Targeting Treg–fibroblast interaction to enhance immunotherapy in steatotic liver disease-related hepatocellular carcinoma

Abstract

Background

Steatotic liver disease-related hepatocellular carcinoma (SLD-HCC), a rising global challenge, is characterised by unique tumour microenvironment (TME) adaptations.

Objective

This study investigates the immune microenvironment and interactions driving immunosuppression and potential resistance to immunotherapy in SLD-HCC.

Design

We employed single-cell transcriptomics, cytometry by time-of-flight (CyTOF) and two independent spatial transcriptomics platforms to study the TME of 22 SLD-HCC and 31 non-SLD-HCC cases. Findings were further validated using multiplex immunohistochemistry in an independent cohort of 103 patients, an HCC model and an immunotherapy-treated patient cohort to evaluate clinical relevance.

Results

Our findings revealed significant alterations in immune and lipid metabolism pathways, particularly in regulatory T cells (Tregs) and cancer-associated fibroblasts (CAFs), suggesting distinct cellular adaptations to a high-fat TME and general immunosuppression. CyTOF revealed a cold, immunosuppressive TME with reduced CD8⁺ T cells and increased Tregs. Spatial transcriptomics further highlighted distinct Treg–CAF clusters localised at tumour margins, suggesting a spatially organised immunosuppressive niche. Mechanistically, tumour necrosis factor superfamily member 14 (TNFSF14)-tumour necrosis factor receptor superfamily member 14 (TNFRSF14)-mediated Treg–CAF interaction was identified as a critical driver of immunotherapy resistance in SLD-HCC. Blocking TNFRSF14 in an HCC model fed with a high-fat diet resulted in reduced Tregs, increased active CD8⁺ and memory CD4⁺ T cells, and a synergistic effect with anti-programmed cell death protein 1 therapy to enhance antitumour immunity and overcome immunotherapy resistance in SLD-HCC.

Conclusion

This study uncovers critical immune and metabolic adaptations in SLD-HCC, identifying TNFSF14-TNFRSF14 signalling as a key driver of immunotherapy resistance. Targeting this signalling axis enhances antitumour immunity and improves immunotherapy efficacy, offering a promising therapeutic strategy for SLD-HCC.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Steatotic liver disease-related hepatocellular carcinoma (SLD-HCC) is characterised by a unique tumour microenvironment (TME) shaped by metabolic and immune adaptations, contributing to immunotherapy resistance. However, the specific immune interactions driving this resistance remain poorly defined.

WHAT THIS STUDY ADDS

• This study identifies tumour necrosis factor superfamily member 14 (TNFSF14)-tumour necrosis factor receptor superfamily member 14 (TNFRSF14)-mediated regulatory T cell and cancer-associated fibroblast interactions as a critical driver of immunosuppression and immunotherapy resistance in SLD-HCC. Blocking TNFRSF14 enhances CD8⁺ T cell responses and synergises with anti-programmed cell death protein 1 therapy to improve antitumour immunity.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Targeting TNFSF14-TNFRSF14 signalling provides a novel strategy to modulate the immunosuppressive TME and enhance immunotherapy efficacy in SLD-HCC. These findings support the development of combination therapies aimed at overcoming immune resistance and improving clinical outcomes.

Introduction

Hepatocellular carcinoma (HCC) is the third leading cause of cancer mortality worldwide, and its global incidence and mortality are predicted to rise by more than 50% by 2040.¹ HCC primarily develops from a background of chronic liver inflammation, with a heterogenous etiopathogenesis that includes chronic viral hepatitis infection, alcoholism and fatty liver diseases, making targeted therapy challenging. Despite recent advancements in immunotherapies for advanced HCC, the objective response rate remains modest,² warranting a deeper understanding of the complex nature of the immune microenvironment in HCC.

Steatotic liver disease (SLD), encompassing alcoholic and non-alcoholic fatty liver disease or steatohepatitis, is detrimental to immune cell functions.³ The overarching term SLD encompasses the various aetiologies of hepatic steatosis, while addressing social stigma associated with the terms ‘non-alcoholic’ or ‘fatty’.⁴ Furthermore, the challenge of retrospectively tracing alcohol intake in patients underscores the need for a more inclusive terminology. Given the rapid rise of SLD globally,⁵ understanding the underlying mechanisms of SLD-related HCC (SLD-HCC) is crucial for developing new immunotherapeutic strategies.

The accumulation of programmed cell death protein 1 (PD-1)⁺ CD8⁺ T cells in the metabolic dysfunction-associated steatohepatitis (MASH) liver, specifically the CXCR6⁺ PD-1⁺ autoaggressive and proinflammatory CD8⁺ T cells, has been associated with disease progression to HCC and poor response to checkpoint inhibition targeting PD-1/programmed cell death ligand 1 (PD-L1) pathways in MASH-HCC.^{3 6} However, comprehensive high-dimensional single-cell and spatial analyses of tumour microenvironment (TME) in this HCC subtype remain underexplored. Our present study focuses on SLD-HCC as the primary disease condition, leveraging cutting-edge single-cell RNA sequencing (scRNA-seq), spatial transcriptomics with proteomics validation to unravel the spatial immune interactome and antitumour immunity in SLD-HCC.

Methods

Patient samples

Patients’ clinical characteristics are described in online supplemental table S1. SLD-HCC (n=22) was defined as patients with at least 5% intrahepatic fat and at least one metabolic risk factor, hence also deemed as metabolic-associated SLD (MASLD), in accordance with the recent Delphi consensus.⁴ Patients who did not fulfil either of these criteria were classified as non-SLD-HCCs (n=31).

Depending on its size, each tumour specimen was systematically sampled 2–5 times with at least 1 cm gap between sectors to account for intratumoural heterogeneity.⁷ Adjacent non-tumour liver tissue that was at least 2 cm away from the tumour edge was also collected. Each tissue sector was subdivided and dissociated with Collagenase IV (Gibco, cat# 17104019) and deoxyribonuclease I (Worthington, cat# LS002147) as described previously.⁷ Due to limited tissue availability, not all samples were included in each multiomics analysis as detailed in online supplemental table S1. Additional validation using multiplex immunofluorescence (mIF) was performed on resected formalin-fixed, paraffin-embedded (FFPE) HCC tissues from 103 patients (online supplemental table S2).

scRNA-sequencing

ScRNA-seq data from 13 tumour sectors from five SLD-HCCs, 21 tumour sectors from six non-SLD-HCCs and 11 non-tumour liver sectors (one from each case) were analysed. Sample collection, library preparation, sequencing methods and data processing have been described in detail previously⁸ and detailed data analysis is provided in online supplemental methods.

