Integrative analysis of multi-omics data identified PLG as key gene related to Anoikis resistance and immune phenotypes in hepatocellular carcinoma

Abstract

Purpose

The extracellular matrix (ECM) plays a pivotal role in the initiation and progression of hepatocellular carcinoma (HCC) by facilitating the proliferation of HCC cells and enabling resistance to Anoikis. ECM also provide structural support that aids in the invasion of HCC cells, thereby influencing the tumor microenvironment. Due to genetic variations and molecular heterogeneity, significant challenges exist in the treatment of HCC, particularly with immunotherapy, which frequently leads to immune tolerance and suboptimal immune responses. Therefore, there is an urgent need for a multi-omics-based classification system for HCC that clarifies the molecular mechanisms underlying the establishment of immune phenotypes and Anoikis resistance in HCC cells. In this study, we employed advanced clustering algorithms to analyze and integrate multi-omics data from HCC patients, with the objective of identifying key genes that possess prognostic potential associated with the Anoikis resistance phenotype. This methodology resulted in the development of a consensus machine learning-driven signature (CMLS), which demonstrates robust predictive capabilities by examining variations in epigenetics, transcription, and immune metabolism, as well as their effects on the core differential gene, plasminogen (PLG).

Results

The integrated multi-omics approach has identified PLG as a critical node within the gene regulatory network associated with Anoikis resistance and immunometabolic phenotypes. As an independent risk factor for poor prognosis in patients with HCC, PLG facilitates Anoikis resistance and enhances the migration of HCC cells. This study provides novel insights into the molecular subtypes of HCC through the application of robust clustering algorithms based on multi-omics data. The constructed CMLS serves as a valuable tool for early prognostic prediction and for screening potential drug candidates that may enhance the efficacy of immunotherapy, thereby establishing a foundation for personalized treatment strategies in HCC.

Conclusions

Our data underscore the pivotal role of PLG in the development of Anoikis resistance and the immunometabolic phenotype in HCC cells. Furthermore, we present compelling experimental evidence that PLG functions as a significant tumor promoter, suggesting its potential as a target for the formulation of tailored therapeutic strategies for HCC.

Introduction

Hepatocellular carcinoma (HCC) is one of the most common tumors and is the world’s third leading cause of cancer death. The relative 5-year survival rate is about 18% [1]. The incidence of HCC varies due to geographical and ethnic differences, with 870,000 new cases reported annually, still ranking sixth in the total number of new cancer cases [2]. Most HCC patients have chronic hepatitis B virus (HBV) or hepatitis C virus (HCV) infections, and alcohol consumption is closely linked to the disease [3]. Liver cancer onset is often hidden, leading to early intrahepatic metastasis, resulting in late-stage diagnosis and limited treatment options. Even after radical hepatectomy or liver transplantation, the 5-year overall survival rate can reach 70%, but with a high recurrence rate of up to 80% [3, 4]. However, the success of the phase III IMbrave150 trial in 2020 marks a new era in the treatment of HCC with immune checkpoint inhibitors (ICI) [3]. The combination therapy of anti-PD-L1 monoclonal antibodies and vascular endothelial growth factor (VEGF) monoclonal antibody, as well as the combination of anti-PD-1 antibodies sintilimab and bevacizumab, have shown significant survival benefits in advanced HCC patients, with superior median overall survival (OS), progression-free survival (PFS), and objective response rate (ORR) compared to sorafenib.

Although immune checkpoint inhibitors have changed the strategy of management and treatment for HCC, most patients with advanced HCC still face drug resistance and contraindications to ICI treatment. This phenomenon may be associated with the significant heterogeneity of HCC. Therefore, a therapy based on molecular subtypes is likely to help overcome this challenge. Examples include biomarker-driven approaches (e.g., targeting the FGF19-FGFR4 pathway) and combination therapies that inhibit compensation suspected to cause treatment resistance (e.g., feedback activation of the EGFR-PAK2-ERK5 pathway as a mediator of lenvatinib resistance) [3]. Hence, there is an urgent need to utilize advanced machine learning algorithms and large-scale multi-omics data to identify biomarkers that can enhance effective prognosis prediction and immunotherapy in HCC patients. This approach aims to achieve precision treatment of HCC by identifying patients who are most likely to benefit from specific therapies at an early stage.

Given an increasing number of studies highlighting the intricate interactions within various tumor omics, we employed a strategy that integrated over ten types of omics data through a random combination of 101 algorithms. Our consensus machine learning-driven signature (CMLS) and consensus groups integrate findings from multiple omics studies to comprehensively elucidate the molecular characteristics, biological behavior, survival prediction, and drug sensitivity of HCC.

We excluded algorithmic combinations that included fewer than five genes to establish a comprehensive consensus for high and low CMLS. These two groups are utilized to evaluate the Anoikis resistance activity of HCC cells, which is indicative of their survival, metastatic potential, as well as their metabolic and immune characteristics. Importantly, a multi-omics approach was employed to identify key molecules that differentiate the high and low CMLS groups, with plasminogen (PLG) identified as the core gene that connects the two groups.

PLG, have a molecular size of 92–94 kDa and exist as beta globulin forms in the blood circulation. They are crucial components of the plasminogen/plasmin activation system and are converted to plasmin by tissue plasminogen activator (tPA) or urokinase-type plasminogen activator (uPA) under normal physiological conditions [5]. The coagulation/fibrinolysis system and the immune system are interconnected and interact in various human physiological and pathological processes. The activation of the coagulation system and the subsequent PLG-mediated fibrinolysis can be seen as part of the innate immune response, initiating a protective inflammatory response against foreign invaders [6, 7]. As a key enzyme in the coagulation/fibrinolysis system, PLG not only regulates fibrinolysis and homeostasis, degrading components of the extracellular matrix and basement membrane, but also plays a role in regulating the coagulation cascade and the complement system. Therefore, PLG bridges two crucial cascades of coagulation and immunity in humans. Research has shown that PLG is involved in various physiological and pathological processes such as tumor cell proliferation, invasion, apoptosis, and tumor angiogenesis. The expression levels of PLG in different types of cancer are related to the prognosis of cancer patients, as well as the prediction of recurrence and survival rates. This clinical value is evident in breast cancer, lung cancer, ovarian cancer, cervical cancer, gastrointestinal cancer, and carcinoma of the urinary system [8, 9]. For instance, a study by Zhao et al. found that PLG was highly expressed in advanced high-grade serous ovarian cancer, and high PLG levels were positively correlated with longer patient survival time (HR = 0.59, 95% CI = 0.37–0.94, P = 0.026) [10]. In non-small cell lung cancer (NSCLC), Maynard et al. analyzed samples from 49 patients with advanced NSCLC using single-cell RNA sequencing (scRNA-seq) technology and found that the fibrinolytic enzyme/fibrinolytic enzyme system was activated in advanced NSCLC and associated with poor overall survival (P < 0.01) [11]. In clear cell renal cell carcinoma, PLG is considered a good prognostic marker [12]. However, there is limited research on PLG in HCC, especially its relationship with Anoikis resistance.

Loss of cell attachment to the extracellular matrix or inappropriate cell adhesion can cause a specific programmed cell death process called Anoikis [13]. Cancer cells must overcome Anoikis caused by loss of cell adhesion if they are to escape tumor tissue to survive, metastasize, or progress in the circulation or elsewhere. At the extracellular matrix (ECM) level, PLG on the cell surface and the plasminogen/plasmin system of peripheral proteins can induce cell retraction and fibronectin fragmentation, leading to cell shedding and morphological/biochemical changes, which are the characteristics of Anoikis [13]. The main factors affecting Anoikis are proteins that increase cell adhesion, biological processes of tight junctions, and cytoskeletal proteins, respectively, in which PLG seems to be the main protease. In recent years, the identification of Anoikis-related genes in a variety of tumors has increased rapidly, and studies have shown that Anoikis resistance is closely related to increased tumor cell migration, invasion, and metastasis.

Results

Establishment of CMLS Computational Framework

The identification of cancer genes plays a crucial role in the development of precision oncology and cancer therapeutics. In the past period of time, scientists have been trying to find cancer genes from the big data. Although thousands of cancer genome sequences have been identified and greatly contributed to the identification of cancer genes, there are still limitations and challenges. For example, the number of identified cancer genes in some tumor types remains low. Moreover, even if the DNA sequence of some genes is not altered, their dysregulation leads to cancer development and progression through mechanisms such as DNA methylation, hypomethylation, and copy number alterations. Therefore, using multi-omics data to develop an interpretable multi-omics data and machine learning approach to prioritize cancer genes from large datasets by integrating different data layers and integrating them into a single framework is particularly important. It is crucial to predict new cancer genes and interact with known cancer drivers in the PPI network to predict cancer genes or identify cancer gene modules.

Firstly, based on the TCGA-LIHC data, we identified 37 genes with the most significant differences related to the Anoikis resistance phenotype of HCC. We then conducted consensus clustering using 10 multi-omics ensemble clustering algorithms, including COCA, NEMO, PINSPlus, SNF, and LRA. By utilizing reference cluster predictors, gap statistical analysis, contour score, and previous research findings, we identified stable HCC prognostic groups: high CMLS and low CMLS. To further pinpoint HCC stable prognostic genes linked to Anoikis resistance by group, we identified hub genes with the highest C-index using various machine learning algorithms (such as stepwise Cox, CoxBoost, ridge regression, RSF, GBM, Survival SVM, LASSO, SuperPC, plsRcox, and Enet). A computational framework for CMLS was established based on the combination of CoxBoost and stepwise Cox. Finally, through Cox regression analysis, we selected 21 genes associated with the prognosis of HCC as cluster candidate genes. Our classification system exhibited a strong correlation OS (P < 0.001). Consequently, we successfully established the connection between CMLS and HCC prognosis related to Anoikis resistance, tumor metabolism, immune microenvironment, immunotherapy response, and potential therapeutic drugs.

