Integrative transcriptome profiling elucidates molecular and immunovascular characteristics of macrotrabecular HCC

Abstract

Background and Aims:

Recent advances in single-cell RNA sequencing (scRNA-seq) technologies have enabled us to clarify gene regulatory networks and immune cell compositions. In this study, the integration of large-scale bulk and single-cell datasets refined HCC classification and shed light on the characteristics of its aggressive subtype.

Approach and Results:

Single-cell analysis of 228,564 live cells from 6 scRNA-seq datasets identified 5 major clusters of HCC cells with high mitotic activity (cluster 1), activated Wnt/β-catenin signaling (cluster 2), elevated glycolysis (cluster 3), and lipogenesis (clusters 4 and 5). Aggressive HCC subtype defined in bulk RNA-seq analysis of 691 tumor samples comprised a combination of cluster 1 with clusters 3, 4, or 5. Gene regulatory network analysis and gene set enrichment analysis highlighted the essential roles of p53 and MYC in aggressive HCC/cluster 1, and cell composition analysis elucidated T cell depletion as an immune resistance mechanism. In a syngeneic mouse model, Trp53 knockout and MYC overexpression caused high mitotic, tumorigenic, and metastatic phenotypes, characterized by a macrotrabecular pattern, vascular encapsulation, and T cell exclusion. Angiogenesis inhibition disrupted macrotrabecular/vascular encapsulation formation, resulting in T cell recruitment, and its combination with immune checkpoint blockade achieved remission.

Conclusions:

Single-cell analysis has deepened our understanding of the molecular mechanism and tumor microenvironment in aggressive HCC. The combination of targeting tumor vasculature and blocking immune checkpoints represents a promising therapeutic strategy for this subtype.

INTRODUCTION

HCC is one of the most common malignancies in the world due to a variety of risk factors including HBV and HCV infection, alcohol-induced cirrhosis and fibrosis associated with the metabolic syndrome.¹ Multiple treatment options such as surgical resection, local ablation, liver transplantation, transarterial chemoembolization (TACE), and systemic therapy provide clinical benefits,² and angiogenesis inhibitors including sorafenib and lenvatinib have dramatically improved patient outcomes.^3–5 The emergence of immune checkpoint inhibitors has further prolonged survival in patients with advanced HCC, but the response rate is limited.^6,7 It is important to clarify how these agents suppress HCC growth and identify surrogate markers for distinguishing responders from non-responders.

Several laboratories have performed comprehensive genome and transcriptome analyses over the past 2 decades,⁸ and have proposed HCC classification schemes based on mutational landscapes, copy number alterations, gene expression patterns, and tumor microenvironments.^9–12 These pioneering efforts have reached the “two-class” model, categorizing HCC into proliferation and non-proliferation classes.^1,13–15 Extending the “two-class” model, we have previously divided HCC samples into 3 molecular subtypes using omics data collected by ourselves and provided from the LIHC-US cohort of the Cancer Genome Atlas Research Network:¹⁶ an aggressive subtype characterized by TP53 mutation and chromosomal instability (MS1), a CTNNB1-mutated subtype (MS2) and a metabolic disease-associated subtype (MS3). In this work, the GSEA program estimated activation of G2/M checkpoint, Wnt/β-catenin signaling, and inflammatory response in MS1, MS2, and MS3, and the CIBERSORT program predicted immune-cold and immune-hot microenvironments in MS2 and MS3, respectively. We also clarified that MS1 corresponded to the S1/2 subclasses of Hoshida classification and the G1/2/3 subgroups of Boyault classification, while MS2 and MS3 corresponded to the G5/6 and G4 subgroups of Boyault classification, respectively.^10,11,16

The MS1 subtype, corresponding to the proliferation class in the consensus “two-class” model, harbors histopathological features of poor differentiation and severe vascular invasion. Current studies have reported that a macrotrabecular (MT) architectural pattern is markedly observed in areas of aggressive HCC.^17,18 The MT structure is frequently coated by endothelial cells and surrounded by vascular spaces, indicating a vascular encapsulation (VE) pattern.¹⁷ Liver cancer with vessels encapsulating tumor clusters (VETC) is significantly associated with intrahepatic metastasis and unfavorable prognosis.^19,20 However, molecular mechanisms underlying these characteristics remain unclear, and effective therapeutic strategies for these subtypes are still undetermined.

Recent advances in single-cell RNA sequencing (scRNA-seq) technologies have enabled us to elucidate gene regulatory networks, individual tumor cell features, and immune cell compositions.²¹ In the present study, integrated analysis of large-scale bulk RNA-seq and scRNA-seq datasets refined our HCC classification and offered a better understanding of key regulators and tumor and immune cell profiles of each subtype. Utilizing these findings, we developed a syngeneic mouse model of MS1, which exhibited high mitotic, tumorigenic, and liver metastatic abilities with the MT/VE structure contributing to immune evasion. This model revealed molecular and biological processes and potential therapeutic targets for this subtype.

METHODS

Ethics approval

This study was conducted ethically in accordance with the World Medical Association Declaration of Helsinki and the ARRIVE guidelines. All experimental protocols were reviewed and approved by the Institutional Review Board (G2017-018, Medical Research Ethics Committee for Life Science of Tokyo Medical and Dental University; G2018-132C5, Medical Research Ethics Committee for Genetic Research of Tokyo Medical and Dental University; A2019-263C2, Institutional Animal Care and Use Committee of Tokyo Medical and Dental University). Written informed consent was obtained from participants.

Bulk and single-cell transcriptomics

Raw count data of RNA-seq analysis in the LICA-FR, LIHC-US, and LIRI-JP cohorts were downloaded from the ICGC Data Portal. Raw count data of the scRNA-seq datasets GSE124395, GSE125449, GSE146115, GSE149614, GSE151530, GSE156625, GSE189903, GSE202642, and GSE242889 were downloaded from the Gene Expression Omnibus site.

