Endothelial Cells Contribute to the Immune-Vascular Milieu Characterizing VETC+ HCC

Abstract

Introduction

Vessel encapsulating tumor cluster (VETC) defines a distinct vascular phenotype of hepatocellular carcinoma (HCC), characterized by tumor nests fully encased within a continuous endothelial lining. This architecture correlates with poor prognosis and enhanced sensitivity to anti-angiogenic therapies. To dissect the morphological, phenotypical, molecular, and functional programs that define VETC-associated ECs in HCC and to identify endothelial pathways potentially driving VETC vascular organization and tumor progression.

Methods

Biological samples (blood, plasma, Peripheral blood mononuclear cells, and tissue) were collected from VETC+ and VETC− HCCs. Quantitative tissue imaging was integrated with EC-specific molecular profiling. Vascular architecture was assessed on whole-slide images, and purified tumor ECs underwent targeted and transcriptome-wide gene expression analyses.

Results

ECs from both VETC+ and VETC− HCCs lacked pro-growth and anti-apoptotic features. VETC+ ECs show overexpression of pro-angiogenic and immune-modulatory factors, in particular vascular endothelial growth factor alpha, FGF and IL10, and a specific gene signature involved in metabolic reprogramming, increased trafficking and migratory capacity. VETC+ transcriptional phenotype is driven by a coordinated hepatic xenobiotic-sensing network in which HNF4A and HNF1A provide the dominant regulatory architecture, while NFE2L2 contributes a distinct regulatory layer governing the antioxidant response elements-dependent arm of the xenobiotic and oxidative-stress response.

Conclusion

VETC+ ECs actively shape the HCC microenvironment through the release of angiogenic and immunomodulatory cues. These findings support the existence of a specialized angiocrine program in VETC+ HCC.

Introduction

Vessel encapsulating tumor cluster (VETC) represents a peculiar vascular phenotype observed in 40–60% of hepatocellular carcinomas (HCCs). VETC is consistently associated with unfavorable prognosis across clinical stages [1–6]. In addition to this, VETC is associated with anti-angiogenic therapies in unresectable HCC [6–8]. At the molecular level, VETC has been linked to the expression of angiopoietin-2 (ANGPT2), vascular endothelial growth factor alpha (VEGFA), fibroblast growth factor 2 (FGF2), and endocan (ESM1) [1, 3, 9]. This phenotype appears to be tightly modulated by the inflammatory milieu of HCC [3]. Specifically, perivascular lymphocytes and tumor-associated macrophages secreting CXCL10 and IL-18, respectively, can suppress VETC formation even in the presence of angiogenic stimuli (3). Consistently, VETC is significantly enriched in HCCs with β-catenin pathway activation devoid of lymphocytes [2, 10]. Nonetheless, rare cases with focal areas where lymphocytes coexist with VETC have been reported [10]. This finding further supports the fine-tuning action of the tumor microenvironment (TME) on VETC. Beyond lymphocytes, recent studies have also highlighted a close spatial and functional interplay between VETC-associated endothelial cells (ECs) and tumor-associated macrophages displaying a foamy morphology and a pro-tumoral transcriptomic profile [11, 12].

Despite the central role of ECs in defining the VETC architecture, no study has specifically focused on their biological features. Here, we characterize the morphological, phenotypic, and molecular properties of VETC-associated ECs and compare them with the endothelial compartment of capillary-type (VETC−) vascularization.

Materials and Methods

Human Tissues

HCC liver tissue specimens were obtained from patients undergoing surgical resection at the Hepatobiliary Unit of IRCCS Humanitas Research Hospital. For each of these cases, an expert liver pathologist collected, directly from the surgical specimen, consecutive serial sections of the tumor, then used as fresh, frozen and formalin-fixed paraffin-embedded (FFPE) material. In addition to this, the tumor was sampled completely. All tumoral FFPE blocks were then stained with CD34 to identify VETC+ cases, according to our experience [2]. For the purpose of this study, we considered as cases only HCC showing 100% VETC+ in all tumoral blocks and VETC− only those without any VETC area. Fresh HCC samples were collected in RPMI 1640 medium (Euroclone, ECB9006L) supplemented with 10% fetal bovine serum (FBS, Ultra-low endotoxin; Euroclone, ECS0186L), 1% penicillin/streptomycin (Euroclone, ECB3001D), and 1% L-glutamine (Euroclone, ECB3004D) until further processing. Human peripheral blood was collected in EDTA vacutainer tubes and processed within 3 h.

Plasma and Peripheral Blood Mononuclear Cell Isolation

Whole blood samples were centrifuged at 800 g for 10 min. Plasma was collected in cryovials and stored at −80°C until further analysis. Peripheral blood mononuclear cells were isolated using Lympholyte Cell Separation Media (Euroclone, DVCL5020R) and stored in 10% Dimethyl sulfoxide/FBS solution (Sigma-Aldrich D4540) at −80°C until further use.