Multiomics Consensus on Anoikis Resistance and Prognosis-related molecular groups of HCC

To identify HCC with Anoikis resistance-related genes, we analyzed 1395 genes that were differentially expressed between normal liver tissue and HCC tissue (Fig. 1A) from the TCGA database. Subsequently, through a cross-analysis with 436 genes associated with Anoikis resistance obtained from Harmonizome, we identified 37 genes linked to Anoikis resistance in HCC (Fig. 1B). These 37 genes were integrated into the computational framework for CMLS. We constructed consistent prognostic models in the training cohort based on the optimal random combination of the 10 clustering algorithms (Fig. 1C). Initially, we utilized the CoxBoost algorithm to identify the most significant genes and then applied the Cox algorithm iteratively to refine the model, resulting in a final model comprising the 21 HCC-related genes (Fig. 1D). The final model, with the highest average C-index, was constructed using a combination of CoxBoost and stepwise Cox algorithm (direction = forward). The mean C-index that kept the highest while building the final model was based on an algorithm consisting of CoxBoost and stepwise Cox (direction = forward). Finally, we calculated the average of each model in all the queue C-index to evaluate the prediction ability of all models. Prognosis based on the model is shown in Fig. 1C. We carried out the single-factor and multiple-factor Cox regression analysis. We aimed to select genes related to Anoikis resistance in HCC, which are associated with a significant OS, and to assess whether they can be used as independent risk predictors of prognosis for HCC patients. Univariate analysis showed that 21 genes were significantly associated with poor prognosis of HCC patients (P < 0.05, Fig. 1D). According to multivariate analysis, the high expression of five genes, SERPINE1, PLG, NQO1, EZH2, and EDIL3, in HCC could be used as independent risk predictors of poor prognosis in HCC patients (P < 0.05, Fig. 1E), and were closely related to Anoikis resistance phenotype. Based on the analysis of multi-omics data, we identified significant prognostic differences between the two groups categorized by high and low CMLS. Notably, patients with HCC in the low CMLS group exhibited a more favorable prognosis (Fig. 1F). Furthermore, PLG was identified as the core gene involved in the interconversion between the two groups. We also performed Protein-protein interaction (PPI) network analysis of SERPINE1, PLG, NQO1, EZH2, and EDIL3 using the STRING database. We found only PLG and SERPINE1 interaction (Fig. 1G).

Fig. 1.

The generation and prognostic value of Consensus Machine Learning-driven Signature (CMLS) and molecular landscape, and validation of hepatocellular carcinoma circulating tumor cells (CSs). (A) Volcano plot of differentially expressed genes between normal liver and HCC from TCGA database; (B) Venn diagram of 1395 differential genes in (A) and 436 genes related to anoikis resistance downloaded from Harmonizome. (C) A comprehensive computational framework generated a combination of 101 machine learning algorithms. The C-index of each model was calculated through the TCGA-LIHC and META-LIHC cohorts and sorted by the average C-index of the validation set. (D) The hub gene was selected through the CoxBoost algorithm. (E) The results of univariate Cox regression analysis of hub genes in training and validation cohorts. (F) Survival analysis of HCC patients with high CMLS and low CMLS. (G) PPI network interaction pattern showed that PLG interacts with SERPINE1. (H) The enrichment of two subtypes for different treatment-related signatures and HCC-related signatures. (I) Regulon activity profiles for 19 TFs (top) and potential regulators associated with chromatin remodeling (bottom) in both isoforms. *P < 0.05; **P < 0.01; ***P < 0.001

Based on the molecular characteristics, fine layering is crucial for accurate treatment of HCC. Utilizing various omics technologies such as transcriptomics and metabolomics has enabled the identification of a series of molecules in HCC. Current research has focused on the genome, transcriptome, and clinical characteristics for molecular classification of HCC. Consequently, efforts have been made to investigate the distinct molecular features of these subtypes. After filtering patient data related to metabolic and immune pathways, we assessed the enrichment of various molecular signatures in the samples using a single-sample gene-set enrichment analysis (ssGSEA) algorithm. We utilized 10 types of omics integrated clustering algorithms to classify HCC patients into two predefined multiomics groups: high CMLS and low CMLS (Fig. 1H). The distribution of the multi-omics data in the two groups is illustrated in Fig. 1H. Pathways such as PI3K_AKT_MTOR, IL6_JAK_STAT3, KRAS, P53, apoptosis, unfolded protein response, DNA repair, and cell cycle were notably enriched in high CMLS. Conversely, metabolic pathways like fatty acid metabolism and bile metabolism were significantly enriched in low CMLS. Furthermore, Wnt/β-Catenin and Epithelial-mesenchymal transition (EMT) pathways were predominantly enriched in high CMLS compared to low CMLS. These findings suggest that high CMLS may represent the proliferative subtype in the current molecular classification of HCC, while low CMLS may represent the nonproliferative subtype.

Given that cancer occurrence is closely related to abnormal gene expression, including changes in the transcription activity of transcription factors, and transcription disorder is one of the core mechanisms of tumors, we further researched the differences in the transcriptome. We analyzed 24 specific transcription factors (TFs) in HCC (Supplementary Table S1), as well as potential regulators associated with chromatin remodeling in cancer (Fig. 1I). Transcription factors FOXA2, FOSB, and FOXA1 in low CMLS were significantly activated, while PRRXY1, KLF5, SOX2, and SP1 showed specific enrichment in high CMLS. JUNB and FOSL2 were activated in both high CMLS and low CMLS. The close correlation between the activities of these transcriptional and regulatory factors and the CMLS groups confirms the biological relevance of the molecular subtypes. At the same time, we also observed that KDM4B, HDAC6, and SIRT5 are potential regulatory factors for chromatin remodeling in low CMLS type HCC, whereas in high CMLS type, SIRT2, SIRT7, HDAC8, and KDM5B are primarily involved. This association with cancer sex chromatin remodeling activity spectrum further highlights the differences between high and low CMLS groups, and the potential differences in the control mode, suggesting that epigenetic transcriptional networks may be important in distinguishing the factors of these molecular subtypes.

PLG as a novel Anoikis-resistance-related gene signature

Based on the CMLS score, the top five genes with the highest score are SERPINE1, PLG, NQO1, EZH2, and EDIL3. Interestingly, the top two genes, PLG and SERPINE1, are the key molecules mainly involved in the coagulation/fibrinolysis system (Fig. 1A and D). Research has confirmed that the coagulation/fibrinolysis system participates in cancer occurrence at multiple levels [5, 6, 8, 10, 14–18]. Therefore, we conducted a Venn-analysis of the 21 genes and coagulation/fibrinolysis molecules. The results shows that PLG and SERPINE1 are closely related to Anoikis resistance phenotype in HCC (Fig. 2A). Studies have shown that SERPINE1 (also known as PAI-1) encodes serine protease inhibitors (serpin), a member of the superfamily [19]. Its primary role is to inhibit the activation of tPA and uPA, which are the main inhibitors, thereby inhibiting the activation of PLG by tPA and uPA, inhibiting fibrinolytic enzyme activation, and preventing the fibrinolytic system from converting into its fibrinolytic enzyme. PPI analysis confirmed that the interaction between PLG and SERPINE1 is shown in Fig. 1E. Given that PLG is the coagulation/fibrinolysis system and its crucial role, we ultimately chose PLG as the core gene for this study.

Fig. 2.

PLG as a novel Anoikis-resistance-related gene signature in HCC. (A) Venn diagram illustrating genes associated with anoikis resistance and components of the coagulation/fibrinolytic system in HCC. (B) Immunohistochemical staining of PLG in tissues from HCC patients and adjacent paracancerous tissues. The images were captured at 100x, 200x, and 400x local magnification. (C) Impact of high and low PLG expression on overall survival prognosis in clinical samples of HCC. (D) Comparison of PLG protein expression in fresh HCC tissue and adjacent tissues from 7 pairs of clinical patients. (E) The mRNA levels of PLG in HCC tissues and adjacent tissues of 13 patients were detected by qRT-PCR. (F) The expression of PLG in the normal liver cell line L02 and HCC cell lines was detected by Western blotting.C

In addition, to be sure, our research group has been devoted to the study of blood coagulation/fibrinolysis of molecular and Anoikis resistance phenotype, emphasize in the discussion at this o ‘clock, so there will be no longer is introduced. Our study suggests PLG is likely to be closely related with HCC Anoikis resistance phenotype of genes, but this conclusion is based on bioinformatics analysis, so we need more clinical data and basic experiment to further prove the conclusion.

In addition, our research group has been dedicated to studying coagulation and fibrinolysis at the molecular level, especially pay special attention to Anoikis resistance. Our study indicates that PLG is likely closely related to Anoikis resistance in HCC cells. However, this conclusion is drawn from bioinformatics analysis, necessitating additional clinical data and basic experiments to validate it further.

