Multi-omics-driven personalized management of advanced HCC

Summary

Hepatocellular carcinoma (HCC) management is challenging due to its complex tumor microenvironment and poor treatment responses. Here, using tumor specimens from a prospective clinical trial of combined transarterial chemoembolization (TACE) with immune checkpoint blockade (ICB), we perform exhaustive multi-omics analysis including spatial proteomics and transcriptomics, single-cell RNA sequencing, and bulk transcriptomics. These analyses reveal that treatment response is associated with enrichment of anti-tumor T cell regions that are regulated by cGAS-STING activation within immune-suppressive epithelial cells. Conversely, fibrotic processes impede these pro-response processes. Based on these insights, we test triple combination therapy consisting of cGAS activation, immune checkpoint blockade, and anti-fibrosis strategies, which shows improved efficacy over dual therapy. To identify patients who would benefit, we construct a predictive model using a group sparse learning algorithm. Our findings provide a blueprint for crafting personalized HCC therapies using next-generation biomarkers.

Graphical abstract

Highlights

• •Multi-omics profiling maps HCC responses after TACE and immune checkpoint blockade
• •cGAS-STING activation in epithelial cells drives T cell recruitment after TACE
• •Fibrotic stroma suppresses cGAS-STING signaling and T cell chemotaxis
• •A restricted-marker GCSS model predicts response to TACE plus ICB

Sheng et al. integrate spatial multi-omics, functional models, and clinical validation to show that fibrotic stroma suppresses epithelial cGAS-STING signaling and T cell recruitment in HCC. Their GCSS model links tumor architecture to response prediction for TACE plus immune checkpoint blockade.

Introduction

Hepatocellular carcinoma (HCC) accounts for nearly 90% of primary liver cancer cases.¹ Early detection of HCC is challenging and treatment options for inoperable patients are often ineffective, making HCC the third leading cause of cancer-related mortality in the world.² Nevertheless, immunotherapy has improved the treatment landscape of HCC. Immune checkpoint blockade (ICB) therapy is a pillar of modern immunotherapy that functions to relieve the immunosuppression of cytotoxic T lymphocytes (CTLs) by blocking immune checkpoint molecules such as programmed cell death protein 1 (PD-1). This enables the killing capacity of CTLs against tumors. In advanced HCC, combining ICB with anti-angiogenics such as bevacizumab or vascular endothelial growth factor (VEGF) receptor (VEGFR) inhibitors modifies tumor vasculature to bolster immune response, resulting in improved overall survival.^3,4,5,6

In addition to systemic therapy, transarterial chemoembolization (TACE) is often used to improve loco-regional control for intermediate-stage HCC with large or multinodular tumors without vascular invasion or extrahepatic spread.^7,8 TACE reduces tumor burden and can improve outcomes for these patients.⁹ One potential synergistic advantage of combining TACE with immunotherapy is that it modifies the immune microenvironment of HCC, including recruitment of antigen-specific T cell and natural killer (NK) cells. This has motivated the exploration of regimens that combine TACE with systemic agents, such as the TACTICS trial of TACE plus sorafenib,¹⁰ and with immune checkpoint inhibitors,^11,12,13 and recently, the EMERALD-1 trial, which evaluated a combination of TACE + durvalumab + VEGF inhibitor in unresectable HCC, showed an improvement in progression-free survival with the addition of ICB + VEGF inhibitor.¹⁴ Despite these advances, many patients still progress after therapy.¹⁵ We hypothesized that features of the tumor microenvironment (TME) may identify patients who respond well to dual therapies and may inform strategies to address non-responders.

To discover such features, we conducted a clinical trial (NCT04174781) of drug-eluting bead TACE (DEB-TACE) and sintilimab (PD-1 antibody) in patients with HCC of Barcelona Clinic Liver Cancer (BCLC) stage A who exceeded the Milan criteria or BCLC stage B. We focused on extensive multi-omics analysis of tumor specimens before and after therapy. Using this approach, we were able to uncover regional response determinants, elucidate distinct mechanisms underlying regional responses and non-responses, optimize clinical trial strategy, and establish a predictive model to guide personalized treatment for HCC.

Results

Trial of TACE + ICB for biomarker discovery

To evaluate the biological determinants of HCC responses following TACE with PD-1 antibody treatment, we conducted a clinical trial of DEB-TACE and sintilimab in patients with locally advanced HCC. Patients with BCLC stage A HCC who exceeded the Milan criteria and those with BCLC stage B HCC were included (Figure 1A; Table S1-1, NCT04174781). All patients had Child-Pugh A liver function and did not have vascular invasion, extrahepatic metastasis, or clinically significant portal hypertension. Sixty-one patients enrolled, all of whom underwent at least one treatment cycle combining DEB-TACE and sintilimab (Figure 1A). The clinical characteristics of the trial patients are summarized in Table S1-1. The majority of the patients were male (52 of 61, representing 85%), with a median age of 58 years (ranging from 26 to 75 years old). Of these patients, 85% (52 individuals) had two or fewer lesions, while 56% (34 individuals) were diagnosed with BCLC stage B HCC. The median size of the target tumor was 7 cm (ranging from 2.1 to 14.1 cm).

Figure 1.

Design and evaluation of spatial multi-omics and HCC clinical trial

(A) Clinical trial design. Enrolled patients received sintilimab at a dose of 200 mg through intravenous infusion at the beginning of the DEB-TACE procedure on day 1 of the first treatment cycle and every 3 weeks thereafter. DEB-TACE procedures were repeated every 4–6 weeks based on tumor response, and sintilimab treatment was continued for a maximum of 3 cycles or until surgical resection, radiologic disease progression, unacceptable toxicity, or withdrawal from the study.

(B) Clinical characteristics of patients enrolled, including treatment response, TNM grading, HBV antigen, and liver function indicators.

(C) Patient enrollment encompassed 60 patients who exceeded the Milan criteria for the clinical trial. However, 10 patients who underwent more than one cycle of combination treatment were excluded, along with three patients who lacked pre-operation biopsy. Additionally, nine patients who lacked post-operation tissues were also excluded from the study, thereby selecting 38 pairs of HCC tissues for IMC analysis.

(D) Representative computed tomography of patients with OR (objective response) and SD (stable disease). The tumor is highlighted in red.

In total, 82% (50 of 61) of the patients underwent just one treatment cycle with sintilimab combined with DEB-TACE, while 18% (11 patients) received two or three treatment cycles. At the data cutoff date, of the 60 patients assessed for effectiveness, 37 (62%) exhibited an objective response (OR) to DEB-TACE combined with sintilimab; 3 patients had a complete response (5%), 34 had a partial response (57%), and 20 experienced stable disease (SD). Overall, the disease control rate was 95% (57/60) with a median follow-up of 26 months. Patients showing complete and partial response were categorized as having an OR, while patients showing SD were classified as such (Figure 1B; Figure S1A). Of the treated patients, adequate biopsy specimens were collected from 38, enabling analysis of biological correlates (Figure 1C). No significant difference was observed between the OR and SD groups for these patients concerning gender (p = 0.35), pathologic grade (p = 0.09), and age (p = 0.21). The representative computed tomographs of the patients with OR and SD are shown in Figure 1D.

Immune-suppressive epithelial cell levels rise in responders

We hypothesized that biological markers known to be active in modulating the HCC TME might be associated with treatment response. We generated a comprehensive spatial multi-omics dataset from patient samples. We excluded patients who underwent multiple rounds of TACE or did not have both pre- and post-treatment samples (Figure 1C). As an initial evaluation, we devised a 40-marker imaging mass cytometry (IMC) panel (39 protein markers and DNA) (Figure 2A) based on a list compiled for the study of cellular neighborhoods (CNs) in HCC.¹⁶ To ensure comprehensive coverage of each slide during IMC scanning, we performed scans across multiple regions (Figure S1B). All antibodies were validated through immunohistochemistry (IHC) before heavy metal conjugation, and each metal-bound antibody was further verified by IMC (Figure S2). The panel included markers for epithelial, endothelial, and stromal cells; various immune cells; as well as the cytokines IL-1β, TNF-α, and IL-6; the proliferation marker Ki-67; and the apoptotic indicator cleaved caspase-3 (Table S2). In total, mass cytometry analysis was done on 459 tumor regions of interest (ROIs) from pre- and post-therapy samples (SD, n = 14; OR, n = 24).

Figure 2.

IMC-derived spatial profiling of the HCC microenvironment before and after combination therapy demonstrates an upregulation of immunosuppressive epithelial cells in responders

(A) Heatmap showing the max-min normalized mean marker expression of 30 clusters.

(B) Segmented images of representative OR and SD regions of interest (ROIs).

(C) t-distributed stochastic neighbor embedding (t-SNE) plots based on the single-cell expression data extracted from IMC images. 18 major clusters of cells from pre- and post-treatment HCC samples were defined according to their markers.

(D) Prevalence of the 18 major clusters as the proportion of total cells in SD-Pre, SD-Post, OR-Pre, and OR-Post, respectively (OR: n = 24 patients, SD: n = 14 patients).

(E) The distribution of key cellular clusters based on IMC results (OR: 24 patients, SD: 14 patients). Statistical analysis was performed with a paired t test (ns: p > 0.05, ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001, ∗∗∗∗p ≤ 0.0001).

Marker expression levels in each cell were quantified and converted into a matrix. Following batch correction, approximately 1,895,221 cells were grouped into 19 cell meta-clusters identified by PhenoGraph (Figures 2A–2C). These meta-clusters were then annotated based on marker expression (Figure 2C; Figures S3A and S3B). Macrophage subsets, including resident macrophages (CD11b^low CD68⁺) and infiltrating macrophages (CD11b⁺ CD68⁺ CD14⁺) were separately identified (Figure 2A). T cell subsets (CD4⁺ or CD8⁺) were further refined based on functional markers (Figure 2A). Additionally, we observed epithelial cells (Pan-Cytokeratin⁺) as well as a subset of epithelial cells co-expressing immune-inhibitory ligands such as PD-L1 and B7-H4 (PD-L1⁺/B7-H4⁺ Pan-Cytokeratin⁺), which we refer to as immune-suppressive epithelial cells (Figure S3C). Fibroblasts were further divided into Collagen I^hi fibroblasts and α-SMA^hi myofibroblasts¹⁷ (Figure 2A). Some clusters contained multiple cell types, such as epithelial cells and T cells, immune-suppressive epithelial cells and resident macrophages, and resident macrophages and fibroblasts, indicating close interactions of neighboring cells. One cluster from lineage-negative cells was also detected (Figure 2A; Figure S3). The frequencies of these cells in evaluated patients are shown in Figure 2D.