Human tissue samples

A total of 104 patients underwent curative resection for HCC at Tokyo Medical and Dental University Hospital between 2013 and 2016. All the patients provided informed consent before enrollment and were anonymously coded in accordance with ethical guidelines.

Statistical analysis

Statistical analysis was performed using Python (Python Software Foundation) and R (R Foundation for Statistical Computing). p<0.05 was considered statistically significant. p values were calculated by the χ ² test to examine the relationship between categorical variables. p values were calculated by the log-rank test to compare patient survival. For parametric data, p values were calculated by ANOVA with the Tukey–Kramer post hoc test, and data are the mean±SD in the figures. For nonparametric data, p values were calculated by the Kruskal–Wallis test with the Steel–Dwass post hoc test, and boxes represent the 25th, 50th, and 75th percentiles in the figures.

Further details regarding the Methods section used in this study are presented in Supplemental Information, including single-guide RNA target sequences and primer sequences (Supplemental Tables S1–S5, http://links.lww.com/HEP/J740).

RESULTS

Transcriptomic and functional characterization of HCC samples

We initially performed a comprehensive bulk analysis of genome and transcriptome data collected in the LICA-FR, LIHC-US, and LIRI-JP cohorts (n=160, 294, and 237, respectively) of the International Cancer Genome Consortium (ICGC) project. After removing batch effects using Harmony integration, we performed dimensionality reduction with uniform manifold approximation and projection (UMAP) for visualization (Supplemental Figure S1A, http://links.lww.com/HEP/J740). We labeled HCC specimens from the LIHC-US cohort as the 3 molecular subtypes, namely MS1, MS2, and MS3, based on our previous clustering analysis using the Ward method,¹⁶ resulting in a similar separation into 3 groups on the UMAP plot (Figure 1A). Clustering analysis with the Leiden algorithm also identified 3 groups (Figure 1A), showing a strong relationship between the results from the 2 clustering methods (p=1.79×10⁻⁸³, χ ² test). Consistent with our prior findings,¹⁶ MS1 and MS2 frequently contained HCC samples with TP53 and CTNNB1 mutations, respectively (Figure 1B and Supplemental Table S6, http://links.lww.com/HEP/J740; TP53: p=1.75×10⁻⁷, χ ² test; CTNNB1: p=3.50×10⁻⁴⁷, χ ² test). Differential gene expression analysis in each molecular subtype unveiled that MS1 had elevated expression of G2/M checkpoint-related genes, including CCNB1, CDC20, FOXM1, and CDK1 (Figure 1C and Supplemental Table S7, http://links.lww.com/HEP/J740​​​​​). Genes controlled by the Wnt/β-catenin signaling pathway (GLUL, AXIN2, and LGR5) and associated with cellular metabolism (GLS2, ALDOB, and ARG1) were overexpressed in MS2 and MS3, respectively (Figure 1C).

FIGURE 1.

Transcriptomic and functional characterization of HCC samples. (A) UMAP plots of 691 HCC samples from 3 cohorts clustered by the Ward method (left) and the Leiden algorithm (right). (B) UMAP plots show the mutation status of TP53 (left) and CTNNB1 (right). TP53 and CTNNB1 mutations are predominantly observed in MS1 (43.1%) and MS2 (81.9%), respectively. (C) Lists (upper) and expression patterns (lower) of representative differentially expressed genes in molecular subtypes; ***p<0.001. (D) Heatmaps of regulon activities estimated by the RTN program (left) and the pySCENIC program (right) for transcription factors. (E) Functional protein–protein interaction networks in MS1 (left) and MS2/3 (right). (F) UMAP plots showing regulon activities estimated by the pySCENIC program for p53 (left) and MYC (right); ***p<0.001. (G) UMAP plots showing enrichment scores estimated by the ssGSEA program for hallmark pathways; **p<0.001 and ***p<0.001. Abbreviations: MS, molecular subtype; NA, not applicable; UMAP, uniform manifold approximation and projection.

We then explored gene regulatory networks in each molecular subtype using the RTN and pySCENIC programs, both of which revealed significant differences between the gene regulatory networks of MS1 and MS2/3 (Figure 1D and Supplemental Table S8, http://links.lww.com/HEP/J740; RTN: p=2.01×10⁻³⁸, χ ² test; pySCENIC: p=3.59×10⁻⁴⁴, χ ² test). Functional protein–protein interaction network analysis using the STRING database illustrated that 66 common regulators in MS1 were centered around TP53 and MYC, with the E2F family involved in cell proliferation and the SOX family involved in cell stemness (Figure 1E). Conversely, 22 common transcription factors in MS2/3 included master regulators of hepatocyte differentiation (FOXA2 and HNF4A) and lipid metabolism (PPAR and RXRA). UMAP plots displayed high regulon activities for p53 in MS1 and for MYC in both MS1 and MS2 (Figure 1F). Single-sample gene set enrichment analysis confirmed the enrichment of the G2M_CHECKPOINT gene set in MS1 and the WNT_BETA_CATENIN_SIGNALING gene set in MS2 (Figure 1G), corroborating the findings from the expression analysis (Figure 1C). High enrichment scores of gene sets associated with p53 and MYC in MS1 (Figure 1G) were also consistent with the results from the gene regulatory network analysis (Figure 1F).