Cytometry by time-of-flight

Immune cells were stained with a panel of antibodies (online supplemental table S4) as previously described⁷ and analysed using a Helios mass cytometer (Fluidigm, USA). Data were downsampled to 10 000 viable CD45+ cells for analysis using in-house developed Extended Poly-dimensional Immunome Characterisation.⁹ Clustering was performed with the Flow‐Self Organizing Maps (FlowSOM) algorithm followed by dimensionality reduction, and t-distributed stochastic neighbour embedding plots visualised using ‘SciAtlasMiner’. Enriched clusters were identified by two-tailed Student’s t-test and validated with manual gating using FlowJo (V.10.5.2; FlowJo, USA).

Multiplex immunofluorescence staining

FFPE tissues were deparaffinised, rehydrated and subjected to heat-induced epitope retrieval. 10% goat serum (DAKO; X0907) was used for blocking. The tissues were stained with respective antibodies (online supplemental table S5; both anti-human and anti-mouse panels) using the Opal 7-colour IHC Kit (Akoya Biosciences) and counterstained with 4′,6-diamidino-2-phenylindole (DAPI) (Thermo Fisher). Images were acquired on Vectra V.3.0 Pathology Imaging System Microscope (Perkin-Elmer) and images analysed using InForm V.2.1 (Perkin Elmer) and Imaris V.9.1.0 (Bitplane). For quantification, the mean number of Tregs and the average distances between Tregs and fibroblasts were quantified from 3 to 5 randomly selected fields of view (FOVs) at tumour margin.

Bulk RNA sequencing

Tissue samples from 11 SLD-HCCs (n=34 tumour sectors) and 23 non-SLD-HCCs (n=91 tumour sectors) as well as murine tumour tissues were mechanically homogenised, and total RNA was isolated using PicoPure RNA Isolation kit (Arcturus, Ambion). Complementary DNA (cDNA) was generated using the SMART-Seq V.4 Ultra Low Input RNA Kit for Sequencing (Clontech, USA). Nextera XT DNA Library Prep Kit (Illumina, USA) was used to generate indexed libraries and subsequently multiplexed for 2×101 bp-sequencing. RNA sequencing was performed on a HiSeq High output platform (Illumina, San Diego, California, USA). Data processing and analysis were previously described.⁷ KEGG_WNT_signaling_pathway signature was obtained from gene set enrichment analysis, and the relative expression levels of these genes were calculated according to the median expression from each tumour sector.

Whole genome sequencing

Tissue samples from 11 SLD-HCCs (n=33 tumour sectors), 23 non-SLD-HCCs (n=90 tumour sectors) and 34 adjacent non-tumour tissues (one from each patient) were mechanically homogenised, DNA was extracted using Qiagen AllPrep kit and sonicated into shorter fragments using the Covaris system. A quality check was performed on an Agilent 2100 Bioanalyser before DNA fragments were end-repaired, ligated, amplified and sequenced using the Illumina sequencing platform (Illumina, San Diego, California, USA). Data processing and analysis was previously described.⁷ Neoantigen was predicted using personalised Variant Antigens by Cancer Sequencing (pVacSeq), with variant calling information obtained from comparison of tumour against non-tumour tissue by MuTect (V.1.1.7). The neoantigens were selected based on VEP: 8-mer to 11-mer epitopes with <500 nM predicted binding affinity to major histocompatibility complex (MHC)-class 1.

Visium

FFPE tissues from seven SLD-HCCs and five non-SLD-HCCs, comprising tumour and non-tumour regions, were subjected to spatial transcriptomics (STs) experiments using the Visium CytAssist Spatial Gene Expression for FFPE, 6.5 mm (10x Genomics, cat# PN-1000520) or 11 mm (10x Genomics, cat# PN-1000522) following the manufacturer’s protocol. Only samples with DV200 ≥50% were selected for ST. 5 µm FFPE tissues were sectioned, dried, deparaffinised and rehydrated. H&E staining was performed and tissues were imaged. Spatial RNA library preparation was performed using the Visium Human Transcriptome Probe Kit V.2 (10x Genomics, California, USA). Probe hybridisation and ligation were performed, followed by transfer onto Spatial Gene Expression Slides containing barcoded capture areas using Visium CytAssist machine. Captured products were extended to incorporate unique molecular identifiers (UMIs) and spatial barcodes. For library construction, cDNA was amplified, barcoded and cleaned up using SPRIselect before sequencing with Illumina NovaSeq 6000 (2 × 150 bp reads).

CosMx spatial molecular imaging

CosMx technology (CosMx Human Universal Cell Characterisation RNA Panel (1000-plex); NanoString, USA) was applied to four FFPE samples (two SLD-HCC, two non-SLD-HCC) with a total of 50 FOVs according to manufacturer’s protocol.¹⁰ Briefly, 5 µm FFPE sections were mounted, baked overnight at 60°C, deparaffinised and subjected to in-situ hybridisation using the Leica BOND RXm system. Tissues were digested, fixed and hybridised overnight with RNA in situ hybridisation (ISH) probes (1000-plex) before staining with a PanCK/CD45/CD3 antibody cocktail and DAPI. The CosMx instrument captured RNA readout and imaging data in 16 cycles of reporter hybridisation-imaging that was repeated eight times to increase RNA detection sensitivity. After the total of nine complete cycles of RNA readout, nine Z-stack images for four channels (three antibodies and DAPI) were captured.

*Detailed ST data analysis for Visium and CosMx was provided as online supplemental methods.