We observed a counterintuitive pattern in immune-suppressive epithelial cells, which increased after the treatment only within the OR group (Figure 2E). In contrast, we identified opposing trends in α-SMA^hi myofibroblasts and Collagen I^hi fibroblasts following the combination treatment. The number of α-SMA^hi myofibroblasts decreased, while Collagen I^hi fibroblasts increased post-treatment. The increase in Collagen I^hi fibroblasts was less pronounced in the OR group (Figure 2E). The combination treatment also revealed opposing patterns in Tregs and CD4/8⁺ T cells, as Tregs experienced a decline, while CD4/8⁺ T cells exhibited a rise. In addition, both resident and CD11c⁺ macrophages demonstrated a frequency increase after the combined therapy (Figure 2E).

These observations suggest that therapy initiates an unexpected adaptive resistance mechanism. A decrease in α-SMA^hi myofibroblasts along with an increase in Collagen I^hi fibroblasts could reflect remodeling of the extracellular matrix that may either support or hinder an antitumor immune response, depending on the context. The observed decrease in Tregs alongside an increase in CD4/8⁺ T cells aligns with the intended immunostimulatory effects of PD-1 blockade, which aims to bolster antitumor immunity. Additionally, the rise in both resident and CD11c⁺ macrophages after the combination therapy underscores the dynamic nature of innate immune responses.

Treatment response is marked by organizational changes in the TME

Beyond tissue composition, we hypothesized that cellular organization may also contribute to or predict for HCC responses to combined therapy. We utilized high-resolution spatial data to identify in situ alterations to CN organization after the combined therapy, which we defined as the 10 closest neighbor cells surrounding a central cell utilizing strategies published before^18,19 (Figure 3A). We analyzed the cellular composition of the neighbors and annotated these neighborhoods based on the primary cellular clusters, resulting in various functional CN units (Figure 3B). Additionally, we employed a Voronoi diagram, which was applied to each IMC image using the CN composition, to visualize the correlation between the defined CN functional units and the original IMC images (Figure 3C).²⁰ The IMC topology map, depicted by the Voronoi diagram with different CNs represented by distinct colors (as shown in Figure 3C), aligned well with the IMC image, confirming that CN assignments were correct.

Figure 3.

Therapeutic response characterization: Co-localization of immunosuppressive epithelial cells with T cells in responders versus fibrosis in non-responders

(A) Schematic representation of cellular neighborhoods (CNs), defined as 10 nearest neighbor cells relative to a center cell.

(B) 15 distinct CNs were identified based on the abundances of original cellular clusters within each CN.

(C) IMC images and corresponding Voronoi diagrams showing CNs and defined tissue areas for representative OR and SD patients. Patients with OR showed enrichment of immune-suppressive territories, while patients with SD exhibited fibroblast-dominant zones.

(D) Boxplots showing the relative abundance of selected CNs between patients with OR and SD before and after treatments. Statistical analysis was performed using paired t tests (ns: p > 0.05, ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001, ∗∗∗∗p ≤ 0.0001).

(E) Barrier score analysis quantifying the spatial proximity of immunosuppressive epithelial cells (ISECs) to the nearest Collagen I⁺ fibroblast. Cumulative distribution function (CDF) curves compare patients with OR and SD before and after treatment. Statistical significance was assessed by Mann-Whitney U test (Pre: p = 4.7 × 10⁻⁹⁸; Post: p = 2.5 × 10⁻⁵³). Only PD-L1⁺ ISECs, but not other epithelial subsets, showed significant differences in fibroblast proximity between responders and non-responders.

(F) Heatmap showing average cell-cell interaction counts in fibroblasts/epithelial-positive areas. y axis: central cell; x axis: neighboring cells. Interactions with |log2(FoldChange)| > 1 and p < 0.05 are labeled.

(G) Differential interaction analysis between immune-suppressive epithelial/immune-positive area and other regions.

Furthermore, we adapted previously described methods to identify important regions within the HCC TME.^21,22 An expansive homotypic cellular cluster, as identified by uniform cellular phenotypes, was demarcated as a discrete spatial domain within the TME. Notable regions identified by spatial clustering included the epithelial region, immune-suppressive epithelial region, fibroblast-enriched region (hereafter referred to as fibrotic region), and the contiguous interfaces—specifically, areas where immune-suppressive epithelial cells co-localized with immune cells, as well as regions where epithelial cells and fibroblasts were in direct proximity (Figure 3C). Similar to the bulk frequency analysis, we found that the immune-suppressive epithelial region increased in the OR group after treatment, while the epithelial region decreased in the OR group. The fibroblast region showed an increase in both the SD and OR groups, although the magnitude of increase was less pronounced in the OR group. Additionally, the immune region experienced an increase in the OR group post-treatment. The epithelial/fibroblast intersection region and the immune-suppressive epithelial/immune region displayed contrasting patterns in the OR group after treatment, with the former decreasing and the latter increasing (Figures 3C and 3D).

To quantitatively assess whether fibrotic stroma preferentially encases specific epithelial subpopulations, we computed a “barrier score” defined as the cumulative distribution of distances from each epithelial cell to the nearest Collagen I⁺ fibroblast. Strikingly, only PD-L1⁺ immunosuppressive epithelial cells (ISECs) exhibited significant differences between patients with OR and SD: in pre-treatment samples, ISECs in patients with SD were located significantly closer to Collagen I⁺ fibroblasts than those in patients with OR (Mann-Whitney U test, p = 4.7 × 10⁻⁹⁸), and this pattern persisted after treatment (p = 2.5 × 10⁻⁵³; Figure 3E). In contrast, non-ISEC epithelial cells showed no significant difference in fibroblast proximity between the two groups. These results indicate that the fibrotic barrier selectively encircles PD-L1⁺ ISECs in non-responders, physically restricting immune cell access to the very tumor cells capable of activating cGAS-STING-mediated antitumor signaling.

We performed interaction analysis to uncover potential mechanisms of resistance in the SD group. In that group, we observed a higher number of cellular interactions involving fibroblasts and epithelial/fibroblast interactions. Conversely, cellular interactions involving infiltrating macrophages, NK cells, and various T cells were notably absent in the epi-fibroblast intersection regions (see Figure 3F). On the other hand, in the immune-suppressive-immune intersection regions, we observed higher cellular interactions involving immune-suppressive epithelial cells and various T cells, while fibroblast interactions were comparatively reduced (see Figure 3G).

This analysis suggests that distinct cellular subsets orchestrate environmental factors that give rise to contrasting regional features. Specifically, immune-suppressive epithelial cells are in a position to mediate cellular interactions with immune cells, particularly various T cells, in the OR group. In contrast, the SD group lacked these cellular interactions. These patterns suggest that immune-suppressive epithelial cells recruit T cells, whereas epithelial cells promote fibroblast enrichment, which may reduce the efficacy of therapeutic responses.

Integrated single-cell and spatial transcriptomics reveals cGAS-STING-linked immune activation underlying TACE + ICB response

Our IMC analysis revealed how CNs respond to TACE + ICB. To discover the regulatory biological processes, we performed single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (Figure S4A) on post-treatment samples from 6 patients with OR and 4 patients with SD (92,785 cells total). We identified 10 primary cell populations (Figures 4A and 4B), including EpCAM⁺ epithelial cells, immune-suppressive epithelial cells (PD-L1⁺/B7-H4⁺ Pan-Cytokeratin⁺), VWF⁺ endothelial cells,²³ ACTA2⁺ fibroblasts,²⁴ and immune subsets (CD3⁺ T cells, CD79B⁺ B cells, CD68⁺ macrophages, GNLY⁺ NK cells,²⁵ and CPA3⁺ mast cells²⁶). T cells comprised 5 subsets (Figures 4C and 4D): CD8⁺ effector, CD8⁺ Tcm, CD8⁺ Tem,²⁷ exhausted CD4⁺, and Tregs. Macrophages included C1QA/B⁺ resident and CCR2⁺/S100A8/9⁺ infiltrating subsets²⁸ (Figures 4E and 4F).

Figure 4.

Integration analysis of scRNA-seq, spatial transcriptomics, and bulk RNA-seq

(A) t-distributed stochastic neighbor embedding (t-SNE) plot of scRNA-seq data identifying 10 major cell types in the tumor microenvironment.

(B) t-SNE plot showing key markers associated with each major cell type.

(C and D) t-SNE plots showing T cell sub-clusters (C) and their associated markers (D).

(E and F) t-SNE plots showing macrophage sub-clusters (E) and their associated markers (F).

(G and H) t-SNE plots showing epithelial cell sub-clusters (G) and their associated markers (H).

(I and J) t-SNE plots showing fibroblast sub-clusters (I) and their associated markers (J).

(K and L) Representative spatial transcriptomic spots corresponding to immune-suppressive epithelial & T cells (K) and epithelial & fibroblasts (L) regions.

(M and N) Selected KEGG enrichment results for genes that are significantly up-regulated in the immune-suppressive epithelial & T cells (M) and epithelial & fibroblasts (N) regions.

(O) Differentially expressed genes between the OR and SD groups from bulk RNA-seq data.

(P) Correlation analysis of pathway-level GSVA scores computed from gene signatures (Figure S5A) in immune-suppressive epithelial and T cell spots.

(Q) SCENIC analysis highlighting the key transcription factors that regulate PD-L1⁺ immunosuppressive epithelial cells.

(R–T) Kaplan-Meier survival analyses using publicly available HCC transcriptomic datasets to validate the prognostic relevance of the cGAS-STING and fibrotic gene signatures identified in this study. Patients were stratified into high- and low-expression groups based on the median score of each signature. Overall survival was compared using the log rank test, and hazard ratios (HRs) with 95% confidence intervals are shown. Datasets analyzed include TCGA-LIHC (R), GSE14520 (S), and GSE76427 (T). For each dataset, the upper panel displays the cGAS_STING signature and the lower panel displays the Fibrotic signature.