EC Isolation

HCC tissues were minced on ice and enzymatically digested with 2.5 mg/mL collagenase IV (Worthington, LS004189), 0.1 mg/mL Deoxyribonuclease I (Type IV from bovine pancreas, Sigma-Aldrich, D5025-150KU), and ROCK inhibitor (Y-27632 dihydrochloride, Abmole Bioscience, M1817) at 37°C for 45 min. The resulting cell suspension was filtered through a 0.22-μm cell strainer, and red blood cells were lysed using Red Blood Cell Lysing Buffer (BD Biosciences, 555899). ECs were isolated using the human CD34 and CD31 MicroBead Kits (Miltenyi Biotec, 130-046-702 and 130-091-935). Purified ECs were seeded in collagen-coated 24-well plates (Merck Life Science, 122-20) and cultured in Endothelial Cell Basal Medium 2 (PromoCell, C-22111).

EC Purity Validation

Once appropriate confluency was reached, EC purity was assessed by flow cytometry. Briefly, 0.5 × 10⁶ cells were resuspended in 50 μL of staining buffer (PBS with 2% FBS and 2 mm EDTA) and stained with Aqua LIVE/Dead Fixable Dye (405 nm, 1:800; BioLegend, Cat. No.: 423102) and surface antibodies CD34-BV650 and CD31-BV711 (BD Biosciences OptiBuild, 745397 and 744078) (online suppl. Fig. 1a; for all online suppl. material, see https://doi.org/10.1159/lic/admag014). Samples were acquired using a BD LSRFortessa flow cytometer and analyzed with BD FACSDiva and FlowJo v10.10 software. In parallel, 1 × 10⁶ HCC ECs were lysed in TRIzol and stored at −80°C for RNA extraction and quantitative PCR (qPCR) validation of VE-cadherin expression (online suppl. Fig. 1b).

Image Analysis

FFPE sections (3–4 μm) were deparaffinized, rehydrated, and subjected to heat-induced antigen retrieval in EDTA buffer (pH 9.0). Endogenous peroxidase was blocked with 3% hydrogen peroxide. Sections were incubated with primary antibodies (anti-ERG, anti-Ki67, anti-p53; Roche) at optimized dilutions for 30–60 min. Detection was performed using a polymer-based HRP system, followed by DAB visualization and hematoxylin counterstaining. Slides were scanned using the Philips IntelliSite Ultra Fast Scanner and imported into QuPath (v0.5.1). For each case, five 1 mm² regions of interest were selected. QuPath’s segmentation tools were used to quantify total cells, immunoreactive nuclei (ERG, Ki67, p53), and ERG-positive nuclear area and perimeter.

qPCR Analysis

Total RNA was extracted using TRIzol reagent (Invitrogen), following the manufacturer’s instructions. Complementary DNA was generated by reverse transcription using the High-Capacity cDNA Archive Kit (Applied Biosystems). Quantitative real-time PCR was performed using SYBR Green PCR Master Mix (Applied Biosystems) on a ViiA™ 7 Real-Time PCR System (Applied Biosystems). Data were analyzed using the 2^−ΔCT method, with SDS software version 2.2.2 (Applied Biosystems), normalizing to GAPDH. Primer sequences are available upon request.

Quantitation of Cytokine Production in Sample Supernatants

Protein release in HCC tumor cell media, HCC-derived EC media, and patient plasma was quantified using LEGENDplex™ panels or ELISA. Specifically, angiopoietin-2 (Ang2) levels were quantified using the Human Angiopoietin-2 ELISA Kit (Thermo Fisher, KHC1641), following the manufacturer’s instructions. Briefly, high-binding 96-well plates were coated with the capture antibody and incubated overnight at 4°C. The next day, wells were blocked with 2% Bovine Serum Albumin in PBS (Sigma-Aldrich, A7030) to prevent nonspecific binding. Supernatants or standard controls were added and incubated for 2 h at room temperature. After washing, plates were incubated with the detection antibody followed by streptavidin-HRP. The reaction was developed with TMB substrate (Invitrogen, 88-7399-88) and stopped using 0.5 m sulfuric acid. Absorbance was measured at 450 nm using a plate reader, and cytokine concentrations were calculated from a standard curve. Cytokine, chemokine, and angiokine secretion in HCC tumor and HCC-EC supernatants was further assessed using LEGENDplex™ Human Angiogenesis Panel 1 (10-plex) and Human Essential Immune Response Panel (13-plex) (BioLegend, Cat. No.: 741215 and 740930, respectively), according to the manufacturer’s instructions. Briefly, 25 μL of mixed capture beads were incubated with 50 μL of either standards or supernatants in a 96-well V-bottom plate for 2 h at room temperature. Following centrifugation and wash steps, 25 μL of detection antibodies were added and incubated for 1 h. Subsequently, 25 μL of streptavidin-phycoerythrin was added to each well and incubated for 30 min. All incubations were performed at room temperature in the dark. After final washes, samples were resuspended in 150 μL of wash buffer. Samples were acquired on a BD FACSymphony™ and analyzed with LEGENDplex™ software. All samples were run in duplicate.