HCC mouse model validation

6-week-old male C57BL/6 mice (InVivos, Singapore) were housed in pathogen-free conditions according to guidelines of Institutional Laboratory Animal Care and Use Committee of the National University of Singapore . The animals were fed with a high-fat diet (HFD) containing 35% lard (TD.03584, Harlan Laboratories) or normal diet for 7 weeks before surgically transplanted with 1×10⁶ Hepa1-6 murine hepatoma cells to the liver.¹¹ 5 days after tumour transplantation, mice were randomly divided into four groups (n=5 mice each) and treated intraperitoneally and biweekly for 2 weeks with rat IgG2a isotype (200 µg/mouse; Bio X Cell, West Lebanon, New Hampshire, USA), Armenian Hamster IgG isotype (2 mg/kg/mouse, eBioscience, San Diego, California, USA), rat anti-mouse PD-1 antibody (200 µg/mouse; Bio X Cell, West Lebanon, New Hampshire, USA) or Armenian Hamster anti-mouse TNFRSF14 antibody (2 mg/kg/mouse, eBioscience, San Diego, California, USA) or its respective combination.

After 2 weeks, the mice were euthanised by CO₂ asphyxiation. The livers were collected and tumour volumes, (length × width²)/2, were measured using a digital calliper. Tumour-infiltrating lymphocytes (TILs) and non-tumour tissue-infiltrating lymphocytes (NILs) were isolated for flow cytometry analysis and tumour tissues were collected for histology, mIF and bulk RNA sequencing analyses.

Flow cytometry

TILs and NILs isolated from HCC-bearing mice were resuspended in cell staining buffer (phosphate-buffered saline with 4% fetal bovine serum) at a concentration of 1×10⁶ cells/mL. Cells were blocked with anti-mouse CD16/32 antibodies, stained with Live/Dead Fixable blue, washed and stained with fluorochrome-conjugated surface antibodies (online supplemental table S7). After incubation, the cells were washed and permeabilised with Intracellular Fixation and Permeabilization Buffer Set (eBioscience, California, USA) before intracellular staining of FoxP3 and GZMB. Flow cytometry data was acquired using a BD LSRFortessa flow cytometer (BD Bioscience, New Jersey, USA) and analysed using FlowJo V.10.10 (BD Bioscience, New Jersey, USA).

Results

Single-cell transcriptomic analysis of the SLD-HCC microenvironment

Using a multiomics approach (figure 1A), we interrogated the TME of SLD-HCC and non-SLD-HCC, defined as ≥5% and <5% intrahepatic fat respectively, by pathological assessment. Alongside steatosis, they also displayed signs of ballooning and inflammation marked by immune infiltration (online supplemental figure S1A). In accordance with the recent Delphi consensus,⁴ all SLD-HCC cases in our cohort were also patients with MASLD, as they harboured at least one metabolic risk factor (online supplemental table S1). Comparison of key clinical parameters showed no significant differences between the two HCC subgroups (online supplemental figure S1B).

Figure 1. Single-cell transcriptomic landscapes of SLD-HCC and non-SLD-HCC. (A) Schematic illustration of the study design. (B) Dot plot showing the top expressed genes in each cell cluster. Dot colour indicates average expression level, and dot size indicates the percentage of cells. (C) UMAP plot showing Louvain clustering of 65 606 single cells. (D) Bar graphs showing the distribution of major cell types between tumour and non-tumour tissues (left) and between patients with SLD-HCC and non-SLD-HCC (middle) and the total cell numbers (right). (E) Dot plot showing the enriched gene ontology pathways, comparing DEG in SLD-HCCs versus non-SLD-HCCs. Dot colour indicates enrichment in SLD-HCCs versus non-SLD-HCCs, and dot size indicates adjusted p value by Benjamini-Hochberg method. (F) Volcano plots depicting representative genes significantly upregulated (red) or downregulated (green) in Tregs, CD8⁺ T cells, fibroblasts and ECs from SLD-HCCs compared with non-SLD-HCCs. DEG analysis using Wilcoxon rank-sum test and log-transformed, with adjusted p value<0.05 considered significant. (A–F) Data from tumours and non-tumour-adjacent liver tissues of patients with SLD-HCC (n=5) and non-SLD-HCC (n=6). CAF, cancer-associated fibroblast; DC, dendritic cell; DEG, differentially expressed gene; ECs, endothelial cells; mIF, multiplex immunofluorescence; NK, natural killer; SLD-HCC, steatotic liver disease-related hepatocellular carcinoma; Treg, regulatory T cell; UMAP, uniform manifold approximation and projection; WGS, whole genome sequencing.

First, we assessed the single-cell transcriptomic landscape of tumour and non-tumour sectors from five SLD-HCCs and six non-SLD-HCCs (online supplemental table S1). Unsupervised clustering of 65 606 single cells identified 27 clusters (online supplemental figure S2A and table S3), which were categorised into 17 major cell types (figure 1B,C, online supplemental figure S2B,C), encompassing immune, fibroblast, endothelial and hepatocyte populations. Cell proportion analysis showed that most hepatocytes were of tumour origin, while other cell types originated from both tumour and non-tumour tissues (figure 1D). Tumour hepatocyte populations were heavily driven by individual patients, highlighting interpatient heterogeneity. In particular, APOA2⁺ hepatocytes mainly originated from SLD-HCCs (figure 1D, online supplemantal figure S2C-S2D) and were enriched in lipid metabolism-related apolipoproteins, consistent with the steatotic HCC phenotype. Other cell types were evenly distributed across patients, reflecting uniformity in cellular representation (figure 1D, online supplemental figure S2E,F).

Next, we isolated the cells from tumour tissues only and compared the phenotypes of individual cell types from SLD-HCCs versus non-SLD-HCCs. Overall, pathway analyses suggested that both CD4⁺ and CD8⁺ T cells in the SLD-HCC TME downregulated pathways involved in immune or inflammatory responses (figure 1E). Conversely, Tregs, endothelial cells (ECs) and cancer-associated fibroblasts (CAFs) showed upregulation of genes and cellular processes involved in lipid or fatty acid processes (figure 1E), suggesting common adaptation to the steatotic microenvironment. For instance, Tregs showed upregulation of genes involved in lipid metabolism including TXNIP, FABP1 and members of the apolipoprotein family such as APOA2, APOC1 and APOA1 (figure 1F). This presumably provides Tregs with survival and functional advantages within the steatotic microenvironment. Conversely, genes associated with antigen presentation (HLA-DRB5 and HLA-DPA1), lymphocyte activation (CREM, IL2RA and ICOS) and immunoglobulin genes (IGKC, IGHG3 and IGLC2) were downregulated in Treg from SLD-HCC compared with non-SLD-HCC (figure 1F). In contrast, CD8⁺ T cells in SLD-HCC downregulated genes involved in both lipid metabolism (FABP5, APOC3 and APOC1) and cytotoxic immune function or antigen presentation such as CST7, CCL5 and HLA-DPB1 (figure 1F).