Similar to the IMC analysis, we discerned two distinct subsets of epithelial cells within the TME: immune-suppressive epithelial cells expressing the pivotal immune checkpoint protein PD-L1 and a separate population of epithelial cells devoid of PD-L1 expression (Figures 4G and 4H). The former subset, characterized by the presence of PD-L1, plays a regulatory role in immune recognition, whereas the latter, which lacks PD-L1, represents a phenotypically different group of epithelial cells within the same tissue context. We classified fibroblasts into two subtypes based on the IMC results: Collagen I^hi fibroblasts and α-SMA^hi myofibroblast cells (see Figures 4I and 4J). Our scRNA-seq landscape encompassed the key cellular components of the HCC TME and are consistent with the IMC results. These findings provide a detailed accounting of the diverse cellular populations in the TME and frame their potential roles in tumor immune responses.

To unravel the mechanisms underlying regional response to TACE and ICB, we performed integrative analysis of scRNA-seq and spatial transcriptomics.²⁹ Using robust cell type decomposition (RCTD), we decomposed spatial transcriptomic spots based on refined cellular clusters from scRNA-seq. After deconvolution, epithelial cells and fibroblast spots were significantly enriched in patients with SD, while immune-suppressive epithelial cells and T cell spots were enriched in patients with OR (Figures S4B–S4E). Notably, double-positive spots indicating co-occurrence of immune-suppressive epithelial cells and T cells were prominent in patients with OR (Figure 4K; Figure S4F), whereas epithelial-fibroblast double-positive spots were enriched in patients with SD (Figure 4L; Figure S4G). These results suggest that immune-suppressive epithelial cells recruit T cells to enhance therapeutic efficacy, while epithelial-fibroblast interactions promote fibrosis and impede treatment response.

To understand the biological determinants of this organization, we conducted Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses on the spots with dual positivity for immune-suppressive epithelial cells and T cells (Figure 4K). In the OR group, immune activation pathways such as chemokine signaling, T cell differentiation/trans-endothelial migration, and NF kappa B were evident. We also observed immune checkpoint signaling following immune activation (see Figure 4M). We also noted activation of cytosolic DNA sensing (cGAS pathway) pathways in the top group (Figure 4M). In contrast, in spots with dual positivity for epithelial cells and fibroblasts (Figure 4L), we observed metabolism-related pathways and oxidative phosphorylation pathways. Particularly noteworthy was the identification of the peroxisome proliferator-activated receptor (PPAR) signaling pathway, which plays a key role in fibrosis,³⁰ in the SD group (see Figure 4N).

To validate integrated single-cell and spatial transcriptomic findings, we performed bulk RNA transcriptomic analysis on post-treatment samples obtained from a separate cohort of 19 patients treated with the same regimen (Table S1-2). Among the top differential genes identified, one was cyclic GMP-AMP synthase (cGAS), a major regulator of double-stranded DNA recognition³¹ (Figure 4O).

To further validate the clinical relevance of these molecular signatures, we performed survival analyses across three independent HCC transcriptomic cohorts. Using the cGAS-STING and fibrotic gene signatures derived from our spatial transcriptomic data, we stratified patients by median signature score and assessed overall survival. In the GSE14520 cohort, high cGAS-STING signature expression was associated with significantly improved overall survival (hazard ratio [HR] = 0.45; 95% confidence interval [CI], 0.29–0.71; p = 0.000354; Figure 4S), whereas high fibrotic signature expression correlated with poor prognosis (HR = 2.27; 95% CI, 1.20–4.28; p = 0.00913; Figure 4S). Similar trends were observed in The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) (Figure 4R) and GSE76427 (Figure 4T) cohorts. These results independently confirm that cGAS-STING pathway activation is associated with favorable outcomes, while fibrotic features are associated with adverse prognosis in HCC.

To determine whether cGAS-STING plays a role in regulating immune suppression in our patient samples, we conducted several analyses. First, we performed correlation analysis for all the pathways identified in Figure 4P, achieved by employing gene signatures specific to each pathway for gene set variation analysis (GSVA). The gene signatures served as input for calculating the GSVA scores. Analysis focused on immune-suppressive epithelial and T cell spots across all pathways outlined in Figure 4L. Subsequently, these GSVA scores were utilized to compute the correlations among the pathways. We noticed that the cGAS-STING pathway correlated positively with the most enriched molecular pathways, including those for T cell recruitment and migration (Figure 4P), with the exception of the ferroptosis pathway (Figure 4P). Second, we used SCENIC analysis (single-cell regulatory network inference and clustering) to identify key transcription factors that form the gene regulatory network within immune-suppressive epithelial cells.³² We detected cGAS-STING-related transcription factors such as IRF3 and NF-κB (Figure 4Q) within immune-suppressive epithelial cells.

Together, these results demonstrate that the spatial co-localization of immune-suppressive epithelial cells and T cells is a defining feature of favorable response to TACE + ICB. This immune-enriched architecture contrasts with the fibrotic epithelial-fibroblast-dominant neighborhoods found in non-responders. Mechanistically, responders exhibited strong enrichment of immune activation pathways—including chemokine signaling, T cell migration, and cytosolic DNA sensing—highlighting the cGAS-STING axis as a key molecular driver. These findings were independently validated in a bulk RNA-seq cohort and further supported by transcriptional regulatory network analysis, positioning cGAS-STING as a central orchestrator of spatial immune reprogramming in the HCC TME.

Fibrotic remodeling mechanically suppresses cGAS-STING signaling to limit T cell recruitment and antitumor immunity

To experimentally validate the key molecular features identified in our integrated single-cell and spatial transcriptomic analyses, we performed western blotting on individual patient tumor samples, which revealed elevated phosphorylation of stimulator of interferon genes (STING) pathway components—including TBK1, IRF3, STING, and RelA—in the OR group but not in the SD group (Figure 5A). These findings suggest that the STING pathway is activated in responsive samples, reinforcing the idea that STING participates in the treatment response. Consistently, multi-color IHC (mIHC) staining also showed more cGAS signal and downstream RelA³³ signal in the OR group, while these signals were absent in the SD group (Figure 5B). Epithelial cells in the OR group displayed high levels of PD-L1 expression (Pan-CK⁺) compared to the absence of PD-L1 expression in epithelial cells in the SD group (Figure 5C). Using mIHC, we observed colocalization of PD-L1 and cGAS expression (Figure 5B). We note that this connection between cGAS-STING pathway and immune checkpoint expression has been observed previously.³⁴

Figure 5.

cGAS-STING pathway is the central regulator for the response group

(A) Representative western blot of p-STING, p-TBK1, p-IRF3, and p-RelA in OR and SD groups showing that cGAS signaling is activated in the OR group, but not in the SD group.

(B) Representative mIHC images add evidence that cGAS is activated in patients with OR. Expression of cGAS, PD-L1, RelA, and Pan-CK, markers of cGAS activity, in cells from tumor tissues are shown for the objective response (OR) and stable disease (SD) groups.

(C) Quantitative assessment of mIHC data showing the frequency of cGAS, PD-L1, dsDNA, Collagen I, and CD45RO in both tumor and fibrotic areas of the OR (n = 24 patients) and SD (n = 14 patients) groups. Each dot represents one high-power field.

(D) Model of how fibrosis and cGAS/dsDNA impacts immune response in the HCC TME.

Scale bars, 30 micrometers.

To further confirm the association between cGAS signaling and PD-L1 expression, we exposed a panel of three human HCC cell lines to the topoisomerase inhibitor epirubicin, which causes double-stranded DNA (dsDNA) breaks. This caused cGAS signaling and upregulation in PD-L1 expression (see Figures S5A and S5B), consistent with a previous report.³⁵ This may partially explain why the single administration of STING agonists has not yielded satisfactory results in clinical trials since upregulation of PD-L1 could theoretically attenuate the anti-tumor immune response.³⁶ Nevertheless, PD-L1 expression does not faithfully predict for immunotherapeutic response in HCC.³⁷

We further analyzed the samples to find actionable hypotheses related to therapeutic vulnerabilities. Spatial transcriptomics revealed distinct features in OR and SD cohorts: regions of immune-suppressive epithelial cells and T cells overlap were pronounced in patients with OR, while epithelial-fibroblast co-localization characterized patients with SD. To elucidate the underlying cellular interactions, we performed ligand-receptor analysis integrating spatial transcriptomics (ST) and scRNA-seq data using NicheNET.³⁸ PD-L1⁺ epithelial cells in patients with OR secreted cytokines and chemokines (CCL5,³⁹ MIF,⁴⁰ and LTB⁴¹) that recruit T cells (Figures S6A and S6C). In patients with SD, epithelial cells transmitted signals to fibroblasts that inhibit PPAR signaling (Figures S6B and S6D), including ATF3,⁴² BCL6,⁴³ LCOR,⁴⁴ EZH2,⁴⁵ NR0B2,⁴⁶ SIRT1,⁴⁷ GPS2,⁴⁸ and NRIP1.⁴⁹ ST analysis confirmed reduced PPAR signaling within the epithelial-fibroblast regions of patients with SD (Figure S6E), with downregulation of PLIN2,⁵⁰ FABP6,⁵¹ PLTP,⁵² UBC,⁵³ and HMGCS1,⁵⁴ indicating suppression of PPAR signaling in non-responders.

To validate NicheNET findings, we evaluated dsDNA levels and cGAS signaling pathway activation using mIHC, finding these to be elevated in the OR group along with PD-L1 expression and increased T cell recruitment (CD45RO) (Figure 5C; Figures S6F–S6G). Given substantial treatment response, we analyzed necrotic regions; since nuclear staining absence renders necrotic areas unsuitable for IMC and cell segmentation (Figure S6H), we used mIHC. Enhanced immune cell infiltration and cGAS activation were observed in perinecrotic regions, but this pattern was attenuated near fibrotic tissue (Figure S6J). T cells also showed resident marker CD103 (Figure S6I). The SD group exhibited high fibrosis with low dsDNA and cGAS signaling and scarce activated T cells (Figure 5C; Figures S6F–S6G), consistent with scRNA-seq and ST pathway analyses linking dsDNA, cGAS signaling, fibrosis, and T cell activation.

To directly assess whether fibrotic stromal components inhibit cGAS-STING activation and downstream T cell chemotaxis in HCC epithelial cells, we established a Transwell co-culture assay (Figure S7A). HepG2 cells were seeded in the lower chamber either alone or together with normal human dermal fibroblasts (NHDF), and Molt-4 T cells were placed in the upper chamber. All groups were stimulated with a cGAS agonist (2′3′-cGAMP, 10 μg/mL), except for a negative control group (HepG2 only, no agonist).