First, we evaluated the expression level of PLG in 95 HCC clinical samples by immunohistochemistry. As shown in Fig. 2B, PLG was highly expressed in HCC tissues (50/95) compared to paired adjacent tissues (45/95). In liver cancer tissues, PLG was expressed in the nucleus and cytoplasm, mainly in the nucleus, indicating that PLG primarily played a carcinogenic role in the nucleus of liver cancer cells. In the tissue adjacent to carcinoma, PLG was mainly expressed in the cytoplasm. To analyze the correlation between PLG and the prognosis of HCC patients, we divided the 95 HCC patients into PLG “high expression” (n = 50) and “low expression” (n = 45) groups. We also collected and statistically analyzed the clinicopathological data of all HCC patients (Table 1). It was found that high PLG expression was closely related to AFP (P = 0.031), tumor size (P = 0.036), number of tumours (P = 0.034), and TNM stage (P = 0.036). However, it was independent of other clinical pathological characteristics such as sex, age, ALT and capsule of tumor (P > 0.05). Follow-up analysis revealed that HCC patients with lower PLG expression had significantly longer survival times than those with high expression (Fig. 2C). Secondly, we quantitatively explored protein imprinting 7 for fresh PLG protein expression in HCC and adjacent tissue levels. As shown in Fig. 2D, the expression of PLG protein in HCC tissues was higher than that in adjacent tissues (P < 0.001). We also confirmed the high expression of PLG in HCC tissues at the mRNA level (Fig. 2E). To analyze PLG’s influence on HCC cells’ malignant behavior, we quantified normal liver cells (L02) and 6 strains of HCC cell lines for PLG expression. PLG expression was highest in HuH-7, Hep-3B, HCCLM3, SMMC-7721 cells, and MHCC97-H cells, while Hep-SKP-1 had the lowest expression (Fig. 2F). Two cell lines, HuH-7 and HCCLM3, were used to generate stable PLG knockdown cells, and Hep-SKP-1 was used to generate stable PLG overexpression cells.

Table 1.

The relationship between PLG expression and clinicopathological features of HCC patients

Variables	Case (n = 95)		PLG expression (n = 95)		χ2	P-value
	number	100%	High expression group (n = 50)	Low expression group (n = 45)
Age(years)					1.068	0.301
≥ 60	39	41.1	23	16
<60	56	58.9	27	29
Gender					0.429	0.512
Male	54	56.8	30	24
Female	41	43.2	20	21
Liver cirrhosis					0.049	0.824
Yes	75	78.9	36	32
No	20	21.1	9	9
ALT(U/L)					1.18	0.277
≥ 40	52	54.7	30	22
<40	43	45.3	20	23
AFP(ng/ml)					4.661	0.031
≥ 200	69	72.6	41	28
<200	26	27.4	9	17
LDH(U/L)					1.265	0.261
≥ 200	47	49.5	22	25
<200	48	50.5	28	20
Tumor size (cm)					4.398	0.036
≥ 3	57	60.0	35	22
<3	38	40.0	15	23
The number of tumors					4.518	0.034
>1	51	53.7	32	19
≤ 1	44	46.3	18	26
Tumor capsular					0.012	0.914
complete	48	50.5	25	23
incomplete	47	49.5	25	22
Degree of tumor differentiation					0.548	0.76
I	6	6.3	4	2
II	63	66.3	33	30
III	26	27.4	13	13
Portal vein carcinoma thrombus					2.248	0.134
Yes	52	54.7	31	21
No	43	45.3	19	24
Extrahepatic metastases					0.021	0.884
Yes	69	72.6	36	33
No	26	27.4	14	12
Recurrence					0.414	0.52
Yes	58	61.1	29	29
No	37	38.9	21	16
Survival status					0.023	0.879
survival	43	45.3	23	20
death	52	54.7	27	25
TNM staging					4.391	0.036
T1 + T2	36	37.9	14	22
T3 + T4	59	62.1	36	23
Pathological type
hepatocellular carcinoma	95	100.0	50	45

PLG facilitated cell proliferation, apoptosis, invasion and migration, and Anoikis-resistance of HCC cells

We first knocked down PLG expression in HuH-7 and HCCLM3 cells and overexpressed PLG in Hep-SKP-1 cells. The knockdown or overexpression of PLG was verified by western blot (Supplementary Figure S2). Subsequently, we observed that knocking down PLG in HuH-7 and HCCLM3 cells led to decreased proliferation (Figs. 3A, 4, 5, 6 and 7B), apoptosis (Fig. 3C), invasion (Fig. 3E), and migration (Figs. 3E, 4, 5, 6 and 7F), while overexpressing PLG promoted proliferation (Figs. 3A, 4, 5, 6 and 7B), apoptosis (Fig. 3C), invasion (Fig. 3E), and migration (Figs. 3E, 4, 5, 6 and 7F) of Hep-SKP-1 cells. Furthermore, we conducted western blot analysis to assess the expression levels of apoptosis-related proteins in HCC cells after knocking down PLG. As depicted in Fig. 3D, knocking down PLG inhibited the expression of Bcl-2 in HuH-7 and HCCLM3 cells but promoted the expression of cleaved-caspase-3/7 and Bax. Conversely, overexpression of PLG inhibited cleaved-caspase-3/7 and Bax expression in Hep-SKP-1 cells while promoting Bcl-2 expression. These findings collectively suggest that PLG may act as a tumor promoter in HCC, enhancing proliferation, invasion, migration, and apoptosis of HCC cells.

Fig. 3.

Validation of the function of PLG in HCC tissues, HuH-7, HCCLM3, and Hep-SKP-1 cell lines. (A-B) Effect of HCC cells lentiviruses sh-PLG or oe-PLG, measured by CCK-8 (A) and colony formation assay (B). (C) The effects of sh-PLG or oe-PLG on the apoptosis of HCC cells were analyzed via flow cytometry. (D) Changes in apoptosis-related proteins after sh-PLG or oe-PLG were analyzed via western blot. (E) Comparison of the number of cell invasion and migration of HCC cells with lentiviruses sh-PLG or oe-PLG. Scale bars, 200 μm. (F) healing assay of HCC cells cells with sh-PLG or oe-PLG. Scale bars, 500 μm. (G) After HCC cells were in suspension culture for 24 h, the effects of sh-PLG or oe-PLG on the anoikis resistance of HCC cells were analyzed via flow cytometry. *P < 0.05; **P < 0.01; ***P < 0.001

Fig. 4.

PLG promotes the HCC process in vivo. (A-C) Gross specimen and weight of the tumor after sacrifice. Knockdown of PLG inhibited the weight (B) and volume (C) of subcutaneous xenograft tumors in NOD/SCID mice. (D) Subcutaneous xenograft tumors were confirmed via H&E staining. Compared with the control group. (E) KEGG pathway enrichment analysis between the sh-CTRL group and the sh-PLG group. (D) Changes in PI3K_AKT_ERK pathway-related proteins after sh-PLG or oe-PLG were analyzed via western blot. *P < 0.05; **P < 0.01; ***P < 0.001

Fig. 5.

Clinical Practice Value of CMLS. (A–B) Comparison between the CMLS and the other 22 published models in the TCGA-LIHC and META-LIHC cohorts. (C) The univariate Cox (proportional hazards or PH model) analysis of hub genes in CMLS. (D) The multivariate Cox analysis of hub genes in CMLS. (E) A comprehensive nomogram constructed based on CMLS and presented in the form of a web calculator for enhanced usability. (F) Calibration curve for the comprehensive nomogram. (G) Net decision curve analyses demonstrating the benefit of the comprehensive nomogram in clinical practice for HCC patients. (H) Comparison of the time-dependent C-index between the comprehensive nomogram and CMLS. *P < 0.05; **P < 0.01; ***P < 0.001

Fig. 6.

The TME-related molecular characteristics of high- and low-CMLS patients. (A) The distribution of metabolic pathways between high- and low-CMLS patients. (B) The distribution of immune suppression signatures between high- and low-CMLS patients. (C) The immune microenvironment characteristics of distribution between high- and low-CMLS patients. (D) The distribution of immune exclusion signatures between high- and low-CMLS patients. (E) The distribution of TCR signals and T-cell exhaustion between high- and low-CMLS patients. (F) The distribution of TMB between high- and low-CMLS patients. (G) The distribution of TNB between high- and low-CMLS patients. (H) The distribution of M0 macrophages between high- and low-CMLS patients. (I) The relationship between CMLS and M0 macrophages. (J–L) Survival analysis combined CMLS with TMB, TNB, and M0 macrophages. *P < 0.05; **P < 0.01; ***P < 0.001

Fig. 7.

The value of CMLS in predicting immunotherapy response in HCC patients. (A) The restricted mean survival (RMS) time difference at 5 years and 10 years after treatment between high- and low-CMLS groups. (B) The long-term survival (LTS) difference after 10 years of treatment between high- and low-CMLS groups. (C) The distribution of CMLS in different immunotherapy response groups. (D-E) The TIDE algorithm predicts the response to immunotherapy between high- and low-CMLS groups. (F) The subclass mapping algorithm predicts the response to immunotherapy between high- and low-CMLS groups. (G) Survival analysis of high- and low-CMLS groups in GSE76427. (H) Survival analysis of high- and low-CMLS group groups (I) Distribution of CMLS in different immunotherapy response groups of GSE91061. *P < 0.05; **P < 0.01; ***P < 0.001

Additionally, to simulate HCC cell Anoikis, we analyzed the effects of PLG on HCC cells Anoikis-resistance using flow cytometry. After 24 h of knocking down PLG and suspension culture, we observed that knocking down PLG promoted Anoikis in HuH-7 and HCCLM3 cells, while overexpression of PLG inhibited Anoikis in Hep-SKP-1 cells (Fig. 3G). These results indicate that PLG also promotes Anoikis resistance of HCC cells.