Single-cell heterogeneity of HCC samples

Next, we examined 6 out of 9 scRNA-seq datasets registered with the Gene Expression Omnibus (GSE124395, GSE125449, GSE149614, GSE151530, GSE156625, and GSE189903) as shown in Supplemental Table S9, http://links.lww.com/HEP/J740, excluding 3 datasets with a high number of dead cells (GSE146115, GSE202642, and GSE242889). We employed Harmony integration for batch correction, and estimated cell types of 228,564 live cells using the CellTypist program with the healthy liver and immune cell models (Figure 2A, Supplemental Figure S1B, http://links.lww.com/HEP/J740, and Supplemental Table S10, http://links.lww.com/HEP/J740). In both models, similar cell types were precisely clustered, and clustering analysis with the Leiden algorithm detected distinct cell groups (Figure 2B, Supplemental Table S11, http://links.lww.com/HEP/J740, and Supplemental Table S12, http://links.lww.com/HEP/J740). From 27,773 hepatocytes and cholangiocytes in 102 tumor tissues, we identified 8 clusters in the UMAP plot (Figure 2B, Supplemental Figure S1B, http://links.lww.com/HEP/J740, and Supplemental Table S13, http://links.lww.com/HEP/J740). Among the 5 major clusters comprising more than 5% of the total tumor cells, cluster 1 exhibited increased expression levels of G2/M checkpoint-related genes including TOP2A, BIRC5, and CDK1 (Figure 2C, Supplemental Figure S1C, http://links.lww.com/HEP/J740, and Supplemental Table S14, http://links.lww.com/HEP/J740), suggesting an accumulation of proliferating cells. In cluster 2, the expression levels of GLUL, SPARCL1, and AQP9 were upregulated. These findings were observed in the CTNNB1-mutated subtype MS2 in both the present and previous bulk analyses (Figure 1C),¹⁶ implying that cluster 2 might contain CTNNB1-mutated HCC cells. Clusters 3 and 4/5 were characterized by overexpression of glycolysis-related genes (ALDOA, PKM1, and ENO1) and lipogenesis-related genes (apolipoproteins and FABP1), respectively, with ALDOB and ARG1 upregulated in MS3 during the current bulk analysis (Figure 1C).

FIGURE 2.

Single-cell heterogeneity of HCC samples. (A) UMAP plots of 228,564 live cells in 6 scRNA-seq datasets annotated by the CellTypist program with the healthy liver (left) and immune cell (right) models. (B) UMAP plots of all cells (left; n=228,564) and epithelial cells (right; n=27,773) clustered by the Leiden algorithm. (C) Lists (upper) and expression patterns (lower) of representative DEGs in clusters. (D) Composition of clusters across 102 HCC tissues. (E) Heatmap showing the Spearman correlation coefficient of cell composition in HCC tissues. (F) UMAP plot with pseudotime trajectories. (G) UMAP plots of 691 bulk samples (left) and 102 pseudo-bulk samples (right) with spatial regions of each molecular subtype defined by kernel density estimation. (H) UMAP plots showing enrichment scores estimated by the ssGSEA program for hallmark pathways; ***p<0.001. (I) Proportions of immune cells in molecular subtypes; *p<0.05, **p<0.01, and ***p<0.001. (J) Schematic representation of HCC classification through bulk and single-cell analyses. Abbreviations: AP, antigen presentation; CTL, cytotoxic T cell; DC, dendritic cell; DEGs, differentially expressed genes; Mono, monocyte; MΦ, macrophage; MS, molecular subtype; scRNA-seq, single-cell RNA sequencing; Tcm, central memory T cell; Tem, effector memory T cell; Th, helper T cell; Trm, resident memory T cell; UMAP, uniform manifold approximation and projection.

Cell composition analysis of clusters 1–5 in each tumor tissue elucidated that no tumor samples consisted solely of cluster 1, which coexisted with either clusters 3, 4, or 5 (Figure 2D). In contrast, 9 tumor specimens were primarily composed of cluster 2. Thus, after excluding 3 minor clusters that did not satisfy the abovementioned criterion, we defined HCC tissues with over 5% of cluster 1 as scMS1, those with over 25% of cluster 2 as scMS2 and those enriched with clusters 3, 4, and 5 as scMS3G (glycolysis), scMS3L1 and scMS3L2 (lipogenesis), respectively (Supplemental Figure S1D, http://links.lww.com/HEP/J740 and Supplemental Table S15, http://links.lww.com/HEP/J740). A correlation heatmap of cluster proportions clarified that scMS2, scMS3G, scMS3L1, and scMS3L2 were independent of each other, with scMS1 clustering dependent on cell clusters other than cluster 1 (Figure 2E). Notably, scMS1 frequently consists of clusters 1 and 3 rather than clusters 1 and 4/5. Pseudotime analysis estimated 2 major trajectories in clusters 1–5: one from cluster 5 toward clusters 3 and 1, and another from cluster 5 toward clusters 4 and 2 (Figure 2F). Pseudo-bulk transcriptome analysis of the 102 single-cell datasets integrated with the 691 bulk datasets revealed that these HCC samples were divided into 3 groups consistent with the bulk subtype classification (Figure 2G; p=2.57×10⁻²¹⁹, χ ² test), and that scMS1 significantly overlapped with bulk MS1, scMS2 with bulk MS2, and scMS3L1 with bulk MS3 (Figure 2G and (Supplemental Table S16 http://links.lww.com/HEP/J740); p=1.39×10⁻⁶, 2.98×10⁻¹⁵, and 6.30×10⁻⁷, χ ² test), demonstrating a strong correlation between the bulk and single-cell subtype classifications (p=9.08×10⁻¹⁹, χ ² test). scMS3G spanned both bulk MS1 and bulk MS3, which might be caused by the similar gene expression signatures between scMS1, predominantly composed of clusters 1 and 3, and scMS3G, primarily composed of cluster 3 (Figures 2D, E), as well as the close connection between clusters 1 and 3 in gene expression trajectories (Figure 2F).