Similar to Tregs, CAFs in SLD-HCC upregulated pathways related to lipid metabolism (APOA1, FABP1 and APOH) (figure 1F), suggesting metabolic reprogramming within the steatotic microenvironment. These fibroblasts also upregulated genes involved in fibrogenesis such as COL4A2, COL4A1, FGB and FGG and downregulated heat-shock protein genes (HSPs) that are associated with the regulation of fibrosis including HSPA1A, HSPA1B and HSPA6¹² (figure 1F). Likewise, the SLD-HCC-associated ECs also upregulated lipid metabolism genes (FABP1, APOA1 and APOC1) and CD36 (figure 1F). These ECs were also enriched in CXCR4 and downregulated HSPs (HSPB1, HSPA6 and HSPA1A) (figure 1F), similar to the SLD-HCC-associated CAFs. Taken together, our scRNA-seq data indicate distinct metabolic reprogramming in the SLD-HCC microenvironment, with enriched lipid metabolism pathways in Tregs, fibroblasts and ECs, reflecting their shared adaptation to the steatotic environment.

CyTOF immunoprofiling of SLD-HCCs reveals cold and immunosuppressive TME

Since our transcriptomic analyses indicated that key subsets enriched in SLD-HCCs are involved in immune modification, we further characterised the TILs in SLD-HCC (n=17) and non-SLD-HCC (n=27) using CyTOF. In total, 1 663 336 single cells were analysed using a panel of antibodies targeting immune markers (online supplemental table S4). Unsupervised clustering identified 49 clusters (online supplemental figures S3A–C), classified into major immune lineages including CD4⁺ T cells, CD8⁺ T cells, B cells, natural killer (NK) cells and antigen-presenting cells (figure 2A).

Figure 2. Immune landscapes profiling of SLD-HCC and non-SLD-HCC. (A) t-SNE plot representing unsupervised clustering of 1 663 336 single immune cells into seven major immune lineages. (B) t-SNE plots showing the density of immune cells. (C) Bar graphs showing the proportions of each major immune lineage. (D) t-SNE plots showing 13 immune subtypes enriched in SLD-HCCs or non-SLD HCCs (left, 50 000 representative cells shown per plot). Heatmap showing the relative expression levels of immune marker proteins in each immune subtype (right). (E) Bar graphs showing the frequencies of each immune subtype within total CD45⁺ cells. P values calculated by two-tailed Mann-Whitney test. (F) Representative mIF staining of CD8⁺ T cells, exhausted CD8⁺ T cells, CD4⁺ T cells, memory CD4⁺ cells and Tregs in SLD-HCCs and non-SLD-HCCs from tissue microarrays. Scale bar denotes 40 µm. (G) Graphs showing mean±SEM of density of total CD8, total CD4, memory CD4 and CD4⁺ FOXP3⁺ Tregs in SLD-HCCs versus non-SLD-HCCs. Cell density reported as count/core, each core=0.785 mm². (H) KEGG_Wnt gene signalling scoring based on bulk RNA-seq data from tumour sectors from SLD-HCCs (n=40) and non-SLD-HCCs (n=91). (I) Neoantigen scores based on whole genome sequencing (WGS) from n=33 versus n=90 tumour sectors from SLD-HCCs and non-SLD-HCCs, respectively. (A–E) Data from tumour and non-tumour tissues of SLD-HCC n=17; non-SLD-HCC n=27. (F and G) Data from independent cohort of 103 patients (SLD-HCC n=41; non-SLD-HCC n=62). (E, H and I) Boxplots show median and the whiskers represent minimum and maximum values with the box edges showing the first and third quartiles. (G–I) P values were calculated by two-tailed Student’s t-test. APC, antigen-presenting cell; mIF, multiplex immunofluorescence; NK, natural killer; PD1, programmed cell death protein 1; PDL1, programmed cell death ligand 1; SLD-HCC, steatotic liver disease-related hepatocellular carcinoma; TNFα, tumour necrosis factor alpha; Treg, regulatory T cell; t-SNE, t-distributed stochastic neighbour embedding.

In general, we observed a higher density of intratumoural CD4⁺ T cells and depletion of CD8⁺ T cells in SLD-HCCs compared with non-SLD-HCCs (figure 2B, C, online supplemental figures S3D, F). Several previously reported immunosuppressive CD4⁺ T cells associated with poorer prognosis, including CD25⁺ FOXP3⁺ Treg and CD4⁺ memory T cells expressing markers such as PD-L1, GITR and PD-1,^{13 14} were significantly enriched in SLD-HCCs (figure 2D,E). We also observed a depletion of exhausted and granzyme B⁺ active CD8⁺ T cells as well as CXCR3⁺ NK cells in SLD-HCCs (figure 2D,E). Collectively, the data indicate an immunosuppressive and ‘cold’ (non-T cell inflamed) TME in SLD-HCCs, corroborating our scRNA-seq findings. Notably, we observed a similar immune phenotype, characterised by increased Tregs and reduced total CD8 T cells, in a sub-analysis restricted to non-viral HCC cases (online supplemental figure S3E,F), indicating that these features are associated with SLD-HCC independent of HBV status. To determine whether this immune landscape was established in the early stages of SLD, we examined the T cell proportions in scRNA-seq data obtained from patients with healthy liver, early stages of MASLD or late stages of MASH (GSE159977). Indeed, we observed a trend of increased Tregs in MASLD and depletion of CD8⁺ T cells in MASH compared with healthy livers (online supplemental figure S4A,B), suggesting that an immunosuppressive microenvironment could emerge from the early phases of SLD.