PLG promotes tumor growth in vivo

To further confirm our in vitro observation results, we explored whether PLG regulates the growth of tumor cells in vivo. First, we established a NOD/SCID mouse subcutaneous tumor model injecting subcutaneously sh-CTRL or sh-PLG HuH-7 transfected cells. The changes in subcutaneous tumor volume in nude mice were observed, and corresponding growth curves were plotted. As expected, PLG knockdown significantly delayed tumor growth in NOD/SCID mice (Fig. 4A). Upon weighing the tumors of sacrificed animals, it was evident that the tumors of nude mice in the sh-Ctrl group were significantly heavier than those in the sh-PG#1 group (Figs. 4B and 8C). Furthermore, subcutaneous tumors were identified through H&E staining (Fig. 4D). Collectively, these results validate that PLG promotes tumor growth in vivo.

To explore how PLG promotes malignant phenotypes, and Anoikis resistance in HCC cells, as well as other underlying mechanisms, we conducted an analysis of HCC cells using RNA sequencing transcriptome. We knocking down PLG and subjected the cells to suspension culture for 24 h to study the transcriptional responses. By analyzing the gene expression profiles of HCC cells, we identified 771 upregulated differentially expressed genes (DEGs) and 605 downregulated DEGs (Supplementary Figure S4). The DEGs were identified through pairwise comparisons between groups. KEGG pathway enrichment analysis revealed that the DEGs were mainly enriched in pathways such as PI3K-AKT, MAPK/ERK, TNF, EMT, and apoptosis, which are closely associated with cancer progression. These findings were consistent with previous bioinformatics analyses. Therefore, we focused on the PI3K/AKT and MAPK/ERK pathways as potential targets for further investigation. To validate our results, we performed western blot analysis on HCC cells with reduced PLG levels and cells subjected to 24 h of suspension culture. We examined the expression levels of proteins related to the PI3K-AKT and MAPK/ERK pathways. The results, depicted in Figs. 4E and 8F, showed that after 24 h of suspension culture, knocking down PLG inhibited the expression of p-PI3K and p-AKT in HuH-7 and p-PI3K in HCCLM3, while promoting the expression of p-ERK1/2. Conversely, overexpression of PLG inhibited the expression of p-ERK1/2 in Hep-SKP-1 cells while promoting the expression of p-PI3K and p-AKT. These results collectively suggest that PLG functions as a tumor promoter by enhancing Anoikis resistance, proliferation, invasion, and migration of HCC cells, possibly through the PI3K-AKT-ERK pathway.

Comparison of prognostic signatures of Anoikis resistance in HCC

In recent years, with the rapid development of sequencing technology, a large number of studies have begun to report the prognostic signature based on gene expression. Through systematic retrieval of relevant literature published over the past five years, we compared the CMLS all-round and other tags, and eventually included 23 different labels in our study (Supplementary Table S2). The features represented by different signatures are closely related to various biological behaviors of cancer, such as metastasis, apoptosis, immune infiltration, and metabolism. The end result that makes us happy is that in the TCGA-LIHC and META-LIHC dataset, we established that CMLS showed better performance than almost all models of C-index (Fig. 5A, B), highlighting its significant prognostic value. This provides an important reference for the precise stratification and individualized treatment of malignant tumors. Considering the clinical application and future development prospects of CMLS, we screened the potential independent prognostic factors related to anoikis resistance in HCC (Fig. 5C, D) through independent prognostic analysis (Fig. 1D, E). We integrated this information into a calculator network and constructed a comprehensive nomogram (https://the-nomogram.shinyapps.io/CMLS-DynNomapp/; Fig. 5E). Calibration curves prove that the accuracy of the nomogram is consistent with the actual situation (Fig. 5F). Decision curve analysis (DCA) shows that the nomogram provides significantly higher clinical benefits to patients compared to pure CMLS (Fig. 5G).

The metabolism and immune characteristics related to CMLS of Anoikis Resistance in HCC

It has been reported in the literature [20] that specific changes in the liver’s metabolic microenvironment under pathological conditions can lead to the occurrence of liver cancer. In particular, significant changes in glucose, lipid, and nucleotide metabolism can be observed in liver tumors. Changes in these metabolic substrates can promote hepatocyte proliferation on the one hand and alter immune activity on the other hand. This metabolic rearrangement mechanism is one of the ten characteristics of HCC, which not only provides HCC cells with a strong ability to proliferate and spread but also predisposes them to develop resistance to chemotherapy and immune drugs.

Research has demonstrated that during tumor progression, cancer cells inhibit mitochondrial oxidative metabolism by increasing glucose uptake and glycolysis, a phenomenon known as the Warburg effect [21]. This metabolic adaptation allows cancer cells to resist Anoikis. Anoikis resistance is a critical mechanism underlying cancer progression and metastasis. Tumor cells acquire resistance to anoikis through various mechanisms, including alterations in integrin expression, activation of pro-survival signaling pathways, and modifications in the tumor microenvironment and immune landscape. These mechanisms enhance the metastatic potential of tumor cells, sustain survival signals, and represent potential targets for anti-metastatic therapeutic interventions [22–26].

Therefore, we conducted a comprehensive enrichment analysis of the metabolism and tumor microenvironment (TME) of HCC using the Immuno-Oncology Biology Research (IOBR) R package to further understand the activated signaling pathways with high and low CMLS specificity.

The findings indicated that the trends observed in the high and low CMLS groups were consistent with existing literature, suggesting that the CMLS model developed in this study is both robust and accurate. Furthermore, the model demonstrates the capacity to not only reflect the Anoikis resistance of HCC cells but also to elucidate its relationship with immune scores and therapeutic efficacy [21–26].

We observed that high CMLS activated prostaglandin biosynthesis, galactose, amino and nucleotide sugars, and pyrimidine metabolism, among others (Fig. 6A). As for low CMLS, they were enriched in pathways such as fatty acid metabolism and catabolism, amino acid synthesis and metabolism, and primary bile acid secretion (Fig. 6A). In addition, changes in the environment’s metabolism can alter the immune response in the liver, leading to immune evasion in tumor cells, and a rearrangement of immune cell metabolism resulting in its dysfunction, impacting the anti-tumor activity of T cells [14]. Using CIBERSORT, we compared immune cell infiltration in high CMLS with that in low CMLS and found that patients with high CMLS had the highest infiltration of Tregs, M0 macrophages, dendritic cells, and neutrophils (Fig. 6B). Quantitative analysis revealed a higher presence of M0 macrophage infiltrates (Fig. 6B and Supplementary Figure S1). Additionally, PPAGs and myeloid-derived suppressor cells (MDSCs) were predominantly enriched in patients with high CMLS. Molecular markers associated with immunosuppression and immune rejection, such as cancer-associated fibroblasts (CAF), EMT pathway, and T cell receptor (TCR) signaling pathway, were also mainly enriched in the high CMLS group, indicating the presence of immunosuppression in this group (Fig. 6C-3E). These findings suggest that HCC characterized by high CMLS levels are more likely to be classified as “cold” tumors. Conversely, for patients with low CMLS are more likely to be classified as “hot” tumors, we observed a high infiltration of naive B and T cells, natural killer cells (NK), monocytes, and anti-inflammatory M2 macrophages (Fig. 6B-3E and Supplementary Figure S1), indicating a state of immune activation. This suggests that HCC characterized by low CMLS levels may be classified as tumors.

Tumor mutation burden (TMB) and tumor neoantigen burden (TNB) are currently recognized biomarkers to evaluate the response of patients to immunotherapy [27–31]. In view of the prognostic value and clinical utility of M0 macrophages in HCC immunotherapy proposed by Xu and Chen et al. [32, 33], we therefore analyzed the differences of TMB and TNB between the two groups. The results showed that TMB was highly enriched in the low CMLS group, while TNB and M0 macrophages were highly enriched in the high CMLS group. This suggests that lower CMLS groups have a better immune checkpoint response inhibition, but that doesn’t mean low immunogenicity in CMLS groups. On the contrary, TNB levels were higher in the high CMLS group, indicating that the more tumor neoantigens on the surface of tumor cells, the more effective the killing effect of immune cells on tumor cells (Fig. 6F-3I). Finally, further survival analysis showed that CMLS can be used as a complement factor for TMB, TNB, and M0 macrophages, thereby distinguishing patient outcomes (Fig. 6J-3L). The 5-year survival prognosis of patients with low CMLS, TMB, TNB (regardless of M0 macrophage infiltration status) in HCC patients is favorable. Low CMLS, high TMB, TNB, and M0 macrophage infiltration indicate a 7.5-year survival prognosis for HCC patients. Additionally, M0 macrophage infiltration type may affect the long-term prognosis of HCC patients. However, in a longer prognosis, for example, 10 years, patients with low CMLS, high TMB, and high TNB may have a better prognosis.