Single-sample gene set enrichment analysis revealed critical correlations of clusters 1, 3, and 5 with p53 signaling and of clusters 1 and 2 with MYC signaling, and discovered the activation of mitosis in cluster 1 and Wnt/β-catenin signaling in cluster 2 (Figure 2H). These findings indicate that both p53 and MYC play crucial roles in MS1 (cluster 1+cluster 3, 4, or 5), consistent with the results from the bulk analysis (Figures 1E–G).

Following the cell composition-based HCC classification, we examined the immune cell profiles of each molecular subtype. MS1 exhibited a low number of helper T cells and cytotoxic T cells and a high number of macrophages, whereas MS2 lacked monocytes and dendritic cells (Figure 2I). These results suggest that T cell depletion and defective antigen presentation cause immune resistance in MS1 and MS2, respectively. In summary, MS1 is enriched with proliferating tumor cells and characterized by TP53 somatic mutation, MYC signaling activation, and T cell depletion (Figure 2J).

Establishment and evaluation of aggressive mouse HCC cells

To create mouse HCC cells recapitulating MS1/cluster 1 for further investigation, we introduced Trp53 mutation in 3H3, a hepatoma cell line of C57BL/6 mouse origin with low tumorigenicity and high immunogenicity,²² using the CRISPR/Cas9 system. We confirmed loss of p53 protein expression (Supplemental Figure S2A, http://links.lww.com/HEP/J740) and a frameshift mutation in Trp53 (Supplemental Figure S2B, http://links.lww.com/HEP/J740) in Trp53-knockout 3H3 cells (3H3P) by Western blot analysis and sequencing analysis, respectively. However, when 3H3P cells were subcutaneously injected into nude mice, there was no tumorigenic potential of them (Supplemental Figure S2C, http://links.lww.com/HEP/J740). Given MYC signaling activation in MS1/cluster 1 (Figures 1E–G and Figure 2F), we generated 3H3P cells ectopically expressing MYC (3H3PM). 3H3PM cells slowly but steadily grew in nude mice (Supplemental Figure S2D, http://links.lww.com/HEP/J740), and 9 subclones were isolated from the transplanted tumors (Supplemental Figure S2E, http://links.lww.com/HEP/J740). It was noteworthy that 8 of the subclones had increased protein expression of MYC as compared to the parental 3H3P cell line (Supplemental Figure S2F, http://links.lww.com/HEP/J740). Because polyploidy was observed in 3H3PM cells but not in 3H3P cells (Supplemental Figure S2G, http://links.lww.com/HEP/J740), indicating abnormal cell division, flow cytometry analysis was performed and demonstrated a decreased number of cells in the G0/G1 phase of the cell cycle in 3H3PM cells (Supplemental Figure S2H, http://links.lww.com/HEP/J740). Quantitative reverse transcription polymerase chain reaction (RT-PCR) analysis revealed that the expression levels of cell cycle regulators, including Aurkb and Plk4, hallmark genes of MS1,¹⁶ were higher in 3H3PM cells than in 3H3 and 3H3P cells (Supplemental Figure S2I, http://links.lww.com/HEP/J740).

Among the 8 subclones, 3H3PM6 and 3H3PM8 clones exhibited rapid tumor progression after subcutaneously inoculated into immunocompetent C57BL/6 mice (Figure 3A). Immunohistochemical analysis of the transplanted tumors elucidated a trabecular stricture comprising more than 6 cells in thickness¹⁷ surrounded by CD31+ endothelial cells (Figure 3B), that is, the MT/VE pattern. Macrophages, particularly arginase-1-positive M2 macrophages, were enriched in the tumors, while CD3+ T cells or CD19+ B cells were not (Figure 3B). To unveil the connection between the MT/VE structure and the tumor immune microenvironment, we performed histological analysis of grafted tumor specimens at earlier stages (Figure 3C). On day 14, 3H3PM8 cells were randomly distributed and not enclosed by endothelial cells, and CD3+ T cells were actively infiltrated. On day 21, however, the tumor cells organized the MT/VE structure, and the numbers of CD3+ and CD8+ T cells were remarkably reduced, while the numbers of F4/80+ and arginase-1+ macrophages were augmented (Figure 3D).

FIGURE 3.

Histopathological and immunological evaluation of mouse HCC with Trp53 knockout and MYC overexpression. (A) Tumorigenicity assay of 3H3PM cells in immunocompetent mice (n=4). Data are the mean±SE. (B) Representative immunohistochemical images of endothelial cell and immune cell markers in tumors derived from 3H3PM6 and 3H3PM8 cells. (C) Representative immunohistochemical images of endothelial cell and immune cell markers in tumors derived from 3H3PM8 cells at early stages. (D) Immune cell infiltration in tumors derived from 3H3PM cells at early stages. (E) Representative photo images of the primary splenic lesion and the liver metastatic lesions of 3H3PM8 cells. (F) Representative immunohistochemical images of endothelial cell and immune cell markers in the primary splenic lesion and the liver metastatic lesions of 3H3PM8 cells. Abbreviation: HE, hematoxylin and eosin staining.

Since the MT/VE structure promotes metastasis of HCC cells,^19,20 we next evaluate the liver metastatic ability of 3H3PM cells. After injection into the mouse spleen for modeling intrahepatic metastasis, 3H3PM8 cells metastasized into the liver within 8 weeks (Figure 3E), while no metastatic lesions were macroscopically observed in the lung or brain, nor were there any signs of distant metastasis such as cyanosis or abnormal behavior. The primary and liver metastatic lesions harbored similar pathological and immunological features to each other, as well as to the subcutaneously transplanted tumors; the MT/VE structure was undergoing formation, and CD3+ T cells were diminished in the liver metastatic lesions (Figure 3E and Supplemental Figure S3A, http://links.lww.com/HEP/J740). Four weeks after being directly inoculated into the liver, 3H3PM cells developed intrahepatic tumors that exhibited the MT/VE pattern, but not aberrant lesions in the lung or brain (Supplemental Figure S3B, http://links.lww.com/HEP/J740).

