), with differentially expressed genes (DEGs) defined by |log₂ fold change| > 2 and adjusted P-value < .05.

For external validation, the GSE116174 dataset containing 64 HCC samples was retrieved from gene expression omnibus database (https://www.ncbi.nlm.nih.gov/geo/, accessed on May 20, 2024). Probe IDs were mapped to gene symbols using platform annotation files (GPL13158), and expression values for probes targeting identical genes were averaged. A comprehensive set of 919 anoikis-related genes was retrieved from GeneCards database (https://www.genecards.org/, accessed on May 20, 2024) using “Anoikis” as the search term, of which 569 genes with relevance scores >0.4 were selected for subsequent analysis. The intersection of DEGs and anoikis-related geness yielded 83 anoikis-related DEGs for downstream analyses.

2.2. Unsupervised clustering analysis of anoikis-related genes

Unsupervised clustering was performed based on the expression of the anoikis-related differential gene set to explore the possible clustering-related subtypes in the TCGA-LIHC dataset. The R package “ConsensuClusterPlus” was used to perform the above operations. The clustering distance was pearson, the clustering method was pam, the sampling proportion was 0.8, and 10 repetitions were performed to ensure the stability of the classification. The Kaplan–Meier (K-M) survival curve was used to distinguish the survival differences between subtypes, and P < .05 was considered statistically significant.

2.3. Biological differences of different subtypes

Gene set variation analysis was conducted for Kyoto Encyclopedia of Genes and Genomes pathway enrichment to investigate biological pathway differences between subtypes. Differential pathway activity was assessed using limma-based analysis of GSVA enrichment scores. The cell type identification by estimating relative subsets of RNA Transcripts (CIBERSORT) algorithm was applied to the FPKM expression matrix of TCGA - LIHC to study the infiltration relationship between different subtypes and 22 types of immune cells, and to analyze the differences in immune cell infiltration among different subtypes. Finally, the log2(FPKM + 1) transformed data were used as input to calculate the TME scores of each HCC sample through the estimation of stromal and immune cells in malignant tumor tissues using expression data (ESTIMATE) algorithm, including aspects such as ImmuneScore, StromalScore, and ESTIMATEScore, aiming to explore the differences in the immune microenvironment of different subtypes. P. adjust < .05 indicates that the difference is statistically significant.

2.4. Construction and validation of prognostic risk model

Following exclusion of samples with missing gene expression data, indeterminate survival status, or survival time of zero, 365 samples were included for model development. Stratified random sampling based on gender, age, and TNM stage was performed to partition the cohort into training (n = 219, 60%) and testing (n = 146, 40%) sets. Baseline characteristics were compared using Chi-square test for categorical variables (gender, tumor stage) and independent samples t test for continuous variables (age), with P < .05 indicating statistical significance.

Univariate Cox proportional hazards regression (Cox) was performed in the training cohort to identify candidate prognostic genes (P < .05), with subsequent verification of the proportional hazards (PH) assumption (P > .05). Least absolute shrinkage and selection operator (LASSO) regression was implemented using the “glmnet” R package (nfolds = 10, alpha = 1) to prevent overfitting, with lambda.min selected as the optimal regularization parameter. Multivariate Cox regression was subsequently performed on LASSO-selected genes, and genes satisfying PH assumptions were retained for model construction. After calculating the risk scores of the training set and the test set, z-score was used for standardization, and the high-risk and low-risk groups were divided, where >0 was the high-risk group and ≤0 was the low-risk group. The K-M survival curve and time-dependent receiver operating characteristic curve (at 1, 2, and 3 years) were drawn respectively to evaluate the performance of the model and analyze the expression of prognostic genes. The prognostic model was externally validated using the GSE116174 dataset following identical risk stratification procedures. P < .05 was considered a statistically significant difference.

2.5. Relationship between different risk groups and clinical characteristics

Associations between risk stratification and clinical characteristics (gender, age [>65 vs ≤65 years], survival status, survival time, TNM stage, and overall stage) were assessed using Chi-square tests, with P < .05 considered significant.