In summary, these data show that CMLS and molecular features can be used to distinguish between “hot” and “cold” HCC. This differentiation is crucial for predicting tumor response and assessing HCC efficacy to immunotherapy. It provides a basis and guidance for making clinical decisions regarding cancer treatment in the future.

The CMLS of Anoikis Resistance in HCC has excellent predictive power for immunotherapy response

In order to determine whether CMLS can be used as a predictive biomarker of ICI reaction, we further evaluated CMLS in the role and mechanism of HCC immunotherapy. Firstly, we analyzed the impact of CMLS on long-term survival in patients undergoing immunotherapy. By comparing patients with high CMLS to those with low CMLS over 5 and 10 years, we used the restricted mean survival (RMS) to explain the delayed clinical effect of immunotherapy. The assessment of patients’ long-term survival (LTS) over 10 years showed a difference between the two groups (Fig. 7A-4B). The analysis results indicated that the low CMLS group had a better prognosis compared to the high CMLS group (P < 0.05), suggesting that HCC patients with low CMLS may benefit more from immunotherapy. Secondly, we further analyzed the relationship between CMLS score, TNB, and the immune response. We observed a correlation between the CMLS score and the immune response. The CMLS score of the effective group (including complete remission (CR) and partial remission (PR)) was significantly lower than that of the ineffective group (progressive disease (PD) and stable disease (SD)) (P < 0.05, Fig. 7C, D; p.fisher = 0.000). A higher TNB indicated a better response in the low CMLS group. Conversely, the high CMLS score group had lower TNB and immune response (P < 0.05, Fig. 7C, D; p.fisher = 0.000). Furthermore, we utilized the tumor immune dysfunction and exclusion (TIDE) algorithm to evaluate the impact on the outcome of immune treatment in HCC patients. The results showed that compared to the high CMLS group (8.65%), the low CMLS group had a higher treatment success rate of 42.54% (p.fisher = 0.003; Fig. 7E). This suggests that HCC patients with low CMLS may exhibit a better response to immune checkpoint inhibitor drugs and derive more benefits from them.

The last and most important point is that we have prognostic information from multiple immunotherapy validation queues to verify once again. As expected, after the immune treatment of HCC patients, the low CMLS group shows better prognosis results (Fig. 7F-4G). Low CMLS is often associated with a better immune treatment outcome (Fig. 7H). These results suggest that the low CMLS group has a good response to immune checkpoint inhibition, and CMLS holds a good predictive value for the efficacy of immunotherapy in HCC patients.

Screening of potential therapeutic drugs related to CMLS of Anoikis Resistance in HCC

According to the results of our analysis above, it is not difficult to find that there is a significant difference in the prognosis of HCC patients with high CMLS and low CMLS. GSEA algorithm found that G2/M Checkpoint, EMT, inflammatory response, and TNF pathway were significantly activated in HCC patients with high CMLS (Fig. 8A). The findings align with our prior transcriptome sequencing results, which demonstrated significant differences between sh-CTRL and sh-PLG in EMT and TNF pathways (Fig. 4E). These results suggest that both high CMLS and low CMLS exhibit pronounced EMT capabilities. Furthermore, PLG, identified as a central molecule, exhibited similar outcomes, indicating its substantial involvement in the EMT process and alterations within the TME. Since we mentioned above that patients with high CMLS have a poor response to immunotherapy, we used the Cancer Therapeutics Response Portal (CTRP) and Profiling Relative Inhibition Simultaneous In Mixture (PRISM) to screen potential therapeutic drugs for patients with high CMLS. We then followed the standard SOP flow (Fig. 8B) to systematically explore the CMLS HCC patients with high potential drugs (P < 0.01, Fig. 5C and D). Among them, Dexamethasone, nutlin-3, parthenolide, UNC0638, bendamustine, cefditoren-pivoxil, and ICI-162,846 were sensitive drugs for HCC patients with high CMLS scores. In the end, we searched in PubMed for these drugs/compounds used in the treatment of liver cancer and the evidence, finding that Dexamethasone can alleviate HCC after arterial embolism chemotherapy embolism syndrome and enhance sensitivity to chemotherapy [34, 35]. Nutlin-3, as p53 and MDM2 inhibitors, can effectively inhibit liver cancer [36, 37]. Parthenolide influences liver cancer cells by affecting autophagy, apoptosis, proliferation, and the curative effect of HCC [38, 39]. For patients with low CMLS, statins (fluvastatin, simvastatin) help improve the treatment of HBV and HCV-related HCC [40, 41], while HCC benefits from Tipifarnib [42, 43].

Fig. 8.

Potential agents for patients with high CMLS. (A) Discovery of pathways significantly activated in the high-CMLS group through the GSEA algorithm. (B) The comprehensive standard operating procedure (SOP) process for screening potential agents. (C-D) The correlation and differential analysis of drug sensitivity for potential drugs screened from the CTRP and PRISM datasets. *P < 0.05; **P < 0.01; ***P < 0.001

Discussion

At present, with the development of bioinformatics and high-throughput sequencing, more and more evidence shows that the heterogeneity of HCC prognosis is closely related to its intrinsic molecular alterations. Many risk stratification models based on different altered molecules and forms have been established. For example, transcriptome studies classified as G1, G2, and G3 subtypes of HCC are associated with TP53, RPS6KA3, AXIN1, ATM mutations, FGF19, and TSC1/TSC2 mutations. This classification is closely related to poorly differentiated, pleomorphic, and polynuclear cells histologically. On the other hand, HCC classified as G4, G5, and G6 subtypes is associated with low cell proliferation and chromosomal stability, the steatohepatitis subtype, and CRP expression, but without CTNNB1 and TP53 mutations [44, 45]. The genomes of these subtypes also show significant differences in prognosis. However, with a deeper understanding of the complex relationship between the TME and treatment outcomes, it is evident that different molecular alterations at the cellular level in HCC may lead to different treatment outcomes and hinder the identification of reliable risk stratification schemes.

Given that gene expression is finely regulated by various genetic and epigenetic processes, we must uncover molecular subtypes and specific regulatory mechanisms based on a comprehensive analysis of multi-omics data to overcome the limitations caused by relying on only one specific omics in the current study [46, 47]. This is the main distinction between our study and other recently published studies [48–51]. Through the utilization of a comprehensive analysis involving 101 random combinations of 10 clustering algorithms, two molecular clusters were delineated, offering potential significance in the precise categorization of patients with HCC. These clusters include the high-CMLS group, characterized by a tendency towards proliferation, and the low-CMLS group, exhibiting nonproliferative traits. The molecular variations and activation of signaling pathways within these groups exhibit notable discrepancies, leading to discernible variances in metabolic pathways and biological characteristics. Additionally, we incorporated several HCC multicenter cohorts to select the best CMLS. Consistent with our expectations, the CMLS that passed our screening demonstrated strong prognostic value in each cohort. Furthermore, we explored the associations between multiple layers of pooled data and OS outcomes to mitigate the impact and constraints of clustering preferences on our analyses. To prevent overfitting the training cohort, we used the average C-index of multiple validation cohorts as the ranking criterion. The results indicated that the CMLS we developed exhibited superior C-index performance compared to nearly all models (Fig. 5A-2B).

Through the GSEA algorithm and IOBR R package, we have the high and low CMLS group are analyzed, in order to highlight the tumor metabolism and the immune correlation between landscape change. Also observed that low CMLS group of TMB concentration is higher, and higher CMLS TNB and M0 macrophages enrichment degree is higher. The subsequent survival analysis and TIDE and the kind of mapping, as we expected, CMLS can be used as TMB, TNB and M0 macrophages of complement factor, low CMLS group can have a better reaction on the immune therapy. At the same time, in a immunotherapy with prognostic information to verify in the queue to verify again also showed that low CMLS often associated with a better immune treatment outcome (Fig. 7F-4H). These data suggest that CMLS immune treatment efficacy in patients with HCC has good predictive value, provide a basis for future cancer treatment clinical decision. Our results showed that HCC patients with high CMLS had a poor prognosis response to immunotherapy, and G2/M Checkpoint, EMT, inflammatory response, and TNF pathways were significantly activated in patients with high CMLS. Therefore, we used CTRP and PRISM to systematically screen potential therapeutic drugs for patients with high CMLS. The screening of Dexamethasone, nutlin-3, parthenolide, UNC0638, bendamustine, cefditoren - pivoxil and ICI − 162,846 as high CMLS group of possible drug candidates. According to the report, Dexamethasone can alleviate the perioperative period in HCC syndrome and syndrome after embolism after tace, and enhance their sensitivity to chemotherapy [34]. Nutlin-3, as a p53-MDM2 inhibitor, can effectively exert anti-tumor effect to inhibit the occurrence of liver cancer [36, 37, 52]. Parthenolide is by influencing the liver cancer cells of autophagy and apoptosis, proliferation and influence the curative effect of HCC [38, 39, 53].