To determine the roles of p53 and MYC in MT/VE formation and T cell depletion, we subcutaneously transplanted 3H3 and 3H3P cells into immunocompetent mice and harvested the tumor tissues before complete regression. Immunohistochemical analysis revealed no clear evidence of MT formation in tumors derived from 3H3 and 3H3P cells (Supplemental Figure S3C, http://links.lww.com/HEP/J740). We utilized RIL-175, an HCC cell line of C57BL/6 mouse origin with Trp53 deletion and Hras ^G12V mutation,²³ and observed that the RIL-175 tumors showed the MT-like pattern, thick trabeculae without surrounding spaces,¹⁷ along with CD3+ T cell infiltration (Supplemental Figure S3C, http://links.lww.com/HEP/J740). Next, we attempted to establish 3H3 cells with MYC overexpression (3H3M), but a few days after MYC introduction, 3H3M cells underwent cell death (Supplemental Figure S3D, http://links.lww.com/HEP/J740). These findings imply that p53 attenuation and MYC activation may be involved in MT and VE formation, respectively, and suggest that loss of p53 may prevent MYC-induced oncogenic stress and cell death, which has been widely acknowledged and repeatedly substantiated by numerous studies, including the seminal work by Hermeking and Eick.²⁴ Taken together, mouse HCC cells with Trp53 knockout and MYC overexpression acquire high mitotic, tumorigenic and liver metastatic properties under normal immune conditions, and the tumors display the MT/VE pattern in the immunosuppressive microenvironment.

Molecular, pathological, and imaging assessment of human MT-HCC tissues

We investigated genetic alterations and signaling pathways involved in MT/VE-HCC employing omics data of 373 HCC samples from the LIHC-US cohort and histological data of them diagnosed by Calderaro et al.¹⁸ Exome-sequencing analysis discovered frequent TP53 somatic mutations (Figure 4A; p<0.001, χ ² test) and gene set enrichment analysis identified significant MYC signaling activation (Figure 4B) in MT-HCC cases compared with other HCC cases (n=61 and 312, respectively). The UMAP plot also showed the enrichment of MT-HCC in MS1 (Figure 4C; p<0.001, χ ² test).

FIGURE 4.

Bioinformatics and clinicopathological assessment of human MT-HCC. (A) Mutation rates of genes specifically altered in the MTM and non-MTM subtypes of HCC (n=61 and 312, respectively). (B) Enrichment scores of signaling pathways and biological processes specifically activated in the MTM subtype. (C) UMAP plot showing the MTM subtype with spatial regions of each molecular subtype defined by kernel density estimation. (D) Representative immunohistochemical images of CD34, p53, MYC, and CD8 in human HCC samples. (E) Kaplan–Meier curves of disease-free (left) and overall (right) survival in patients harboring HCC with and without the MT/VE pattern, p53 mutation, MYC expression, and T cell depletion. Abbreviations: HE, hematoxylin and eosin staining; MT, macrotrabecular; MTM, macrotrabecular massive; NA, not applicable; UMAP, uniform manifold approximation and projection; VETC, vessels encapsulating tumor clusters.

We next conducted an immunohistochemical staining study of 104 human HCC tissues surgically resected in our institution (Figure 4D). Immunohistochemical assessment demonstrated a close correlation between the MT pattern and p53 nuclear accumulation, indicating a change in p53 signaling and a potential involvement of p53 mutation (p<0.001, χ ² test), and histopathological analysis of MT-HCC revealed encapsulation by CD34+ endothelial cells (p<0.001, χ ² test), namely the VE structure, and depletion of CD8+ T cells (p=0.022, χ ² test). Patients harboring HCC with p53 nuclear accumulation and patients with MYC-overexpressing HCC had poor overall survival, and patients with MT/VE-HCC and patients with CD8+ T cell-depleted HCC had both discouraging disease-free and overall survival (Figure 4E). In clinicopathological factors, p53 nuclear accumulation was strikingly associated with high serum α-fetoprotein (AFP) concentration, vascular invasion, and poor differentiation (Supplemental Table S17, http://links.lww.com/HEP/J740). The strong relationships between MYC overexpression and high TNM stage, between MT/VE formation and high serum AFP and des-γ-carboxy prothrombin (DCP) concentration, and between T cell exclusion and poor differentiation were also observed (Supplemental Table S18–S20, http://links.lww.com/HEP/J740).

The imaging characteristics of MT/VE-HCC by contrast-enhanced CT were heterogeneous enhancement, intratumoral arteries, intratumoral necrosis, and large tumor size (>50 mm), as shown in Supplemental Table S21, http://links.lww.com/HEP/J740, consistent with the previous report on the predictors for this pathological subtype.²⁵ A scoring system was generated, with a range of 0 to 4 points based on these 4 CT findings and named the MT/VE imaging score. The proportion of MT/VE-HCC cases as the score increased (Supplemental Figure S4A, http://links.lww.com/HEP/J740), and the receiver operating characteristics curve demonstrated an AUC of 0.895 with a 95% CI of 0.833–0.958 (Supplemental Figure S4B, http://links.lww.com/HEP/J740). These data indicated a good diagnostic ability of the MT/VE imaging score in identifying MT/VE-HCC.