2.6. Analysis of immune cell infiltration between different risk groups

CIBERSORT algorithm was applied to quantify immune cell infiltration differences between high-risk and low-risk groups. ESTIMATE scores were compared between risk groups, and Pearson correlation analysis was performed to evaluate relationships between risk scores, prognostic genes, and immune cell infiltration levels. Statistical significance was set at P. adjust < .05.

3. Results

3.1. Screening of DEGs related to anoikis in HCC

A total of 1394 DEGs were obtained by limma differential analysis, including 1016 up-regulated genes and 378 down-regulated genes (Fig. 1A). The intersection of DEGs and 569 anoikis genes retrieved from the GeneCards database yielded 83 DEGs related to anoikis (Fig. 1B).

Figure 1.

Identification of anoikis-related differentially expressed genes. (A) Volcano plot of DEGs in TCGA-LIHC. X-axis: log₂(fold change); Y-axis: −log₁₀(adjusted P-value). Red dots: significantly upregulated genes; blue dots: significantly downregulated genes; gray dots: nonsignificant genes. Dashed lines indicate significance thresholds (P.adjust= 0.05) and expression fold change thresholds. (B) Venn diagram showing the intersection between TCGA-LIHC DEGs (blue circle) and anoikis-related genes (red circle). DEGs = differentially expressed genes, TCGA = the cancer genome atlas.

3.2. Identification of anoikis-related subtypes

Consensus clustering analysis of the 83 anoikis-related genes identified optimal clustering stability at K = 2, delineating 2 molecular subtypes designated C1 and C2 (Figure 2A–C). K-M survival analysis revealed significant survival disparities between subtypes, with C2 demonstrating superior survival outcomes compared to C1 (P = .012) (Fig. 2D).

Figure 2.

(A) Consensus cumulative distribution function (CDF) curves. X-axis: consensus index ranging from 0 to 1, representing the stability of sample clustering; Y-axis: cumulative distribution function (CDF) value. Curves of different colors represent the CDF curves when the number of clusters K ranges from 2 to 10. (B) Relative change in the area under the CDF curve. X-axis: number of clusters K (ranging from 2 to 10); Y-axis: relative change value of the area under the CDF curve. (C) Consensus matrix heatmap when K = 2. (D) (K–M) survival curves of C1 and C2 subtypes. X-axis: time, with the number of at-risk patients at different time points marked below the X-axis; Y-axis: survival probability. Red curve: C1 subtype; blue curve: C2 subtype.

3.3. Analysis of enriched pathway differences in different subtypes

A total of 126 signaling pathways were enriched, including 58 up-regulated and 68 down-regulated pathways. Most of the P.adjust values were very small, and the differences were highly statistically significant. The top 5 up-regulated and down-regulated signaling pathways in C1 vs C2 are shown in Table 1.

Table 1.

KEGG pathway enrichment analysis of C1 vs C2.

Enriched pathway	logFC	P. adjust
Up-regulated
1. Pathogenic_Escherichia_Coli_Infection	0.200	1.07E-17
2. Vibrio_Cholerae_Infection	0.174	1.97E-17
3. Spliceosome	0.251	1.00E-12
4. Purine_Metabolism	0.084	1.61E-11
5. RNA_Polymerase	0.204	1.85E-10
Down-regulated
1. Adipocytokine_Signaling_Pathway	−0.219	8.38E-26
2. Primary_Bile_Acid_Biosynthesis	−0.526	7.05E-25
3. Glycine_Serine_and_Threonine_Metabolism	−0.435	8.06E-25
4. Lysine_Degradation	−0.287	8.83E-25
5. PPAR_Signaling_Pathway	−0.341	5.08E-24

3.4. Analysis of the immune microenvironment of different subtypes

The CIBERSOR algorithm analysis found that there were differences in the infiltration of immune cells between the C1 and C2 subtypes. Among them, regulatory T cells (Tregs) (P. adjust < .0001), resting natural killer (NK) cells, M0 macrophages, monocytes, resting dendritic cells, resting mast cells (P. adjust < .001), activated NK, M2 macrophages (P. adjust < .05) had different infiltration levels in the 2 subtypes (Fig. 3A).

Figure 3.