Next, we performed mIF (online supplemental table S5) on FFPE tissue comprising tumour cores from an independent cohort of 103 patients with SLD-HCC (n=41) or non-SLD-HCC (n=62) (online supplemental table S2). Consistently, we observed significant enrichment of total CD4⁺ T cells, memory CD4⁺ T cells and Tregs, and a reduction in CD8⁺ T cells in SLD-HCCs (figure 2F,G). The data were consistent in non-viral HCC cases only (online supplemental figure S3G), further supporting a predominant role of steatosis. As immune exclusion, marked by depleted intratumoural cytotoxic CD8⁺ T cells, has been linked to the Wnt signalling pathway,¹⁵ we examined KEGG_Wnt signalling scoring in our bulk RNA-seq data (online supplemental table S1). Indeed, SLD-HCCs showed higher scores for the Wnt signalling pathway, consistent with immune-excluded tumours (figure 2H). SLD-HCCs also displayed a lower neoantigen score (figure 2I), in agreement with a previous report where the number of predicted MHC class I-associated neoantigens correlated with T cell cytolytic activity.¹⁶

Overall, SLD-HCCs are characterised by an immune-excluded microenvironment with depletion of cytotoxic CD8⁺ T cells and enrichment of immunosuppressive memory CD4⁺ T cells and Tregs.

Spatial transcriptomic analysis identifies the unique spatial architecture of SLD-HCC

To unravel the spatial organisation of key SLD-HCC-associated cell populations as revealed by scRNA-seq and CyTOF analyses above, we performed STs on FFPE tissue samples from seven SLD-HCCs and five non-SLD-HCCs using Visium spatial gene expression assay (10x Genomics) (figure 3A; online supplemental table S1). Visium is a spot-based transcriptomics assay where each 55 µm spot captures transcripts from multiple cells. In total, we analysed 126 099 spots across these 12 tissue sections with a median depth of 9698 UMI/spot and 3708 genes/spot (online supplemental figure S5A). Normalised data across all tissue sections showed homogenous distribution of spots across all patient samples in the uniform manifold approximation and projection (figure 3B, online supplemental figure S5B). To estimate the cellular composition in Visium data, we scored each spot based on the top 100 genes from all cell types identified in our scRNA-seq analysis (figures1C 3A). Spots with higher scores for immune cells also exhibited higher scores for fibroblasts and ECs, suggesting that they were in close proximity within the tissue (figure 3C, online supplemental figure S5C). In contrast, hepatocyte-associated spots showed lower scores for immune cells, fibroblasts and ECs (figure 3C, online supplemental figure S5C).

Figure 3. Spatial architecture of SLD-HCC and non-SLD-HCC. (A) Data analysis workflow for Visium spatial transcriptomics (ST). (B) UMAP plot illustrating the distribution of ST data after Harmony integration. n=126 099 spots from n=7 SLD-HCCs and n=5 non-SLD-HCCs. (C) UMAP plots showing gene signature scorings for major immune and non-immune subsets from all Visium data. (D) Representative H&E images on FFPE tissues comprising tumours, non-tumours and margin areas from patients with SLD-HCC and non-SLD-HCC (left); spatial deconvolution using SpaCET, indicating tumour and non-tumour regions (right). Box, field of view (FOV)=6.5 x 6.5 mm. (E) Spatially aware clustering using Banksy, showing the different domains within each tissue sample. (F) Heatmap showing relative gene signature scoring used to estimate the cellular composition within each domain. DC, dendritic cell; FFPE, formalin-fixed, paraffin-embedded; NK, natural killer; scRNA seq, single-cell RNA sequencing; SLD-HCC, steatotic liver disease-related hepatocellular carcinoma; Treg, regulatory T cell; UMAP, uniform manifold approximation and projection.

To understand the cellular interactions within their spatial context, individual tissue samples were analysed using SpaCET¹⁷ and Banksy.¹⁸ SpaCET leverages the gene signatures from The Cancer Genome Atlas (TCGA) data set to map the location of tumour, which is consistent with the histological annotation based on H&E images (figure 3D). Separately, we used the spatially-aware clustering algorithm Banksy to segment each tissue into different domains based on similarities in gene expression (figure 3E, online supplemental figure S5D). We annotated each Banksy domain and observed a higher score for immune cells in domains where the fibroblast score was higher, particularly at the domains lining the tumour margin of SLD-HCCs (figure 3F). The SpaCET deconvolution method based on the TCGA data set consistently showed colocalisation of fibroblast with immune cells at the tumour margin of SLD-HCCs (online supplemental figure S5E), whereas non-SLD-HCCs exhibited a distinct spatial organisation, with domains with high scores for fibroblasts and immune cells dispersed throughout the tissue (figure 3D,F, online supplemental figure S5D,E).

Taken together, our Visium ST data showed potential interactions between immune and fibroblast populations, particularly at the margin of SLD-HCCs.

Cell–cell interaction analysis reveals Treg–fibroblast interaction at the SLD-HCC tumour margin

Visium ST analysis identified potential interactions between cell populations, but its limited resolution (55 µm, spot-based) necessitated a more refined assay to specify the immune subset(s) interacting with fibroblasts. To address this, we first performed cell–cell interaction analysis on our scRNA-seq data, which demonstrated a relative increase in interaction strengths in Tregs and CD4⁺ T cells in SLD-HCC (online supplemental figure S6A). In contrast, the relative interaction strength was reduced in plasmacytoid dendritic cell (pDC), macrophages, Kupffer cells and CD8⁺ T cells within SLD-HCCs (online supplemental figure S6A).

To further validate these cellular interactions, we performed CosMx Spatial Molecular Imager,¹⁰ which allows ST interrogation at a single-cell resolution on FFPE tissues from two SLD-HCCs and two non-SLD-HCCs. Tissue sections were stained for CD45 (immune cells), CD3 (T cells), pan-cytokeratin (epithelial cells) and DAPI (nuclei) (figure 4A) and 50 FOVs that encompassed tumour core, tumour margin and non-tumour regions were selected across the four tissues (online supplemental figure S7A). Cell segmentation identified 152 214 cells across all FOVs, and Leiden clustering classified these into 25 cell types, annotated based on their top expressed genes (figure 4A–C; online supplemental table S6). Overall, 52 118 623 transcripts were detected from 993 genes, with a mean of 270 transcripts per cell.