RNA Sample Preparation and Sequencing

ECs were purified from HCC tissues. Cells were stained with CD45-APC-H7 (BioLegend, 368516), CD31-PE (BioLegend, 303106), CD34-APC (BioLegend, 343608), and CD140b-PE-Cy7 (Miltenyi Biotec, 130-131-905), and subsequently sorted using a FACS Melody cell sorter (BD Biosciences, USA). RNA was extracted from sorted ECs using the TRIzol reagent (Invitrogen), following the manufacturer’s protocol. Residual genomic DNA was removed by DNase treatment with the RQ1 RNase-Free DNase (Promega, USA), and RNA integrity was assessed with the Agilent TapeStation 4200 (Agilent Technologies, USA). Only samples with RIN ≥7 were processed further. Libraries were prepared using the SMARTer Stranded Total RNA-seq Kit v3 – Pico Input Mammalian (Takara Bio, Japan). Library quality and fragment size distribution were validated with the Agilent TapeStation 4200 prior to sequencing. Sequencing was performed on the NextSeq2000 platform (Illumina, USA) to generate paired-end reads (2 × 150 bp), with a targeted coverage of 50 million reads per sample.

Data Processing and Quality Control

Raw sequencing data, available as binary base call files, were converted to FASTQ format using Illumina’s software bcl-convert (https://www.illumina.com/). Initial quality control of raw reads was performed using FastQC [13] in order to assess sequencing quality, adapter content, and potential systematic biases. After trimming for adapter content, processed reads were aligned to the GRCh38 reference genome using the STAR aligner [14], producing coordinate-sorted and indexed BAM files. Gene and transcript expression levels were quantified using RSEM [15], and expression estimates were reported as raw counts. Finally, a post-alignment quality control was also performed using Picard tools (Picard Toolkit by Broad Institute, GitHub repository available online at: https://broadinstitute.github.io/picard/) and samtools [16] to assess RNA-seq-specific metrics, alignment quality, and read-level statistics.

Comprehensive Analysis of Gene Expression

Differential gene expression analysis between ECs from VETC+ and VETC− HCC samples was performed using the DESeq2 R package [17], which implements methods for testing differential expression via negative binomial generalized linear models. Differentially expressed genes were defined as those with an adjusted p value <0.05 and an absolute log2-fold change (log2 FC) ≥0.8. Gene Ontology (GO) overrepresentation analysis and GO-based Gene Set Enrichment Analysis (GSEA) were performed using the clusterProfiler package available on Bioconductor [18], through the functions enrichGO and gseGO, respectively. To assess the robustness of pathway enrichment results across gene-set databases, GSEA was performed using the fgsea package [19] against eight collections from the Molecular Signatures Database (MSigDB) [20, 21]: Hallmark (H), C2:CP:KEGG, C2:CP:Reactome, C2:CP:WikiPathways, C2:CP:BioCarta, C5:GO:BP, C3:TFT:GTRD, and C6 Oncogenic Signatures. The DESeq2 Wald statistic was used as the gene-level ranking metric, as it integrates effect size with its standard error and is recommended for GSEA ranking [19]. Gene sets with a minimum size of 10 and maximum size of 500 genes were tested, and significance was assessed at a Benjamini-Hochberg-adjusted p value (FDR) <0.05. Normalized enrichment scores (NES) were reported for all enriched pathways. To confirm robustness to ranking metric choice, GSEA was additionally performed using log2 FC as the ranking statistic, and concordance was assessed by Spearman correlation of NES values and Jaccard overlap of significant pathway sets across the two rankings. To evaluate leading-edge consistency across collections, leading-edge genes were extracted from the top 30 significantly enriched pathways per collection (FDR <0.05, leading edge ≥3 genes), and pairwise Jaccard similarity was computed across all pathways. Hierarchical clustering (average linkage on 1 – Jaccard distance) was applied to identify coherent gene modules, and the number of clusters (k = 4) was selected by visual inspection of the dendrogram and biological interpretability of the resulting modules.

Transcription Factor Activity Inference

Transcription factor (TF) activity inference was performed using four methods based on three distinct methodological principles, applied to the DESeq2 Wald-statistic-ranked gene signature from the VETC+ vs. VETC− comparison. First, the decoupleR R package [22] was used with three statistical estimators: (i) the univariate linear model (run_ulm), which tests each TF independently; (ii) the multivariate linear model (run_mlm), which regresses all TFs simultaneously and controls for collinear regulons; and (iii) the permutation-based weighted sum (run_wsum, 1,000 permutations). For all three estimators, TF-target interaction information was obtained from the CollecTRI curated regulatory network [23], with a minimum regulon size of 5. TFs were ranked by activity score, and significance was assessed by the method-specific p value.