Immune microenvironment characterization of molecular subtypes. (A) Differences in immune cell infiltration between C1 and C2 subtypes. X-axis: 22 different types of immune cell subsets; Y-axis: relative expression of immune cells; red boxplots: C1 subtype; blue boxplots: C2 subtype. (B) Differences in tumor microenvironment scores between C1 and C2 subtypes. X-axis: 3 types of tumor microenvironment scores, namely StromalScore, ImmuneScore, and ESTIMATEScore; Y-axis: Score, representing the enrichment degree of stromal cells and immune cells in the tumor microenvironment. *P. adjust < .05, **P. adjust < .01, ***P. adjust < .001, ****P. adjust < .0001.

The ESTIMATE algorithm analysis results showed that the immune score (P. adjus < .0001) and ESTIMATE score (P. adjus < .001) of the immune microenvironment of the C1 subtype were higher than those of the C2 subtype (Fig. 3B).

3.5. Construction of a prognostic risk model related to cell anoikis

In this study, 219 samples were randomly selected as the training dataset, and 146 samples as the test dataset. A balance test of clinical characteristics was conducted between the 2 groups: no statistically significant differences were observed in sex (χ² = 0.042, P = .837), age (t =−1.013, P = .312), or tumor TNM stage (χ² = 2.537, P = .281). These results indicated that the baseline characteristics of the 2 groups were well-balanced, which were eligible for model training and validation (Table 2).

Table 2.

Baseline characteristics between training and test sets.

Variables	Training set (n = 219)	Test set (n = 146)	Statistical test	P-value
Gender	Female: 32.0%/ Male: 68.0%	Female: 33.6%/ Male: 66.4%	χ2 = 0.042	.837
Age	59.1 ± 13.7	60.5 ± 12.9	t = −1.013	.312
Tumor stage	I–II: 77.0%/ III: 21.5%/ IV: 1.4%	I–II: 70.5%/ III: 28.8%/ IV: 0.8%	χ2 = 2.537	.281

Firstly, based on the expression profiles of 83 genes and survival data in the training dataset, univariate Cox analysis was performed to identify 21 anoikis-related genes (Fig. 4A). Among these genes, BIRC5 failed to satisfy the PH assumption in the univariate analysis (PH P < .05). Secondly, LASSO regression analysis with 10-fold cross-validation was applied to the included genes to obtain the optimal model. The lambda value was selected by minimizing the partial likelihood deviance, which was determined to be 0.0497. The corresponding model contained 7 genes with nonzero coefficients, which were further included in the multivariate Cox analysis (Figs. 4B, 4C). Finally, multivariate Cox analysis was performed, and 4 genes were ultimately included, all of which met the PH assumption (PH P > .05). The risk score formula of the constructed prognostic model was as follows:

Figure 4.

Construction and validation of the prognostic risk model. (A) Forest plot of univariate Cox regression. The forest plot shows the hazard ratios (HR) and 95% confidence intervals (CI) of 21 genes. (B) LASSO coefficient profiles. X-axis: -ln (lambda), where lambda is the regularization parameter; Y-axis: regression coefficients, representing the contribution of each gene to the prognostic risk. Curves of different colors represent different genes. As -ln (lambda) decreases, the curves continuously converge until the coefficient value is 0, indicating that the gene is eliminated. (C) Cross-validation for optimal lambda selection. X-axis: ln (lambda); Y-axis: partial likelihood deviance. The lambda value was determined when the likelihood deviance was minimized, and the red dot indicates the minimum likelihood deviance (lambda = 0.0497). (D, F) Time-dependent ROC curves for training and testing cohorts. X-axis: false positive rate; Y-axis: true positive rate. The curves show the predictive efficacy at 1, 2, and 3 years in the training dataset, with the corresponding area under the curve (AUC) and 95% confidence interval marked. (E, G) (K–M) survival curves for risk stratification. X-axis: follow-up time (years); Y-axis: survival probability. Red: high-risk group; light blue: low-risk group.(H, I) Risk score distribution, survival status, and gene expression profiles in the training dataset and test dataset. Upper part: Risk score distribution. X-axis: samples sorted in ascending order of risk scores; Y-axis: risk score (RiskScore). Cyan-blue represents the low-risk group (Low), and red represents the high-risk group (High). Middle part: Scatter plot of survival time and outcomes. Y-axis: survival time (years); green dots represent surviving patients (Alive), blue dots represent deceased patients (Dead), and the sample order is consistent with that in the upper plot. Lower part: Heatmap of characteristic gene expression. It shows the expression levels of 4 genes, namely SFN, BUB1, BSG, and HMOX1. The color from blue to red indicates the expression level from low to high, and the sample order is consistent with the risk score sorting.