In the HepG2-only + agonist condition, we observed robust activation of the cGAS-STING pathway, with >80% of the epithelial cells staining positive for phospho-STING and cGAS (Figure S7B). In contrast, the HepG2 + fibroblast + agonist group showed markedly reduced activation (∼35% positive cells; p < 0.001, unpaired t test). Chemokine analysis of culture supernatants showed significantly lower CXCL10 levels in the fibroblast co-culture group (HepG2-only + agonist: 650 ± 40 pg/mL vs. HepG2 + fibroblast + agonist: 290 ± 30 pg/mL; Figure S7C), while the levels in the no-agonist group remained low (85 ± 15 pg/mL).

Functionally, T cell migration was substantially impaired in the presence of fibroblasts. Flow cytometry of cells collected from the lower chamber revealed an 85% reduction in Molt-4 migration when fibroblasts were present (HepG2-only + agonist: 10.05% ± 1.30%; HepG2 + fibroblast + agonist: 1.50% ± 0.42%; Figures S7D–S7E). These findings support a model in which fibrotic stroma limits cGAS-STING activation in tumor epithelial cells, leading to impaired chemokine production and reduced T cell recruitment (Figure S7F).

To further explore the mechanism by which fibrotic stroma suppresses cGAS-STING signaling, we cultured HepG2 cells with cGAS agonist in Matrigel substrates with varying stiffness by modulating matrix concentration (Figure S7G). This system mimics the physical conditions of fibrotic (high stiffness) versus normal (low stiffness) liver tissue. We observed that HCC cells grown in higher-concentration Matrigel (Matrigel-to-cell ratio of 3:1) exhibited reduced cGAS-STING activation, as evidenced by the decreased phosphorylation of p65, TBK1, IRF3, and STING compared to cells cultured in low-concentration Matrigel (ratio of 1:1) (Figure S7H). These findings suggest that matrix stiffness may intrinsically suppress DNA sensing pathway activation in tumor cells, providing a mechanical explanation for how fibrosis impairs innate immune signaling in the HCC microenvironment.

In summary, our integrated analyses across patient tissues and functional models uncover a mechanism wherein fibrotic stromal components—and the associated increase in tissue stiffness—mechanically suppress cGAS-STING signaling in tumor epithelial cells. This inhibition diminishes downstream chemokine production and impairs T cell recruitment, thereby restricting antitumor immune activation. These findings were validated in both spatially resolved transcriptomic analyses and functional co-culture systems and further confirmed in Matrigel stiffness assays. Our results reveal that beyond molecular crosstalk, the biophysical properties of the TME actively modulate innate immune sensing pathways, offering insights into how fibrosis may blunt immune-based therapies in HCC (Figure 5D).

Anti-fibrotic therapy enhances immunotherapy by restoring cGAS-STING-mediated immune activation

The observation that fibrosis limited the efficacy of TACE + anti-PD-1 treatment suggests that anti-fibrotic agents might enhance therapy. We evaluated this concept in a mouse HCC model induced by cMyc and Nras expression¹⁶ since in murine models of HCC, activation of both Myc and Ras oncogenes has been implicated in the promotion of hepatic fibrosis.^55,56 Treatments were conditioned starting on day 14 with elafibranor, a PPAR agonist known to have anti-inflammatory and anti-fibrotic effects. Elafibranor has shown promising results in reducing liver inflammation, oxidative stress, and fibrosis in preclinical liver disease models.⁵⁷ On day 21, we mimicked TACE by treating with the STING agonist cGAMP and anti-PD-1 therapy (sintilimab). After 30 days, we harvested the liver with tumors for further analysis (Figure 6A).

Figure 6.

Enhanced efficacy of triple combination therapy for hepatocellular carcinoma

(A) Design of mouse model treatment study. An orthotopic liver cancer model was established through tail vein plasmid injection. Elafibranor, 2′,3′-cGAMP, and PD-1 monoclonal antibody were administered at specific doses and time points. Tumors were collected and analyzed at a designated time point to evaluate the efficacy of the triple combination therapy.

(B) Tumor response in the orthotopic liver cancer model to control (Ctrl), elafibranor, 2′,3′-cGAMP & PD-1, and elafibranor combined with 2′,3′-cGAMP & PD-1. Qualitatively, combination regimens were more effective at reducing tumor nodules.

(C) Quantitation of liver surface tumors (n = 6 mice per group). Elafibranor + 2′,3′-cGAMP & PD-1 showed the fewest tumors and was statistically better than all other treatments.

(D) Quantitation of CD8⁺ cell and Collagen I⁺ cell density (n = 5 mice per group). Elafibranor + 2′,3′-cGAMP & PD-1 showed the highest immune response and reduced degree of fibrosis. Group means were compared by one-way ANOVA with Tukey's multiple comparisons test (ns: p > 0.05, ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001, ∗∗∗∗p ≤ 0.0001).

Both elafibranor and the dual combination therapy of TACE + anti-PD-1 demonstrated reductions in tumor burden, with the dual combination therapy showing a greater reduction. However, the triple combination therapy involving elafibranor, TACE, and anti-PD-1 exhibited the least tumor burden (Figures 6B and 6C). Treatment with elafibranor alone resulted in reduced fibrosis (Figure 6D).

The dual combination therapy of TACE + anti-PD-1 showed an increase in CD8⁺ T cells (Figure 6C). Notably, the triple combination therapy led to reduced fibrosis and increased CD8⁺ T cells (Figure 6D; Figure S8A).

Based on the success of the triple combination therapy in our mouse experiments, we conducted a separate registered exploratory study (NCT05213221) in an independent cohort of 30 patients with advanced HCC, testing TACE combined with envafolimab, an anti-PD-L1 antibody, and lenvatinib. Because elafibranor is not approved for HCC, we used lenvatinib as a clinically available multi-kinase inhibitor with reported anti-fibrotic activity (Figure S8B). The exploratory triple combination cohort showed superior clinical outcomes compared with the original TACE plus anti-PD-1 trial cohort (Figure S8C). In the triple combination cohort, fibrosis levels were lower, cGAS signaling was highly activated, and PD-L1 expression in epithelial cells was more pronounced (Figure S8D).

These results demonstrate that fibrosis suppresses cGAS-STING signaling and impairs CD8⁺ T cell-mediated immunity in HCC, thereby limiting the efficacy of TACE-based immunotherapy. Targeting fibrosis with agents such as elafibranor or lenvatinib may enhance immune activation by relieving mechanical suppression of DNA sensing in tumor epithelial cells. Across the mouse model and exploratory human cohort, combining an anti-fibrotic strategy with TACE-related immune activation and checkpoint blockade reduced fibrosis, restored cGAS-STING signaling, increased T cell infiltration, and improved therapeutic outcomes. These findings uncover a mechanism wherein fibrotic remodeling acts as a physical and immunologic barrier to innate immune sensing and suggest that incorporating anti-fibrotic strategies may help overcome immune resistance in HCC.

Graphical Classification Based on Spatial Similarity model predicts response to HCC combination treatment

An enticing clinical strategy would be to predict which patients are likely to have a poor response to therapy and then add an anti-fibrotic agent for that population. We hypothesized that it might be possible to generate a predictive model using IMC data obtained from needle biopsies. We used pre-treatment IMC images and associated clinical outcomes from the TACE + ICB trial as the training dataset for our model (Figure 7A).

Figure 7.

Prediction of combinatorial treatment outcomes using GCSS with a restricted marker panel

(A) Workflow of the GCSS model. Cell cluster locations are mapped, a spatial similarity matrix is generated to quantify the spatial relationships between clusters, and the model predicts responder (OR) vs. non-responder (SD) classification.

(B) Spatial similarity matrix showing pairwise spatial relationships between cellular clusters. Darker red indicates higher spatial similarity; blue indicates lower or negative correlations.

(C) Model architecture: this panel illustrates the schematic diagram of our model, which was designed by following the approaches of Yan et al. (2023) and Xu et al. (2023).

(D) AUC values for each algorithm (GCSS, GCSS-6 marker, Lasso, and SVM) in distinguishing the OR group from the SD group.

(E) Critical cell-cell interaction edges identified by the GCSS model, representing communication pathways associated with the OR and SD groups.

(F) Minimized marker panel of six markers (CD4, CD8, CD68, Pan-CK, PD-L1, and Collagen I) derived from critical interaction edges in (E), with representative IHC staining.

(G) Receiver operating characteristic (ROC) curve evaluating the minimized marker panel performance in an independent HCC cohort for predicting treatment response.

Scale bars, 40 micrometers.

The model was constructed by dividing segmented cells within each ROI into separate images (N = 29), each representing the spatial localization of a specific cell type. Pairwise spatial similarity between cell types was calculated using structure similarity matrix (SSIM) measurement (Figure 7B). We employed GCSS (Graphical Classification Based on Spatial Similarity),⁵⁸ which integrates spatial similarity information with network structure for efficient classification. We compared GCSS with SVM, Lasso, and MRI-based deep learning benchmarks^59,60 (Figure 7C). Using the clinical trial cohort, GCSS achieved higher sensitivity and accuracy than SVM and Lasso (85% vs. 75% and 70%, respectively) (Figure 7D) and superior area under the curve (AUC) compared with MRI-based deep learning.

Incidentally, GCSS also provided confirmatory mechanistic data regarding cellular interactions within each group. In the SD group, we observed interactions between epithelial/fibroblast and CD45RA⁺ CD8⁺ T cells, effector CD8⁺ T cells and epithelial/fibroblast cells, as well as neutrophil and epithelial cell interactions. In the OR group, we found interactions between NK cells and infiltrating macrophages, effector CD8⁺ T cells and immune suppressive epithelial cells, and GATA3⁺ CD4⁺ T cells and resident macrophages (Figure 7E). These interactions are consistent with our integrated spatial multi-omics analysis and prior mechanistic exploration.

We used dominant patterns in the model to create a minimized mIHC panel for predicting treatment response consisting of six markers, including CD4, CD8, CD68, Collagen I, Pan-CK, and PD-L1 (Figure 7F). These markers represented the crucial cell interactions responsible for different treatment responses within the OR and SD groups (Figure 7E). This approach is intended to enable clinical feasibility since it reduces assay costs and can be done more rapidly.