Second, VIPER [24] was applied using the analytic rank-based enrichment statistic with the DoRothEA regulon [25] filtered to high-confidence interactions (confidence levels A, B, and C), and a minimum regulon size of 10. The DoRothEA regulon was selected as a cell type-agnostic, literature- and ChIP-seq-curated network to avoid the tissue-specific biases inherent in data-derived networks. TF-target interactions were converted to VIPER regulon format using the df2regulon function from the dorothea R package. Third, ChIP-seq-based enrichment (ChEA3) [26] was used to provide orthogonal, ChIP-seq-based evidence of TF-DNA binding. The top 500 upregulated genes (ranked by the Wald statistic) were submitted to the ChEA3 API, which integrates six independent evidence sources: three experimental ChIP-seq compendia (Literature ChIP-seq, ReMap, ENCODE) and three co-expression databases (ARCHS4, GTEx, Enrichr Queries). TFs were ranked by the integrated mean rank across libraries, and individual library ranks were additionally reported to distinguish direct binding evidence from co-expression-based inference. Resulting TF activity scores were visualized using the ggplot2 R package [27]. Cross-method concordance was assessed by comparing TF ranks for NFE2L2 and related regulators across all four methods.

Results

VETC+ and VETC− ECs Show Similar Morphological and Phenotypic Features

To assess EC morphology, we quantified EC abundance, nuclear area, and nuclear perimeter using ERG – a highly specific EC marker [28] – stained whole-slide images from 20 HCCs (10 VETC+ and 10 VETC−, Fig. 1a, b) using QuPath 0.5.1. EC abundance (VETC−: median = 295.72 cells per mm², SD = 143.08, IQR = 188.95; VETC+: median = 192.60 cells per mm², SD = 97.95, IQR = 131.74; p = NS) (Fig. 1), nuclear area (VETC−: median = 25.25 µm², SD = 28.93, IQR = 22.19; VETC+: median = 25.13 µm², SD = 28.54, IQR = 19.19; p = NS) (Fig. 1d), and perimeter (VETC−: median = 25.91 μm, SD = 27.60, IQR = 19.93; VETC+: median = 24.93 μm, SD = 27.88, IQR = 16.24; p = NS) (Fig. 1e) were comparable between VETC+ and VETC− tumors.

Fig. 1.

Morphological and phenotypic features of EC in VETC+ and VETC− HCC. a, b The nuclei of endothelial cells (EC), highlighted by ERG, show a different morphological distribution in VETC+ and VETC− vascularization. In VETC+, they surround clusters of neoplastic cells, while in VETC−, they are aligned along hepatocyte trabeculae, recapitulating the path of normal sinusoids (c, d, sinusoidal pattern) (ERG immunostain, 40×). c–e The abundance, area and perimeter of EC nuclei are similar in VETC+ and VETC− cases.

Fig. 2.

Angiogenic and immune-modulatory capacities of EC in VETC+ and VETC− HCC. a Gene fold change of EC derived from VETC+ and VETC− HCC for vegf α (n = 9/8 VETC+/VETC−), esm1 (n = 10/8), ang2 (n = 11/8) and ctnnb1 (n = 12/10). b Assessment FGF, VEGF, TNF, IL-6, CXCL-8, Ang1, EGF, CD31 and PIGF levels in EC culture media (n = 5/6). c ANG2 protein level quantified via ELISA in plasma and EC media derived from VETC+ and VETC−. Error bars indicate the standard deviation of the mean. *p < 0.05, **p < 0.01, ***p < 0.001, unpaired multiple t test. d Assessment of IL-10, IL-4, IL-2, CXCL-10, IL-1β, MCP-1, IL-17, IFN-γ, and IL-12p70 in EC culture media (n = 5/5; except IL-2 n = 5/3). Error bars indicate the standard deviation of the mean. *p < 0.05, **p < 0.01, ***p < 0.001, unpaired multiple t test.

Next, we evaluated EC Ki67 and p53 expression in EC in the same cases. ECs displayed uniformly low Ki67 indices (1% vs. 1.5%; p = NS) and retained wild-type p53 expression across both vascular phenotypes.

VETC+ ECs Display a Pro-angiogenic and Immune-Modulatory Shift

To investigate the molecular features of ECs, we first measured the expression level of genes previously associated with the VETC phenotype, VEGFA, ESM1, and ANGPT2 [1, 3, 9], using qPCR. Given the enrichment of β-catenin-activated HCCs within the VETC+ group [1, 2, 9], we also assessed CTNNB1 expression. Among these genes, only VEGFA showed a significant difference between VETC+ and VETC− ECs, whereas ESM1, ANGPT2, and CTNNB1 levels were comparable (Fig. 2a).

To further explore the angiogenic capacity of ECs, we used a 10-plex panel of angiogenesis-related proteins. FGF (p = 0.06) and VEGF (p = 0.12) protein levels tended to be higher in EC culture media obtained from VETC+ cases (Fig. 2b), while no changes were observed in the HCC media (online suppl. Fig. 2a). The plasma levels of ANG2 were significantly higher in patients with VETC+ HCC, while no differences were observed in both EC and HCC media (Fig. 2c and online suppl. Fig. 2a). Taking into consideration the different structure of the immune cell compartment between VETC+ and VETC− HCC [3, 12], we sought to evaluate whether that difference could be explained by a deregulation of cytokines or chemokines release. By applying a 10-plex panel of immune-related proteins, we observed that IL-10 levels were higher (p = 0.06) in HCC-EC media from VETC+ (Fig. 2d). No differences were observed in HCC media obtained from VETC+ and VETC− cases (online suppl. Fig. 2b). Together, these findings indicate that VETC+ ECs, but not their VETC− counterparts, contribute to a pro-angiogenic and immune-modulatory milieu.