$${\displaystyle \begin{array}{l}{\displaystyle \begin{array}{lllll}& {\displaystyle \mathrm{R}\mathrm{i}\mathrm{s}\mathrm{k}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=(\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{l}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{f}\mathrm{S}\mathrm{F}\mathrm{N}\times 0.006609104)}& {\displaystyle +(\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{l}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{f}\mathrm{B}\mathrm{U}\mathrm{B}1\times 0.204168345)}& {\displaystyle +(\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{l}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{f}\mathrm{B}\mathrm{S}\mathrm{G}\times 0.002336105)}& {\displaystyle +(\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{l}\mathrm{e}\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}\mathrm{f}\mathrm{H}\mathrm{M}\mathrm{O}\mathrm{X}1\times 0.002439738)}\end{array}}\end{array}}$$

In the training dataset, the time-dependent area under the curve values at 1, 2, and 3 years were 0.79, 0.71, and 0.65, respectively (Fig. 4D), while those in the test dataset were 0.70, 0.70, and 0.71, respectively (Fig. 4F). In both the training and test datasets, patients with high-risk scores exhibited significantly lower survival probabilities and earlier death times compared with those with low-risk scores (Figs. 4E, 4G). The distributions of patients with different risk scores and survival statuses in the training and test datasets are illustrated in Figures 4H and 4I, respectively.

3.6. External validation of the prognostic model

Application of the risk model to the GSE116174 validation cohort (n = 64) stratified patients into high-risk (n = 31, 48.4%) and low-risk (n = 33, 51.6%) groups. The model achieved time-dependent area under curves of 0.75 (95% confidence interval [CI]: 0.56–0.92), 0.72 (95% CI: 0.57–0.88), and 0.66 (95% CI: 0.51–0.80) for 1-, 2-, and 3-year survival, respectively (Fig. 5A). Kaplan–Meier analysis confirmed significantly higher mortality in high-risk (58.1%) versus low-risk (27.3%) groups (χ² = 8.526, P = .0035) (Fig. 5B).

Figure 5.

(A) Time-dependent ROC curve in the validation dataset. X-axis: false positive rate; Y-axis: true positive rate. The curve shows the predictive efficacy at 1, 2, and 3 years in the validation dataset, with the corresponding area under the curve (AUC) marked. (B) Risk-stratified (K–M) survival curve in the validation dataset. X-axis: follow-up time; Y-axis: survival probability. Red: high-risk group; dark blue: low-risk group.

3.7. Association between risk stratification and clinical characteristics

Comparison of high-risk (n = 132) and low-risk (n = 233) groups revealed no significant differences in demographic characteristics: gender (62.1% vs 70.4% male, χ² = 2.257, P = .133) or age (66.7% vs 59.7% ≤65 years, χ² = 1.475, P = .224). However, significant associations were observed for T stage (χ² = 32.585, P < .001) and overall stage (χ² = 28.566, P < .001), with advanced stages (T3 + T4, III + IV) more prevalent in the high-risk group. Mortality was significantly higher in high-risk (48.5%, 64/132) compared to low-risk (28.3%, 66/233) patients (χ² = 14.066, P < .001) (Fig. 6).

Figure 6.

Donut chart showing the distribution of clinical characteristics in the high- and low-risk groups. The donut chart illustrates the differences in the composition ratios of each clinical characteristic between the 2 groups, with different colors representing different categories.