To validate the predictive capability of the minimal panel, we collected 36 pre-treatment specimens from patients with HCC who underwent same treatment from an external hospital (Table S1-3). Using mIHC input into the GCSS model, we achieved over 80% accuracy in predicting treatment response (Figure 7G), similar to results on IMC data (compare Figure 7D). These results suggest that the mIHC panel will be effective for predicting treatment responses for HCC to TACE and ICB at minimal cost (Figure 7G). This result was achieved despite a relatively small sample size, suggesting robustness in the method.

Discussion

Immunotherapy can improve outcomes in unresectable HCC, but some patients do not respond. This study identified TME features that influence response to combined TACE and ICB therapy. Using spatial multi-omics, we revealed key interactions between Collagen I^hi fibroblasts, immune-suppressive epithelial cells, and T cells. In responders, TACE-induced dsDNA breaks activated the cGAS-STING pathway in immune-suppressive epithelial cells, upregulating PD-L1 and enhancing cytokine production that contributes to T cell recruitment. However, the immunoregulatory role of PD-L1⁺ epithelial cells is complex and context dependent. Non-responders exhibited prominent fibrotic stroma acting as a physical and molecular barrier limiting T cell infiltration and immune activation. This suggests that triple combination therapy including an anti-fibrotic agent may be beneficial, a hypothesis confirmed in our mouse model. We developed a GCSS-based assay to predict HCC responses to TACE and ICB treatments.

While combination TACE therapies have shown promise, this study offers a mechanism-guided approach. In our exploratory clinical study, elafibranor was replaced with lenvatinib, a multi-kinase inhibitor approved for HCC, based on pragmatic considerations: elafibranor is not approved for clinical use in HCC. The checkpoint inhibitor was envafolimab, consistent with the registered NCT05213221 protocol. Lenvatinib has been reported to exert modest anti-fibrotic effects through inhibition of VEGFR/FGFR signaling, and we observed reduced fibrosis in patients receiving triple therapy. However, lenvatinib has multiple mechanisms of action including direct tumor inhibition and anti-angiogenic effects. Therefore, clinical benefit in the triple combination group may reflect synergistic effects rather than anti-fibrotic activity alone. We incorporated fibrosis and immune marker profiling to track treatment impact on the TME, but dedicated future studies with selective anti-fibrotic agents will be necessary to isolate the anti-fibrotic component.

In our preclinical experiments, we used cGAMP as a substitute for TACE to simulate its immunostimulatory effects. Our spatial multi-omic profiling showed that TACE-associated cGAS-STING activation in responders triggers chemokine production and T cell recruitment. Given the technical limitations of reproducing embolization in small animals, cGAMP was administered to mimic this key signaling axis. Although this approach confirmed the contribution of STING activation to treatment response, cGAMP does not recapitulate the full procedural and microvascular impact of TACE. Future studies could incorporate locoregional delivery methods to better mirror the procedural complexity of TACE in vivo.

This study demonstrates that direct measures of TME composition and organization can yield predictive models and mechanistic insights. Prior efforts utilizing gene signatures, single-cell RNA sequencing, microbiome analysis, radiomics, and clinical markers have yielded inconsistent associations with treatment response.³⁷ Similarly, studies evaluating neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR) in patients with HCC receiving anti-PD-1 antibodies, and high-dimensional immune profiling of peripheral blood mononuclear cells (PBMCs), have not identified reliable biomarkers for predicting ICB response in HCC.^37,61

This study expands on our previous work identifying distinct HCC TME regions,¹⁶ uncovering that excessive fibrosis hinders dsDNA signal transmission and prevents cGAS activation. This aligns with Yao et al.‘s findings on a tumor immune barrier composed of SPP1⁺ macrophages and cancer-associated fibroblasts (CAFs) that restricts immune cell infiltration.⁶² We propose that the effect of fibrotic stroma is highly context dependent, governed by spatial organization, matrix stiffness, and cell-cell crosstalk. Relieving fibrosis using elafibranor or lenvatinib enhanced T cell infiltration and treatment efficacy, confirming fibrosis as a modifiable resistance mechanism. However, certain fibroblast-rich regions in responders exhibited immune cell recruitment, suggesting that select fibroblast subsets may play supportive roles. Future studies should classify fibroblast heterogeneity and its context-specific impact on immune function.

Limitations of the study

Utilizing spatial multi-omics data integration, we discovered that features associated with response and non-response coexisted within the TME prior to treatment. Successful combination therapy enhances anti-tumor features via the cGAS pathway. Based on this mechanistic understanding, we established a prediction model with potential to guide treatment decisions, though clinical trials are needed for validation. Despite the strengths of our prospective trial design and integrated spatial multi-omics approach, several limitations warrant consideration. First, the sample size remains modest (n = 61 total, with paired spatial proteomics in 38 cases and single-cell analyses in a subset). To mitigate this, we validated key observations in independent transcriptomic cohorts. Future studies incorporating broader disease stages and etiologies will be critical to assess generalizability. Second, our study focused on BCLC stage A and B HCC, where TACE-based therapies are most commonly indicated. Our findings are most directly applicable to early and intermediate-stage HCC; whether the fibrosis-mediated suppression of cGAS-STING signaling extends to advanced disease requires further investigation. Third, the IMC panel did not include hepatocyte-specific markers such as albumin or HepPar1; tumor epithelial cells were annotated using Pan-Cytokeratin, E-cadherin, and immune-checkpoint markers. Future IMC panels incorporating hepatocyte lineage markers will be useful. Fourth, in vivo therapeutic validation relied on the c-Myc/Nras-driven murine HCC model. Additional models with diverse oncogenic drivers and stromal architectures will be needed to confirm the generalizability of the fibrosis-cGAS-STING axis.

Our study provided extensive spatial characterization of varying responses to TACE and ICB therapy. Future research should include patients with HCV or NASH-related HCC to verify our findings in these populations.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Jianpeng Sheng (shengjp@zju.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

The transcriptomic datasets generated in this study have been deposited in the National Genomics Data Center (NGDC), China National Center for Bioinformation (CNCB), under BioProject PRJCA017273. Raw single-cell RNA sequencing data have been deposited in the Genome Sequence Archive for Human (GSA-Human) under accession numbers HRA004695 and HRA007308, and raw bulk RNA sequencing data have been deposited under accession number HRA007307. Raw spatial transcriptomics data have been deposited in GSA-Human under accession numbers HRA004737 and HRA020344. The corresponding processed bulk and single-cell transcriptomic data are available in OMIX under accession number OMIX: OMIX018501, and the processed spatial transcriptomics data are available under accession number OMIX: OMIX006713. Imaging mass cytometry (IMC) data are available in OMIX under accession numbers OMIX: OMIX002774 and OMIX003997.

Please note that the single-cell and spatial transcriptomics analysis in this study does not involve the original code. To access the original codes used, please refer to the STAR Methods section and respective references provided. The GCSS codes used in this analysis have been uploaded to GitHub at the following link: https://github.com/shaoweinuaa/GCSS. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

The authors would like to thank Infinity Scope Inc. for IMC and mIHC help, Tianxing Yan for designing the graphical abstract and the cover image, and Yiwen Chen for assisting in trial coordination. This work was supported by the Beijing Natural Science Foundation (grants L234002 and L2510026 to J. Sheng), the Science and Health Joint Project of Shapingba District (grant 2025SQKWLHZD005 to Q.L.), the Key Science and Technology Project of Chongqing Municipal Education Commission (2026) (project no.: KJZD-K202600109 to Q.L.), the National Natural Science Foundation of China (grant 81871925 to X.B. and grant 82188102 to T. Liang), the National Key Research and Development Program (grant 2019YFC1316000 to T. Liang) and the National Natural Science Foundation Basic Science Centre of China (Study of Tumor Material and Energy Dynamics, grant 8218810 to T. Liang). We also thank Dr. Zhiwei Zhou and Dr. Kenneth D. Westover, who contributed to manuscript proofreading and revision but are not listed as authors.