Molecular mechanisms of MT/VE structure formation

To clarify the molecular mechanism of MT/VE formation in MS1, we subsequently performed RNA-seq analysis of 3H3P cells and 3H3PM subclones (Figure 5A and Supplemental Tables S22 and S23, http://links.lww.com/HEP/J740), and determined genes differentially expressed between 3H3P and 3H3PM cells and between 3H3PM1 and 3H3PM3 subclones (non-tumorigenic in C57BL/6 mice) and 3H3PM6 and 3H3PM8 subclones (highly tumorigenic in C57BL/6 mice). Gene set enrichment analysis revealed MYC signaling activation and cell cycle progression in 3H3PM cells (Figure 5B), consistent with the abovementioned findings (Supplemental Figures S2G–I, http://links.lww.com/HEP/J740 and Supplemental Table S24, http://links.lww.com/HEP/J740), and identified that hypoxic conditions and glycolytic metabolism might be involved in the highly tumorigenic group of 3H3PM cells (Figure 5B and Supplemental Table S25, http://links.lww.com/HEP/J740). Quantitative RT-PCR analysis confirmed overexpression of several genes, including the immune checkpoint molecule Cd274, encoding PD-L1, and the stem cell marker Dclk1.²⁶ The expression level of Angptl2 was more than 100-fold higher in 3H3PM cells than in 3H3 and 3H3P cells. Il11 was overexpressed in the highly tumorigenic group as compared to the non-tumorigenic group (Figure 5C). Among 3H3 cells, 3H3P cells, highly tumorigenic 3H3PM cells, and non-tumorigenic 3H3PM cells, there were no significant differences in mutation landscapes of frequently mutated genes in HCC such as Ctnnb1, Arid1a, and Arid2.

FIGURE 5.

Vascular encapsulation and immune cell exclusion induced by angiogenic factors in mouse MT/VE-HCC. (A) Volcano plots of genes differentially expressed between 3H3P and 3H3PM cells (left) and between the non-tumorigenic (3H3PM1 and 3H3PM3) and highly tumorigenic (3H3PM6 and 3H3PM8) groups (right). (B) Enrichment plots of gene sets positively associated with 3H3PM cells (left two) and the highly tumorigenic group (right two). (C) Quantitative RT-PCR analysis of genes upregulated in 3H3PM cells. (D) Tumorigenicity assay of the Angptl2-KO and Il11-KO 3H3PM8 cells in immunocompetent mice (n=8). Data are the mean±SE. (E) Representative immunohistochemical images of endothelial cell and immune cell markers in tumors derived from the Angptl2-KO and Il11-KO 3H3PM8 cells. (F) T cell infiltration in tumors derived from the Angptl2-KO and Il11-KO 3H3PM8 cells. Abbreviations: FDR, false discovery rate; HE, hematoxylin and eosin staining; MT, macrotrabecular; NES, normalized enrichment score; RT-PCR, reverse transcription polymerase chain reaction; VE, vascular encapsulation.

Two angiogenesis-related genes Angptl2 ²⁷ and Il11 ²⁸ were promising candidates because VE formation required the recruitment of endothelial cells. To explore the roles of the 2 genes in MT/VE-HCC development, we further derived the Angptl2-KO and Il11-KO subclones from 3H3PM8 cells using the CRISPR/Cas9 system and examined tumorigenic activity in immunocompetent mice (Figure 5D). Transplanted tumors of the Angptl2-KO and Il11-KO 3H3PM8 cells were significantly smaller than those of the control cells. In the Angptl2-KO and Il11-KO tumor tissues, the MT/VE structure was heavily distorted, and the number of CD3+/CD8+ T cells was notably elevated (Figures 5E, F).

We surveyed the expression profiles of angiogenesis-related genes, including ANGPTL2 and IL11, in the LIHC-US dataset. We discovered that the mRNA expression level of IL11 was specifically upregulated in the MS1 subtype of our classification and the G3 subgroup of Boyault classification, both of which are characterized by frequent TP53 mutations, high mitotic activity, and MT-HCC enrichment (Figure 3),^10,18 while no other angiogenesis-related gene expression was associated with the subtypes (Supplemental Figure S5, http://links.lww.com/HEP/J740).

To determine which cell types express ANGPTL2 and IL11 using the scRNA-seq datasets, we re-performed Harmony integration using the gene set including ANGPTL2 and IL11 (Supplemental Figure S6, http://links.lww.com/HEP/J740). In the updated UMAP plot, ANGPTL2 expression was observed in the ALB+ cluster (epithelial cells), ACTA2+ cluster (fibroblasts), and PLVAP+ cluster (endothelial cells). Similarly, Il11 was weakly expressed but detectable in the same 3 clusters. Although scRNA-seq analysis has limitations in accurately quantifying genes with low expression levels compared to bulk RNA-seq analysis or RT-PCR analysis, these results indicate that the 2 angiogenesis-related genes are expressed in cancer cells at least.

We also investigated the regulatory mechanism of IL11 expression in MT/VE-HCC cells. RT-PCR analysis showed that isoform 1 of the Il11 gene was predominantly expressed in 3H3PM cells (Supplemental Figures S7A, B, http://links.lww.com/HEP/J740). Chromatin immunoprecipitation analysis of MYC in 3H3PM cells revealed that MYC protein bound to the promoter region of Il11 isoform 1 (Supplemental Figure S7C, http://links.lww.com/HEP/J740). Therefore, MYC might directly contribute to the increased expression level of Il11 in 3H3PM cells.

Vulnerability of MT/VE-HCC to angiogenesis inhibition and immune checkpoint blockade

We noticed the mild inhibitory effects of Angptl2 and Il11 knockout in the context of severe T cell infiltration in 3H3PM cells. Considering the possibility of T cell exhaustion in the tumor tissues, consistent with the increased expression level of PD-L1 (Figure 5C), we intraperitoneally administered anti-PD-1 antibody into tumor-bearing C57BL/6 mice. Tumor progression of 3H3PM8 cells was not significantly prevented by anti-PD-1 therapy due to T cell exclusion. In contrast, treatment with anti-PD-1 antibody successfully diminished tumor development of the Il11-KO 3H3PM8 cells (Supplemental Figure S8A, http://links.lww.com/HEP/J740), although the number of CD3+ T cells was unchanged (Supplemental Figures S8B, C, http://links.lww.com/HEP/J740). These observations suggest that the combination of angiogenesis inhibition and PD-1 inhibition may attenuate tumorigenesis by disrupting the MT/VE structure and reinvigorating immune surveillance.