3.8. Analysis of the immune microenvironment of different risk groups

The CIBERSORT algorithm analysis found that there were differences in the infiltration of immune cells between the high-risk and low-risk groups. Among them, follicular helper T cells, M0 macrophages, resting mast cells (P.adjust < .0001), resting CD4 + memory T cells, activated CD4 + memory T cells, regulatory T cells, resting NK cells (P.adjust < .01), activated NK cells, M2 macrophages (P.adjust < .05) had different infiltration levels in the 2 groups (Fig. 7A). No significant differences in ESTIMATE scores were observed between risk groups (P.adjust> .05)(Fig. 7B).

Figure 7.

(A) Differences in immune cell infiltration between the high- and low-risk groups. X-axis: 22 different types of immune cell subsets; Y-axis: relative expression of immune cells; red boxplots: high-risk group (High); cyan-blue boxplots: low-risk group (Low). (B) Differences in tumor microenvironment scores between the high- and low-risk groups. X-axis: 3 types of tumor microenvironment scores, namely StromalScore, ImmuneScore, and ESTIMATEScore; Y-axis: Score, representing the enrichment degree of stromal cells and immune cells in the tumor microenvironment.(C) Correlation analysis between risk score and immune cell types, showing Pearson correlation coefficients (r) between RiskScore and the abundance of 22 immune cell types. The color gradient from blue to red represents negative to positive correlations, respectively. (D) Correlation analysis between prognostic genes and immune cell types, with color and significance labeling rules consistent with Figure C. *P.adjust < .05, **P.adjust < .01, ***P. adjust < .001.

In the correlation analysis between risk scores and immune cells, the results showed that the risk score was positively correlated with M0 macrophages, regulatory T cells, etc, and negatively correlated with M1 macrophages, resting CD4 + memory T cells, etc (P. adjust <.05)(Fig. 7C).

The results of the correlation analysis between prognostic genes and immune cell infiltration indicated that: SFN was positively correlated with regulatory T cells, M0 macrophages, etc (P. adjust < .05);BUB1 was positively correlated with follicular helper T cells, M0 macrophages, etc, and negatively correlated with mast cells, resting NK cells, etc (P. adjust <.05); BSG was positively correlated with regulatory T cells and negatively correlated with M1 macrophages (P. adjust < .05); HMOX1 was positively correlated with neutrophils (P. adjust < .05)(Fig. 7D).

4. Discussion

HCC represents a highly aggressive malignancy with dismal clinical outcomes, largely attributable to substantial genomic heterogeneity driving diverse biological behaviors and variable treatment responses.^[6,7] Anoikis resistance constitutes a critical factor facilitating tumor metastasis and invasion, rendering the characterization of anoikis-related molecular heterogeneity essential for prognostic stratification and therapeutic optimization. This study integrates transcriptomic profiling with anoikis-related gene signatures to identify clinically relevant molecular subtypes and develop a robust 4-gene prognostic model, offering novel insights into HCC molecular stratification and personalized therapeutic strategies.

4.1. Biological significance and causal mechanisms of anoikis-related subtypes

The identification of C1 and C2 molecular subtypes, characterized by distinct survival outcomes, pathway enrichment patterns, and immune microenvironment landscapes, fundamentally reflects differential intrinsic capacities for anoikis resistance. These molecular phenotypes emerge from complex causal networks linking pathway dysregulation with immune microenvironment remodeling.

The C1 subtype exhibited a pro-inflammatory and high anoikis resistance phenotype. Inflammatory pathways such as pathogenic Escherichia coli infection and Vibrio cholerae infection were upregulated, prompting tumor cells to secrete pro-inflammatory cytokines, particularly tumor necrosis factor-α (TNF-α) and interleukin-6 (IL-6). The release of TNF-α and IL-6 activated the NF-κB and STAT3 pathways, triggering a molecular cascade of anoikis resistance.^[8,9] On the one hand, it regulated the expression of apoptosis-related molecules such as Bcl-2 and Caspase-3, inhibiting the programmed death of tumor cells after loss of adhesion. On the other hand, the phosphorylated activation of STAT3 regulated the expression of cell adhesion-related molecules such as focal adhesion kinase and integrin β1, enhancing the nonspecific binding between tumor cells and the matrix. Meanwhile, it promoted the epithelial-mesenchymal transition (EMT) process, enabling cells to acquire stronger invasive ability and survival capacity after matrix detachment.^[10,11] In addition, the significant upregulation of spliceosome and RNA polymerase pathways strongly suggested the presence of extensive abnormalities in transcriptional and posttranscriptional regulation. Abnormal splicing events can generate oncogenic isoforms to drive tumor progression,^[12,13] which is also the core molecular reason for the stronger invasiveness of the C1 subtype. The high infiltration of Tregs and M0 macrophages in the C1 subtype is the result of active remodeling by tumor cells through the secretion of factors such as CCL22 and TGF-β via the NF-κB pathway.^[14,15] Immune-suppressive cells not only impair antitumor immunity but also activate the PI3K/AKT pathway by secreting immune-suppressive factors,^[16] forming a positive feedback loop of immune suppression and anoikis resistance, which ultimately leads to extremely poor prognosis of the C1 subtype.