In silico Analysis Reveals Increased Trafficking and Migratory Capacity in VETC+ ECs

To investigate more comprehensively the specific functions exerted by VETC+ ECs, we performed RNA-seq analysis. We derived a VETC+ HCC ECs gene signature by identifying genes differentially expressed between ECs from VETC+ and VETC− samples. We identified a total of 68 differentially expressed genes, comprising 52 upregulated and 16 downregulated genes (Fig. 3a; online suppl. Table 1). Next, we employed overrepresentation analysis and GO-based GSEA of differentially expressed genes to obtain a biological functional interpretation. GO overrepresentation showed significant enrichment of pathways related to endocytosis, hemostasis, coagulation and immune regulation. Metabolic pathways, including xenobiotic metabolism, lipid and steroid metabolic processes, and triglyceride homeostasis, were also enriched, indicating substantial metabolic reprogramming of VETC+ ECs (Fig. 3b). GSEA of GO Cellular Components showed that VETC+ ECs were enriched for components associated with vesicle trafficking, secretory granules, clathrin-coated vesicles, endosomal and Golgi membranes, and vesicle lumina, indicating enhanced intracellular transport and secretory activity. Multiple plasma membrane subdomains (apical, basolateral, and external membrane regions) and cell projection structures (cell projection membrane, and cluster of actin-based cell projections) were also positively enriched, consistent with dynamic remodelling of the endothelial surface and increased migratory capacity. The Weibel-Palade body, a hallmark organelle of quiescent endothelium, appeared negatively enriched, highlighting a loss of canonical endothelial identity and suggesting a transition toward a pathologically EC phenotype (Fig. 3c). Together, these results define VETC+ ECs as metabolically reprogrammed and functionally activated, with enhanced intracellular transport, secretory activity, and migratory potential.

Fig. 3.

VETC+ EC gene signature and biological function interpretation. a Heatmap showing VETC+ and VETC− HCC EC DEGs. Rows represent genes, and columns represent individual samples. Gene expression values are shown as z scores, scaled across each row. Samples and genes were hierarchically clustered based on Euclidean distance and complete linkage. Red indicates higher expression, while blue indicates lower expression relative to the row mean. The top annotation bar indicates sample condition, with azure for VETC+ and purple for VETC− samples. b, c Bubble plots of overrepresentation analysis for the 52 genes significantly upregulated in VETC+ vs. VETC− HCC samples using ORA (b) and GSEA of Gene Ontology Cellular Component (c). Each bubble represents a biological process enriched among the upregulated genes. The size of the bubble indicates the number of genes associated with each term, while the color reflects the significance level (adjusted p value). DEG, differentially expressed gene.

Cross-Collection Pathway Enrichment Identifies NFE2L2/Xenobiotic-Response Program in VETC+ ECs

To characterize the transcriptional programs associated with the VETC+ endothelial phenotype beyond a single gene-set collection, we performed GSEA using the DESeq2 Wald statistic as the ranking metric against eight MSigDB collections: Hallmark (H), C2:CP:KEGG, C2:CP:Reactome, C2:CP:WikiPathways, C2:CP:BioCarta, C5:GO:BP, C3:TFT:GTRD, and C6 oncogenic signatures (online suppl. Table 2). This multi-collection approach tests whether the biological themes identified in any single collection are reproducible across independently curated gene-set universes. To confirm robustness to the choice of ranking metric, we additionally performed GSEA using log2 FC; NES values across the two rankings were strongly correlated (Spearman ρ = 0.81; online suppl. Fig. 3a, b). Across all eight collections, a total of 580 gene sets reached FDR <0.05 (online suppl. Table 2). The most prominent biological theme was a coordinated xenobiotic-biotransformation and oxidative-stress response program, which was enriched across six of the eight collections tested. Representative gene sets from this theme, including the top-ranked Hallmark pathway (HALLMARK_XENOBIOTIC_METABOLISM; NES = 2.60, FDR = 7.5 × 10⁻¹⁷), Phase I and phase II functionalization pathways from KEGG and Reactome (NES = 2.63–2.80, FDR <10⁻¹¹), and xenobiotic-response gene sets from GO:BP and WikiPathways (NES = 2.49–2.66, FDR <10⁻⁹) are shown in Figure 4a. Interestingly, two explicitly NFE2L2-defined gene sets showed concordant positive NES: the WikiPathways WP_NRF2_PATHWAY showed a borderline enrichment (NES = 1.53, FDR = 0.051), while the C6 perturbation signature derived from NFE2L2-knockout fibroblasts (NFE2L2.V2) was significantly enriched (NES = 1.67, FDR = 1.2 × 10⁻³). Among the NFE2L2-downstream gene sets, NFE2L2.V2, derived from genes downregulated upon knockout of NFE2L2 in embryonic fibroblasts, shared five genes with the VETC+ EC signature: vitronectin (VTN), PCSK9, GPNMB, CYP4F3, and CXCL12 (Fig. 4b). All five genes have established roles in vascular remodelling, extracellular matrix organization, and tumor-endothelial interactions [29–34], providing a biologically meaningful link between NFE2L2 transcriptional activity and the VETC phenotype at the individual-gene level. These results provide regulator-level evidence suggesting NFE2L2 as a TF whose target-gene program is activated in VETC+ ECs.