Author contributions

J. Sheng conceptualized the study and designed the experiments. J. Sheng and Q.L. drafted the manuscript. L.W. performed data analysis, while J.Z. and J. Song conducted the experimental work. W.S. and Q.L. developed the prediction model and performed imaging analysis. Y.Z. and H.S. provided bioinformatic support. J.T., Y.J., and T. Li contributed to the experimental procedures. T. Liang and X.B. organized the clinical trial and obtained clinical samples, with C.G. assisting in trial coordination. W.Z. organized and provided the external validation cohort. T. Liang, X.B., and J. Sheng supervised the project and acquired funding.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
IMC antibody: anti-CD45	Cell Signaling Technology	Clone D9M8I; antibody Cat# 13917S; metal isotope 89 Y; metal source Macklin Cat# Y820645
IMC antibody: anti-CD14	Abcam	Clone EPR3653; antibody Cat# ab226121; metal isotope 141 Pr; metal source Fluidigm Cat# 201141A
IMC antibody: anti-FoxP3	Cell Signaling Technology	Clone D6O8R; antibody Cat# 12653S; metal isotope 142 Pr; metal source Fluidigm Cat# 201142A
IMC antibody: anti-CD16	Abcam	Clone EPR16784; antibody Cat# ab256582; metal isotope 143 ND; metal source Fluidigm Cat# 201143A
IMC antibody: anti-CD69	Abcam	Clone EPR21814; antibody Cat# ab234512; metal isotope 144 ND; metal source Fluidigm Cat# 201144A
IMC antibody: anti-CD4	Abcam	Clone EPR6855; antibody Cat# ab181724; metal isotope 145ND; metal source Fluidigm Cat# 201145A
IMC antibody: anti-CD8a	BioLegend	Clone C8/144B; antibody Cat# 372902; metal isotope 146 ND; metal source Fluidigm Cat# 201146A
IMC antibody: anti-Collagen I	Abcam	Clone EPR7785; antibody Cat# ab215969; metal isotope 147 Sm; metal source Fluidigm Cat# 201147A
IMC antibody: anti-Cleaved Caspase-3	Cell Signaling Technology	Clone 5A1E; antibody Cat# 9664S; metal isotope 148 ND; metal source Fluidigm Cat# 201148A
IMC antibody: anti-CD31	Cell Signaling Technology	Clone 89C2; antibody Cat# 3528S; metal isotope 149 Sm; metal source Fluidigm Cat# 201149A
IMC antibody: anti-E-cadherin	BioLegend	Clone DECMA-1; antibody Cat# 147302; metal isotope 150 ND; metal source Fluidigm Cat# 201150A
IMC antibody: anti-B7-H4	Thermo Fisher Scientific	Clone H74; antibody Cat# 14-5949-82; metal isotope 151 Eu; metal source Fluidigm Cat# 201151A
IMC antibody: anti-VISTA	Cell Signaling Technology	Clone D1L2G; antibody Cat# 64953S; metal isotope 152 Sm; metal source Fluidigm Cat# 201152A
IMC antibody: anti-CD7	Abcam	Clone EPR4242; antibody Cat# ab230834; metal isotope 153 Eu; metal source Fluidigm Cat# 201153A
IMC antibody: anti-CD169	Abcam	Clone HSn 7D2; antibody Cat# ab18619; metal isotope 154 Sm; metal source Fluidigm Cat# 201154A
IMC antibody: anti-CD103	Abcam	Clone EPR4166(2); antibody Cat# ab271889; metal isotope 155 Gd; metal source Fluidigm Cat# 201155A
IMC antibody: anti-PD-L1	Cell Signaling Technology	Clone E1L3N; antibody Cat# 13684S; metal isotope 156 Gd; metal source Fluidigm Cat# 201156A
IMC antibody: anti-LAG3	Cell Signaling Technology	Clone D2G4O; antibody Cat# 15372S; metal isotope 158 Gd; metal source Fluidigm Cat# 201158A
IMC antibody: anti-CD68	BioLegend	Clone C8/144B; antibody Cat# 372902; metal isotope 159 Tb; metal source Fluidigm Cat# 201159A
IMC antibody: anti-CD11b	Abcam	Clone EPR1344; antibody Cat# ab209970; metal isotope 160 Gd; metal source Fluidigm Cat# 201160A
IMC antibody: anti-CD20	Abcam	Clone IGEL/773; antibody Cat# ab213033; metal isotope 161 Dy; metal source Fluidigm Cat# 201161A
IMC antibody: anti-CD11c	Abcam	Clone EP1347Y; antibody Cat# ab216655; metal isotope 162 Dy; metal source Fluidigm Cat# 201162A
IMC antibody: anti-CD15	BD Biosciences	Clone HI98; antibody Cat# 563872; metal isotope 163 Dy; metal source Fluidigm Cat# 201163A
IMC antibody: anti-Granzyme B	Abcam	Clone EPR20129-217; antibody Cat# ab219803; metal isotope 164 Dy; metal source Fluidigm Cat# 201164A
IMC antibody: anti-PD-1	Cell Signaling Technology	Clone D4W2J; antibody Cat# 86163S; metal isotope 165 Ho; metal source Fluidigm Cat# 201165A
IMC antibody: anti-Ki67	BD Biosciences	Clone B56; antibody Cat# 550609; metal isotope 166 Er; metal source Fluidigm Cat# 201166A
IMC antibody: anti-GATA-3	Cell Signaling Technology	Clone D13C9; antibody Cat# 5852S; metal isotope 167 Er; metal source Fluidigm Cat# 201167A
IMC antibody: anti-HLA-DR	Novus Biologicals	Clone EDHu-1; antibody Cat# NB110-40686; metal isotope 168 Er; metal source Fluidigm Cat# 201168A
IMC antibody: anti-CD45RA	BioLegend	Clone HI100; antibody Cat# 304143; metal isotope 169 Tm; metal source Fluidigm Cat# 201169A
IMC antibody: anti-CD3ε	Cell Signaling Technology	Clone D7A6E; antibody Cat# 85061S; metal isotope 170 Er; metal source Fluidigm Cat# 201170A
IMC antibody: anti-CD45RO	BioLegend	Clone UCHL1; antibody Cat# 304239; metal isotope 173 Yb; metal source Fluidigm Cat# 201173A
IMC antibody: anti-CD57	BD Biosciences	Clone NK-1; antibody Cat# 555618; metal isotope 174 Yb; metal source Fluidigm Cat# 201174A
IMC antibody: anti-IL-2 Receptor alpha (CD25)	Abcam	Clone EPR6452; antibody Cat# ab215378; metal isotope 175 Lu; metal source Fluidigm Cat# 201175A
IMC antibody: anti-Pan-Cytokeratin	BioLegend	Clone AE-1/AE-3; antibody Cat# 914204; metal isotope 176 Yb; metal source Fluidigm Cat# 201176A
IMC antibody: anti-Alpha Smooth Muscle Actin	BioLegend	Clone 1A4; antibody Cat# 904601; metal isotope 194 Pt; metal source Fluidigm Cat# 201194
IMC antibody: anti-Vimentin	Cell Signaling Technology	Clone D21H3; antibody Cat# 5741S; metal isotope 198 Pt; metal source Fluidigm Cat# 201198
mIHC antibody: anti-cGAS	ABclonal	Cat# A23846
mIHC antibody: anti-Pan-CK	Abcam	Cat# ab215838
mIHC antibody: anti-RelA	Abcam	Cat# ab207297
mIHC antibody: anti-dsDNA	Abcam	Cat# ab273137
mIHC antibody: anti-PD-L1	Proteintech	Cat# 66248-1-Ig
mIHC antibody: anti-Collagen I	Abcam	Cat# ab138492
mIHC antibody: anti-CD45RO	Abcam	Cat# ab216024
mIHC antibody: anti-CD8	Abcam	Cat# ab217344
Immunofluorescence antibody: anti-PD-L1	Abcam	Cat# ab205921
Immunofluorescence antibody: anti-dsDNA	Abcam	Cat# ab273137
Immunofluorescence antibody: anti-gamma-H2AX	ABclonal	Cat# AP0687
AF488 anti-mouse secondary antibody	Cell Signaling Technology	Cat# 4408
AF647 anti-rabbit secondary antibody	Cell Signaling Technology	Cat# 4414
Western blot antibody: anti-phospho-p65	Target/dilution only in source files	Dilution 1:1000; catalog/RRID not reported
Western blot antibody: anti-phospho-TBK1	Target/dilution only in source files	Dilution 1:1000; catalog/RRID not reported
Western blot antibody: anti-phospho-IRF3	Target/dilution only in source files	Dilution 1:1000; catalog/RRID not reported
Western blot antibody: anti-phospho-STING	Target/dilution only in source files	Dilution 1:1000; catalog/RRID not reported
Western blot antibody: anti-GAPDH	Target/dilution only in source files	Dilution 1:2000; catalog/RRID not reported
Biological samples
Human HCC tumor specimens from the TACE plus sintilimab trial	This paper	ClinicalTrials.gov: NCT04174781; FFPE and fresh HCC tissue specimens
Human HCC tumor specimens from the exploratory triple-combination cohort	This paper	ClinicalTrials.gov: NCT05213221; TACE plus envafolimab plus lenvatinib cohort
External mIHC validation cohort HCC specimens	This paper	36 pretreatment HCC specimens; see Tables S1-3
Chemicals, peptides, and recombinant proteins
Sintilimab (anti-PD-1 antibody)	Innovent Biologics/Eli Lilly (TYVYT)	TYVYT/sintilimab injection; anti-PD-1 monoclonal antibody
Envafolimab (anti-PD-L1 antibody)	Alphamab Oncology/3D Medicines/Simcere (ENWEIDA; KN035)	ENWEIDA/KN035/envafolimab; anti-PD-L1 antibody used in NCT05213221 triple-combination cohort
Lenvatinib	Eisai/Merck (LENVIMA)	LENVIMA/lenvatinib mesylate
Epirubicin	Commercial epirubicin	CAS 56420-45-2; final concentration 1 μg/mL; exact supplier not reported in manuscript
2′,3′-cGAMP	Commercial 2′,3′-cGAMP	STING agonist; 10 μg/mL in vitro; 5 mg/kg in mice; CAS 1441190-66-4; exact supplier not reported in manuscript
PD-1 monoclonal antibody for mouse treatment	Exact supplier/clone not reported	5 mg/kg intraperitoneal injection every other day from day 21
Elafibranor	Elafibranor/GFT505	PPAR agonist; 20 mg/kg/day by gavage; exact supplier not reported in manuscript
Matrigel matrix	Corning (Matrigel matrix product line)	Matrigel matrix; Matrigel-to-cell ratios 1:1 and 3:1; exact catalog not reported in manuscript
Triton X-100	Beyotime	Cat# P0096
SuperBlock blocking buffer	Thermo Fisher Scientific	Cat# 37515
Phalloidin-iFluor 555	Abcam	Cat# ab176756
RIPA lysis buffer	Beyotime Biotechnology	Cat# P0013C
Phenylmethanesulfonyl fluoride (PMSF)	Beyotime Biotechnology	Cat# ST505
Phosphatase Inhibitor Cocktail	Beyotime Biotechnology	Cat# P1081
BCA reagent	Beyotime Biotechnology	Cat# P0012
NuPAGE LDS sample buffer (4×)	Thermo Fisher Scientific	Cat# NP0007
PVDF membrane	Millipore	Cat# IPVH00010
Intercalator-Ir	Fluidigm	Cat# 201192B; 1.25 uM
Critical commercial assays
Opal Polaris 7-color Manual IHC Kit	Akoya Biosciences	Opal Polaris 7-Color Manual IHC Kit; Cat# NEL861001KT
Maxpar ×8 Antibody Labeling Kit	Fluidigm/Standard BioTools	Maxpar ×8 antibody labeling kits; isotope-specific metal catalog numbers are listed in antibody rows
CXCL10 ELISA assay	Supplier not reported	CXCL10 quantification in culture supernatants; kit catalog not reported
Transwell co-culture assay inserts/system	Supplier/pore size not reported	Molt-4 cells in upper chamber; HepG2 +/− NHDF in lower chamber; cGAMP stimulation
10× Genomics single-cell RNA-seq reagents/platform	10× Genomics	Chromium droplet-based scRNA-seq platform
10× Genomics Visium spatial transcriptomics platform	10× Genomics	Visium Spatial Gene Expression platform
Deposited data
Associated BioProject	This paper; BioProject, NGDC	BioProject: PRJCA017273
Raw single-cell RNA sequencing data	This paper; GSA-Human, NGDC	GSA-Human: HRA004695; HRA007308
Processed single-cell transcriptomic data	This paper; OMIX, NGDC	OMIX: OMIX018501
Raw bulk RNA sequencing data	This paper; GSA-Human, NGDC	GSA-Human: HRA007307
Processed bulk transcriptomic data	This paper; OMIX, NGDC	OMIX: OMIX018501
Raw spatial transcriptomics data	This paper; GSA-Human, NGDC	GSA-Human: HRA004737; HRA020344
Processed spatial transcriptomics data	This paper; OMIX, NGDC	OMIX: OMIX006713
Imaging mass cytometry data	This paper; OMIX, NGDC	OMIX: OMIX002774; OMIX003997
TCGA-LIHC transcriptomic dataset	The Cancer Genome Atlas	TCGA-LIHC
GSE14520 transcriptomic dataset	Gene Expression Omnibus	GEO: GSE14520
GSE76427 transcriptomic dataset	Gene Expression Omnibus	GEO: GSE76427
Experimental models: Cell lines
Huh-7 human HCC cell line	Laboratory of T.L.	RRID:CVCL_0336
PLC/PRF/5 human HCC cell line	Laboratory of T.L.	RRID:CVCL_0485
SK-HEP-1 human liver cancer cell line	Laboratory of T.L.	RRID:CVCL_0525
HepG2 human liver cancer cell line	Laboratory of T.L.	RRID:CVCL_0027
Molt-4 human T cell leukemia cell line	Infinity Scope Inc.	RRID:CVCL_0013
Normal human dermal fibroblasts (NHDF)	Infinity Scope Inc.	Normal human dermal fibroblasts used in HepG2 co-culture; catalog/RRID not reported
Experimental models: Organisms/strains
C57BL/6 mice, female, 6 weeks old	Model Animal Research Center of Nanjing University	C57BL/6; female, 6 weeks old
Recombinant DNA
PT3-EF1a-C-Myc plasmid	Wen et al.63	Dose 10 μg/20 g body weight; repository plasmid ID not reported
PT/Caggs-NRas-V12 plasmid	Wen et al.63	Dose 10 μg/20 g body weight; repository plasmid ID not reported
pCMV-SB11 plasmid	Wen et al.63	Dose 2 μg/20 g body weight; repository plasmid ID not reported
Software and algorithms
TissueNet pretrained model for cell segmentation	Greenwald et al.64/DeepCell	TissueNet pretrained model
Harmony R package	Korsunsky et al.65	Version 0.1.0
Rphenograph R package	Levine et al.66	Version 0.99.1
Cell Ranger Single-Cell Software Suite	10× Genomics	Cell Ranger software
Seurat R package	Stuart et al.67	Version 4.3.0
Space Ranger	10× Genomics	Space Ranger software
Robust Cell Type Decomposition (RCTD)	Cable et al.29/spacexr/RCTD	RCTD/spacexr
clusterProfiler R package	Wu et al.68/Bioconductor	clusterProfiler
NicheNet	Browaeys et al.38/saeyslab	NicheNet
nf-core/rnaseq pipeline	Ewels et al.69/nf-core	nf-core/rnaseq
edgeR R/Bioconductor package	Robinson et al.70/Bioconductor	edgeR; RRID:SCR_012802
ComBat batch correction	Johnson et al./sva package	sva:ComBat
GraphPad Prism	GraphPad Software	GraphPad Prism; RRID:SCR_002798
Graphical Classification based on Spatial Similarity (GCSS)	This paper	https://github.com/shaoweinuaa/GCSS
LASSO and L1-SVM comparison models	This paper/referenced methods	Comparison models; implementation details not reported in manuscript
Instruments
HistoCore MULTICUT system	Leica Biosystems	HistoCore MULTICUT
DMi8 optical microscope	Leica Microsystems	Leica DMi8
Vectra Polaris Quantitative Pathology Imaging System	Akoya Biosciences	Vectra Polaris
TCS SP8 confocal microscope	Leica Microsystems	Leica TCS SP8
ChemiScopeTouch imaging system	Clinx Science Instruments	ChemiScopeTouch
Imaging mass cytometry acquisition platform	Fluidigm/Infinity Scope Inc.	IMC acquisition platform; exact instrument model not reported