Closely related to the prognosis of HCC patients is the BCLC staging. Patients with early-stage HCC (BCLC 0 or BCLC A) are suitable for surgical resection, ablation, or liver transplantation [54]. The median survival time for these patients can reach 6 years and more than 10 years, respectively. Patients with intermediate-stage HCC (BCLC B) have a median survival of 26–30 months. For patients with advanced HCC (BCLC C), with a poor prognosis where the overall survival (OS) is about 19 months [55]. In the latest first-line and subsequent treatment guidelines for liver cancer, various systemic therapy options have emerged, significantly altering the treatment landscape and prognosis of HCC. However, due to the occult onset of HCC and its advanced stage at the time of diagnosis, less than 30% of patients are suitable for radical treatment at the initial diagnosis. Systemic anti-tumor therapy only plays a crucial role in treating advanced liver cancer, and most HCC patients have potential cirrhosis, which significantly impacts their tolerance to surgery, local, and systemic therapy [3, 45]. Therefore, identifying the patients most likely to benefit from a specific therapy through screening is crucial. Although many potential biomarkers have been identified to predict the prognosis of HCC treatment, their adoption in clinical practice is extremely rare.

The purpose of our study is to investigate the potential value of screening for HCC cells’ resistance to Anoikis genes in relation to HCC and prognosis. The CoxBoost algorithm was employed to identify the significant genes, and then the stepwise Cox algorithm was used to filter out the 21 most valuable genes with prognostic significan. Notably, during the construction of the final model, the algorithm that yielded the highest mean C-index was based on a combination of CoxBoost and stepwise Cox (direction = forward). Multivariate analysis revealed that five genes highly expressed in HCC were independent risk predictors of poor prognosis in HCC patients (P < 0.05, Fig. 1E), and these genes were closely associated with the Anoikis resistance phenotype. These genes included SERPINE1, PLG, NQO1, EZH2, and EDIL3. To investigate potential interactions or synergies among these genes, we further analyzed the PPI, which indicated an interaction between PLG and SERPINE1 (Fig. 1G). PLG, a fibrinolytic enzyme, is a key member of the blood coagulation system and the fibrinolytic enzyme activation system. On the other hand, SERPINE1, also known as PAI-1 (Plasminogen activator inhibitor-1), is a member of the serpin superfamily, primarily inhibiting the conversion of PLG to plasmin by tPA and uPA. Given our group’s focus on studying the relationship between the coagulation/fibrinolysis system and tumors [14, 15, 17, 18, 56], we also examined the interaction between these 21 genes associated with the Anoikis resistance phenotype and the components of the coagulation/fibrinolysis system. Interestingly, only PLG and SERPINE1 showed prognostic value in HCC (Fig. 5A).

There are two physiological activators of PLG, uPA and tPA. The intravascular conversion of PLG into plasmin is primarily induced by tPA. This serine protease is constitutively active, but its ability to convert PLG increases up to 1,000-fold in the presence of fibrin [57]. In contrast to tPA, uPA is fibrin-independent and is mainly involved in ECM remodeling and pericellular proteolysis, which is activated upon binding to its specific receptor, the urokinase-type plasminogen activator receptor (uPAR) [58]. SERPINE1 (PAI-1) is a 379-amino acid serine protease inhibitor [59]. Under normal physiological conditions, a balance between coagulation and fibrinolysis is maintained intravascularly by regulating the levels of plasminogen activators (tPA and uPA) and inhibitors (PAI-1, PAI-2, and α2-antiplasmin) [9, 58]. Considering the established relationship between PLG and SERPINE1, it is evident that PLG plays a crucial role in the plasminogen/plasmin activation system and the coagulation/fibrinolysis system. Given our team’s previous research on the role and mechanism of SERPINE1 in HCC [17], we directed our focus towards PLG, particularly its role in HCC.

Based on our previous bioinformatics analysis, we found that the high expression of PLG in HCC was closely related to Anoikis resistance and poor prognosis of HCC cells. Therefore, we conducted further research. Initially, immunohistochemical (IHC) verified that PLG was highly expressed in HCC tissues and correlated with poor prognosis of patients. Simultaneously, we selected the PLG-expressing HuH-7 and HCCLM3 cells, as well as the low-expressing Hep-SKP-1 cells (Considering that SMMC-7721 cells are internationally suspected to be contaminated with HeLa cells, and Hep-3B cells in our laboratory are also contaminated, these two cell lines were not selected by us). As anticipated, upon knocking down PLG expression, HuH-7 and HCCLM3 cells exhibited reduced proliferation, migration and invasion abilities, but enhanced the apoptosis abilities. Conversely, overexpression of PLG inhibited apoptosis in Hep-SKP-1 cells and enhanced the proliferation, migration, and invasion of HCC cells (Fig. 6).

Next, to simulate tumor cells surviving in a after detachment from the extracellular matrix, we subjected HCC cells to suspension culture. After PLG knockdown and 24 h of suspension culture, Anoikis in HCC cells in the sh-PLG group increased compared to the control group (Fig. 6G). This reaffirms the role of PLG as an oncogene in HCC, promoting Anoikis resistance in HCC cells.

To further validate our in vitro observations, we investigated whether PLG regulates tumor cell growth in vivo. Initially, we developed a NOD/SCID mouse subcutaneous tumor model and monitored the changes in subcutaneous tumor volume in nude mice on a weekly basis. After 5 weeks, the tumor tissue was excised and weighed. As anticipated, the suppression of PLG markedly slowed down tumor growth in NOD/SCID mice (Fig. 7A and C). These findings affirm that PLG fosters tumor growth in vivo. Finally, we utilized transcriptome study technology to analyze potential pathways and found that PLG may promote HCC cells through the PI3K-AKT-ERK pathway, enhancing Anoikis resistance, proliferation, invasion, and migration abilities, thereby acting as a tumor-promoting factor.

In summary, our selection of modeling molecules based on multiple cohorts of SPRGs greatly improved the stability and prognostic value of our model genes. Additionally, we demonstrated that PLG acts as an oncogenic factor in HCC and promotes Anoikis resistance in HCC cells. However, our study has several limitations. For instance, the cohorts we included varied in size and sequencing platform, and these differences could not be completely eliminated even with the use of correction algorithms. The clinical value of CMLS was only validated in a single-center, small-sample cohort in our study; therefore, further validation is needed in a larger, prospective, multicenter cohort. Finally, the relationship and mechanism between PLG and immune microenvironment changes and immunotherapy effects also need to be elucidated. A study found that in gastric cancer [60], high expression of PLG infiltration may potentially inhibit immunity. PLG expression, with the exception of NK cells, central memory T cell (Tcm), and gamma delta T cell (Tgd), negatively correlated with almost all types of immune cell infiltrate. We are excited about these results, which are already part of our future research plans.

Conclusion

In this study, we aimed to investigate the epigenetic and molecular alterations associated with Anoikis resistance in HCC to elucidate the underlying mechanisms. To achieve this objective, we employed 101 algorithms to categorize the samples into two groups: high CMLS and low CMLS, which were found to be closely related to Anoikis resistance through a comprehensive analysis of multi-omics data from HCC (Fig. 1). Furthermore, the CMLS model we developed demonstrated strong robustness (Fig. 5) and significant prognostic value compared to existing models, thereby providing a crucial reference for precise stratification and individualized treatment of malignant tumors. Additionally, we identified the pivotal role of PLG as a core gene differentiating the two groups. In recent years, PLG has emerged as a significant gene with prognostic implications in cancer, playing a crucial role in the proliferation, Anoikis, and metastasis of tumor cells [8–12]. Functional experiments (Figs. 2, 3 and 4) further demonstrated the significant involvement of PLG in the malignant biological behavior and Anoikis resistance of HCC. Following the knockdown of PLG expression, we observed a reduction in the proliferation, migration, and invasion of HCC cells, while apoptosis of these cells increased, suggesting that PLG may enhance Anoikis resistance in HCC cells. Furthermore, Anoikis resistance is recognized as a critical mechanism underlying the progression and metastasis of HCC. A substantial body of literature has indicated that resistance to Anoikis is closely associated with alterations in energy conversion, the immune landscape, and the metabolic microenvironment of tumor cells [21–26].

Our study presents a significant divergence from other recently published research [61, 62] due to its comprehensive analysis of multi-omics data related to HCC. The high and low CMLS models we developed exhibit robust predictive capabilities concerning glucose metabolism, immune evasion, and the tumor microenvironment, which are consistent with findings from previous studies. High CMLS is typically associated with proliferative subtypes that are linked to molecular alterations, the activation of signal transduction pathways, and immunosuppression, often manifesting as cold tumor phenotypes. Conversely, low CMLS is more likely to be nonproliferative and associated with metabolic pathways such as fatty acid and bile metabolism, and is classified as “hot tumors.

Our findings collectively provide substantial experimental evidence that PLG functions as a crucial regulator of anoikis resistance and immune evasion in HCC, thereby identifying it as a potential therapeutic target. PLG augments the metastatic potential of HCC cells through multiple mechanisms, including the activation of pro-survival signaling pathways and alterations to both the tumor microenvironment and the immune landscape. Evaluating variations in PLG levels prior to ICI therapy may improve treatment outcomes for cancer patients who are at an increased risk of therapeutic failure.