Fig. 4.

Oncogenic pathways and potential transcriptional regulators of VETC+ EC. a Cross-collection convergence of the NFE2L2/xenobiotic-response signal in VETC+ endothelial cells. Heatmap of normalized enrichment scores (NES) for fourteen NFE2L2-related and xenobiotic-response gene sets across five MSigDB collections (Hallmark, KEGG, Reactome, WikiPathways, GO:BP) and the C6 oncogenic signature collection. Color encodes NES (red, positive enrichment in VETC+ endothelial cells; white, NES near zero; grey, gene set not present in that collection). Asterisks denote Benjamini-Hochberg-adjusted significance: *FDR <0.05, **FDR <0.01, ***FDR <0.001. The top two rows correspond to explicitly NFE2L2-defined gene sets curated from independent sources: WP_NRF2_PATHWAY (WikiPathways; NES = 1.53, FDR = 0.051) and NFE2L2.V2 (C6 oncogenic perturbation signature; NES = 1.67, FDR = 1.2 × 10⁻³). The remaining rows are NFE2L2-downstream xenobiotic-response pathways whose leading-edge genes are enriched for canonical antioxidant response element (ARE)-regulated targets (see text). GSEA was performed using the DESeq2 Wald statistic as the ranking metric; results are robust to log2 fold change ranking. b Boxplots showing variance-stabilized expression (VST) of five genes shared by the VETC+ EC signature and NFE2L2.V2 oncogenic signature, across two experimental conditions (VETC+ vs. VETC−). Each dot represents an individual sample. p values are from Mann-Whitney U tests. Boxplots display median and interquartile range. c Cross-method transcription factor activity ranks for NFE2L2 and related regulators. Activity ranks across six inference methods based on three distinct principles: regulon-based linear models (dR_MLM, dR_ULM, dR_wsum; CollecTRI regulon), regulon-based rank enrichment (VIPER; DoRothEA A–C regulon), and ChIP-seq-based enrichment (ChEA3 integrated and Literature ChIP-seq). Color encodes log₁₀(rank); numbers indicate absolute rank; grey cells indicate absent TFs. NFE2L2 ranks 3rd in the multivariate model that controls for co-regulatory overlap, and 83rd by Literature ChIP-seq (top 5%). BACH1, NFE2L2’s direct ARE competitor, ranks 7th by Literature ChIP-seq, providing orthogonal evidence for active ARE-mediated transcription. HNF4A and HNF1A consistently rank highest, supporting a multi-TF regulatory architecture.

Computed Pairwise Similarity Analysis Links the Enriched Program to Canonical NFE2L2/Antioxidant Response Element Target Genes

To assess whether the enrichment signal reflects a coherent transcriptional program rather than independent pathway-label artifacts, we extracted the leading-edge genes from the top 30 significantly enriched pathways per collection and computed pairwise Jaccard similarity across 163 pathways from all eight collections (online suppl. Fig. 3c, d). Hierarchical clustering revealed one dominant module encompassing 146 of 163 pathways across five independent MSigDB collections (Hallmark, KEGG, Reactome, WikiPathways, and GO:BP). The 1,096 unique leading-edge genes defining this module were dominated by phase I cytochrome P450 enzymes (CYP3A4, CYP1A2, CYP2C9, CYP2C8, CYP2D6), phase II conjugation enzymes (UGT1A1, glutathione-S-transferases), and hepatic transporters (ABCB11, FABP1, APOA1, APOE). These gene families are responsive to NFE2L2 through antioxidant response elements (AREs), but are also co-regulated by hepatic nuclear receptors including HNF4A and NR1I2 (PXR). As shown in the next paragraph, this is consistent with the multi-TF regulatory architecture identified by TF analysis.