Experimental model and study participant details

Ethics

The experimental protocol was established according to the ethical guidelines of the Helsinki Declaration. The Research Ethics Committee of the First Affiliated Hospital, College of Medicine, Zhejiang University approved all animal experiments. The human clinical protocols and use of associated tissue specimens were approved by the Ethics Committee of the First Affiliated Hospital, Zhejiang University School of Medicine. Written informed consent was obtained from each participant or legal guardian. Sample tissues from patients diagnosed with HCC were isolated by professional surgeons. All samples were diagnosed by pathological examination. We performed H&E staining for every tissue sample, and the resulting stained samples were examined with an optical microscope (Leica DMi8, Germany) by a pathology expert to select the ROIs. We chose at least 3 ROIs for each tissue to be scanned by IMC. For experimental models, all mice used in this study were female and 6 weeks old. All mice were housed in the specific pathogen-free (SPF) facility of the First Affiliated Hospital, Zhejiang University School of Medicine, with approval from the Institutional Animal Care & Use Committee (IACUC). The primary clinical trial of TACE + sintilimab was registered at ClinicalTrials.gov (NCT04174781). The exploratory clinical study of envafolimab + lenvatinib + TACE in advanced HCC was registered at ClinicalTrials.gov (NCT05213221).

mRECIST evaluation

The response evaluation was conducted according to the modified Response Evaluation Criteria in Solid Tumors (mRECIST) guidelines⁷¹ in solid tumors by contrast-enhanced computed tomography at the end of first treatment cycle, then every 3 months postoperatively for 2 years, and every 6 months thereafter. An illustrative case was depicted in Figure 1D.

Pathological evaluation

Liver tumor specimens measuring 3 cm or smaller were uniformly procured in their entirety. For tumors exceeding 3 cm, dissection was performed at intervals of 0.5 cm across the greatest dimension, with careful selection of sections that typify areas of necrosis and adjacent viable tumor for analysis. Concurrently, tissue samples from the tumor margin and adjacent hepatic parenchyma were harvested for comparative purposes. A pathologic assessment was then conducted to ascertain the proportion of viable neoplastic tissue through detailed microscopic examination.

Animal study

In brief, C57BL/6 mice were purchased from the Model Animal Research Center of Nanjing University (China). To establish spontaneous HCC, mice were administered plasmids carrying PT3-EF1a-C-Myc, PT/Caggs-NRas-V12, and pCMV-SB11 in PBS at doses of 10 μg/20 g, 10 μg/20 g, and 2 μg/20 g body weight (0.1 mL/g body weight intravenous), respectively, as previously reported.⁶³ Elafibranor was dissolved in a mixture of 10% DMSO and 90% corn oil and administered via gavage at a dose of 20 mg/kg/day starting from day 14. 2′,3′-cGAMP and a PD-1 monoclonal antibody were administered at a dose of 5 mg/kg via intraperitoneal injection every other day from day 21. On day 30, tumors were collected and analyzed.

Method details

Multiplex immunohistochemistry

Multiplex immunohistochemistry was carried out using Opal Polaris™ 7-color Manual IHC Kit. Briefly, sections were blocked after deparaffinization. Tissues were then incubated with primary antibodies, followed by secondary antibody incubation and fluorescent linking. DAPI staining was performed for visualization of opal slides using the Vectra Polaris Quantitative Pathology Imaging Systems (Akoya, USA). The primary antibodies used in the multi-color immunohistochemistry (mIHC) included cGAS (Abclonal cat#:A23846), Pan-CK (Abcam cat#:ab215838), RelA (Abcam cat#: ab207297), dsDNA (Abcam cat#:ab273137), PD-L1 (Proteintech cat#:66248-1-Ig), Collagen I (Abcam cat#:ab138492), CD45RO (Abcam cat#:ab216024), and CD8 (Abcam cat#:ab217344).

IMC section preparation

The formalin-fixed paraffin-embedded (FFPE) tissue samples from HCC patients were cut into 4-μm sections using the HistoCore MULTICUT system (Leica, Germany). These sections were then heated at 68°C for 1 h. Dewaxing was carried out by incubating the sections twice in xylene at 68°C for 10 min each time. Following this, the sections were rehydrated in 95%, 85%, and 75% ethanol at room temperature for 5 min each. For antigen retrieval, the sections were subjected to heat-mediated treatment in sodium citrate solution at 100°C for 30 min. Once the sections had cooled naturally to room temperature, they were washed twice with PBS-TB (PBS containing 0.5% Tween 20 and 1% BSA) for 5 min each. SuperBlock (37515, Thermo Fisher Scientific) was then used to block the sections for 30 min at room temperature. After three washes with PBS-TB, the sections were incubated with an antibody cocktail overnight at 4°C. For most antibodies, metal labeling was performed using the Maxpar ×8 antibody labeling kits (Fluidigm). In summary, the antibodies were conjugated to an ×8 polymer, which in turn loaded a specific metal. The final antibody panel is detailed in Table S2. Following overnight incubation with the antibody cocktail at 4°C, the sections were washed three more times with PBS-TB. To label the cell nuclei, the sections were incubated with an Intercalator-Ir (201192B, Fluidigm) solution in PBS-TB (1.25 μM) for 30 min at room temperature, followed by two washes with PBS-TB and one wash with ddH₂O.

IMC single cell segmentation

We utilized the pre-trained TissueNet, which was introduced in a previous work,⁶⁴ to perform cell segmentation on the IMC images. The TissueNet requires two channels of imaging data for its operation. The first channel corresponds to the nuclear channel, typically labeled with DAPI, which provides information about the locations of cell nuclei. The second channel represents either the membrane or cytoplasmic channel, indicating the cell boundaries. In our study, we specifically chose the membrane channels to serve as the second channel.