We next examined the susceptibility of MT/VE-HCC to the anti-angiogenic multikinase inhibitor lenvatinib and anti-PD-1 therapy, both of which are approved by the Food and Drug Administration for unresectable HCC treatment. Pharmacological angiogenesis modulation using lenvatinib suppressed tumor growth of 3H3PM8 cells by destroying the MT/VE structure and recruiting CD3+ T cells into the cluster, but incompletely (Figures 6A–C), which was consistent with the pathological findings of the Angptl2-KO and Il11-KO tumor specimens (Figures 5D–F). While there was no difference in tumor volumes between treatment with and without anti-PD-1 antibody, the combination of lenvatinib and anti-PD-1 antibody eliminated MT/VE-HCC and achieved remission (Figure 6A).

FIGURE 6.

Susceptibility of MT/VE-HCC to targeting tumor vasculature and blocking immune checkpoints. (A) Tumorigenicity assay of 3H3PM8 cells in immunocompetent mice treated with angiogenesis inhibition and immune checkpoint blockade (n=8). The anti-angiogenic agent lenvatinib (3 mg/kg/day, orally) and anti-PD-1 antibody (200 μg/head/week, intraperitoneally) were administered into tumor-bearing mice. Data are the mean±SE. (B) Representative immunohistochemical images of endothelial cell and immune cell markers in tumors derived from 3H3PM8 cells in mice treated with lenvatinib and anti-PD-1 antibody. (C) T cell infiltration in tumors derived from 3H3PM8 cells in mice treated with lenvatinib and anti-PD-1 antibody. (D) Treatment course and changes in serum DCP levels in patient 1 pathologically and radiologically diagnosed with MT/VE-HCC. (E) HE staining and silver staining of a biopsy sample obtained from the primary lesion. (F) Changes in contrast-enhanced CT findings of the primary lesion before and after treatment with Atezo plus Bev and TACE. Red arrowheads show the disappearance of intratumoral arteries with heterogeneous enhancement of the primary lesion and the shrinkage of the residual tumor lesion. (G) Schematic representation of mechanisms underlying MT/VE organization and immune evasion and therapeutic strategies for MT/VE-HCC. Abbreviations: Atezo, atezolizumab; Bev, bevacizumab; DCP, des-γ-carboxy prothrombin; HE, hematoxylin and eosin; Len, lenvatinib; MT, macrotrabecular; NS, not significant; TACE, transarterial chemoembolization; Veh, vehicle; VE, vascular encapsulation; VETC, vessels encapsulating tumor clusters.

We evaluated the clinical relevance of the image-scoring system (Supplemental Figure S4, http://links.lww.com/HEP/J740) and therapeutic strategy (Figures 6A–C) developed for MT/VE-HCC in this study. Patient 1 was suspected to have MT/VE-HCC (Figures 6D–F), meeting all 4 imaging characteristics of MT/VE-HCC. Histological examination of a biopsy sample confirmed the diagnosis by revealing the MT/VE pattern through not only hematoxylin and eosin staining but also silver staining (Figure 6E). The patient was subsequently treated with atezolizumab plus bevacizumab, which led to substantial tumor shrinkage and decreased serum DCP levels. Although a residual tumor lesion was detected after the 14th cycle, treatment with TACE followed by atezolizumab plus bevacizumab resulted in its regression. Similarly, patients 2 and 3 were diagnosed with MT/VE-HCC based on the imaging score (Supplemental Figure S9, http://links.lww.com/HEP/J740). Surgical resection after lenvatinib treatment significantly reduced primary tumor sizes and serum DCP levels. While patient 2 developed lung metastases and patient 3 experienced intrahepatic recurrence, remission was attained in both cases with atezolizumab plus bevacizumab treatment. These clinical cases demonstrate the effectiveness of our image-scoring system for the diagnosis of MT/VE-HCC and highlight the potential of the combination of targeting tumor vasculature and blocking immune checkpoints as a promising therapeutic option for this subtype.

DISCUSSION

In this study, the single-cell analysis identified 5 major clusters of HCC cells (Figure 2B): cluster 1, with high mitotic activity; cluster 2, with activated Wnt/β-catenin signaling; cluster 3, with enhanced glycolysis; and clusters 4 and 5, with enhanced lipogenesis. We demonstrated that the 3 molecular subtypes of HCC established from bulk analysis, namely the aggressive subtype (MS1), the CTNNB1-mutated subtype (MS2), and the metabolic disease-associated subtype (MS3), correspond to combinations of these clusters. Specifically, MS1 comprises a combination of cluster 1 with clusters 3, 4, or 5; MS2 corresponds solely to cluster 2; and MS3 consists of tumors mainly containing clusters 3, 4, and 5 (Figure 2J). Similarly to this research, 2 teams have recently integrated multiple scRNA-seq datasets to stratify liver cancer. Li et al²⁹ divided HCC specimens into 3 categories based on fatty acid degradation using 2 scRNA-seq datasets (GSE149614 and GSE151530) and addressed possible differences in responses to sorafenib, TACE, and immunotherapy. Chen et al³⁰ proposed 5 HCC classes with distinct tumor microenvironments, including fibroblasts, endothelial cells, and immune cells, and therapeutic responses using 3 scRNA-seq datasets (GSE125449, GSE149614, and GSE156625). In contrast to their novel HCC classifications, our comprehensive analysis of 691 bulk samples from 3 cohorts and 27,773 tumor cells of 102 HCC tissues from 6 scRNA-seq datasets consistently supports the "three-molecular-subtype" model previously established,¹⁶ and additionally provides profound insights into the tumor cell compositions of each molecular subtype.