The C2 subtype exhibited a metabolic reprogramming and low anoikis resistance phenotype. The enrichment of metabolic pathways such as adipocytokine and PPAR signaling pathways had a direct causal relationship with anoikis sensitivity. Activation of the PPAR pathway could inhibit the expression of EMT-related transcription factors, while maintaining mitochondrial function to reduce oxidative stress damage.^[17,18] Enrichment of the primary bile acid biosynthesis pathway regulated the expression of cell adhesion molecules through farnesoid X receptor (FXR), enhancing the adhesion ability of tumor cells to the matrix and increasing their sensitivity to anoikis.^[19,20] The high infiltration of NK cells and monocytes in the C2 subtype was closely related to metabolic pathway regulation. The secretion of CXCL10 induced by PPAR pathway activation specifically recruited NK cells, and the IFN-γ released by NK cells could further inhibit the EMT process of tumor cells,^[17] forming a negative feedback loop of innate immune activation and inhibition of anoikis resistance. This bidirectional regulation between the immune microenvironment and anoikis is an important biological basis for the significantly better prognosis of the C2 subtype than the C1 subtype.

4.2. Regulatory mechanisms of prognostic genes

The 4 prognostic genes identified in this study: SFN, BUB1, BSG, and HMOX1: regulate core pathways of anoikis resistance while mediating tumor immune microenvironment remodeling, serving as critical molecular biomarkers for HCC prognosis.

SFN encodes 14–3-3σ protein and exerts dual regulatory functions in anoikis resistance and immune modulation. Mechanistically, SFN promotes tumor cell survival through 3 principal pathways: inhibition of c-Cbl-mediated EGFR ubiquitination and degradation, sustaining ERK1/2 signaling activation, upregulating Bcl-2, and suppressing Caspase-3 activation and PARP cleavage^[21]; phosphorylation-dependent inhibition of GSK-3β (Ser9) promoting β-catenin nuclear translocation and Wnt/β-catenin pathway activation, inducing EMT and reducing cell-cell adhesion^[22]; and direct binding to AKT preventing PHLPP2-mediated dephosphorylation, resulting in constitutive AKT activation and sorafenib resistance.^[23] These coordinated mechanisms confer anchorage-independent growth capacity and metastatic potential. In immune modulation, SFN promotes CCL17/CCL22 secretion recruiting Tregs and enhances Treg survival and suppressive function, constructing an immunosuppressive microenvironment that reciprocally activates tumor cell AKT signaling, forming a “anoikis resistance-immune evasion” positive feedback network.^[21,23]

The BUB1 (Budding Uninhibited by Benzimidazoles 1) is a key gene encoding a serine/threonine protein kinase. Overexpression of BUB1 is closely associated with rapid proliferation, migration, and invasion of tumor cells. In HCC, BUB1 is related to the downregulation of antigen-presenting molecules, upregulation of immune checkpoint molecules, and decreased infiltration of dendritic cells, which constructs an immunosuppressive microenvironment, leading to immune therapy resistance and poor prognosis.^[24] In this study, the regulatory role of BUB1 in HCC associated with specific T cell subsets (such as follicular helper T cells and CD4+memory T cells) remains to be further investigated.