IMC preprocessing and cell type identification

Marker expression was first hyperbolic arc-sine transformed and capped at the 1^st and 99^th percentiles as minimum and maximum values for each channel. Max-min normalization was applied for each channel separately, followed by R package Harmony (version 0.1.0)⁶⁵ to align batch effects. Then we used Rphenograph (version 0.99.1), an R implementation of Phenograph⁶⁶ and 100 nearest neighbors to cluster cells with two rounds. First, all lineage-related markers including CD14, FOXP3, CD16, CD4, CD8, CollagenI, CD31, CD163, CD68, CD11b, CD11c, CD20, CD3, CD15, CD57, Pankeratin, αSMA, CD45, CD7, Vimentin, and HLA-DR were used to identify the major clusters. Then T cell subsets were further re-clustered using the following channels including FOXP3, CD69, CD4, CD8, Caspase3, B7H4, VISTA, PD-L1, LAG3, Granzyme B, PD-1, KI67, GATA3, TNFα, CD45RA, CD3, CD25, CD45RO, ECadherin, CD7, and CD103; and epithelial cells were further re-clustered using Pankeratin, B7H4, PD-L1, VISTA, LAG3, and PD-1. The cluster means of each channel were displayed as heatmap and used for annotation of cell types (CT).

Identification of cellular neighborhoods (CN)

The cellular neighborhood (CN) for each cell was captured using windows consisting of the 10 nearest neighboring cells as measured by Euclidean distance between X/Y coordinates. These windows were then clustered by their compositions concerning all cell types using kmeans clustering. With k = 15, each cell was then allocated to the neighborhood (CN) that its surrounding window was. CN assignment was validated by overlaying their Voronoi allocation graphs on the original tissue IMC images.

Downstream spatial analysis

Spatial context-based expression of Pankeratin, Collagen I, CD45RO, B7H4, PD-L1, VISTA, LAG3, and PD-1 for each cell was calculated as the mean expression of the cell itself and its 10-nearest neighbors. ComBat⁷² was used to adjust for batch effects. Based on the spatial context-based marker expression, six types of areas were identified. Epithelial, fibroblast, and immune areas were defined as areas with expression of Pankeratin, Collagen I, and CD45RO greater than the 50% quantile across all profiles, respectively. ImmuneSuppressiveEpithelial areas were defined as epithelial areas with any B7H4, PD-L1, VISTA, LAG3, or PD-1 expression greater than the 85% quantile. ImmuneSuppressiveEpithelial&Immune areas were defined as the overlap between ImmuneSuppressiveEpithelial and immune areas, while Fibroblast&Epithelial areas were defined as overlap regions between fibroblast and epithelial areas. Interaction counts between center and neighbor cells were defined as, for each cell of type A (center), the average number of neighbors of type B (neighbor) per region of interest (ROI) in selected areas. Differential interaction analyses were performed between the selected area and all other areas.

Single-cell cell dissociation

Tissue was digested in culture medium supplemented with 0.6 mg/mL collagenase IV and 0.01 mg/mL DNase I at 37°C for 1 h. Digested tissues were passed through a 0.4-μM cell strainer to obtain single-cell suspensions, followed by centrifugation at 300g for 5 min and resuspension in 36% Percoll. After another round of centrifugation at 500g for 5 min, cells were resuspended in 10 mL of blood lysis buffer for 10 min at RT to remove red blood cells.

Single-cell RNA-Seq data processing

For the droplet-based 10× Genomics data, the Cell Ranger Single-Cell toolkit provided by 10× Genomics was applied to align reads and generate the unique molecular identifier (UMI) matrix for each sample with Grch38 as the reference genome. Seurat (version 4.3.0) was used for downstream analysis.⁶⁷ Cells with fewer than 200 and above 6,000 detected genes were filtered out, as well as cells with a high proportion of mitochondrial gene counts per cell (>25%). Batch effects across different individuals were removed by first identifying anchors using the FindIntegrationAnchors() function, followed by IntegrateData() to obtain a batch-corrected space. For visualization, the dimensionality was reduced using TSNE implemented by RunTSNE() function. Major Clusters were identified by a shared nearest neighbor (SNN) modularity optimization-based clustering algorithm using FindNeighbors() and FindClusters() with default settings. Cells of each major cell type including epithelial, macrophage, T, and fibroblast cells were extracted from the raw matrix and further clustering was performed to obtain the subpopulation structures. FindMarkers() was applied to identify cluster-specific higher expressed genes. These genes were used for cluster annotation together with known lineage-specific markers.

In vitro experiments

Human HCC cell lines Huh-7, PLC/PRF/5 (PLC), SK-HEP-1, and HepG2 were maintained in the laboratory of T.L. at the First Affiliated Hospital, Zhejiang University School of Medicine. The human T cell line Molt-4 and normal human dermal fibroblasts (NHDF) were provided by Infinity Scope Inc. All cell lines were routinely tested negative for mycoplasma contamination.

In the experimental group, Epirubicin was dissolved in a complete culture medium to the final concentration of 1 μg/mL. then three cell lines including Huh-7, PLC, and SK-HEP-1 were cultured in the prepared medium and a control group was established by treating three cell lines with vehicle control, The cells were further used for immunofluorescence staining after incubation in the drug-containing medium at 37°C for 36 h.

Immunofluorescent imaging

Cells were cultured in a suitable culture dish using DMEM and incubated at 37°C until they reached the desired density. Cells were suspended in a buffer containing 0.2% Triton X-100 (P0096, Beyotime) and then fixed by incubating in 4% paraformaldehyde for 10 min at room temperature. After blocking with SuperBlock (37515, Thermofisher), primary antibodies PD-L1 (ab205921, Abcam), dsDNA (ab273137, Abcam), γ-H2AX(AP0687, Abclonal) and Phalloidin-iFluor 555 (ab176756, Abcam) followed by AF488 anti-mouse secondary antibody (4408, Cell Signaling Technology) and AF647 anti-rabbit secondary antibody (4414, Cell Signaling Technology) incubation. Then, slides were stained with DAPI for 5 min at RT. All immunofluorescent images were scanned via a four-laser scanning confocal microscope (Leica TCS SP8, Germany).

Western Blot

For western blot, tissues were lysed with RIPA lysis buffer (Beyotime Biotechnology, P0013C) containing phenylmethanesulfonyl fluoride (Beyotime Biotechnology, ST505) and Phosphatase Inhibitor Cocktail (Beyotime Biotechnology, P1081)for 30 min on ice, after which the mixture was centrifuged at 12,000×g for 15 min and the upper layer containing soluble proteins collected. Protein concentration was measured using a

bicinchoninic acid (BCA) reagent (Beyotime Biotechnology, P0012). Lysates were heated to 100°C in NuPAGE LDS sample buffer (4×) (Thermo Fisher Scientific, NP0007) for 3–5 min. Proteins were separated by SDS–PAGE then transferred to a PVDF membrane (Millipore, IPVH00010) which was blocked in 5% milk in TBST. PVDF membrane was incubated with suitable concentration of primary antibody at 4°C overnight, and then incubated with corresponding secondary antibody at room temperature for 2 h, and corresponding bands visualized using a ChemiScopeTouch (Clinx Science Instruments, China).

Spatial transcriptome preparation

The fresh livers were immediately frozen by optimal cutting temperature (OCT) after being cut from the patients, and then 10-μm slices were cut and pasted on the chip. After the tissues had been fixed, it was permeabilized. Next, the RNA in it was reverse-transcribed and amplified to obtain cDNA, followed by purification and library construction. Finally on the machine for sequencing.

ST data analysis and region identification

ST data (10× Genomics Visium) were processed with SpaceRanger and subsequently normalized by SCTransform provided in Seurat. Robust Cell Type Decomposition (RCTD)28 was used to estimate the proportion of each cell type in every spot. Spots with cell proportions over the 50% quantile of the corresponding cell type were defined as ImmuneSuppressiveEpithelial, epithelial, fibrosis, and T cell positive spots, respectively. ImmuneSuppressiveEpithelial&T cell positive spots were defined as spots positive for both ImmuneSuppressiveEpithelial and T cell signatures, while Epithelial&Fibroblast positive spots were defined as spots positive for both epithelial and fibroblast signatures.

Differential expression and cell-cell communication analysis for ST

FindMarkers() function from Seurat package was used to identify significantly highly expressed genes within Immune Suppressive Epithelials & T Cells and Epithelials & Fibroblasts positive spots (adjusted p-value <0.01) and clusterProfiler⁶⁸ was applied to find enriched KEGG pathways of differential expressed genes of selected regions. NicheNet³⁸ was applied to identify cell-cell communication events between selected cell subclusters and pathways.

Bulk RNAseq analysis

Raw reads were processed for quality controls and expression measures using the nf-core/rnaseq pipeline⁶⁹ with Grch38 as the reference genome. Differential expression analysis was performed using edgeR⁷⁰ and genes with FDR<0.05 were selected as significant.

Graph classification based on the spatial similarity of cells

We first divide the segmented cells in each ROI into N = 29 cell images according to the annotation results, where each cell image refers to the spatial localization for one cell type. Then, the pairwise spatial similarity between different cell types for each ROI i.e., M_i∈Rˆ(29 × 29) can be calculated via the measurement of structure similarity matrix (SSIM). After that, a network classification method introduced in⁵⁸ is applied to predict the drug efficacy on HCC and identify the distinguishing intersections among different cell types. In addition, we also compared the adopted network classification model⁵⁸ with two sparse learning algorithms i.e., LASSO⁷³ and L1-SVM,⁷⁴ where the LASSO is a regression analysis method that performs both variable selection and classification tasks, while the L1-SVM is a classic classification technique in machine learning that incorporates an L1 regularization term in the optimization process. Compared to traditional support vector machine (SVM), L1-SVM is better suited to handling high-dimensional sparse data.

Quantification and statistical analysis

Raw data obtained from FACS were analyzed in GraphPad Prism. Statistical tests were selected according to the comparison type, data distribution, and variance assumptions. For two-group comparisons, two-tailed Student’s t-tests or Mann-Whitney U tests were used as indicated in the corresponding figure legends. Paired t-tests were used for paired pre- and post-treatment comparisons. For spatial transcriptomic differential expression, Seurat FindMarkers() was used with an adjusted p-value threshold of <0.01, followed by pathway enrichment using clusterProfiler. For bulk RNA-seq analyses, edgeR was used and genes with FDR <0.05 were considered significant. For cell-cell interaction analyses, interactions with |log2(fold change)| > 1 and p-value <0.05 were highlighted. The data are represented as the means ± SDs unless otherwise stated. A p-value <0.05 was considered statistically significant. For comparisons among more than two groups, one-way ANOVA with Tukey's multiple comparisons test was used.

Published: October 2, 2026