Materials and methods

Acquisition and preprocessing of multi-omics data and multi-center cohort data for HCC

First, in March 2024, “TCGAbiolinks” from the TCGA R package (https://portal.gdc.cancer.gov) was used to download liver (LIHC) queue HCC multiple omics data, which included the whole transcriptome expression and clinical outcomes of patients. Transcriptome profiles of mRNAs were obtained using the TCGAbiolinks software package. The upregulated Resistant 436 related gene expression data was downloaded from the Harmonizome database (https://maayanlab.cloud/Harmonizome/). Additionally, complete information from 7 other HCC queues was collected from the Gene Expression database (Gene Expression Omnibus, http://www.ncbi.nlm.nih.gov/geo, GEO. GSE68465 and GSE72094). The classification system is significantly associated with overall survival (OS) (P < 0.001). To ensure robust downstream analysis, patients with less than 1 month of overall survival (OS) were filtered out. High-throughput sequencing of transcriptomes was converted to transcripts per kilobase million (TPM) to enhance sample comparability. Additionally, all microarray expression profiles were deduplicated and normalized to closely resemble gene expression from microarrays. The expression matrix and clinical information from the official website were downloaded for microarray data. The data were processed using the robust limma software package for background correction, log2 transformation, and quantile normalization. Boxplot functions were used to visualize and assess the uniformity of the distribution of expression abundance values in the samples.

Multi-omics consensus integration analysis, along with the specific molecular characteristics and stability of CMLS groups

In order to comprehensively analyze multi-layer data using multi-omics data and machine learning methods, we first matched omics information of different dimensions by sample ID. The R package Venn diagram was used to draw the Venn diagram of common differentially expressed genes from TCGA-LIHC and Harmonizome. Then, the use of multiple omics integration and the Visualization of cancer subtypes (MOVICS) package “getElites” function to filter genetic traits. We set the “Methods” parameter to “cox”, and to better match CMLS groups with their corresponding clinical information, we extracted factors closely associated with OS based on Cox regression survival analysis, where P < 0.001. Finally, we applied “getMOIC” function for clustering analysis. We conducted a series of more than ten of the most advanced learning (CIMLR, ConsensusClustering, SNF, iClusterBayes, PINSPlus, moCluster, NEMO, IntNMF, COCA, and LRA) to establish the model as the final classification “Methodslist” input parameters, and used the MOVICS package to provide the default parameters of the corresponding clustering results. Finally, we integrated the clustering results of these 10 different algorithms using the “getConsensusMOIC” function, with the “distance” parameter set to “Euclidean” and the parameter set to “average” to improve the robustness of the clustering. Two stable prognostic groups of HCC, high CMLS and low CMLS, were identified from previous studies.

We calculated enrichment scores for multiple treatment-related signatures using gene set variation analysis (GSVA). We constructed transcriptional regulatory networks (regulons) through transcriptional regulatory network reconstruction and Regulatory Analysis (RTN) R package. We analyzed the relationship between the transcriptional regulatory network and the prognosis of the patients. This included 24 HCC-specific TFS and 71 candidate regulators related to chromatin remodeling in cancer that we retrieved from PubMed. We then performed GSVA enrichment to detect tumor cell metabolism and the immune microenvironment. We assessed subtype stability by validating the clustering results using subtype-specific biomarkers in the validation cohort and comparing the agreement of the consensus clustering with the NTP and PAM classifiers.

Establishing a Consensus on the prognosis of a learning-driven prognostic signature

We chose our TCGA-HCC queue as the training set and the META-HCC queue as the validation set. Additionally, we utilized the sva package to address mass effect. To evaluate the relationship between CMLS, immunotherapy, and outcomes, and to enhance comparability across cohorts by applying z-scores to the data beforehand, we developed CMLS with high accuracy by integrating 101 machine learning algorithms. These algorithms included stepwise Cox, CoxBoost, ridge regression, RSF, GBM, survival-SVMs, LASSO, SuperPC, plsRcox, and Enet. During the model-building phase, we utilized the GEO cohort (GSE68465) as the training set for initial model construction. The Lasso, Ridge, and Enet models were implemented using the glmnet package and the “cv.glmnet” function. The survivalsvm model was implemented using the “survivalsvm” function in the survivalSVM software package. The gbm model was implemented using the GBM package. The superpc model was implemented using the SuperPC package. The “gbm” function was employed for 10-fold cross-validation to fit the GBM. The plsRcox model was implemented using the “cv.plsRcox” function of the plsRcox software package. The RSF model was implemented using the randomForestSRC package and the “RFSRC” function (with ntree = 1000 and nodesize = 5). For the CoxBoost model, the “optimCoxBoostPenalty” function was used to determine the optimal punishment (shrinkage) value, followed by a 10-fold cross-validation. Subsequently, the CoxBoost model was fitted using the “CoxBoost” function. The complexity of the statistical models was evaluated based on the Akaike information criterion (AIC), and the survival package was utilized for stepwise Cox analysis. The regularization parameter lambda was determined through 10-fold cross-validation, and the trade-off parameter alpha was set between 0 and 1 (interval = 0.1). When alpha = 0, Ridge was executed; when alpha = 1, Lasso was performed. For any other value of alpha, Enet was executed. Finally, we calculated the average C-index of all models in the queue to evaluate the predictive ability of the models.

CMLS prognostic value and the comprehensive analysis of the immune response to treatment

We based our model on the prognosis of multiple omics data and various algorithms to build the model shown in Fig. 1C, selected by Cox regression analysis in HCC with a phenotype and prognosis closely related to genetic SPRGs. Simultaneously, based on the grading of samples, they can be divided into high and low CMLS groups. The prognostic significance of SPRGs and CMLS in HCC was evaluated. We utilized the STRING database for further analysis of the SERPINE1, PLG, NQO1, EZH2, and EDIL3 gene protein interaction network (Protein-Protein Interaction Network, PPI). The C-index in each queue was assessed to predict prognosis. Time-dependent C-index curves and calibration curves were drawn to describe accuracy, and decision curves were used to calculate the clinical benefit for patients. A nomogram was constructed using a multivariable Cox regression factor.

For metabolic and immune cell abundance between patients with high and low CMLS, signatures related to metabolic pathways, immunotherapy response, immune microenvironment, and immune rejection were collected using the IOBR package. These signatures were then standardized to obtain an enrichment score for each sample. Immune cell abundance was assessed for each patient using the CIBERSORT algorithm. Additionally, the joint CMLS restructuring patients were compared based on differences in TMB, TNB, and M0 distribution of macrophages. The delayed response survival of HCC patients to immunotherapy was evaluated using the TIP algorithm, subclass mapping, and TIDE algorithm to estimate the immunotherapy response of patients with high and low CMLS.

To screen potential drugs sensitive to patients with high CMLS

The activation status of oncogenic pathways between patients with high and low CMLS was analyzed using the GSEA algorithm. Expression data of human cancer cell lines (CCL) were obtained from the Broad Institute Cancer Cell Line Encyclopedia (CCLE). The CTRP v. 2.0 (https://portals.broadinstitute.org/ctrp) and PRISM repurpose data set (19 q4; https://depmap.org/portal/prism/) were utilized for human cancer cell line (CCL) susceptibility data. The R package to predict the corresponding sensitivity. The analysis involved comparing the high and low CMLS GSEA algorithm results between patients with activated cancer pathways. The area under the dose-response curve (AUC) value was used as a measure of drug sensitivity.

Collection of clinical samples

Pairs of HCC and adjacent non-cancerous tissues from Lanzhou University Second Hospital Surgical Oncology were studied. PLG (1:500, sc-376324; Santa Cruz) protein expression was quantified by Western blot in 7 pairs of fresh HCC tissues and adjacent tissues, while PLG mRNA expression was detected in 13 pairs of HCC tissues and adjacent tissues. For 95 HCC and adjacent carcinoma tissues, formalin-fixed and paraffin-embedded samples were used for IHC detection of PLG protein expression. All patients included in the study signed informed consent forms, and the study protocol was approved by the Ethics Committee of Lanzhou University Second Hospital. No economic compensation was provided.

Cell culture

Normal liver cells (L02) were purchased from Guangzhou Jennio Biological (https://www.jennio-bio.com/), while HCC cell lines HuH-7, Hep-3B, HCCLM3, SMMC-7721, MHCC97-H, and Hep-SKP-1 were acquired from Shanghai Fuheng Science and Technology Co., Ltd. (https://www.fuhengdio.cn/). L02 cells were cultured in Roswell Park Memorial Institute (RPMI) 1640 medium (Gibco, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (FBS) (Gibco) and 1% penicillin-streptomycin (Beyotime, Shanghai). The other cell lines were cultured in Dulbecco’s Modified Eagle Medium (DMEM) (Gibco) supplemented with 10% FBS and 1% penicillin-streptomycin. All cell lines were maintained in a 37 °C incubator with 5% CO₂ under humidified conditions.

Cell transfection

Shanghai Gemma Gene synthesized small PLG hairpin RNA (shRNA) and negative control shRNA (sh-Ctrl). The coding sequence of PLG was amplified and cloned into the hU6-MCS-CBh-gcGFP-IRES-puromycin plasmid (Compass Biotechnology Co., Ltd, Beijing, China) and packaged as lentivirus. The lentiviruses were transfected into candidate HCC cell lines. Positive cells were screened using puromycin (2 µg/mL) for 2 weeks. The efficiency of PLG overexpression was evaluated through quantitative reverse transcription polymerase chain reaction (qRT-PCR) and western blot analysis. The target sequences of PLG shRNA are provided in Supplementary Table S3.