The BSG (CD147) gene encodes a transmembrane glycoprotein and plays a crucial role in matrix metalloproteinase regulation, cell adhesion, and immune responses. BSG enhances cell adhesion ability, resists anoikis, and promotes anchorage-independent growth and metastasis by activating the integrin-FAK-PI3K-Ca²⁺ pathway.^[25] Meanwhile, as a “Warburg oncogene,” BSG activates the Akt/mTOR pathway through MCT1-mediated lactate efflux, and upregulates fatty acid synthase (FASN) and acetyl-CoA carboxylase 1 to drive lipid synthesis, providing energy support for tumor cells.^[26] In addition, BSG induces the secretion of MMP-2/MMP-9 to promote extracellular matrix degradation and enhance invasive ability,^[25] thereby facilitating tumor progression through the triple mechanism of “anoikis resistance-metabolic reprogramming-invasive metastasis.” This study found that BSG was positively correlated with regulatory T cells and negatively correlated with M1 macrophages, which is supported by existing literature. The interaction between BSG and CD98 maintains Foxp3 stability and enhances immunosuppressive function, while the specific mechanism underlying the negative correlation between BSG and M1 macrophage infiltration remains to be further investigated.^[27,28]

The HMOX1 gene is a key gene encoding heme oxygenase-1 (HO-1). In anoikis resistance, HMOX1 is activated through the ATF4/Nrf2 pathway, producing carbon monoxide (CO) and biliverdin to exert antioxidant effects, maintain cell survival after detachment from the matrix, and promote metastatic colonization.^[29] Meanwhile, high expression of HMOX1 mediates sorafenib resistance by upregulating ABC transporters such as ABCB1 and ABCG2, and is closely associated with the formation of an immunosuppressive microenvironment.^[29] HMOX1 expression is positively correlated with M2-type macrophage infiltration, promoting the polarization of TAMs toward an immunosuppressive phenotype.^[30] However, its association with neutrophils in HCC requires further validation.

4.3. Clinical value and limitations

The 4-gene prognostic model constructed in this study showed good predictive efficacy in the training set, test set, and validation set, with excellent reproducibility, and possessed dual clinical values of prognostic evaluation and therapeutic guidance. In terms of prognostic evaluation, this model can complement the traditional TNM staging, realize precise molecular stratification of HCC patients, and predict the risk of metastasis and recurrence earlier and more accurately. In terms of individualized treatment, the model can effectively identify patient groups with the same TNM stage but significantly different prognoses, providing incremental information for assisting individualized treatment decisions. In addition, the 4 core genes in the model are all potential targeted therapy targets, which can provide directions for the development of specific inhibitors. This study also has certain limitations: first, it lacks in vitro and in vivo experimental verification, and the regulatory role of core genes in HCC anoikis resistance and immune microenvironment has not been confirmed by cell and animal models; second, the sample size of the validation dataset is relatively small, and the generalization of the model needs to be verified by a larger sample size of multi-center datasets.

5. Conclusion

The identification of different subtypes of anoikis in HCC and the construction of a prognostic model provide new insights and directions for the treatment and prognosis research of HCC.

Acknowledgments

Thank you for successfully completing the thesis with the support and assistance of all authors.

Author contributions

Conceptualization: Xifeng Zhang, Feihua Chen, Rongxian Qiu, Xiangyang Ye, Zhenting Hu.

Data curation: Xifeng Zhang, Zhenting Hu.

Formal analysis: Xifeng Zhang, Rongxian Qiu, Xiangyang Ye.

Funding acquisition: Xifeng Zhang, Feihua Chen.

Investigation: Xifeng Zhang, Feihua Chen, Rongxian Qiu, Xiangyang Ye, Zhenting Hu.

Methodology: Xifeng Zhang, Feihua Chen, Rongxian Qiu, Xiangyang Ye, Zhenting Hu.

Project administration: Xifeng Zhang, Xiangyang Ye.

Resources: Xifeng Zhang, Feihua Chen, Xiangyang Ye.

Software: Xifeng Zhang.

Supervision: Xifeng Zhang, Rongxian Qiu.

Validation: Xifeng Zhang.

Visualization: Xifeng Zhang.

Writing – original draft: Xifeng Zhang.

Writing – review & editing: Xifeng Zhang.