Figure 4. Cellular interaction network in SLD-HCC and non-SLD-HCC. (A) Representative immunofluorescence (IF) image matched with ST data generated from CosMx unsupervised clustering. Each cluster is colour-coded. PanCK, pan-cytokeratin. Each FOV=0.7 x 0.9mm. (B) UMAP plot illustrating all clusters from CosMx with n=152 214 total cells, n=50 FOVs from two SLD-HCCs and two non-SLD-HCCs. (C) Heatmap showing relative expression levels of selected genes representing each CosMx cluster. (D) Bar graphs showing the absolute number (top) and proportion within total cells (bottom) of each identified cell type across the two SLD-HCC and two non-SLD-HCC tissue samples. (E) Neighbourhood enrichment scores showing interaction strength between Tregs and other cell types from CosMx ST data. Two-sided p values calculated by pairwise Mann-Whitney test. (F) Neighbourhood enrichment scores showing interaction strength between Tregs and fibroblasts at margin domains from deconvoluted Visium ST data (n=7 SLD-HCCs and n=5 non-SLD-HCCs). Two-sided p values calculated by pairwise Mann-Whitney test. (G) Representative IF images of margin areas from SLD-HCC and non-SLD-HCC tissues stained for CD4, FoxP3 (Treg) and αSMA (fibroblast). DAPI was used for nuclear staining. Scale bar denotes 20 µm. (H) Comparison of mean number of Tregs between SLD-HCC and non-SLD-HCC, quantified from three to five randomly selected FOVs per tissue at tumour margin. (I) The distance between Treg and nearest fibroblast in the same FOVs from (H) was compared between SLD-HCCs and non-SLD-HCCs. (F, H and I) Boxplots show median and the whiskers represent minimum and maximum values with the box edges showing the first and third quartiles. (H and I) mIF data was obtained from six SLD-HCCs and six non-SLD-HCCs, and analysis was performed using Mann-Whitney U test. Graphs show mean±SEM. CAFs, cancer-associated fibroblasts; DAPI, 4',6-diamidino-2-phenylindole; DC, dendritic cell; EC, endothelial cell; FOVs, fields of views; LSEC, Liver sinusoidal endothelial cells; mIF, multiplex immunofluorescence; NK, natural killer; pDC, plasmacytoid dendritic cell; SLD-HCC, steatotic liver disease-related hepatocellular carcinoma; ST, spatial transcriptomic; Treg, regulatory T cell; UMAP, uniform manifold approximation and projection.

Cell populations including CD4⁺ and CD8A⁺ T cells, myeloid and DC populations (CD74, CIITA and HLA-DPA1), plasma cells (IGHG1, IGHG2 and JCHAIN) and fibroblasts (COL1A1, COL1A2 and COL5A1) were identified (figure 4C, online supplemental figure S8A and table S6), consistent with our scRNA-seq clusters. Hepatocytes were found to be the most abundant cell type (figure 4D). The Treg population expressed higher levels of ICOS and PDCD1 (figure 4C; online supplemental table S6), both of which are linked to a highly immunosuppressive Treg phenotype.^{19 20} We also observed PD-1⁺ exhausted CD8⁺ T cells and CD163^hi tumour-associated immunosuppressive macrophages (figure 4C; online supplemental table S6), consistent with a general immunosuppressive TME of HCC tumours. Of particular interest, we observed two distinct fibroblast populations; one was enriched with collagen genes, and the other, characterised by the expression of proinflammatory genes CCL2, CXCL12 and IL1R1 (figure 4C; online supplemental table S6). The latter resembles the activated fibroblast, also known as CAF, previously implicated in cancer progression and therapy.²¹

Next, we employed neighbourhood enrichment analysis on CosMx data to validate the potential cell–cell interactions identified in CellChat and Visium analyses. We aimed to identify key immune cells potentially interacting with this activated fibroblast (CAFs) and found Tregs as the only population showing significant interaction (online supplemental figure S9A). Conversely, focusing on Tregs, we similarly observed significant increased interaction with CAFs in SLD-HCCs compared with non-SLD-HCCs, along with a trend toward increased interactions with naïve CD8⁺ T cells (figure 4E, online supplemental figure S9B).

To validate this interaction on the Visium data, we used cell type deconvolution based on scRNA-seq data and consistently found a significant correlation between Tregs and CAFs at the tumour margins in SLD-HCCs versus non-SLD-HCCs (figure 4F, online supplemental figure S9C). For further validation, we conducted mIF of Tregs and fibroblasts in six SLD-HCC and six non-SLD-HCC FFPE samples (figure 4G). Focusing on FOVs near the tumour margin, we found enriched Tregs (figure 4H) and a shorter average distance between Tregs and fibroblasts (figure 4I), demonstrating the potential Treg-CAF interactions at the tumour leading edge in SLD-HCCs.

TNFSF14-TNFRSF14 mediated Treg–fibroblast interaction is associated with immunotherapy resistance in SLD-HCC

Next, we assessed the ligand–receptor (L–R) pairs responsible for Treg–fibroblast interaction by performing COMMOT (COMMunication analysis by Optimal Transport) analysis, which evaluates L–R pairs while accounting for spatial distances and signalling directionality.²² From the CosMx data, we identified a total of 11 common signalling pathways enriched in SLD-HCC for Treg–fibroblast and fibroblast–Treg interactions (figure 5A). Separately, we employed NICHES (Niche Interactions and Communication Heterogeneity in Extracellular Signaling)²³ analysis on Visium data, which identified 12 L–R pairs that were exclusively enriched in the margins of SLD-HCC tumours (figure 5B, online supplemental figure S10A). Comparing these two analyses, we found three common L–R pairs (online supplemental figure S10B), with only tumour necrosis factor superfamily member 14 (TNFSF14)-tumour necrosis factor receptor superfamily member 14 (TNFRSF14) showing a significantly stronger COMMOT score comparing SLD-HCCs to non-SLD-HCCs (online supplemental figure S10C). Visualisation of TNFSF14-TNFRSF14 expression on Visium data showed a spatial enrichment along the tumour leading edge in SLD-HCCs but further away from the tumour border in non-SLD-HCCs (figure 5C, online supplemental figure S10D).