Gene regulatory network analysis and gene set enrichment analysis inferred that p53 and MYC could play pivotal roles in MS1/cluster 1 (Figures 1E, G and Figure 2F), and cell composition analysis estimated that T cell depletion could serve as an immune resistance mechanism in MS1 (Figure 2H). To validate these findings, we created a syngeneic mouse model of HCC with Trp53 knockout and MYC overexpression, which successfully replicated aggressive human HCC with not only tumorigenic and liver metastatic abilities but also chromosomal instability and mitotic activity, and moreover, pathologically displayed the distinctive MT/VE pattern and T cell exclusion. Boyault et al¹⁰ have proposed that the proliferation class of HCC is composed of 3 subgroups with different gene expression signatures and that the G3 subgroup harbors TP53 mutation and exhibits increased mitosis. Calderaro et al¹⁸ examined the close correlation between molecular subtypes and histological phenotypes and concluded that the G3 subgroup was enriched with MT-HCC. Previous studies have reported that the MT structure is frequently accompanied by the VE pattern,¹⁷ and that the VE formation promotes intrahepatic metastasis and contributes to dismal prognosis.²⁰ In addition, immunovascular analysis indicated that the immune-low/angiogenic subgroup included HCC with the G3-like gene expression signature and the MT/VE pattern.³¹ These studies corroborate the pathological findings of the syngeneic mouse model of MS1.

Among genes upregulated in 3H3PM cells with high tumorigenic and liver metastatic abilities under normal immune conditions, we highlighted ANGPTL2 and IL11 as angiogenic factors potentially involved in MT/VE formation (Figures 5C–F). Angiopoietin-like proteins are widely expressed in various kinds of tissues including the liver, vascular system, and hematopoietic system, and play important roles in lipid metabolism, inflammation, and vascularization.²⁷ A transgenic mouse model of Angptl2 expression under the control of a skin keratinocyte-specific promoter showed an increased number of blood vessels with excessive leakiness.³² ANGPTL2 expression correlates with intrahepatic metastasis in HCC patients, and overexpression of ANGPTL2 in HCC cell lines drives liver and lung metastasis.³³ IL11, a member of the IL6 family of cytokines, is secreted from hepatocytes in response to liver damage, presumably for organ protection, but promotes steatosis, fibrosis, and inflammation as well.³⁴ Recent studies have described that IL11 expression in breast cancer cells promotes high intratumoral vascular density,²⁸ and that IL11 protects human umbilical vascular endothelial cells from immune-mediated injury,³⁵ which raises the possibility that IL11 directly induces tumor angiogenesis. IL11 released from cancer cells also exerts paracrine immunosuppressive effects on cytotoxic CD4+ T cells.³⁶ In the present study, we discovered that both Angptl2 and Il11 knockout impaired MT/VE organization in mouse p53-deficient and MYC-activated HCC, similar to the previous finding that Angpt2 knockdown destroyed the VE structure in tumor tissues derived from the β-catenin-mutated mouse HCC cell line Hepa1-6.¹⁹ It is of great interest that IL11 is specifically overexpressed in the MS1 subtype of our classification as well as the G3 subgroup of Boyault classification (Supplemental Figure S5, http://links.lww.com/HEP/J740), which is principally composed of MT-HCC,¹⁷ implying subtype-specific molecular mechanisms of MT/VE formation such as IL11 in the MS1/G3 subtype and ANGPT2 in the β-catenin-mutated subtype.

Although immune checkpoint inhibitors have provided a revolutionary approach to the treatment of various types of cancer, anti-PD-1 monotherapy failed to improve overall survival compared with conventional therapy of HCC.^37,38 However, the combination of the anti-PD-L1 antibody atezolizumab and the anti-VEGF antibody bevacizumab was superior to sorafenib with respect to HCC patient prognosis, suggesting substantial synergistic effects of the 2 different classes of drugs on liver cancer production, progression, and metastatic spread.^6,7 Four biological mechanisms explain why anti-angiogenic therapy reinforces immunotherapy³⁹: (1) attenuation of cytotoxic T cell trafficking and effector function, (2) impairment of dendritic cell maturation and antigen presentation, (3) proliferation and recruitment of immunosuppressive cells including regulatory T cells, myeloid-derived suppressor cells, and M2-like tumor-associated macrophages, and (4) aberrant vasculature formation preventing immune effector cell infiltration. Given the findings in the current study (Figures 5 and 6), we propose that combination therapy of angiogenesis inhibition and immune checkpoint blockade recruits active T cells through the distortion of the MT/VE structure and reverts T cell exhaustion, respectively, resulting in tumor remission (Figure 6G). This hypothesis is consistent with the clinical data that such combination therapy is highly effective for the proliferation class containing MT/VE-HCC,^6,7 but not for the non-proliferation class, which includes β-catenin-mutated and metabolic disease-associated liver cancer.^40,41

In conclusion, the bioinformatics analysis of large-scale bulk and single-cell datasets and the pathological analysis of the syngeneic mouse model revealed that MT/VE formation protected tumor cells from T lymphocytes in MS1. Using the preclinical model, we discovered the synergistic anticancer effects of anti-angiogenic therapy and immune checkpoint blockade on this subtype by disrupting the MT/VE structure and preventing T cell exhaustion. As we presented clinical cases of MT/VE-HCC, sequential therapy by targeting tumor vasculature and blocking immune checkpoints may offer significant benefits to patients suffering from MT/VE-HCC (Figures 6D–F and Supplemental Figure S9, http://links.lww.com/HEP/J740).

Supplementary Material

hep-83-231-s001.pdf ^{(1.4MB, pdf)}