Multi-Method TF Inference Validates NFE2L2 within a Hepatic Xenobiotic-Sensing Network

To identify upstream transcriptional regulators driving the VETC+ gene expression program, we performed TF activity inference using four methods based on three distinct methodological principles: regulon-based linear models (decoupleR univariate and multivariate, with CollecTRI regulon), regulon-based rank enrichment (VIPER with DoRothEA A–C regulon), and experimental ChEA3. Among the regulon-based methods, the top-ranked TFs consistently included HNF1A, HNF4A, CEBPA, and FOXA2, master hepatic nuclear receptors and endodermal TFs (Fig. 4c). NFE2L2 ranked 3rd genome-wide in the decoupleR multivariate linear model (score = 3.80, p = 1.45 × 10⁻⁴), which controls for collinear regulons by regressing all TFs simultaneously. This suggests that NFE2L2 has a regulatory contribution independent of HNF4A (rank 2) or HNF1A (rank 1). NFE2L2 ranked 16th by the univariate model (p = 3.53 × 10⁻⁷) and 17th by weighted sum (p = 0.002). VIPER analysis with the DoRothEA regulon placed NFE2L2 at rank 106 (NES = 0.90; top 17%), with the lower rank reflecting VIPER’s known sensitivity to regulon overlap among co-regulatory TFs. Nevertheless, VIPER recovered the complete ARE-binding module: BACH1 (rank 20), MAFK (rank 37), and MAFG (rank 60), NFE2L2’s direct ARE competitor and obligate heterodimerization partners, respectively, constituting the expected molecular signature of active NFE2L2/ARE signaling.

To provide evidence based on a fundamentally different principle, we performed ChEA3 enrichment, which identifies TFs by direct experimental ChIP-seq-binding data rather than curated regulons. NFE2L2 ranked 83rd in the Literature ChIP-seq library (top 5%) and 170th in the ReMap ChIP-seq library (top 11%). BACH1 ranked 7th in the Literature ChIP-seq library: since it functions exclusively as a transcriptional repressor at AREs, its top ranking among upregulated genes constitutes direct ChIP-seq-based evidence that ARE-containing promoters are transcriptionally active in VETC+ ECs, consistent with NFE2L2-mediated displacement of BACH1. We note that HNF4A and HNF1A consistently ranked above NFE2L2 across all methods. We accordingly interpret the VETC+ transcriptional phenotype as driven by a coordinated hepatic xenobiotic-sensing network in which HNF4A and HNF1A provide the dominant regulatory architecture, while NFE2L2 contributes a distinct, independently validated regulatory layer that specifically governs the ARE-dependent arm of the xenobiotic and oxidative-stress response.

Discussion

VETC represents a distinct vascular phenotype of HCC, characterized by the encapsulation of tumor nests within a continuous layer of ECs. Although this vascular phenotype is associated with poor prognosis [1–6] and a better response to anti-angiogenic treatment [6–8], the functional state of its hallmark, namely ECs, has remained completely unexplored. In this study, we provide the first comprehensive characterization of VETC+ ECs and provide evidence supporting their active contribution to HCC progression and possibly disclosing clues for novel treatment of VETC+ HCC.

ECs within VETC+ HCC were indistinguishable, at the morphological and phenotypic level, from those in VETC− cases. On the other hand, VETC+ but not VETC− ECs showed a peculiar pro-tumoral molecular and functional profile. Our data identify VEGF signaling as a defining feature of VETC+ ECs. Consistently, functional profiling revealed a trend toward increased secretion of key pro-angiogenic factors, namely VEGF and FGF, restricted to conditioned media obtained from VETC+ ECs. By contrast, we observed that ANG2 levels were significantly elevated in serum but not in EC-conditioned media of VETC+ HCC. These findings are in keeping with the notion that ANG2 is derived from HCC cells [1]. In line with this, transcriptomic analyses revealed negative enrichment of Weibel-Palade bodies, the main endothelial storage organelles for ANG2 [35], providing a possible explanation for the absence of ANG2 release by ECs. These data also suggest that circulating ANG2 may serve, at the clinical level, as a surrogate serum biomarker for VETC to predict the response to anti-angiogenic treatments [7, 8]. Beyond classical angiogenic mediators, transcriptomic and functional analyses revealed that VETC+ ECs upregulate a set of factors with established pro-tumoral and immunomodulatory functions. Notably, VTN, which is known to enhance VEGF-VEGFR2 interactions [29, 30], could promote endothelial migration and neovascularization, providing a mechanistic link between VEGFA upregulation and sustained endothelial activation in VETC. Additionally, both VTN and PCSK9 have been previously implicated in endothelial-to-mesenchymal transition [31, 32], a plasticity that may underlie the unique vascular permeation characteristic of VETC, which occurs independently of classical epithelial-mesenchymal transition [1]. In parallel, VETC+ ECs appear to actively shape an immunosuppressive TME. The downregulation of CXCL12, together with the upregulation of IL-10 and GPNMB, is consistent with impaired anti-tumor immunity. Indeed, endothelial-derived CXCL12 has a central regulatory role on immune cells [33], while IL-10 secreted by ECs limits lymphocyte trafficking [33] and endothelial expression of GPNMB promotes CD8⁺ T cell exhaustion [34]. These findings align with current understanding of the biological processes characterizing ECs across different cancers, which include not only angiogenesis but also immunomodulation and extracellular matrix organization [36]. Taken as a whole, our data support the existence of a specialized angiocrine program in VETC+ HCC, whereby ECs actively modulate tumor behavior and the surrounding microenvironment through paracrine cues rather than serving solely as structural support [37].