Figure 5. Ligand–receptor interactomes in SLD-HCC. (A) Heatmap showing relative COMMOT scores of enriched L–R pathways from CosMx data. (B) Heatmap showing relative expression levels of L–R pairs, determined by NICHES analysis. A representative Visium map highlighting the tumour margin domains was shown (upper right). Specific enriched L–R pairs from clusters enriched at the tumour margin domains were shown (boxed, bottom right). (C) Representative images from Visium data showing tumour fraction scoring, tissue segmentation into tumour, margin and non-tumour regions as well as relative expression of TNFSF14-TNFRSF14 in SLD-HCCs versus non-SLD-HCCs. (D) COMMOT analysis on Visium data showing distinct TNFSF14-TNFRSF14 strength and directionality in SLD-HCCs versus non-SLD-HCCs. Gene expression intensity is marked by size and directionality by the pointed end of the arrows. Tumour (T) and non-tumour (NT) regions are separated by dashed red lines. (E) Representative IF images showing expression of TNFSF14-TNFRSF14 at tumour margins in SLD-HCCs and non-SLD-HCCs. Scale bar denotes 100 µm. (F) COMMOT scores comparing the strength of the TNFSF14-TNFRSF14 interaction between SLD-HCCs (n=7) versus non-SLD-HCCs (n=5) on Visium data. Treg–CAFs and CAFs–Treg interactions from margin domains were analysed. (G) COMMOT scores comparing the strength of the TNFSF14-TNFRSF14 interaction between three responders and five non-responders to immunotherapy in SLD-HCC (n=5) versus non-SLD-HCC (n=3). Treg–CAFs and CAFs–Treg interactions from margin domains were analysed. (B–D) Visium FOV=6.5 x 6.5 mm. (F and G) Boxplots show median and the whiskers represent minimum and maximum values with the box edges showing the first and third quartiles. P value determined by two-tailed Mann-Whitney test. CAFs, cancer-associated fibroblasts; COMMOT, COMMunication analysis by Optimal Transport; DAPI, 4',6-diamidino-2-phenylindole;FOV, fields of view; IF, immunofluorescence; L–R, ligand–receptor; NA, not applicable; NICHES, Niche Interactions and Communication Heterogeneity in Extracellular Signaling; SLD-HCC, steatotic liver disease-related hepatocellular carcinoma; TNFSF14, tumour necrosis factor superfamily member 14; TNFRSF14, tumour necrosis factor receptor superfamily member 14; Treg, regulatory T cell.

We next performed COMMOT on the deconvoluted Visium data and consistently observed higher scores for TNFSF14-TNFRSF14 interaction specifically between Tregs and fibroblasts in SLD-HCCs (figure 5D). Furthermore, COMMOT analysis for TNFSF14-TNFRSF14 showed a directionality from tumour towards non-tumour areas, particularly at the tumour margin, in the SLD-HCC. In contrast, random directionality was observed in the non-SLD-HCC tumours (figure 5E, online supplemental figure S10E). mIF analysis further validated this finding, showing enriched expression of TNFSF14-TNFRSF14 at the margin in SLD-HCCs (figure 5F).

To investigate the clinical significance of this interaction, we compared COMMOT scores for TNFSF14-TNFRSF14 between Tregs and fibroblasts in patients with SLD-HCC, who showed different responses to immunotherapy (both anti-PD-1 monotherapy and anti-PD-L1+ anti-VEGFA combination therapy; online supplemental table S1). We observed significantly higher TNFSF14-TNFRSF14 COMMOT score in non-responders compared with responders in SLD-HCCs, but this difference was not observed in the patients with non-SLD-HCCs (figure 5G and 10F).

Blocking TNFSF14-TNFRSF14 signaling enhance response to immunotherapy in SLD-HCC

To further validate the role of TNFRSF14 in immunotherapy resistance, we established a murine model by feeding a HFD for 7 weeks to induce a fatty liver microenvironment, followed by orthotopic transplantation of murine hepatoma cells¹¹ (figure 6A). Histological analyses confirmed that this dietary regimen successfully induced hepatic steatosis in these mice (online supplemental figure S11A). Following transplantation, tumours were allowed to establish for 5 days followed by biweekly treatment for 2 weeks with isotype control, anti-PD-1, anti-TNFRSF14 or a combination of anti-PD-1 and anti-TNFRSF14 antibodies (figure 6A). Combination treatment significantly reduced tumour burden, with two of the five treated mice exhibiting complete tumour regression (figure 6B,C).

Figure 6. Blocking Treg–CAFs interaction enhances antitumour immunity and response to ICB. (A) Mice were fed with HFD for 7 weeks and established HCC model by orthotopically injecting Hepa1–6 cells into the liver. The mice were randomly assigned to four treatment groups and treated biweekly for 2 weeks from day 5 to 19 (n=4–5 mice per group). (B) The tumour volume ((length × width²)/2 in mm³) of each mouse was measured. Two-tailed p values were calculated from one-way ANOVA test with Tukey’s multiple comparisons test. (C) Representative images of livers were shown. (D) Proportion of T cell populations (%) in tumour tissues from mice with respective treatment conditions. Two-tailed p values calculated from unpaired Student’s t-test. (E) Dot plot of enriched Gene Ontology (GO) pathways from differentially expressed genes (DEGs) in tumour tissues of mice under respective treatment conditions compared with control. Dot colour represents the log-transformed adjusted p value, while dot size indicates fold enrichment. (F) Tumour volume ((length×width²)/2 in mm³) of mice on normal diet was measured. Two-tailed p values calculated from one-way ANOVA test with Tukey’s multiple comparisons test. (A–E) n=4 mice for isotype antibodies control and anti-PD-1 treated group; n=5 mice for anti-TNFRSF14 and anti-PD-1+ anti-TNFRSF14 combination treatment group. (B, D and F) Graphs show mean±SEM. ANOVA, analysis of variance; CAFs, cancer-associated fibroblasts; HCC, hepatocellular carcinoma; HFD, high-fat diet; ICB, immune checkpoint blockade; NK, natural killer; PD-1, programmed cell death protein 1; TNFRSF14, tumour necrosis factor receptor superfamily member 14; Treg, regulatory T cell.