Construction of the Anoikis model

For the model’s setup, after the methyl-2-hydroxyethyl acrylate (Poly-HEMA) is completely dissolved and is evenly distributed onto a 6-well plate (Corning, NY, USA). Excess solution is absorbed, and this process is repeated twice to create a layer of condensed water film at the bottom of the plate. The plate is then placed under ultraclean and ultraviolet light for 30 min until the water film is completely dry. Cells are suspended at a concentration of 1 × 10^6 in the 6-well plates and cultured for 24 h. The suspended cells are then collected for further experiments. Poly-HEMA was procured from Sigma Aldrich (Shanghai) Trading Co., LTD.

qRT-PCR

The solution, according to the manufacturer, involves using Trizol reagent (Invitrogen, Carlsbad, CA, USA)) to extract total RNA from HCC cells and tissue samples. The SweScript One-Step RT-PCR Kit (Servicebio Technology Co., Ltd, Wuhan, China) was utilized to amplify the total RNA through reverse transcription. The quantification was performed using Applied Biosystems ABI 7500 and 2×Universal Blue SYBR Green qPCR Master Mix (Servicebio Technology). GAPDH served as the internal reference gene. The 2^-ΔΔCt method was employed to quantify fold change. All experiments were conducted in triplicate. Primer sequences can be found in Supplementary Table S4.

Western blotting and immunohistochemistry

By adding protease inhibitors (Beyotime, Shanghai, China) radiation immune precipitation (RIPA) (Beyotime) total protein extracted from HCC cells and tissue samples. Protein concentrations were determined using the BCA assay kit (Beyotime). At 8–12% SDS/PAGE gel running on total protein, and transferring them to polyvinylidene fluoride (PVDF) membrane (Millipore Corp. Billerica, MA, USA). Blocks were blocked using QuickBlock™ protein-free Western blocking solution (Beyotime) and incubated with primary antibodies overnight at 4 °C. Information on the primary antibodies used is provided in Supplementary Table S5. Goat anti-rabbit IgG HRP (ImmunoWay Biotechnology, Beijing, China; RS0002) and goat anti-HRP mouse IgG (RS0001) as the secondary antibodies, at a 1:10 dilution ratio. From the Surgical Oncology department at the Second Hospital of Lanzhou University, paraffin sections of HCC tissue from patients undergoing surgery were subjected to immunohistochemical staining to determine PLG protein levels (1:500).

Cell proliferation assay

According to the manufacturer’s instructions, the cell count kit 8 (CCK-8) (Yeasen Biotechnology, Shanghai, China) was used to quantify the proliferation of HCC cells. Cells from the sh-Ctrl and sh-PLG groups were seeded at a concentration of 1 × 10^4 cells per well in 96-well plates. After 24 h, 48 h, and 72 h, 10 µL of CCK-8 solution was mixed with the cells and incubated at 37 °C for 1 h to assess proliferation activity. Absorbance was measured at 450 nm using a BioTek microplate reader.

Transwell assay

Transwell chambers without and with Matrigel^® Matrix (Corning) to measure cell migration and invasion, respectively. The cells were suspended in DMEM medium without FBS, and 1 × 10^4 cells were inoculated into the Transwell Chambers in the upper chamber. Subsequently, 700 µL of DMEM medium containing 10% FBS was added to the lower chamber. After 24 h, the cells were fixed with 4% paraformaldehyde for 20 min and stained with 0.1% crystal violet for 20 to 30 min.

Wound healing assay

To vaccinate 6 different groups of cells in the orifice plate and cultivate them until they reach 90% confluence. The wound was scraped using a 200 µL pipette tip and washed with PBS to remove cell debris. Subsequently, the cells were cultured for 24 h in medium without FBS. Using an Olympus microscope, images of the scratches were taken at 0 and 24 h. The ImageJ Software (National Institutes of Health, Bethesda, MD, USA) was used to assess the relative area of wound closure.

Apoptosis assay

Annexin V-FITC/PI apoptosis detection kit (Yeasen Biotechnology), and PE Annexin V/7-AAD apoptosis detection kit (BD Biosciences, Franklin Lakes, NJ, USA) were used to determine the apoptosis rate of adherent and suspended cells. The cells were washed twice with PBS buffer, then 100 µL of buffer was added to a test tube. Subsequently, 5 µL of Annexin V-FITC or PE Annexin V and 10 µL of PI or 7-AAD dye were added for staining in the dark for 15 min at room temperature. The buffer was then topped up to 500 µL. The percentage of apoptosis was measured using flow cytometry (Beckman Cytoflex).

Subcutaneous tumor assay

HuH-7 cells were transfected with shRNA against PLG and the corresponding lentiviral infection interference. HuH-7 cells transfected with sh-Ctrl and sh-PLG 1:1 were mixed with PBS and Matrigel suspension reagent. Briefly, 1 × 10^6 cells were injected subcutaneously into Male NOD/SCID mice (Jiangxi Jicui Pharmaceutical Kang Biotechnology Co., Limited), aged 5–6 weeks. Tumor volumes were measured every 7 days and calculated by the formula (length×width^2)/2. After 5 weeks, mice were sacrificed by deep anesthesia with ethyl ether. Tumor tissues were harvested for subsequent experiments. Animal experiments were reviewed and approved by the Animal Use Ethics Committee of the Second Hospital of Lanzhou University. All animals were cared for in accordance with the Guide for the Care and Use of Laboratory Animals.

Hematoxylin and eosin (H&E) staining and immunohistochemistry (IHC)

H&E staining of 5 μm paraffin sections was performed using a standard protocol. The primary antibody used was anti-PLG (1:100). Two experienced pathologists independently reviewed all sections. IHC scores were determined using an H-score. Staining intensity and the percentage of positive cells were scored as previously described. The final immunoreactivity score was calculated by multiplying the staining intensity score by the percentage of positive cells. The expression level of PLG was classified as “low expression” if the immunoreactivity score was 0–4 and “high expression” if the score was 5–9.

RNA sequencing

Using Trizol reagent, total RNA was isolated from the sh-Ctrl group and sh-PLG cells. The samples were rapidly frozen in liquid nitrogen for storage and then sent to Majorbio Bio-Pharm Technology Co. Ltd. (Shanghai, China) for transcriptome data analysis.

Statistical analysis

The Limma software package was used to analyze differential expression analysis, and multi-omics clustering was completed by the MOVICS software package. The survminer package was utilized to determine the cut-off value of CMLS scores, and a double classification method was employed to classify CMLS scores. In each cohort, HCC patients could be distinguished into high CMLS and low CMLS groups based on the log-rank statistics of the maximum selection, thereby reducing the calculated batch effect. For the comparison between the two groups, if the variable follows a normal distribution, the Student’s t-test is used; if the variable does not follow a normal distribution, the Wilcoxon rank-sum test is used. For comparisons involving more than two groups, parametric and nonparametric variables were tested using one-way ANOVA and the Kruskal-Wallis test, respectively. Two-sided Fisher’s exact test was performed on contingency tables. Statistical analyses were conducted using SPSS Statistics 25 software (IBM Corporation, Armonk, NY, USA) (version 24.0) and R software (R Foundation for Statistical Computing, Vienna, Austria) (R v.4.1.0). P values < 0.05 were considered statistically significant.

We used Student’s t-test to examine the difference in means between two groups, and two-tailed ANOVA to examine the difference between multiple groups. A Chi-square test was used for correlation analysis between gene expression and clinicopathological features. Cox regression analysis was used to analyze the influencing factors affecting the prognosis of patients. Functional analysis was performed using the SPSS software package and GraphPad Prism (GraphPad Software, La Jolla, CA, USA) (version 8.0). P values < 0.05 were considered statistically significant.

Abbreviations

HCC

Hepatocellular carcinoma

NK

Natural killer cells

OS

Overall survival

TMB

Tumor mutation burden

TME

Tumor microenvironment

VEGF

Vascular endothelial growth factor

CMLS

Consensus Machine Learning-driven Signature

CR

Complete remission

DCA

Decision curve analysis

ECM

Extracellular matrix

EMT

Epithelial-mesenchymal transition

HBV

Hepatitis B virus

HCV

Hhepatitis C virus

ICI

Immune checkpoint inhibitor

IOBR

Immuno-Oncology Biology Research

LTS

Long-term survival

MDSCs

Myeloid-derived suppressor cells

NSCLC

Non-small cell lung cancer

ORR

Objective response rate

PD

Progressive disease

PFS

Progression-free survival

PLG

Plasminogen

PPI

Protein-protein interaction

PR

Partial remission

SD

Stable disease

ssGSEA

Single-sample gene-set enrichment analysis

Tcm

Central memory T cell

TFs

Transcription factors

Tgd

Gamma delta T cell

TNB

Tumor neoantigen burden

tPA

Tissue plasminogen activator

uPA

Urokinase-type plasminogen activator

Author contribution

Xueyan Wang, Hao Chen: Conceptualization, Writing – Review & Editing. Xueyan Wang, Lei Gao, Haiyuan Li: Writing - Original Draft. Yanling Ma, Baohong Gu, Xuemei Li: Visualisation. Lin Xiang, Yuping Bai: Investigation. Bofang Wang, Chenhui Ma: Supervision.

Funding

This research was supported by National Natural Science Foundation of China (No. 82160129), Bethunen advanced Digestive Cancer Treatment Clinical Research Program (071124023), Cuiying Scientific and Technological Innovation Program of Lanzhou University Second Hospital (CY2023-ZD-01), Key Project of Science and Technology in Gansu province (22ZD6FA054), Medical Innovation and Development Project of Lanzhou University (lzuyxcx-2022-160, lzuyxcx-2022-45, lzuyxcx-2022-88), Gansu Province Innovation Driven Assistance Project (GXH20230817-14).

Declarations

Conflict of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.