Pathway-based analyses pointed to activation of stress- and metabolism-related transcriptional programs in VETC+ ECs. In particular, a cross-collection analysis identified a dominant xenobiotic-biotransformation module whose recurrent gene content includes multiple gene families responsive to NFE2L2 through AREs. Two explicitly NFE2L2-defined gene sets were independently enriched, providing regulator-level evidence linking the enriched program directly to NFE2L2 transcriptional activity. The convergence of angiogenic (VTN, PCSK9), metabolic (CYP4F3), and immune- (GPNMB, CXCL12) modulatory genes, characterizing the VETC signature, within this program suggests that NFE2L2 may be associated with the angiocrine programs observed in VETC-associated ECs. A deep in silico analysis of upstream transcriptional regulators revealed that the VETC+ transcriptional phenotype is driven by a coordinated hepatic xenobiotic-sensing network in which HNF4A and HNF1A provide the dominant regulatory architecture, while NFE2L2 contributes a distinct regulatory layer governing the ARE-dependent arm of the xenobiotic and oxidative-stress response. Together, these converging lines support the prioritization of NFE2L2 as a transcriptional regulator linking metabolic reprogramming, endothelial activation, and immune modulation in VETC vasculature.

Our study has some limitations, admittedly. First, while we identified NFE2L2 as a possible regulator through multi-method bioinformatic inference spanning regulon-based, rank-enrichment, and ChIP-seq-based approaches, our results remain descriptive, and functional validation via gene silencing or overexpression models is required to definitively establish causality. Second, the sample size for secretome analysis was relatively small, which may have limited the statistical power for some analytes that showed clear upregulation trends (e.g., FGF, IL-10). Although we performed transcriptomic profiling on highly purified ECs, the use of bulk RNA-sequencing on in vitro-cultured cells might not fully capture the spatial heterogeneity or the complete cross talk occurring within the complex in vivo TME. Finally, the study focused only on mRNA and lacked any validation at the protein level. Nonetheless, we think that our results may lay the foundation for deeper analysis.

In conclusion, our study provides a comprehensive characterization of VETC+ ECs in HCC, revealing that they are transcriptionally and functionally reprogrammed to support HCC progression. VETC+ ECs display a coordinated activation of angiogenic, metabolic, and immunomodulatory programs, with NFE2L2 emerging as a possible relevant player in metabolic reprogramming, endothelial activation, and immune modulation in VETC vasculature.

Statement of Ethics

The study was approved by the Ethics Committee of the Humanitas Clinical and Research Center (COMITATO ETICO TERRITORIALE LOMBARDIA 5 – approval: 52/24), and written informed consent was obtained from all patients.

Conflict of Interest Statement

None of the authors report any potential conflict of interest.

Funding Sources

The research leading to these results has received funding from AIRC under IG 2020 – ID. 25087 project – P.I. Di Tommaso Luca. C.K.Y.N. is supported by the Italian Foundation for Cancer Research – AIRC Start-Up Grant No.: 30787. The sponsor (Fondazione AIRC per la Ricerca sul cancro) supports research activities on cancer, but it has no role in this specific study (design, execution, and analysis), nor was it involved in the conception, planning, writing, and decision to publish the paper.

Author Contributions

Study conceptualization: B. Durante, A. Lleo, G. Torzilli, F. Procopio, L. Viganò, S. Piscuoglio, I. Simeone, L. Di and Tommaso; data curation: B. Durante, S. Zanella, F. de Napoli, B. Esposito, C. De Carlo, R. Akpinar, C. Richichi, F. Lazzaroni, B. Franceschini, C. Soldani, F. Grizzi, F. Pasqualini, A. Lleo, G. Torzilli, F. Procopio, L. Viganò, S. Piscuoglio, C.K.Y. Ng, F. Airi, E. Paraboschi, I. Simeone, and L. Di Tommaso; formal analysis: B. Durante, S. Zanella, F. de Napoli, B. Esposito, C. De Carlo, C.K.Y. Ng, F. Airi, E. Paraboschi, I. Simeone, and L. Di Tommaso; investigation: B. Durante, S. Zanella, I. Simeone, and L. Di Tommaso, methodology: B. Durante, S. Zanella, F. Lazzaroni, C. Soldani, C.K.Y. Ng, F. Airi, E. Paraboschi, I. Simeone, and L. Di Tommaso; project administration: B. Durante, I. Simeone, and L. Di Tommaso.

Funding Statement

The research leading to these results has received funding from AIRC under IG 2020 – ID. 25087 project – P.I. Di Tommaso Luca. C.K.Y.N. is supported by the Italian Foundation for Cancer Research – AIRC Start-Up Grant No.: 30787. The sponsor (Fondazione AIRC per la Ricerca sul cancro) supports research activities on cancer, but it has no role in this specific study (design, execution, and analysis), nor was it involved in the conception, planning, writing, and decision to publish the paper.

Data Availability Statement

The authors confirm that the data supporting the findings of this study are available within the article and its supplementary materials. Further enquiries can be directed to the corresponding author.