Importantly, combination therapy led to a robust intratumoural immune response characterised by a reduced proportion of intratumoural Tregs and increased levels of active CD8⁺ T cells and active memory CD4⁺ T cells (figure 6D, online supplemental figure S11B). This effect was more pronounced within the tumour, while non-tumour liver tissues showed only a trend toward increased active memory CD4⁺ T cells (online supplemental figure S11C). As expected, PD-1 blockade alone also showed elevated active CD8⁺ T cells, but combination therapy elicited a markedly stronger immune response, correlating with its superior tumour control compared with the other treatment groups (figure 6D). To further elucidate the underlying mechanisms, we performed bulk RNA sequencing of tumour tissues from treated mice. Anti-TNFRSF14, alone or in combination with anti-PD-1, enriched immune activation pathways, whereas anti-PD-1 monotherapy was associated with signatures of lipid metabolism and collagen biosynthesis, including interleukin-13 signalling, a known driver of fibrosis²⁴ (figure 6E and online supplemental figure S11D). These findings suggest that TNFRSF14 blockade enhances antitumour immunity and alleviates liver pathology. Consistently, histological analysis showed reduced steatosis in the anti-TNFRSF14 and combination groups (online supplemental figure S11A). Spatial proximity analyses confirmed that anti-TNFRSF14 treatment disrupted Treg–fibroblast interactions, as evidenced by increased distance between them following monotherapy or combination treatment (online supplemental figure S11E), supporting the functional relevance of TNFSF14-TNFRSF14 signalling in mediating Treg–CAF crosstalk. Furthermore, this therapeutic effect was specific to the SLD-HCC, as mice without prior HFD feeding did not exhibit enhanced response to anti-PD-1 treatment (figure 6F and online supplemental figure S11F).

These findings suggest that targeting TNFSF14-TNFRSF14 signalling enhances T cell-mediated immunity, remodels the steatotic-fibrotic tumour milieu and potentially overcome immunotherapy resistance in SLD-HCC.

Discussion

Our current study employed a multiomics approach—including CyTOF, single-cell and STs—to delineate the TME of SLD-HCC. Single-cell transcriptomic analyses revealed common lipid metabolic reprogramming in Tregs, ECs and CAFs, reflecting their adaptation to the steatotic milieu. CyTOF analysis reflected a cold, immunosuppressive TME, characterised by depleted CD8+ T cells and enrichment of CD4+ memory T cells and Tregs. Notably, many patients in our cohort had chronic hepatitis B. While our subanalyses support that the immunosuppressive features are driven predominantly by steatosis and not viral status (online supplemental figures S3E,G), we acknowledge that the independent contribution of HBV requires validation in larger external cohorts. Furthermore, the key mechanisms were validated in an HFD-induced murine HCC model, rather than in a normal-diet control model. This study also acknowledges the limitations of retrospective alcohol consumption data, as well as the potential influence of coexisting aetiologies and interpatient heterogeneity within the SLD-HCC group. While a recent study investigating the TME in MASH-HCC identified TAMs and MDSCs as key mediators of immunosuppression,²⁵ our current study further unravels the intricate spatial architecture of SLD-HCCs and revealed Treg–CAF interactions via TNFSF14/TNFRSF14, as associated with poor response to immunotherapy and potential therapeutic targets in SLD-HCC.

Tregs are known to maintain immunosuppression in the TME and promote tumour progression in several cancers, including HCC.¹⁴ However, their role in MASH remains controversial. One study found adoptive Treg transfer, instead of suppressing inflammation, paradoxically worsened steatosis in a MASH mouse model,²⁶ while another implicated Tregs and neutrophils in MASH-driven carcinogenesis,²⁷ suggesting their complex role in MASH-HCC progression. Conversely, fibroblasts play a central role in orchestrating liver fibrosis, a hallmark of advanced MASH. CAFs are involved in tumour progression through complex crosstalk with other cell types in the TME, and targeting CAFs to augment immunotherapy has shown promise in preclinical models and clinical trials.²⁸ CAFs were also shown to recruit and retain CD4+ CD25+ T cells, promoting Treg differentiation via CXCL12 in breast cancer.²⁹ However, little is known about the potential interactions between Tregs and CAFs in the context of MASH or MASH-driven HCC. Here, we observed close Treg–CAF interactions in SLD-HCC, marked by shared lipid metabolic reprogramming, suggesting a mutual adaptation to the steatotic microenvironment and their role in driving the disease phenotype.

TNFSF14 (also known as LIGHT) and its receptor TNFRSF14 (also known as Herpes virus entry mediator, HVEM), play a crucial role in immunomodulation and inflammation in various diseases, including MASH.³⁰ HVEM expression in melanoma has been shown to mediate functional inhibition of tumour antigen-specific CD8+ T cells via BTLA, highlighting it as a potential checkpoint target.³¹ Furthermore, blockade of the HVEM/BTLA pathway in murine cancer models has been demonstrated to enhance antitumour immune responses and inhibit tumour growth.³² Here, we show the enhanced interaction between Tregs and CAFs is driven by TNFSF14-TNFRSF14 and is linked to a poorer response to immunotherapy in SLD-HCC. Moreover, blocking TNFRSF14 in the HCC murine model established within a fatty liver microenvironment significantly enhanced the intratumoural immune response and improved antitumour effects with anti-PD-1 blockade. These findings underscore that immunotherapy efficacy is potentially dependent on the unique spatial architecture and interactome within the TME of SLD-HCCs.

Overall, our comprehensive multiomics profiling reveals key immune-stromal interactions driving immune evasion and immunotherapy resistance in SLD-HCC. These findings deepen our understanding of SLD-HCC biology and lay the groundwork for future precision therapies.

Data availability statement

Data are available in a public, open access repository. All data relevant to the study are included in the article or uploaded as supplementary information.